Chemical-Guided Machine Learning Framework for Reactive Simulations of NOx Mitigation in Alcohol-Enhanced Ammonia–Methane Combustion

Chemical-Guided Machine Learning Framework for Reactive Simulations of NOx Mitigation in Alcohol-Enhanced Ammonia–Methane Combustion | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Chemical-Guided Machine Learning Framework for Reactive Simulations of NOx Mitigation in Alcohol-Enhanced Ammonia–Methane Combustion Amirali Shateri, Zhiyin Yang, Lei Xing, Yuying Yan, Zhen Wu, Jianfei Xie This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7800991/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract As carbon-free energy carrier, combustion of ammonia suffers from low burning rates, poor flame stability, and excessive nitrogen oxide (NOx) emissions. Although blending with methane alleviates some drawbacks, NOx formation remains a critical barrier. To address these challenges, we propose a hybrid framework combining reactive force field molecular dynamics (MD) simulations with machine learning (ML). MD simulations at 2,000–3,000 K were performed for ammonia-methane blended combustion with 0–10% addition of ethanol or methanol. Adding alcohols suppressed the NOx formation by altering charge redistributions and redirecting nitrogen intermediates into stabilising pathways. At 3,000 K, 10% ethanol and methanol reduced NOx by ~ 39.6% and ~ 30.1%, respectively. Both chemical and physical descriptors derived from MD were used to train ML models and extrapolated to unseen conditions (> 10% alcohol) with < 5% error in ethanol-rich mixtures. This framework reduces reliance on costly simulations while providing mechanistic insights and predictive capability of designing alternative fuels. Chemical Engineering NOx emission Ammonia–methane combustion Alcohol additives Machine learning Reactive force field Molecular dynamics Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Growing worldwide demand for sustainable energy sources promotes the use of fuels with low carbon emissions and carbon-neutral options. Ammonia (NH₃) represents a viable option for fuel applications since it burns carbon-neutrally and can integrate with existing infrastructure [ 1 – 2 ]. The slow reaction kinetics combined with low laminar burning velocity and high ignition temperature make NH₃ difficult to use as a standalone fuel. Ammonia-methane fuel blends have emerged as promising candidates for low-emission energy systems, especially in high-temperature combustion applications [ 3 ], to reduce the environmental impact of combustion processes by lowering harmful emissions, particularly nitrogen oxides (NOx) [ 4 ]. Studies on the combustion dynamics of NH₃/CH₄ mixes are conducted in conjunction with focuses on enhanced burner design. Singh et al. [ 5 ] used both computational models and experimental testing in a self-recuperative burner to show how different ammonia levels influence the behaviour of OH, HNO and HCO radicals, which in turn alters flame stability and emission levels of pollutants. Chu et al. [ 6 ] conducted a study on the ignition delay time of NH₃/CH₄ mixtures at temperatures ranging from 1,200 to 2,200 K, indicating consistent ignition performance at specific pressures with NH₃-dominant mixtures, thereby suggesting enhanced flame control capabilities for advanced combustion systems. When incorporating alcohol-based additives into fuel mixtures, it can effectively reduce pollutant emissions because of their oxygen content together with high octane ratings, including hydrocarbons (HC), carbon monoxide (CO), and NOx. Yu et al. [ 7 ] investigated how the methanol affected the kerogen pyrolysis using reactive force filed molecular dynamics (ReaxFF-MD) simulations. The collective findings from these studies demonstrate how alcohol additives can influence the hydrocarbon pyrolysis process and suggest encouraging applications for fuel optimization in energy solutions [ 8 ]. Because of the complicated nature of the interactions, traditional modelling approaches are not sufficient in determining the NOx formation in ammonia–methane–alcohol systems. This has led to a greater need for tools that can handle correlations between multiple variables and make accurate predictions in a wide range of cases. Machine learning (ML) is a promising solution in this context, facilitating data-driven investigation of combustion behaviour and informing analysis derived from molecular-scale simulations [ 8 ]. Models such as artificial neural network (ANN) have been widely used to accurately predict NOx emissions, enabling real-time control of combustion parameters and effective mitigation of emissions [ 9 ]. Similarly, use of ensemble models, such as random forest and extreme gradient boosting, has gained popularity due to their robust capacity, with accuracy rates typically being over 90% in the case of gas turbines [ 10 ]. Under the employment of feature optimization methods, their performance can be improved by reducing variable selection and conserving computational capacity, ensuring that predictions are always grounded in chemical and thermodynamic reality [ 11 – 13 ]. This study investigates molecular mechanism of NOx formation in ammonia–methane combustion systems using ReaxFF-MD simulations under various alcohol ratios. At 3,000 K, adding 10% ethanol or methanol reduced NOx emission by 39.6% or 30.1%, confirming the strong mitigation potential of alcohol-enhanced ammonia–methane combustion. To extend the predictive capability, we predict NOx emissions at off-design alcohol ratios (up to 12%) that lie outside the training zone by integrating atomistic-level simulation descriptors (e.g. bond energies, charge dynamics, and species evolution) into a data-driven framework. This approach enables the models not only to capture the hidden chemical patterns that drive NOx formation but also to generalize reliably to unseen fuel compositions. The novel combination of chemically rich descriptors with algorithmic diversity allows us to identify the most effective models for extrapolative combustion analysis. This hybrid chemical-guided data-driven framework therefore overcomes the limitations of both traditional simulations (computationally expensive) and previous ML approaches (interpolation-only, coarse inputs), providing a novel and scalable route to NOx mitigation in alcohol-enhanced ammonia–methane combustion. 2. Results 2.1. Pure MD simulations 2.1.1. Charge equilibration energy and NOx emissions Charge equilibration energy in ReaxFF simulations reveals the electron redistribution during chemical processes. Figure 1 displays how this parameter changes during combustion at 2,000 K and 3,000 K, highlighting the role of alcohol additives in forming dynamic electronic structures. At 2,000 K in Fig. 1 (a), mixtures that contain methanol show significantly lower charge equilibration energies compared to other mixtures. The system demonstrates a more stable electron distribution, which can be primarily attributed to the larger number of electronegative oxygen atoms present in alcohol molecules. The single C atom and pronounced polarity of methanol lead to early polar interactions that accelerate the electron delocalisation and subsequently decrease the excessive charge build-up on reactive intermediates [ 18 ]. The base fuel (C1, 0% Alc.) demonstrated the maximum equilibration energy, suggesting a less favourable charge distribution and probably slower radical reactions. The influence of alcohol additives becomes even more pronounced at 3,000 K in Fig. 1 (b). At this elevated temperature, ethanol-containing mixtures, particularly C6 (10% ethanol), achieved the maximum equilibration energy among all cases, underscoring their high reactivity and delayed stabilisation of the charge network. Conversely, methanol-rich systems, such as C10, show more moderate equilibration trends, indicating a superior capacity for sustaining intermediate species without excessive charge localisation. The increased thermal energy in these cases amplifies polar bond interactions to accelerate charge redistribution processes, especially in oxygen-rich molecular environments. 2.1.2. NOx emissions Figure 2 indicates direct connection between charge equilibration dynamics and NOx formation trends. At 2,000 K in Fig. 2 (a), NOx emissions are highest in methanol-containing blends, particularly C7 (5%) and C9 (10%), corresponding closely to the steepest charge gradient observed in those mixtures. The connection between electron-rich conditions caused by polar additives and early NO formation demonstrates that rapid NH 3 oxidation drives this process [ 19 ]. At 500 ps, the base fuel C1 produces a NOx level of 90.79 ppm; conversely, C5 (10% ethanol) obtains the lowest NOx concentration at 88.31 ppm. C3 (5% ethanol) and C7 (5% methanol) produce higher NOx emissions, 100.76 ppm and 110.54 ppm respectively; C9 (10% methanol) produces 94.46 ppm. Using 10% methanol in fuel results in lower NOx emissions compared with the base fuel although NOx emissions rise with other alcohol additives. At 3,000 K in Fig. 2 (b), however, the NOx profiles reverse. The blend with the highest ethanol content demonstrates minimal NOx output, illustrating a change in dominant chemical reaction routes. Here, the faster oxidation kinetics enabled at high temperature and coupled with the stabilising characteristic of ethanol's molecular structure, appear to suppress NO formation. Methanol-containing mixtures, on the other hand, remain more reactive over time, with higher NOx levels sustained at the final 20 ps (see inset in Fig. 2 (b)). The base fuel (C2) produces 109.95 ppm of NOx at 500 ps while C6 (10% ethanol) results in the lowest NOx level of 66.46 ppm, which is followed by C10 (10% methanol) with 76.84 ppm, then C4 (5% ethanol) with 81.63ppm, and C8 (5% methanol) with 89.99 ppm. Methanol promotes NOx formation via enhanced reactivity and rapid oxygen donation, primarily through NO₂ and HNO 3 intermediates. In addition, alcohol-derived radicals such as OH become more active at high temperatures to interact with nitrogen-containing intermediates, thereby stopping the forming routes of NOx [ 20 – 21 ]. NOx concentration inversions caused by variations in temperatures demonstrate how various alcohol types and concentrations affect this phenomenon. Methanol leads to early NOx production when the temperature is moderate. However, ethanol reduces NOx output at high temperatures even though it contains more energy. At 3,000 K, 10% ethanol reduces NOx emissions by approximately 39.5% compared to the base fuel, while 10% methanol achieves a 30.1% reduction. These findings underscore ethanol's superior role in suppressing NOx at high temperatures. This transition in NOx trends from 2,000 K to 3,000 K reflects a fundamental shift in dominant chemical mechanisms. 2.1.3. Dominant reaction pathway To elucidate how alcohol additives modulate nitrogen–oxygen reaction networks during combustion, Table 1 compiles the forward and reverse frequencies of the most active NOx-associated reactions across all ten fuel systems (C1–C10). The reactions were identified based on their participation in NO, NO₂, HNO, HNO₂, and HNO₃ interconversion loops, critical mechanisms governing NOx formation and removal [ 27 ]. As observed, the base case (C1 at 2000 K) exhibits strong engagement of NO₂ ⇌ HNO₃ and HNO₃-related formation pathways, reflecting the dominance of higher-order nitrogen oxides under lower-temperature oxidative conditions. At 3000 K (C2), the reaction network shifts toward simpler cycles such as HNO₂ ⇌ NO and NO ⇌ HNO, indicating increased fragmentation of NOx species and reduced stability of nitric acid derivatives at elevated temperatures. The addition of alcohols introduces pronounced shifts in pathway dominance and reversibility. Ethanol-blended systems (C3, C5, C4, C6) continue to feature NO₂–HNO₃ cycling, particularly at 2000 K, where NO₂ ⇌ HNO₃ remains among the top reversible pathways. This suggests ethanol's role in temporarily stabilizing NO₂ into higher-order oxides like HNO₃, a phenomenon that contributes to reduced free NO concentrations. At 3000 K, ethanol-enhanced systems (notably C6) show elevated activity in NO ⇌ HNO₂ and HNO₂ ⇌ NO + NO reactions, reinforcing the interpretation that ethanol increases cycling efficiency within intermediate species, thus delaying direct NO buildup. In contrast, methanol-enriched systems (C7, C9 at 2000 K and C8, C10 at 3000 K) reveal a marked shift in reaction priorities. The most frequent reactions are centred on NO ⇌ HNO and NO ⇌ HNO₂, with strong reverse reaction rates that exceed or match forward activity. This indicates methanol’s robust effect in diverting reactive nitrogen intermediates away from NO₂ production. Notably, in C10, the NO ⇌ HNO₂ reverse flux reaches 46, among the highest of all cases, highlighting methanol’s superior capacity to suppress NO₂ accumulation via intermediate stabilization. These trends are consistent with methanol’s high oxygen content and rapid radical propagation, which facilitate