Foresight Scenario Modelling for Global Wheat Production Under Shared Socioeconomic Pathways: Testing AI Emulators of CGE Models

Foresight Scenario Modelling for Global Wheat Production Under Shared Socioeconomic Pathways: Testing AI Emulators of CGE Models | 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 Article Foresight Scenario Modelling for Global Wheat Production Under Shared Socioeconomic Pathways: Testing AI Emulators of CGE Models Xinxin Wang, Saeed Moghayer, Daan Korporaal, Maryam Hajialibeigi, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9405456/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 While Computable General Equilibrium (CGE) models such as MAGNET are very useful to informing robust policy decisions, their high computational demands often hinder the exploration of large ensembles of future scenarios. Although artificial intelligence (AI) has been applied to biophysical crop modelling, there is limited evidence regarding its capacity to accurately replicate the core behavioral logic of economic equilibrium models. To address this methodological gap, this study evaluates whether a data-driven AI emulator can serve as a robust surrogate for a structural CGE model, using global wheat production as an illustration application case. We trained the emulator using a historical panel dataset derived from MAGNET, incorporating a set of key macroeconomic drivers, and benchmarked its performance against MAGNET projections under diverse Shared Socioeconomic Pathways (SSPs). The results indicate that the AI emulator reproduces historical production patterns with high accuracy and closely replicates the MAGNET model’s behavior in near term scenario projections. However, distinct divergences emerge in long-term projections, where the AI model follows historical constraints while the CGE framework embeds forward-looking assumptions such as technical progress. Ultimately, this research demonstrates that hybrid CGE-AI frameworks can successfully complement structural economic logic with computational agility, offering a scalable methodology for rapid scenario stress-testing. Scientific community and society/Agriculture Scientific community and society/Social sciences Computable General Equilibrium (CGE) Artificial Intelligence Emulation Hybrid Modelling Shared Socioeconomic Pathways (SSPs) Economic Foresight Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Ensuring global food security under growing socio-economic pressures and escalating climate risks is one of the most pressing challenges (Lobell et al., 2008; Xuhui et al., 2020). Wheat, as one of the world’s principal staple crops (Shiferaw et al., 2013), provides a substantial share of daily caloric intake for billions of people worldwide (Carriquiry et al., 2022; Erenstein et al., 2022; Lobell et al., 2008). Projections of future wheat production are therefore central to informing robust policy decisions, designing sustainable food systems, and mitigating risks associated with climate change and market shocks (Hultgren et al., 2022; Lobell et al., 2008). Traditionally, projections of wheat production rely on well-established agro-economic models, including partial and computable general equilibrium (CGE) frameworks such as MAGNET(Modular Applied GeNeral Equilibrium Tool) (Moghayer et al., 2023; van Meijl et al., 2020). These structural models explicitly embed production functions, supply-demand linkages, trade balances, and policy feedbacks, allowing researchers and policymakers to explore how macro drivers—such as Gross domestic product (GDP) growth, population dynamics, and policy interventions—translate into production outcomes and trade flows under various scenarios, including the widely used Shared Socioeconomic Pathways (SSPs)(O’Neill et al. 2017). While robust and theoretically consistent, these models are computationally intensive and time-consuming to run, especially when exploring large ensembles of policy or climate scenarios to quantify uncertainty and stress-test resilience. As noted by van Meijl et al. (2020), running MAGNET scenarios under diverse SSPs requires full equilibrium recalibration, which is computationally intensive when testing alternative futures. Similarly, Antle et al. (2017) emphasize that multi-scenario CGE applications, particularly under climate uncertainty, require substantial computational effort due to the need for repeated model convergence and calibration. In addition, CGE models like MAGNET are built on large nonlinear systems calibrated to social accounting matrices and solved under equilibrium constraints. Although they capture key nonlinear economic mechanisms, their functional forms and parameterizations may not fully reflect the complex, high-dimensional nonlinear relationships observed in empirical data. Advances in artificial intelligence methods such as machine learning offer a promising complementary pathway to accelerate foresight and scenario exploration in global food systems. Machine learning is particularly well-suited for learning highly nonlinear relationships from data, capturing interactions that are difficult to encode in structural models, and generating large numbers of simulations at negligible computational cost. While there has been recent interest in using machine learning to emulate biophysical crop models or climate impact modules (Eyring et al., 2024; Mansfield et al., 2020; Wang et al., 2022), there remains limited evidence on whether machine learning can accurately replicate the core behavioral logic of equilibrium agro-economic systems for staple crops under realistic socio-economic drivers. In this study, we investigated whether machine learning approach can serve as a robust surrogate for a structural CGE-style model of global wheat production. Using a panel dataset that captures the nonlinear relationships among essential economic drivers, such as GDP, population, harvested area, prices, and