hydrogen transfer reactions that reduce NO directly to inert or less reactive species. Table 1 Dominant NOx reaction pathways and their forward and reverse frequencies across C1–C10 fuel systems. Reaction Systems Forward Reaction Frequency C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 NO2 ⇌ HNO3 64 0 50 0 52 0 52 0 58 0 NO ⇌ HNO 0 0 0 36 0 42 0 42 0 28 NO ⇌ HNO2 0 27 0 9 8 42 0 36 0 48 NO2 ⇌ NO 14 16 26 22 18 20 36 24 20 22 NO + NO ⇌ HNO2 0 20 0 16 0 0 0 39 0 24 HNO2 ⇌ NO 0 21 6 0 0 0 14 0 12 0 HNO3 ⇌ NO2 0 0 0 0 18 0 0 0 0 0 NO ⇌ N#N 0 6 0 8 0 10 0 9 0 4 HNO2 + HNO ⇌ NO + NO 0 0 0 0 0 12 0 16 0 20 Reverse Reaction Frequency NO2 ⇌ HNO3 46 0 48 0 48 0 50 0 52 46 NO ⇌ HNO2 0 26 0 22 6 30 0 32 0 46 NO ⇌ HNO 0 12 0 28 0 44 0 30 0 20 NO ⇌ NO2 12 14 10 16 14 18 16 14 6 12 HNO3 ⇌ NO2 0 0 0 0 9 0 20 0 0 0 HNO2 ⇌ NO 0 15 4 0 0 36 4 0 10 0 NO2 + NO ⇌ HNO2 3 15 0 0 0 0 0 18 0 3 NO ⇌ N#N 0 2 0 6 0 6 0 0 0 4 NO ⇌ HNO2 + HNO 0 9 3 9 0 0 0 15 0 0 HNO2 + HNO ⇌ NO + NO 3 12 0 4 9 8 4 12 4 16 2.2. ML predictions ML models are applied to predict NOx emissions from ammonia–methane–alcohol combustion based on previous MD simulation data. The goal is twofold: we first conduct model performance assessments with different architectures and then demonstrate the ability to make extrapolative predictions for off-design alcohol blends at 2%, 7%, and 12%. Four algorithms were considered in the predictions, including RFR, GBR, SVR, and FCNN, along with four distinct architectural configurations in each model. The comparison of predictions between these models allows for a deeper understanding of how algorithm design affects the predictive accuracy, stability, and feature sensitivity. The predictive robustness of trained ML models was assessed by comparing their outputs against MD-derived NOx values over time across ten combustion cases (C1–C10) that examined different alcohol concentrations (0%, 5%, 10%) and both ethanol and methanol blends at temperatures of 2,000 K and 3,000 K. Tables 2 display a direct comparison of NOx emissions between ML model predictions and MD simulation results in the base fuel systems, C1 at 2,000 K and C2 at 3,000 K, which complement the previously discussed performance metrics and learning curves. The analysis reports ten separate time steps ranging from 50 to 500 ps to consider the complete trajectory span. The RFR model demonstrates superior accuracy through its lowest recorded average absolute and relative errors at every time step. RFR exhibited the lowest error margins, with percentage errors consistently below 1.6% at 2,000 K and under 1% for most time steps at 3,000 K. For instance, RFR predicted 86.896 ppm vs. 87.883 ppm in MD at 250 ps, reporting an absolute error of only 0.987 ppm (1.12%). FCNN also maintained strong accuracy with errors generally under 2.5%, occasionally outperforming RFR at later time steps, such as at 500 ps in C2 in Table 2 , where it only had an error of 0.14 ppm. SVR and GBR showed good performance but displayed minor over- or under-predictions of NOx during longer reaction periods (e.g., 300–500 ps) which aligns with findings from predicted–actual plots (Fig. 4 in Extended data) and sensitivity analysis (Fig. 6 in Extended data). For example, at 300 ps in C1, the reported error by SVR reached 5.707 ppm (6.15%), which was closely followed by GBR. This quantitative evaluation supports the observed trends from the performance plots specifically the actual vs. predicted correlations in (Fig. 4 in Extended data), the learning curves in (Fig. 5 in Extended data), and the NOx evolution patterns predicted by the four ML models, reinforce the established performance hierarchy: RFR > FCNN2 > GBR4 ≈ SVR1, with RFR2 and FCNN2 emerging as the most effective configurations. While the results align well with the behaviour observed in physics-based simulations, this consistency suggests that ML models, particularly ensemble approaches can capture key patterns in NOx evolution within ammonia–methane combustion systems when trained appropriately. Table 2 Comparison of NOx (ppm) emissions between ML model predictions and MD simulation results at 2,000 K (System C1, 0% Alcohol) and 3,000 K (System C2, 0% Alcohol). 2.3. Extrapolative predictions for untrained alcohol ratios Performing extrapolative prediction is crucial for real-world combustion research, as it enables low-cost virtual screening of fuel blends without requiring additional expensive and time-consuming simulations. It reflects the model’s capacity to generalize beyond known data while remaining chemically consistent. Figure 3 shows the ML-predicted NOx profiles for untrained alcohol concentrations, e.g., 2%, 7%, and 12% of both ethanol and methanol at 2,000 K and 3,000 K. These systems (C11–C22) were not included in the original training data, meaning they represent a direct test of the model’s extrapolation capability. At 2,000 K in Fig. 3 (a), C17 (2% methanol) yields the highest NOx production, surpassing 110 ppm at 500 ps. In contrast, C15 (2% ethanol) closely tracks the baseline behaviour of C1, stabilising around 90–95 ppm. As concentration increases, this disparity narrows: C21 (7% methanol) maintains higher NOx than C19 (7% ethanol), while at 12% concentration, both methanol in C13 and ethanol in C11 converge around 95 ppm. These results align with earlier MD findings, reaffirming that methanol’s aggressive early oxidation at lower temperatures fosters NOx, while ethanol moderates the release of N-containing radicals. The trend shifts greatly at 3,000 K in Fig. 3 (b). All systems exhibit reduced NOx formation, but now ethanol-rich blends dominate the suppression. C12 (12% ethanol) falls below 70 ppm in the end, marking the lowest NOx level among all tested extrapolated cases. Meanwhile, C14 (12% methanol) levels off just under 80 ppm. Across mid-level concentrations, C20 (7% ethanol) also outperforms C22 (7% methanol). The low-concentration region again favours ethanol: C16 (2% ethanol) marginally outperforms C18 (2% methanol). These trends reinforce two interpretations of the underlying chemical reaction mechanisms. First, methanol is more reactive at low temperatures, likely accelerating NH 3 oxidation through earlier radical generation, which in turn drives NO formation. However, as temperature increases, ethanol’s slower and more controlled radical propagation favours routes that stabilise intermediates or divert nitrogen away from NO/NO 2 loops. Second, the model’s ability to recover these patterns, despite having been trained solely on 0%, 5%, and 10% alcohol ratios, demonstrates the capacity of chemical-guided ML. Table 3 compares the predicted values with synthesised ReaxFF MD datasets to validate the extrapolation results for 12% alcohol scenarios. These datasets were predicted for C11, C13 at 2,000 K and C12, C14 at 3,000 K, representing 12% ethanol and 12% methanol, respectively. The results are summarised in terms of absolute and percentage errors for each 100 ps period across 500 ps. The average percentage error for ethanol is relatively low, increasing from 3.845% at 100 ps, to 4.49% at 500 ps. The prediction remains stable, but the trend shows an incremental increase in error with time, reflecting the complexity of extrapolation. The percentage error begins at 2.02% when measured at 100 ps and increases to 12.78% at 400 ps. The rising error trend arises from methanol's increased reactivity, leading to more significant departures from MD results. The percentage error for 3,000 K cases begins at 6.58% at 100 ps and decreases to 5.97% at 500 ps. It suggests the model's predictions stabilize as the combustion dynamics approach the steady state at higher temperatures. In addition, a similar error pattern with methanol is observed. The error starts at 3.26% at 100 ps but increases significantly to 17.74% at 400 ps. The substantial error rise during intermediate timesteps underscores the difficulties in extrapolating results for high methanol concentrations and elevated temperatures. It can be concluded that increased prediction errors in methanol are due to its stronger reactivity that creates complex reaction pathways, making predictive process more challenging for ML models. Table 3 Quantitative comparison of RFR model predictions for 12% alcohol concentrations at 2,000 K (cases C11 and C13) and 3,000 K (cases C12 and C14). Output System Time (ps) Alc. % Alc. type MD ML Abs. error Percentage error 2000 K NOx (ppm) C11 100 12% Ethanol 91.061 94.908 3.847 4.225 C11 200 12% Ethanol 105.141 98.170 6.971 6.630 C11 300 12% Ethanol 109.811 99.385 10.426 9.494 C11 400 12% Ethanol 101.188 96.625 4.563 4.509 C11 500 12% Ethanol 83.330 87.818 4.488 5.386 C13 100 12% Methanol 87.706 89.726 2.020 2.303 C13 200 12% Methanol 88.135 95.456 7.321 8.306 C13 300 12% Methanol 101.925 103.652 1.727 1.695 C13 400 12% Methanol 87.059 99.838 12.779 14.679 C13 500 12% Methanol 88.026 92.682 4.656 5.289 3000 K NOx (ppm) C12 100 12% Ethanol 91.837 98.422 6.584 7.169 C12 200 12% Ethanol 70.915 82.390 11.475 16.181 C12 300 12% Ethanol 69.590 76.338 6.749 9.698 C12 400 12% Ethanol 70.112 74.099 3.987 5.687 C12 500 12% Ethanol 60.390 66.358 5.968 9.883 C14 100 12% Methanol 97.138 100.400 3.262 3.359 C14 200 12% Methanol 83.728 90.587 6.860 8.193 C14 300 12% Methanol 74.588 83.787 9.199 12.333 C14 400 12% Methanol 65.553 77.181 11.628 17.738 C14 500 12% Methanol 71.582 74.921 3.339 4.664 While the extrapolation to 12% alcohol is successfully achieved with acceptable accuracy, the extrapolation to higher ratios of alcohol (e.g., 15–20%) is difficult. The increased reactivity and complexity of chemical pathways at higher concentrations make it difficult to maintain high predictive accuracy. When the alcohol ratio is increased to 15–20%, mechanisms of the fuel and air mixture and reaction can change significantly, leading to big differences between the ML predicted results and the actual combustion outcomes. This limitation likely stems from a lack of representative training samples capturing the altered charge distributions, energy profiles, and species interaction dynamics unique to these higher alcohol blends. Further work is needed to enhance the extrapolation process, particularly to 15–20% alcohol levels, which may need more training data or better feature chemical reactions to capture non-linear effects at those levels. These non-linearities arise from threshold-driven changes in radical formation, bond dissociation patterns, and secondary reactions that do not scale linearly with alcohol concentration. For example, higher hydroxyl content at > 15% alcohol may shift dominant NOx formation pathways or trigger new intermediate stabilization loops not seen at lower concentrations. 3. Conclusions This study employed ReaxFF-MD simulations to assess the effects of alcohol additives in ammonia–methane combustion. Alcohol enrichment significantly altered the charge redistribution, intermediate stability, and NOx pathways. At 2,000 K, methanol promoted electron withdrawal and increased NOx formation, whereas ethanol suppressed emissions. At 3,000 K, 10% ethanol showed the most significant effect, a 39.5% decrease of NOx compared to the base case, while methanol produced a more moderate reduction. Mechanistic analysis revealed distinct pathways: ethanol favoured NO 2 -HNO 3 loops and nitrate decomposition, while methanol promoted HNO- and N 2 O-mediated conversion. ML models trained on atomistic descriptors confirmed the predictive capability of the hybrid framework. Among the tested models, RFR2 model achieved the highest accuracy, FCNN2 ranked second, GBR4 exhibited moderate performance, and SVR1 was the weakest. Importantly, the RFR model reliably extrapolated to unseen alcohol ratios up to 12%, maintaining prediction errors below 5% in ethanol-rich mixtures. Overall, the integrated MD-ML framework provides both mechanistic insight and strong predictive capability, enabling scalable exploration of ammonia–methane–alcohol blends for cleaner combustion. Future work will aim to enhance the extrapolative robustness of the ML models by integrating noise-reduced, chemically valid training data [ 22 ] and incorporating reaction-aware physical priors [ 23 ] to better capture transition dynamics. In addition, emerging frameworks such as optimal transport for reaction state mapping [ 24 ], physics-informed neural networks (PINN) for reaction–diffusion systems [ 23 ], and language-model-driven chemical embedding techniques [ 25 ] are alternatives to generalize the current approach toward broader chemical regimes, including transition state estimation and real-time reactive force field tuning [ 26 ]. 4. Online Methods 4.1. Reactive force field molecular dynamics (ReaxFF MD) Reactive force field (ReaxFF) molecular simulations provide a robust framework for modelling chemical reactions at the atomic scale by dynamically accounting for bond formation and breaking. This modelling technique bridges the gap between quantum mechanical computations, which are precise but computationally difficult, and classical molecular dynamics (MD) simulations, which are effective but non-reactive. ReaxFF enables the modelling of extensive systems under extreme pressures and temperatures with a high level of precision, facilitating the physical capture of complex chemical reactions. It has been widely applied in combustion studies, offering valuable insights into reaction mechanisms, intermediate species formation, and product distributions [ 14 ]. The ReaxFF can be expresses as a function of the bond order: $$\:{E}_{system}={E}_{bond}+{E}_{over}+{E}_{under}+{E}_{lp}+{E}_{angle}+{E}_{tors}+{E}_{vdW}+{E}_{Coul}$$ 1 The total potential energy in ReaxFF is represented as the sum of various energy contributions, including bond energy, penalties for over- and under-coordination, lone-pair stabilization, valence and torsional angle energies, and non-bonded interactions such as Coulombic and van der Waals forces [ 15 – 16 ]. By parameterizing these energy terms based on quantum mechanical calculations, ReaxFF can simulate chemical processes in large systems with both computational efficiency and chemical accuracy. 4.1.1. System configuration To investigate the influence of alcohol additives on ammonia–methane combustion, ten computational cases (C1–C10) were configured. Each case employed an equivalence ratio of λ = 0.7, a moderately fuel‑rich condition typical of practical combustors where perfect stoichiometry (λ = 1) is rarely achieved. Operating under fuel‑rich conditions promotes the build‑up of reactive intermediates and enables detailed examination of NOx formation pathways which is the primary objective of this study. For physical consistency, the mass density was fixed at 0.34 g/cm³ in every simulation. Although the total number of molecular remained constant at 800, the cubic domain length was adjusted case‑by‑case to satisfy the density constraint, ensuring that any changes in system behaviour originate solely from variations in fuel composition or temperature. The baseline mixture comprised 422 fuel molecules—methane (CH₄) and ammonia (NH₃) in a 1:1 ratio—together with 378 oxygen molecules. Alcohol co‑fuels, ethanol (C₂H₆O) or methanol (CH₄O), were introduced by substituting 5% or 10% of the original CH₄-NH₃ base fuel while preserving the overall molecule count. Distinct C–H bond energies and oxygen contents in selected alcohols are expected to affect their radical pools and chain‑branching behaviour. Each composition was evaluated at 2000 K and 3000 K to capture temperature‑dependent kinetics. 4.2. Machine learning framework 4.2.1. Data structure and preprocessing Figure 4 shows the data processing pipeline that links MD simulations to ML models applied in this paper. ReaxFF-based MD simulations created the baseline dataset from simulations performed at temperatures 2,000 K and 3,000 K for 10 chemical cases, which are labelled C1 to C10. Each case simulated a 500 ps reactive trajectory and produced a series of thermochemical and molecular descriptors, which were designed to characterise the evolving behaviour of ammonia–methane–alcohol mixtures. The resulting dataset includes 26 input features ( X ), encompassing key thermophysical observables such as temperature, pressure, and density, as well as energy terms including total, kinetic, and potential energies. In addition, features derived from the ReaxFF were included to represent mechanistic quantities of the direct chemical relevance. These consist of individual energy contributions: v_ea (atomic), v_eb (bond), v_elp (lone pair), v_ev (valence angle), v_epen (penalty), v_ecoa (conjugation), v_ehb (hydrogen bond), v_et (torsion), v_eco (conjugated), v_ew (van der Waals), v_ep (Coulomb), and v_eqeq (charge equilibration). Element-specific charge distribution properties were also computed: v_qC, v_qH, v_qN, and v_qO, representing the mean partial charge for each atom type, which are essential for capturing electron redistribution during combustion. Chemical identity was represented by the mass fraction of alcohol feature, which distinguishes 0%, 5%, and 10% mixed fuels, and binary indicators for ethanol and methanol, to account for molecular structure effects. These enable the models to learn not only compositional trends but also the functional impact of hydroxyl content and C–H/O–H bond energetics on various alcohol additives. A rigorous data cleaning step was employed prior to training the models. This included the detection and removal of anomalous entries owing to rare bond breakage events, unphysical charge spikes, and species mislabelling during dynamic equilibration. Outliers were eliminated using interquartile-based cutoff values, and no missing values were encountered, as molecular dynamics simulations deterministically generate complete trajectories with time-resolved data for all atoms and properties. Preprocessing was followed by 80:20 (e.g., 80% of data used for training and 20% for validation) stratified train-test splitting in a manner that ensured proportional representation of all the compositions and temperatures within the training and testing sets [ 8 , 17 ]. 4.2.2. ML algorithms Four ML algorithms were chosen in this work to model the intricate interaction between combustion descriptors and NOx emissions: RFR, GBR, FCNN, and SVR. Their capacity to capture non-linear dependencies, resilience against overfitting and proven success in thermochemical prediction tasks guided selection of these models. All the algorithms have different strengths that are well-suited to the nature of combustion-related data. RFR, as an ensemble decision tree, performs exceptionally well with high-dimensional data and produces variable importance information, which is critical in understanding the influence of physical parameters. GBR augments the same by incorporating gradient-based optimization in the construction of trees and thereby is well suited to pick up weak interactions in noisy or structured data. FCNNs, on the other hand, are capable of learning intricate, continuous functions due to their hierarchical representation power, particularly useful in tracking multi-variable chemistries' dynamics. SVR becomes a standard regression benchmark because it excels with non-linear data by using the kernel-based similarity functions to capture non-linearity. Four distinct versions of each algorithm were created to explore the effect of design choices on their performance by adjusting internal parameters. To evaluate the robustness and generalizability of tree-based algorithms (i. e., RFR, GBR), a systematic approach was taken to adjust both tree count and depth parameters. Neural networks (NN) configurations included various hidden layer counts and neuron distributions ranging from basic one or two layered networks to deep structures with up to five hidden layers. SVR configurations were altered by experimenting with different kernel functions (linear, polynomial, or radial basis) and regularisation parameters. 4.3. Chemical-guided extrapolation To extend the predictive capability of different ML models beyond the training range, a chemical-guided extrapolation strategy was employed. This approach aimed to infer NOx emissions for untrained alcohol ratios, specifically 2%, 7%, and 12%, by leveraging both data-driven models and domain-specific chemical insights derived from ReaxFF simulations. As illustrated on the right panel of Fig. 4 , the post-processing workflow builds upon the trained model’s outputs at known alcohol concentrations (0%, 5%, and 10%). The key idea was to avoid the need for new MD simulations at unseen compositions, and instead to estimate the target variables by interpolating or extrapolating trends observed in the training data. This hybridisation of ML and physical chemistry offers a practical route for off-design predictions without incurring the computational cost of generating new datasets. Specifically, for in-range predictions such as 2% and 7%, a linear interpolation scheme was adopted between the adjacent training points (e.g., 0%–5%, 5%–10%). This approach assumes smooth transitions in NOx behaviour across small alcohol increments, which was consistent with the gradual trends captured by ReaxFF simulations. For out-of-range prediction at 12%, which is a value not seen during model training, a scaling-based extrapolation technique was applied. This was informed by the observed gradient between the 5% and 10% alcohol systems and adjusted proportionally beyond the upper bound. To mathematically implement this extrapolation, a scaling coefficient is introduced and is expressed below: $$\:{{Feature}_{12\%}}^{\:}={{Feature}_{10\%}+\left(\frac{12-10}{10-5}\right)\times\:\left({{Feature}_{10\%}-{{Feature}_{5\%}}^{\:}}^{\:}\right)}^{\:}$$ 2 This transformation was applied independently to each input feature before feeding it through the trained models, allowing the prediction of NOx emissions for novel alcohol levels. This method is described in Fig. 4 as post-processing step to generate unseen alcohol ratios, and it bridges the gap between chemical-guided trends and data-driven modelling. The underlying rationale for this strategy is based on the smooth, thermochemically governed nature of alcohol substitution in reactive systems. Then, ML models, which were trained on clean, labelled inputs, were able to capture the nonlinear coupling between chemical features and NOx formation, reinforcing the robustness of predictions. By combining these two pillars, e.g., data-driven learning and chemically grounded extrapolation, the hybrid method demonstrated in this study offers a reliable alternative to conventional simulation-heavy routes for combustion design. This novel approach enables fast, cost-effective insights into emission reduction under off-design conditions, showcasing a vital step towards adaptive, intelligent optimisation in combustion chemistry. References Lü M, Long W, Wei F, Dong D, Li C, Dong P, Tian H, Chen X, Chen S, Wang Y, Wang P (2024) Assessment of carbon-free fuel ammonia combustion with low methanol blends in reducing GHG emissions including N2O. J Clean Prod 463:142755 Lü M, Long W, Wei F, Dong D, Li C, Dong P, Tian H, Chen X, Chen S, Wang Y, Wang P (2024) Assessment of carbon-free fuel ammonia combustion with low methanol blends in reducing GHG emissions including N2O. J Clean Prod 463:142755 Guo Y, Shi H, Liu H, Xie Y, Guan Y (2023) Reactive molecular dynamics simulation and chemical kinetic modeling of ammonia/methane co-combustion. Fuel Shateri A, Jalili B, Saffar S, Jalili P, Ganji DD (2023) Numerical study of the effect of ultrasound waves on the turbulent flow with chemical reaction. Energy Singh AS, Vijrumbana Y, Reddy VM (2023) Experimental and computational (Chemical Kinetic + CFD) analyses of Self-Recuperative annular tubular porous burner for NH3/CH4 -air Non-Premixed combustion. Chem Eng J Chu X, Gao XL, Ma P, Xiao Z, Xie H, Zhang C, Z., Wang X (2024) High-temperature auto-ignition characteristics of NH3-H2-CH4. Fuel 365:131228 Yu M, Zhan J-H, Li X, He W, Liu X (2023) Effect of methanol on the pyrolysis behaviour of kerogen by ReaxFF molecular dynamics simulations. Molecular Simulation Shateri A, Yang Z, Xie J (2025) Machine Learning-Based Molecular Dynamics Studies on Predicting Thermophysical Properties of Ethanol–Octane Blends. Energy & Fuels Mao G-P, Shi T, Mao C, Wang P (2023) Prediction of NOx emission from two-stage combustion of NH3–H2 mixtures under various conditions using artificial neural networks. Int J Hydrog Energy Hoque KE, Hossain T, Haque AM, Miah MAK, Haque MA (2024) NOx Emission Predictions in Gas Turbines through Integrated Data-Driven Machine Learning Approaches. J Energy Resour Technology-Transactions Asme Pachauri N (2024) An emission predictive system for CO and NOx from gas turbine based on ensemble machine learning approach. Fuel Adams D, Oh DH, Kim DW, Lee CH, Oh M (2020) Prediction of SOx–NOx emission from a coal-fired CFB power plant with machine learning: Plant data learned by deep neural network and least square support vector machine. J Clean Prod 270:122310 Shateri A, Yang Z, Xie J (2024) Utilizing Artificial intelligence to identify an Optimal Machine learning model for predicting fuel consumption in Diesel engines. Energy and AI Mao Q, Yang MD, Wei C, Liu X, Jiang X, Ren Y, Luo KH, van Duin AC T (2023) Classical and reactive molecular dynamics: Principles and applications in combustion and energy systems. Prog Energy Combust Sci 97:101084 Chowdhury A, van Duin ACT (2017) Extension of the ReaxFF Combustion Force Field toward Syngas Combustion and Initial Oxidation Kinetics. J Phys Chem A 121(5):1051–1068 van Duin ACT, Dasgupta S, Lorant F, Goddard WA (2001) ReaxFF: A Reactive Force Field for Hydrocarbons. J Phys Chem A 105(41):9396–9409 Ihme M, Chung WT, Mishra A (2022) Combustion machine learning: Principles, progress and prospects. Prog Energy Combust Sci 91:101010 Zapiain DM, de Venkatraman O, A., Wilson M (2024) Accelerating Charge Estimation in Molecular Dynamics Simulations Using Physics-Informed Neural Networks: Corrosion Applications Borodin A, Höfft O, Kempter V, Ferro Y, Allouche A-R (2004) Electron delocalization by polar molecules: interaction of Na atoms with solid ammonia films studied with MIES and density functional theory. J Chem Phys 121(8):3717–3721 Watson GMG, Versailles P, Bergthorson JM (2017) NO formation in rich premixed flames of C1–C4 alkanes and alcohols. 36(1):627–635 Gao Y, Guo Y, Xie Y, Qin H, Guan Y (2024) Kinetic mechanism and NOx emission characteristics of the ammonia/alcohol co-combustion explored by reactive molecular dynamics calculation and kinetic numerical simulation. Combust Flame 273:113953 Toniato A, Schwaller P, Cardinale A, Geluykens J, Laino T (2021) Unassisted noise reduction of chemical reaction datasets. Nat Mach Intell 3(6):485–494 Rao C, Ren P, Wang Q, Buyukozturk O, Sun H, Liu Y (2023) Encoding physics to learn reaction–diffusion processes. Nat Mach Intell 5(7):765–779 Duan, C., Liu, G. H., Du, Y., Chen, T., Zhao, Q., Jia, H., … & Kulik, H. J. (2025). Optimal transport for generating transition states in chemical reactions. Nature Machine Intelligence, 7(4), 615–626 Jablonka KM, Schwaller P, Ortega-Guerrero A, Smit B (2024) Leveraging large language models for predictive chemistry. Nat Mach Intell 6(2):161–169 Gong, S., Zhang, Y., Mu, Z., Pu, Z., Wang, H., Han, X., … Xiang, L. (2025). A predictive machine learning force-field framework for liquid electrolyte development. Nature Machine Intelligence, 1–10 Shateri A, Yang Z, Sherkat N, Xie J (2026) Alcohol additives to enhance ammonia-methane combustion efficiency and reduce emissions: A reactive force field analysis. Fuel 405:136565 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7800991","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":526064265,"identity":"aa39fcec-4ab1-4e9a-9d6a-fab1d851c7d2","order_by":0,"name":"Amirali Shateri","email":"","orcid":"","institution":"University of Derby - Derby Campus","correspondingAuthor":false,"prefix":"","firstName":"Amirali","middleName":"","lastName":"Shateri","suffix":""},{"id":526064266,"identity":"04fe200a-4e68-4056-aafd-44f1d31018ca","order_by":1,"name":"Zhiyin Yang","email":"","orcid":"","institution":"University of Derby","correspondingAuthor":false,"prefix":"","firstName":"Zhiyin","middleName":"","lastName":"Yang","suffix":""},{"id":526064267,"identity":"f599ece0-30cc-42b6-8e9f-12fad0371fbe","order_by":2,"name":"Lei Xing","email":"","orcid":"","institution":"University of Surrey","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Xing","suffix":""},{"id":526064268,"identity":"cb956b00-f083-4d24-9013-fd08094653b3","order_by":3,"name":"Yuying Yan","email":"","orcid":"","institution":"University of Nottingham","correspondingAuthor":false,"prefix":"","firstName":"Yuying","middleName":"","lastName":"Yan","suffix":""},{"id":526064269,"identity":"04518dd3-97ca-4564-864c-94f18c3ad0d8","order_by":4,"name":"Zhen Wu","email":"","orcid":"","institution":"Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Zhen","middleName":"","lastName":"Wu","suffix":""},{"id":526064270,"identity":"29178297-40b4-4811-b4fa-a10ba5cbc8ce","order_by":5,"name":"Jianfei Xie","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyklEQVRIiWNgGAWjYBACxgYGhgMMFcwMDBJQEQl8yhFazpCiBaKvjRQtzA08hgc+zrOWM5/dY8Dwo4YhcWYDQYexJRycuS3dWObOGQPGnmMMibMJuqmB+cBh3m2HE2dI5Bgw8DYwJM4jrIWx4fDfOYfrQVoY/xKnBWgLUFeCBFALM8gWwg5rBvql51i64QyZYwWHZY5JGBP0vmF7j/GHHzXW8hLSzRsfvqmxkZ1xgJCWZjiTw+AAUREpj2CyPyCsfBSMglEwCkYkAABPNz65or3u5AAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-2217-3506","institution":"University of Derby","correspondingAuthor":true,"prefix":"","firstName":"Jianfei","middleName":"","lastName":"Xie","suffix":""}],"badges":[],"createdAt":"2025-10-07 15:31:09","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7800991/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7800991/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":93104390,"identity":"2b4be29a-5598-42ae-9d50-46a3878c70db","added_by":"auto","created_at":"2025-10-09 06:19:43","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1701409,"visible":true,"origin":"","legend":"","description":"","filename":"MainText.docx","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/e64c98232314165a442847c2.docx"},{"id":93104091,"identity":"7c297a64-c0ad-4276-aef3-2d610c4b4749","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7800991.json","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/193281e7fb147c6b7c53a7e1.json"},{"id":93104094,"identity":"5173165f-52dd-4799-a1b7-fed22795fe3b","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":117104,"visible":true,"origin":"","legend":"","description":"","filename":"rs78009910enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/05e176e17f1d1721b6ac8431.xml"},{"id":93104386,"identity":"cab1fda1-5b0a-43f0-b0be-b8ee1e9a05ed","added_by":"auto","created_at":"2025-10-09 06:19:41","extension":"jpeg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1042600,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/ee84d0466818444aa424a0cf.jpeg"},{"id":93104099,"identity":"b804a041-2f91-493d-9ce7-bae1883f5e74","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":945892,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/16487adafd5a3c263f17b88a.jpeg"},{"id":93104107,"identity":"5c82ea49-6e7f-4b4f-a66f-a9f36efad12d","added_by":"auto","created_at":"2025-10-09 06:11:42","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1126561,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/b766643695bc9d2c3847c047.jpeg"},{"id":93104101,"identity":"2bcaa0c5-c5c4-420b-a5a0-10a9e54fbd39","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":978914,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/0aabe3b3affc71417b6d4ec2.jpeg"},{"id":93104103,"identity":"0cd0b653-56b7-4e53-b0c1-8e3a05c51c36","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":693352,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/ed94b78bc7497cf8a5de9e0d.png"},{"id":93104105,"identity":"4a97915c-d3df-4ec5-b245-ed9ad4560f24","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":263846,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/de66ce9d1cc62b26566f0552.png"},{"id":93104387,"identity":"9ff65e80-8fdf-42ad-b9d0-b9ae7a5d6926","added_by":"auto","created_at":"2025-10-09 06:19:41","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":229933,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/ea211525f2ad6338353d3c81.png"},{"id":93104102,"identity":"4b11b7f8-8a8d-47ff-baf6-e187ed3764f8","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":231080,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/4d6315ec78bbcf3c490c1301.png"},{"id":93104104,"identity":"0b7f98cb-7bef-42ee-bc90-5ead35c83200","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":226519,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/9471ab16822b926b43207a3f.png"},{"id":93104109,"identity":"f965b2df-aa42-494a-9294-1b43df35e56f","added_by":"auto","created_at":"2025-10-09 06:11:42","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":174877,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/68c3acca88136e5b30a5c348.png"},{"id":93104106,"identity":"a2a8310c-00c2-4694-af11-d5401a1a7d82","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"xml","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":115889,"visible":true,"origin":"","legend":"","description":"","filename":"rs78009910structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/b3c4c859965385c3e9f3e609.xml"},{"id":93104388,"identity":"74e5df44-8166-4f90-99ce-c22c42de6b75","added_by":"auto","created_at":"2025-10-09 06:19:42","extension":"html","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":123146,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/073d1156806be695279501cd.html"},{"id":93104092,"identity":"d733b1a5-f6a4-4250-a973-53c5ff349ffd","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1042600,"visible":true,"origin":"","legend":"\u003cp\u003eTemporal evolution of charge equilibration energy for all fuel blends (0%, 5%, 10% Alc.) at: (a) 2,000 K and (b) 3,000 K. Insets magnify the final 20ps to highlight the equilibrium behaviour.\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/e1f550ee08d16a145461a15a.jpeg"},{"id":93104093,"identity":"925b686d-77b3-44fc-9e5f-bf442df2de49","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":945892,"visible":true,"origin":"","legend":"\u003cp\u003eNOx emissions over time for: (a) 2,000 K and (b) 3,000 K. Trends are directly linked to alcohol type (ethanol and methanol) and concentration (0%, 5%, 10% Alc.). Insets show stabilisation levels at the final 20ps of the simulation.\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/77358d3f0f3021fe45dda8ab.jpeg"},{"id":93104095,"identity":"43b49e38-924c-49aa-8dd7-a73c0fb0483e","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":978914,"visible":true,"origin":"","legend":"\u003cp\u003eML-predicted NOx emissions for extrapolated alcohol concentrations (2%, 7%, and 12%) of methanol-ethanol in ammonia–methane combustion at: (a) 2,000 K and (b) 3,000 K. The colour codes are: dark red (C15, C16) for 2% ethanol, dark green (C17, C18) for 2% methanol, orange (C19, C20) for 7% ethanol, purple (C21, C22) for 7% methanol, dark blue (C11, C12) for 12% ethanol, and grey (C13, C14) for 12% methanol.\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/4470a029561dc4817325b1de.jpeg"},{"id":93104096,"identity":"4fc4f4ac-5158-4a47-96b7-893003a08850","added_by":"auto","created_at":"2025-10-09 06:11:41","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1094749,"visible":true,"origin":"","legend":"\u003cp\u003eML workflow integrating chemical-guided extrapolation for NOx prediction from ReaxFF-MD derived combustion data. NOx was calculated based on the normalized counts of NOx-related species (i.e., NO, NO₂, N₂O).\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/1aeea6c3d47c9660aab32060.png"},{"id":93105272,"identity":"293be18a-e1a1-48d9-8619-1a533172f808","added_by":"auto","created_at":"2025-10-09 06:27:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4636560,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7800991/v1/c4e62f9b-9126-4c97-bece-c70a1ae4f0d9.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eChemical-Guided Machine Learning Framework for Reactive Simulations of NOx Mitigation in Alcohol-Enhanced Ammonia–Methane Combustion\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eGrowing worldwide demand for sustainable energy sources promotes the use of fuels with low carbon emissions and carbon-neutral options. Ammonia (NH₃) represents a viable option for fuel applications since it burns carbon-neutrally and can integrate with existing infrastructure [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The slow reaction kinetics combined with low laminar burning velocity and high ignition temperature make NH₃ difficult to use as a standalone fuel. Ammonia-methane fuel blends have emerged as promising candidates for low-emission energy systems, especially in high-temperature combustion applications [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], to reduce the environmental impact of combustion processes by lowering harmful emissions, particularly nitrogen oxides (NOx) [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eStudies on the combustion dynamics of NH₃/CH₄ mixes are conducted in conjunction with focuses on enhanced burner design. Singh et al. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] used both computational models and experimental testing in a self-recuperative burner to show how different ammonia levels influence the behaviour of OH, HNO and HCO radicals, which in turn alters flame stability and emission levels of pollutants. Chu et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] conducted a study on the ignition delay time of NH₃/CH₄ mixtures at temperatures ranging from 1,200 to 2,200 K, indicating consistent ignition performance at specific pressures with NH₃-dominant mixtures, thereby suggesting enhanced flame control capabilities for advanced combustion systems. When incorporating alcohol-based additives into fuel mixtures, it can effectively reduce pollutant emissions because of their oxygen content together with high octane ratings, including hydrocarbons (HC), carbon monoxide (CO), and NOx. Yu et al. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] investigated how the methanol affected the kerogen pyrolysis using reactive force filed molecular dynamics (ReaxFF-MD) simulations. The collective findings from these studies demonstrate how alcohol additives can influence the hydrocarbon pyrolysis process and suggest encouraging applications for fuel optimization in energy solutions [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eBecause of the complicated nature of the interactions, traditional modelling approaches are not sufficient in determining the NOx formation in ammonia\u0026ndash;methane\u0026ndash;alcohol systems. This has led to a greater need for tools that can handle correlations between multiple variables and make accurate predictions in a wide range of cases. Machine learning (ML) is a promising solution in this context, facilitating data-driven investigation of combustion behaviour and informing analysis derived from molecular-scale simulations [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Models such as artificial neural network (ANN) have been widely used to accurately predict NOx emissions, enabling real-time control of combustion parameters and effective mitigation of emissions [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Similarly, use of ensemble models, such as random forest and extreme gradient boosting, has gained popularity due to their robust capacity, with accuracy rates typically being over 90% in the case of gas turbines [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Under the employment of feature optimization methods, their performance can be improved by reducing variable selection and conserving computational capacity, ensuring that predictions are always grounded in chemical and thermodynamic reality [\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis study investigates molecular mechanism of NOx formation in ammonia\u0026ndash;methane combustion systems using ReaxFF-MD simulations under various alcohol ratios. At 3,000 K, adding 10% ethanol or methanol reduced NOx emission by 39.6% or 30.1%, confirming the strong mitigation potential of alcohol-enhanced ammonia\u0026ndash;methane combustion. To extend the predictive capability, we predict NOx emissions at off-design alcohol ratios (up to 12%) that lie outside the training zone by integrating atomistic-level simulation descriptors (e.g. bond energies, charge dynamics, and species evolution) into a data-driven framework. This approach enables the models not only to capture the hidden chemical patterns that drive NOx formation but also to generalize reliably to unseen fuel compositions. The novel combination of chemically rich descriptors with algorithmic diversity allows us to identify the most effective models for extrapolative combustion analysis. This hybrid chemical-guided data-driven framework therefore overcomes the limitations of both traditional simulations (computationally expensive) and previous ML approaches (interpolation-only, coarse inputs), providing a novel and scalable route to NOx mitigation in alcohol-enhanced ammonia\u0026ndash;methane combustion.\u003c/p\u003e"},{"header":"2. Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.1. Pure MD simulations\u003c/h2\u003e\n \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e\n \u003ch2\u003e2.1.1. Charge equilibration energy and NOx emissions\u003c/h2\u003e\n \u003cp\u003eCharge equilibration energy in ReaxFF simulations reveals the electron redistribution during chemical processes. Figure \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e displays how this parameter changes during combustion at 2,000 K and 3,000 K, highlighting the role of alcohol additives in forming dynamic electronic structures. At 2,000 K in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e(a), mixtures that contain methanol show significantly lower charge equilibration energies compared to other mixtures. The system demonstrates a more stable electron distribution, which can be primarily attributed to the larger number of electronegative oxygen atoms present in alcohol molecules. The single C atom and pronounced polarity of methanol lead to early polar interactions that accelerate the electron delocalisation and subsequently decrease the excessive charge build-up on reactive intermediates [\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e]. The base fuel (C1, 0% Alc.) demonstrated the maximum equilibration energy, suggesting a less favourable charge distribution and probably slower radical reactions. The influence of alcohol additives becomes even more pronounced at 3,000 K in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e(b). At this elevated temperature, ethanol-containing mixtures, particularly C6 (10% ethanol), achieved the maximum equilibration energy among all cases, underscoring their high reactivity and delayed stabilisation of the charge network. Conversely, methanol-rich systems, such as C10, show more moderate equilibration trends, indicating a superior capacity for sustaining intermediate species without excessive charge localisation. The increased thermal energy in these cases amplifies polar bond interactions to accelerate charge redistribution processes, especially in oxygen-rich molecular environments.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e\n \u003ch2\u003e2.1.2. NOx emissions\u003c/h2\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e indicates direct connection between charge equilibration dynamics and NOx formation trends. At 2,000 K in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e(a), NOx emissions are highest in methanol-containing blends, particularly C7 (5%) and C9 (10%), corresponding closely to the steepest charge gradient observed in those mixtures. The connection between electron-rich conditions caused by polar additives and early NO formation demonstrates that rapid NH\u003csub\u003e3\u003c/sub\u003e oxidation drives this process [\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e]. At 500 ps, the base fuel C1 produces a NOx level of 90.79 ppm; conversely, C5 (10% ethanol) obtains the lowest NOx concentration at 88.31 ppm. C3 (5% ethanol) and C7 (5% methanol) produce higher NOx emissions, 100.76 ppm and 110.54 ppm respectively; C9 (10% methanol) produces 94.46 ppm. Using 10% methanol in fuel results in lower NOx emissions compared with the base fuel although NOx emissions rise with other alcohol additives. At 3,000 K in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e(b), however, the NOx profiles reverse. The blend with the highest ethanol content demonstrates minimal NOx output, illustrating a change in dominant chemical reaction routes. Here, the faster oxidation kinetics enabled at high temperature and coupled with the stabilising characteristic of ethanol\u0026apos;s molecular structure, appear to suppress NO formation. Methanol-containing mixtures, on the other hand, remain more reactive over time, with higher NOx levels sustained at the final 20 ps (see inset in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e(b)). The base fuel (C2) produces 109.95 ppm of NOx at 500 ps while C6 (10% ethanol) results in the lowest NOx level of 66.46 ppm, which is followed by C10 (10% methanol) with 76.84 ppm, then C4 (5% ethanol) with 81.63ppm, and C8 (5% methanol) with 89.99 ppm. Methanol promotes NOx formation via enhanced reactivity and rapid oxygen donation, primarily through NO₂ and HNO\u003csub\u003e3\u003c/sub\u003e intermediates. In addition, alcohol-derived radicals such as OH become more active at high temperatures to interact with nitrogen-containing intermediates, thereby stopping the forming routes of NOx [\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e]. NOx concentration inversions caused by variations in temperatures demonstrate how various alcohol types and concentrations affect this phenomenon. Methanol leads to early NOx production when the temperature is moderate. However, ethanol reduces NOx output at high temperatures even though it contains more energy. At 3,000 K, 10% ethanol reduces NOx emissions by approximately 39.5% compared to the base fuel, while 10% methanol achieves a 30.1% reduction. These findings underscore ethanol\u0026apos;s superior role in suppressing NOx at high temperatures. This transition in NOx trends from 2,000 K to 3,000 K reflects a fundamental shift in dominant chemical mechanisms.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e\n \u003ch2\u003e2.1.3. Dominant reaction pathway\u003c/h2\u003e\n \u003cp\u003eTo elucidate how alcohol additives modulate nitrogen\u0026ndash;oxygen reaction networks during combustion, Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e compiles the forward and reverse frequencies of the most active NOx-associated reactions across all ten fuel systems (C1\u0026ndash;C10). The reactions were identified based on their participation in NO, NO₂, HNO, HNO₂, and HNO₃ interconversion loops, critical mechanisms governing NOx formation and removal [\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e]. As observed, the base case (C1 at 2000 K) exhibits strong engagement of NO₂ ⇌ HNO₃ and HNO₃-related formation pathways, reflecting the dominance of higher-order nitrogen oxides under lower-temperature oxidative conditions. At 3000 K (C2), the reaction network shifts toward simpler cycles such as HNO₂ ⇌ NO and NO ⇌ HNO, indicating increased fragmentation of NOx species and reduced stability of nitric acid derivatives at elevated temperatures. The addition of alcohols introduces pronounced shifts in pathway dominance and reversibility. Ethanol-blended systems (C3, C5, C4, C6) continue to feature NO₂\u0026ndash;HNO₃ cycling, particularly at 2000 K, where NO₂ ⇌ HNO₃ remains among the top reversible pathways. This suggests ethanol\u0026apos;s role in temporarily stabilizing NO₂ into higher-order oxides like HNO₃, a phenomenon that contributes to reduced free NO concentrations. At 3000 K, ethanol-enhanced systems (notably C6) show elevated activity in NO ⇌ HNO₂ and HNO₂ ⇌ NO\u0026thinsp;+\u0026thinsp;NO reactions, reinforcing the interpretation that ethanol increases cycling efficiency within intermediate species, thus delaying direct NO buildup.\u003c/p\u003e\n \u003cp\u003eIn contrast, methanol-enriched systems (C7, C9 at 2000 K and C8, C10 at 3000 K) reveal a marked shift in reaction priorities. The most frequent reactions are centred on NO ⇌ HNO and NO ⇌ HNO₂, with strong reverse reaction rates that exceed or match forward activity. This indicates methanol\u0026rsquo;s robust effect in diverting reactive nitrogen intermediates away from NO₂ production. Notably, in C10, the NO ⇌ HNO₂ reverse flux reaches 46, among the highest of all cases, highlighting methanol\u0026rsquo;s superior capacity to suppress NO₂ accumulation via intermediate stabilization. These trends are consistent with methanol\u0026rsquo;s high oxygen content and rapid radical propagation, which facilitate hydrogen transfer reactions that reduce NO directly to inert or less reactive species.\u003c/p\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDominant NOx reaction pathways and their forward and reverse frequencies across C1\u0026ndash;C10 fuel systems.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eReaction\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eSystems\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"10\"\u003e\n \u003cp\u003eForward Reaction Frequency\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC1\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC2\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC3\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC4\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC5\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC6\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC7\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC8\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC9\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC10\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO2 ⇌ HNO3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO ⇌ HNO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO ⇌ HNO2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e48\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO2 ⇌ NO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO\u0026thinsp;+\u0026thinsp;NO ⇌ HNO2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eHNO2 ⇌ NO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eHNO3 ⇌ NO2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO ⇌ N#N\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eHNO2\u0026thinsp;+\u0026thinsp;HNO ⇌ NO\u0026thinsp;+\u0026thinsp;NO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"10\"\u003e\n \u003cp\u003e\u003cstrong\u003eReverse Reaction Frequency\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO2 ⇌ HNO3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e46\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO ⇌ HNO2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e46\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO ⇌ HNO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO ⇌ NO2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eHNO3 ⇌ NO2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eHNO2 ⇌ NO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO2\u0026thinsp;+\u0026thinsp;NO ⇌ HNO2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO ⇌ N#N\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eNO ⇌ HNO2\u0026thinsp;+\u0026thinsp;HNO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eHNO2\u0026thinsp;+\u0026thinsp;HNO ⇌ NO\u0026thinsp;+\u0026thinsp;NO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e2.2. ML predictions\u003c/h2\u003e\n \u003cp\u003eML models are applied to predict NOx emissions from ammonia\u0026ndash;methane\u0026ndash;alcohol combustion based on previous MD simulation data. The goal is twofold: we first conduct model performance assessments with different architectures and then demonstrate the ability to make extrapolative predictions for off-design alcohol blends at 2%, 7%, and 12%. Four algorithms were considered in the predictions, including RFR, GBR, SVR, and FCNN, along with four distinct architectural configurations in each model. The comparison of predictions between these models allows for a deeper understanding of how algorithm design affects the predictive accuracy, stability, and feature sensitivity.\u003c/p\u003e\n \u003cp\u003eThe predictive robustness of trained ML models was assessed by comparing their outputs against MD-derived NOx values over time across ten combustion cases (C1\u0026ndash;C10) that examined different alcohol concentrations (0%, 5%, 10%) and both ethanol and methanol blends at temperatures of 2,000 K and 3,000 K. Tables \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e display a direct comparison of NOx emissions between ML model predictions and MD simulation results in the base fuel systems, C1 at 2,000 K and C2 at 3,000 K, which complement the previously discussed performance metrics and learning curves. The analysis reports ten separate time steps ranging from 50 to 500 ps to consider the complete trajectory span. The RFR model demonstrates superior accuracy through its lowest recorded average absolute and relative errors at every time step. RFR exhibited the lowest error margins, with percentage errors consistently below 1.6% at 2,000 K and under 1% for most time steps at 3,000 K. For instance, RFR predicted 86.896 ppm vs. 87.883 ppm in MD at 250 ps, reporting an absolute error of only 0.987 ppm (1.12%). FCNN also maintained strong accuracy with errors generally under 2.5%, occasionally outperforming RFR at later time steps, such as at 500 ps in C2 in Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, where it only had an error of 0.14 ppm. SVR and GBR showed good performance but displayed minor over- or under-predictions of NOx during longer reaction periods (e.g., 300\u0026ndash;500 ps) which aligns with findings from predicted\u0026ndash;actual plots (Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e in Extended data) and sensitivity analysis (Fig. 6 in Extended data). For example, at 300 ps in C1, the reported error by SVR reached 5.707 ppm (6.15%), which was closely followed by GBR. This quantitative evaluation supports the observed trends from the performance plots specifically the actual vs. predicted correlations in (Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e in Extended data), the learning curves in (Fig. 5 in Extended data), and the NOx evolution patterns predicted by the four ML models, reinforce the established performance hierarchy: RFR\u0026thinsp;\u0026gt;\u0026thinsp;FCNN2\u0026thinsp;\u0026gt;\u0026thinsp;GBR4\u0026thinsp;\u0026asymp;\u0026thinsp;SVR1, with RFR2 and FCNN2 emerging as the most effective configurations. While the results align well with the behaviour observed in physics-based simulations, this consistency suggests that ML models, particularly ensemble approaches can capture key patterns in NOx evolution within ammonia\u0026ndash;methane combustion systems when trained appropriately.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e Comparison of NOx (ppm) emissions between ML model predictions and MD simulation results at 2,000 K (System C1, 0% Alcohol) and 3,000 K (System C2, 0% Alcohol).\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1759989803.png\" width=\"1173\" height=\"618\"\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e2.3. Extrapolative predictions for untrained alcohol ratios\u003c/h2\u003e\n \u003cp\u003ePerforming extrapolative prediction is crucial for real-world combustion research, as it enables low-cost virtual screening of fuel blends without requiring additional expensive and time-consuming simulations. It reflects the model\u0026rsquo;s capacity to generalize beyond known data while remaining chemically consistent. Figure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e shows the ML-predicted NOx profiles for untrained alcohol concentrations, e.g., 2%, 7%, and 12% of both ethanol and methanol at 2,000 K and 3,000 K. These systems (C11\u0026ndash;C22) were not included in the original training data, meaning they represent a direct test of the model\u0026rsquo;s extrapolation capability. At 2,000 K in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e(a), C17 (2% methanol) yields the highest NOx production, surpassing 110 ppm at 500 ps. In contrast, C15 (2% ethanol) closely tracks the baseline behaviour of C1, stabilising around 90\u0026ndash;95 ppm. As concentration increases, this disparity narrows: C21 (7% methanol) maintains higher NOx than C19 (7% ethanol), while at 12% concentration, both methanol in C13 and ethanol in C11 converge around 95 ppm. These results align with earlier MD findings, reaffirming that methanol\u0026rsquo;s aggressive early oxidation at lower temperatures fosters NOx, while ethanol moderates the release of N-containing radicals. The trend shifts greatly at 3,000 K in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e(b). All systems exhibit reduced NOx formation, but now ethanol-rich blends dominate the suppression. C12 (12% ethanol) falls below 70 ppm in the end, marking the lowest NOx level among all tested extrapolated cases. Meanwhile, C14 (12% methanol) levels off just under 80 ppm. Across mid-level concentrations, C20 (7% ethanol) also outperforms C22 (7% methanol). The low-concentration region again favours ethanol: C16 (2% ethanol) marginally outperforms C18 (2% methanol).\u003c/p\u003e\n \u003cp\u003eThese trends reinforce two interpretations of the underlying chemical reaction mechanisms. First, methanol is more reactive at low temperatures, likely accelerating NH\u003csub\u003e3\u003c/sub\u003e oxidation through earlier radical generation, which in turn drives NO formation. However, as temperature increases, ethanol\u0026rsquo;s slower and more controlled radical propagation favours routes that stabilise intermediates or divert nitrogen away from NO/NO\u003csub\u003e2\u003c/sub\u003e loops. Second, the model\u0026rsquo;s ability to recover these patterns, despite having been trained solely on 0%, 5%, and 10% alcohol ratios, demonstrates the capacity of chemical-guided ML.\u003c/p\u003e\n \u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e compares the predicted values with synthesised ReaxFF MD datasets to validate the extrapolation results for 12% alcohol scenarios. These datasets were predicted for C11, C13 at 2,000 K and C12, C14 at 3,000 K, representing 12% ethanol and 12% methanol, respectively. The results are summarised in terms of absolute and percentage errors for each 100 ps period across 500 ps. The average percentage error for ethanol is relatively low, increasing from 3.845% at 100 ps, to 4.49% at 500 ps. The prediction remains stable, but the trend shows an incremental increase in error with time, reflecting the complexity of extrapolation. The percentage error begins at 2.02% when measured at 100 ps and increases to 12.78% at 400 ps. The rising error trend arises from methanol\u0026apos;s increased reactivity, leading to more significant departures from MD results. The percentage error for 3,000 K cases begins at 6.58% at 100 ps and decreases to 5.97% at 500 ps. It suggests the model\u0026apos;s predictions stabilize as the combustion dynamics approach the steady state at higher temperatures. In addition, a similar error pattern with methanol is observed. The error starts at 3.26% at 100 ps but increases significantly to 17.74% at 400 ps. The substantial error rise during intermediate timesteps underscores the difficulties in extrapolating results for high methanol concentrations and elevated temperatures. It can be concluded that increased prediction errors in methanol are due to its stronger reactivity that creates complex reaction pathways, making predictive process more challenging for ML models.\u003c/p\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003ctable id=\"Tab3\" border=\"1\" class=\"fr-table-selection-hover\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eQuantitative comparison of RFR model predictions for 12% alcohol concentrations at 2,000 K (cases C11 and C13) and 3,000 K (cases C12 and C14).\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eOutput\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eSystem\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eTime (ps)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAlc. %\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eAlc. type\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMD\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eAbs. error\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003ePercentage error\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\" rowspan=\"10\"\u003e\n \u003cp\u003e2000 K\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" rowspan=\"10\"\u003e\n \u003cp\u003eNOx (ppm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e91.061\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e94.908\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e3.847\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e4.225\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e200\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e105.141\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e98.170\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e6.971\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e6.630\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e300\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e109.811\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e99.385\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e10.426\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e9.494\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e400\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e101.188\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e96.625\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e4.563\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e4.509\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e83.330\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e87.818\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e4.488\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e5.386\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e87.706\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e89.726\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e2.020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e2.303\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e200\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e88.135\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e95.456\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e7.321\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e8.306\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e300\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e101.925\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e103.652\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e1.727\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e1.695\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e400\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e87.059\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e99.838\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e12.779\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e14.679\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e88.026\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e92.682\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e4.656\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e5.289\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"10\"\u003e\n \u003cp\u003e3000 K\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" rowspan=\"10\"\u003e\n \u003cp\u003eNOx (ppm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e91.837\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e98.422\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e6.584\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7.169\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e200\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e70.915\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e82.390\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e11.475\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e16.181\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e300\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e69.590\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e76.338\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e6.749\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9.698\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e400\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e70.112\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e74.099\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e3.987\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5.687\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eEthanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e60.390\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e66.358\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e5.968\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9.883\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e97.138\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e100.400\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e3.262\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3.359\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e200\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e83.728\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e90.587\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e6.860\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8.193\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e300\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e74.588\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e83.787\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e9.199\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12.333\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e400\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e65.553\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e77.181\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e11.628\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e17.738\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003e12%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMethanol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e71.582\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e74.921\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e3.339\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4.664\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003cp\u003eWhile the extrapolation to 12% alcohol is successfully achieved with acceptable accuracy, the extrapolation to higher ratios of alcohol (e.g., 15\u0026ndash;20%) is difficult. The increased reactivity and complexity of chemical pathways at higher concentrations make it difficult to maintain high predictive accuracy. When the alcohol ratio is increased to 15\u0026ndash;20%, mechanisms of the fuel and air mixture and reaction can change significantly, leading to big differences between the ML predicted results and the actual combustion outcomes. This limitation likely stems from a lack of representative training samples capturing the altered charge distributions, energy profiles, and species interaction dynamics unique to these higher alcohol blends. Further work is needed to enhance the extrapolation process, particularly to 15\u0026ndash;20% alcohol levels, which may need more training data or better feature chemical reactions to capture non-linear effects at those levels. These non-linearities arise from threshold-driven changes in radical formation, bond dissociation patterns, and secondary reactions that do not scale linearly with alcohol concentration. For example, higher hydroxyl content at \u0026gt;\u0026thinsp;15% alcohol may shift dominant NOx formation pathways or trigger new intermediate stabilization loops not seen at lower concentrations.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"3. Conclusions","content":"\u003cp\u003eThis study employed ReaxFF-MD simulations to assess the effects of alcohol additives in ammonia\u0026ndash;methane combustion. Alcohol enrichment significantly altered the charge redistribution, intermediate stability, and NOx pathways. At 2,000 K, methanol promoted electron withdrawal and increased NOx formation, whereas ethanol suppressed emissions. At 3,000 K, 10% ethanol showed the most significant effect, a 39.5% decrease of NOx compared to the base case, while methanol produced a more moderate reduction. Mechanistic analysis revealed distinct pathways: ethanol favoured NO\u003csub\u003e2\u003c/sub\u003e-HNO\u003csub\u003e3\u003c/sub\u003e loops and nitrate decomposition, while methanol promoted HNO- and N\u003csub\u003e2\u003c/sub\u003eO-mediated conversion.\u003c/p\u003e\u003cp\u003eML models trained on atomistic descriptors confirmed the predictive capability of the hybrid framework. Among the tested models, RFR2 model achieved the highest accuracy, FCNN2 ranked second, GBR4 exhibited moderate performance, and SVR1 was the weakest. Importantly, the RFR model reliably extrapolated to unseen alcohol ratios up to 12%, maintaining prediction errors below 5% in ethanol-rich mixtures. Overall, the integrated MD-ML framework provides both mechanistic insight and strong predictive capability, enabling scalable exploration of ammonia\u0026ndash;methane\u0026ndash;alcohol blends for cleaner combustion.\u003c/p\u003e\u003cp\u003eFuture work will aim to enhance the extrapolative robustness of the ML models by integrating noise-reduced, chemically valid training data [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] and incorporating reaction-aware physical priors [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] to better capture transition dynamics. In addition, emerging frameworks such as optimal transport for reaction state mapping [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e], physics-informed neural networks (PINN) for reaction\u0026ndash;diffusion systems [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], and language-model-driven chemical embedding techniques [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] are alternatives to generalize the current approach toward broader chemical regimes, including transition state estimation and real-time reactive force field tuning [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e"},{"header":"4. Online Methods","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e4.1. Reactive force field molecular dynamics (ReaxFF MD)\u003c/h2\u003e\u003cp\u003eReactive force field (ReaxFF) molecular simulations provide a robust framework for modelling chemical reactions at the atomic scale by dynamically accounting for bond formation and breaking. This modelling technique bridges the gap between quantum mechanical computations, which are precise but computationally difficult, and classical molecular dynamics (MD) simulations, which are effective but non-reactive. ReaxFF enables the modelling of extensive systems under extreme pressures and temperatures with a high level of precision, facilitating the physical capture of complex chemical reactions. It has been widely applied in combustion studies, offering valuable insights into reaction mechanisms, intermediate species formation, and product distributions [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The ReaxFF can be expresses as a function of the bond order:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{E}_{system}={E}_{bond}+{E}_{over}+{E}_{under}+{E}_{lp}+{E}_{angle}+{E}_{tors}+{E}_{vdW}+{E}_{Coul}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe total potential energy in ReaxFF is represented as the sum of various energy contributions, including bond energy, penalties for over- and under-coordination, lone-pair stabilization, valence and torsional angle energies, and non-bonded interactions such as Coulombic and van der Waals forces [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. By parameterizing these energy terms based on quantum mechanical calculations, ReaxFF can simulate chemical processes in large systems with both computational efficiency and chemical accuracy.\u003c/p\u003e\u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\u003ch2\u003e4.1.1. System configuration\u003c/h2\u003e\u003cp\u003eTo investigate the influence of alcohol additives on ammonia\u0026ndash;methane combustion, ten computational cases (C1\u0026ndash;C10) were configured. Each case employed an equivalence ratio of λ\u0026thinsp;=\u0026thinsp;0.7, a moderately fuel‑rich condition typical of practical combustors where perfect stoichiometry (λ\u0026thinsp;=\u0026thinsp;1) is rarely achieved. Operating under fuel‑rich conditions promotes the build‑up of reactive intermediates and enables detailed examination of NOx formation pathways which is the primary objective of this study. For physical consistency, the mass density was fixed at 0.34 g/cm\u0026sup3; in every simulation. Although the total number of molecular remained constant at 800, the cubic domain length was adjusted case‑by‑case to satisfy the density constraint, ensuring that any changes in system behaviour originate solely from variations in fuel composition or temperature. The baseline mixture comprised 422 fuel molecules\u0026mdash;methane (CH₄) and ammonia (NH₃) in a 1:1 ratio\u0026mdash;together with 378 oxygen molecules. Alcohol co‑fuels, ethanol (C₂H₆O) or methanol (CH₄O), were introduced by substituting 5% or 10% of the original CH₄-NH₃ base fuel while preserving the overall molecule count. Distinct C\u0026ndash;H bond energies and oxygen contents in selected alcohols are expected to affect their radical pools and chain‑branching behaviour. Each composition was evaluated at 2000 K and 3000 K to capture temperature‑dependent kinetics.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e4.2. Machine learning framework\u003c/h2\u003e\u003cdiv id=\"Sec14\" class=\"Section3\"\u003e\u003ch2\u003e4.2.1. Data structure and preprocessing\u003c/h2\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e shows the data processing pipeline that links MD simulations to ML models applied in this paper. ReaxFF-based MD simulations created the baseline dataset from simulations performed at temperatures 2,000 K and 3,000 K for 10 chemical cases, which are labelled C1 to C10. Each case simulated a 500 ps reactive trajectory and produced a series of thermochemical and molecular descriptors, which were designed to characterise the evolving behaviour of ammonia\u0026ndash;methane\u0026ndash;alcohol mixtures. The resulting dataset includes 26 input features (\u003cem\u003eX\u003c/em\u003e), encompassing key thermophysical observables such as temperature, pressure, and density, as well as energy terms including total, kinetic, and potential energies. In addition, features derived from the ReaxFF were included to represent mechanistic quantities of the direct chemical relevance. These consist of individual energy contributions: v_ea (atomic), v_eb (bond), v_elp (lone pair), v_ev (valence angle), v_epen (penalty), v_ecoa (conjugation), v_ehb (hydrogen bond), v_et (torsion), v_eco (conjugated), v_ew (van der Waals), v_ep (Coulomb), and v_eqeq (charge equilibration). Element-specific charge distribution properties were also computed: v_qC, v_qH, v_qN, and v_qO, representing the mean partial charge for each atom type, which are essential for capturing electron redistribution during combustion. Chemical identity was represented by the mass fraction of alcohol feature, which distinguishes 0%, 5%, and 10% mixed fuels, and binary indicators for ethanol and methanol, to account for molecular structure effects. These enable the models to learn not only compositional trends but also the functional impact of hydroxyl content and C\u0026ndash;H/O\u0026ndash;H bond energetics on various alcohol additives. A rigorous data cleaning step was employed prior to training the models. This included the detection and removal of anomalous entries owing to rare bond breakage events, unphysical charge spikes, and species mislabelling during dynamic equilibration. Outliers were eliminated using interquartile-based cutoff values, and no missing values were encountered, as molecular dynamics simulations deterministically generate complete trajectories with time-resolved data for all atoms and properties. Preprocessing was followed by 80:20 (e.g., 80% of data used for training and 20% for validation) stratified train-test splitting in a manner that ensured proportional representation of all the compositions and temperatures within the training and testing sets [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section3\"\u003e\u003ch2\u003e4.2.2. ML algorithms\u003c/h2\u003e\u003cp\u003eFour ML algorithms were chosen in this work to model the intricate interaction between combustion descriptors and NOx emissions: RFR, GBR, FCNN, and SVR. Their capacity to capture non-linear dependencies, resilience against overfitting and proven success in thermochemical prediction tasks guided selection of these models. All the algorithms have different strengths that are well-suited to the nature of combustion-related data. RFR, as an ensemble decision tree, performs exceptionally well with high-dimensional data and produces variable importance information, which is critical in understanding the influence of physical parameters. GBR augments the same by incorporating gradient-based optimization in the construction of trees and thereby is well suited to pick up weak interactions in noisy or structured data. FCNNs, on the other hand, are capable of learning intricate, continuous functions due to their hierarchical representation power, particularly useful in tracking multi-variable chemistries' dynamics. SVR becomes a standard regression benchmark because it excels with non-linear data by using the kernel-based similarity functions to capture non-linearity. Four distinct versions of each algorithm were created to explore the effect of design choices on their performance by adjusting internal parameters. To evaluate the robustness and generalizability of tree-based algorithms (i. e., RFR, GBR), a systematic approach was taken to adjust both tree count and depth parameters. Neural networks (NN) configurations included various hidden layer counts and neuron distributions ranging from basic one or two layered networks to deep structures with up to five hidden layers. SVR configurations were altered by experimenting with different kernel functions (linear, polynomial, or radial basis) and regularisation parameters.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e4.3. Chemical-guided extrapolation\u003c/h2\u003e\u003cp\u003eTo extend the predictive capability of different ML models beyond the training range, a chemical-guided extrapolation strategy was employed. This approach aimed to infer NOx emissions for untrained alcohol ratios, specifically 2%, 7%, and 12%, by leveraging both data-driven models and domain-specific chemical insights derived from ReaxFF simulations. As illustrated on the right panel of Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the post-processing workflow builds upon the trained model\u0026rsquo;s outputs at known alcohol concentrations (0%, 5%, and 10%). The key idea was to avoid the need for new MD simulations at unseen compositions, and instead to estimate the target variables by interpolating or extrapolating trends observed in the training data. This hybridisation of ML and physical chemistry offers a practical route for off-design predictions without incurring the computational cost of generating new datasets. Specifically, for in-range predictions such as 2% and 7%, a linear interpolation scheme was adopted between the adjacent training points (e.g., 0%\u0026ndash;5%, 5%\u0026ndash;10%). This approach assumes smooth transitions in NOx behaviour across small alcohol increments, which was consistent with the gradual trends captured by ReaxFF simulations. For out-of-range prediction at 12%, which is a value not seen during model training, a scaling-based extrapolation technique was applied. This was informed by the observed gradient between the 5% and 10% alcohol systems and adjusted proportionally beyond the upper bound. To mathematically implement this extrapolation, a scaling coefficient is introduced and is expressed below:\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:{{Feature}_{12\\%}}^{\\:}={{Feature}_{10\\%}+\\left(\\frac{12-10}{10-5}\\right)\\times\\:\\left({{Feature}_{10\\%}-{{Feature}_{5\\%}}^{\\:}}^{\\:}\\right)}^{\\:}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThis transformation was applied independently to each input feature before feeding it through the trained models, allowing the prediction of NOx emissions for novel alcohol levels. This method is described in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e as post-processing step to generate unseen alcohol ratios, and it bridges the gap between chemical-guided trends and data-driven modelling. The underlying rationale for this strategy is based on the smooth, thermochemically governed nature of alcohol substitution in reactive systems. Then, ML models, which were trained on clean, labelled inputs, were able to capture the nonlinear coupling between chemical features and NOx formation, reinforcing the robustness of predictions. By combining these two pillars, e.g., data-driven learning and chemically grounded extrapolation, the hybrid method demonstrated in this study offers a reliable alternative to conventional simulation-heavy routes for combustion design. This novel approach enables fast, cost-effective insights into emission reduction under off-design conditions, showcasing a vital step towards adaptive, intelligent optimisation in combustion chemistry.\u003c/p\u003e\u003c/div\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eL\u0026uuml; M, Long W, Wei F, Dong D, Li C, Dong P, Tian H, Chen X, Chen S, Wang Y, Wang P (2024) Assessment of carbon-free fuel ammonia combustion with low methanol blends in reducing GHG emissions including N2O. J Clean Prod 463:142755\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eL\u0026uuml; M, Long W, Wei F, Dong D, Li C, Dong P, Tian H, Chen X, Chen S, Wang Y, Wang P (2024) Assessment of carbon-free fuel ammonia combustion with low methanol blends in reducing GHG emissions including N2O. J Clean Prod 463:142755\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuo Y, Shi H, Liu H, Xie Y, Guan Y (2023) Reactive molecular dynamics simulation and chemical kinetic modeling of ammonia/methane co-combustion. \u003cem\u003eFuel\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShateri A, Jalili B, Saffar S, Jalili P, Ganji DD (2023) Numerical study of the effect of ultrasound waves on the turbulent flow with chemical reaction. Energy\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSingh AS, Vijrumbana Y, Reddy VM (2023) Experimental and computational (Chemical Kinetic\u0026thinsp;+\u0026thinsp;CFD) analyses of Self-Recuperative annular tubular porous burner for NH3/CH4 -air Non-Premixed combustion. Chem Eng J\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChu X, Gao XL, Ma P, Xiao Z, Xie H, Zhang C, Z., Wang X (2024) High-temperature auto-ignition characteristics of NH3-H2-CH4. Fuel 365:131228\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYu M, Zhan J-H, Li X, He W, Liu X (2023) Effect of methanol on the pyrolysis behaviour of kerogen by ReaxFF molecular dynamics simulations. \u003cem\u003eMolecular Simulation\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShateri A, Yang Z, Xie J (2025) Machine Learning-Based Molecular Dynamics Studies on Predicting Thermophysical Properties of Ethanol\u0026ndash;Octane Blends. Energy \u0026amp; Fuels\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMao G-P, Shi T, Mao C, Wang P (2023) Prediction of NOx emission from two-stage combustion of NH3\u0026ndash;H2 mixtures under various conditions using artificial neural networks. Int J Hydrog Energy\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHoque KE, Hossain T, Haque AM, Miah MAK, Haque MA (2024) NOx Emission Predictions in Gas Turbines through Integrated Data-Driven Machine Learning Approaches. J Energy Resour Technology-Transactions Asme\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePachauri N (2024) An emission predictive system for CO and NOx from gas turbine based on ensemble machine learning approach. Fuel\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAdams D, Oh DH, Kim DW, Lee CH, Oh M (2020) Prediction of SOx\u0026ndash;NOx emission from a coal-fired CFB power plant with machine learning: Plant data learned by deep neural network and least square support vector machine. J Clean Prod 270:122310\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShateri A, Yang Z, Xie J (2024) Utilizing Artificial intelligence to identify an Optimal Machine learning model for predicting fuel consumption in Diesel engines. Energy and AI\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMao Q, Yang MD, Wei C, Liu X, Jiang X, Ren Y, Luo KH, van Duin AC T (2023) Classical and reactive molecular dynamics: Principles and applications in combustion and energy systems. Prog Energy Combust Sci 97:101084\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChowdhury A, van Duin ACT (2017) Extension of the ReaxFF Combustion Force Field toward Syngas Combustion and Initial Oxidation Kinetics. J Phys Chem A 121(5):1051\u0026ndash;1068\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003evan Duin ACT, Dasgupta S, Lorant F, Goddard WA (2001) ReaxFF: A Reactive Force Field for Hydrocarbons. J Phys Chem A 105(41):9396\u0026ndash;9409\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIhme M, Chung WT, Mishra A (2022) Combustion machine learning: Principles, progress and prospects. Prog Energy Combust Sci 91:101010\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZapiain DM, de Venkatraman O, A., Wilson M (2024) \u003cem\u003eAccelerating Charge Estimation in Molecular Dynamics Simulations Using Physics-Informed Neural Networks: Corrosion Applications\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBorodin A, H\u0026ouml;fft O, Kempter V, Ferro Y, Allouche A-R (2004) Electron delocalization by polar molecules: interaction of Na atoms with solid ammonia films studied with MIES and density functional theory. J Chem Phys 121(8):3717\u0026ndash;3721\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWatson GMG, Versailles P, Bergthorson JM (2017) NO formation in rich premixed flames of C1\u0026ndash;C4 alkanes and alcohols. 36(1):627\u0026ndash;635\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGao Y, Guo Y, Xie Y, Qin H, Guan Y (2024) Kinetic mechanism and NOx emission characteristics of the ammonia/alcohol co-combustion explored by reactive molecular dynamics calculation and kinetic numerical simulation. Combust Flame 273:113953\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eToniato A, Schwaller P, Cardinale A, Geluykens J, Laino T (2021) Unassisted noise reduction of chemical reaction datasets. Nat Mach Intell 3(6):485\u0026ndash;494\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRao C, Ren P, Wang Q, Buyukozturk O, Sun H, Liu Y (2023) Encoding physics to learn reaction\u0026ndash;diffusion processes. Nat Mach Intell 5(7):765\u0026ndash;779\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDuan, C., Liu, G. H., Du, Y., Chen, T., Zhao, Q., Jia, H., \u0026hellip; \u0026amp; Kulik, H. J. (2025). Optimal transport for generating transition states in chemical reactions. Nature Machine Intelligence, 7(4), 615\u0026ndash;626\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJablonka KM, Schwaller P, Ortega-Guerrero A, Smit B (2024) Leveraging large language models for predictive chemistry. Nat Mach Intell 6(2):161\u0026ndash;169\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGong, S., Zhang, Y., Mu, Z., Pu, Z., Wang, H., Han, X., \u0026hellip; Xiang, L. (2025). A predictive machine learning force-field framework for liquid electrolyte development. Nature Machine Intelligence, 1\u0026ndash;10\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShateri A, Yang Z, Sherkat N, Xie J (2026) Alcohol additives to enhance ammonia-methane combustion efficiency and reduce emissions: A reactive force field analysis. Fuel 405:136565\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"University of Derby","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"NOx emission, Ammonia–methane combustion, Alcohol additives, Machine learning, Reactive force field, Molecular dynamics","lastPublishedDoi":"10.21203/rs.3.rs-7800991/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7800991/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAs carbon-free energy carrier, combustion of ammonia suffers from low burning rates, poor flame stability, and excessive nitrogen oxide (NOx) emissions. Although blending with methane alleviates some drawbacks, NOx formation remains a critical barrier. To address these challenges, we propose a hybrid framework combining reactive force field molecular dynamics (MD) simulations with machine learning (ML). MD simulations at 2,000\u0026ndash;3,000 K were performed for ammonia-methane blended combustion with 0\u0026ndash;10% addition of ethanol or methanol. Adding alcohols suppressed the NOx formation by altering charge redistributions and redirecting nitrogen intermediates into stabilising pathways. At 3,000 K, 10% ethanol and methanol reduced NOx by ~\u0026thinsp;39.6% and ~\u0026thinsp;30.1%, respectively. Both chemical and physical descriptors derived from MD were used to train ML models and extrapolated to unseen conditions (\u0026gt;\u0026thinsp;10% alcohol) with \u0026lt;\u0026thinsp;5% error in ethanol-rich mixtures. This framework reduces reliance on costly simulations while providing mechanistic insights and predictive capability of designing alternative fuels.\u003c/p\u003e","manuscriptTitle":"Chemical-Guided Machine Learning Framework for Reactive Simulations of NOx Mitigation in Alcohol-Enhanced Ammonia–Methane Combustion","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-09 06:11:36","doi":"10.21203/rs.3.rs-7800991/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c0d82e7f-0e66-4138-ae08-d782f53a85ac","owner":[],"postedDate":"October 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":56000978,"name":"Chemical Engineering"}],"tags":[],"updatedAt":"2025-10-09T06:11:36+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-09 06:11:36","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7800991","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7800991","identity":"rs-7800991","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}