trade flows, and cross-crop substitution effects, we used machine learning to emulate and reproduce wheat production outcomes consistent with theoretical supply, demand, and trade relationships. Beyond benchmarking predictive performance under historical and short-term conditions, we evaluated the emulator’s behaviour under Shared Socioeconomic Pathway (SSP) scenarios, comparing AI-based projections with those generated by the MAGNET CGE model to assess robustness across contrasting long-term development trajectories. The emulator’s performance, robustness under pseudo-unseen scenarios, and interpretability are further examined using explainable artificial intelligence tools (van der Velden, 2024). The aim of this study was to assess the extent to which data-driven AI emulators can replicate the behaviour and policy-relevant outcomes of structurally grounded CGE models in global wheat production scenarios under alternative socio-economic and climate futures. In particular, the study aimed to (i) assess the emulator’s ability to reproduce historical and near-term CGE-based wheat production outcomes, (ii) examine divergences between AI-based and CGE-based projections under alternative SSP scenarios, and (iii) demonstrate how explainable AI tools can be used to interpret emulator behaviour in a manner that is consistent with established economic reasoning. By pursuing these aims, the study clarified when and how AI-based emulators can complement structural economic models in short term and long term SSP-based food security foresight, while delineating the boundaries of their applicability for policy-relevant analysis 2. Results This section reports the outcomes of the MAGNET and AI emulator, focusing on changes in wheat production across scenarios. Results are presented at both global and regional scales. 2. 1 AI Model performance: The AI model achieved a root mean square error (RMSE) between actual and predicted wheat production of 1,301 thousand tons and a high coefficient of determination (R² = 0.99) under five-fold cross-validation, indicating strong generalization within the training dataset (Table 1). On the test set, the RMSE was 1603 with R 2 = 0.98, and on the 2019 data, RMSE was 1,301 with R² = 0.99. This closely matches the predictive accuracy of the MAGNET model for the 2019 (RMSE = 1,012; R² = 0.99, Table 2). Table 1 : AI Model Performance Metrics Model Dataset RMSE R 2 Score AI Train 989 0.99 Test 1603 0.98 2019 1301 0.99 MAGNET 2019 1012 0.99 Explainable AI showed that export volume, production of other grains, and total harvested area were the top three predictors (Figure 1). These variables capture both direct and indirect production dynamics: higher export capacity and larger multi-crop production areas signal more robust agricultural systems that likely support greater wheat yields. The magnitude of the SHAP values for export indicates that high export volumes are consistently associated with substantially higher predicted wheat production, aligning with the economic principle that surplus-oriented producers scale output beyond domestic needs. Similarly, harvested area exhibits strong positive SHAP contributions, reflecting the role of land allocation in sustaining wheat productivity. Socioeconomic variables also exert measurable impacts: Population and GDP show consistently positive SHAP values, suggesting that countries with larger populations and stronger economies tend to have higher predicted wheat production, likely due to their capacity to mobilize agricultural inputs, infrastructure, and technology. Notably, specific country indicators (e.g., the United States of America, the United Kingdom, Spain) also appear among the top features, indicating persistent regional elements—such as institutional, climatic, or policy environments. This unobserved heterogeneity acts similarly to the region-specific productivity parameters or fixed effects commonly used in panel data econometrics and CGE calibration. Collectively, the SHAP analysis confirms that the AI model relies on a coherent blend of agronomic, economic, and geographic determinants to accurately estimate wheat production across diverse national contexts. Partial dependence and ICE analyses were conducted to interpret how the most important predictors shape the AI model’s estimation of national wheat production (Figure 2). The partial dependence plots (dashed lines) represent the average model-implied response of wheat production to a given predictor, holding the joint distribution of other variables constant. The ICE curves (light blue lines) show the corresponding response for individual countries, thereby revealing cross-country heterogeneity around the average trend. For the top three predictors (Export, Other grains production, and Harvested area), the plots indicate clear, positive relationships with distinct curve shapes. The partial dependence plot for Export shows a sharp, near-linear increase in predicted production as export volume rises, particularly up to ~15,000 units(1000 tone per unit), after which the curve begins to level off. This pattern suggests a strong positive association between export capacity and production, but also a saturation effect where additional exports yield diminishing marginal gains. For Other grains production, the relationship is initially steep but flattens beyond ~50,000–100,000 units(1000 tone per unit), indicating that diversification into other cereals supports wheat output up to a point, after which production synergies stabilize. Similarly, Harvested area displays a steadily increasing but concave relationship, where gains in predicted production become less pronounced once cultivated area surpasses ~5,000–7,000 units (1000 ha per unit), reflecting land constraints and diminishing returns to expansion. For the next three predictors (Other grains export, Population, and GDP), the partial dependence plots reveal more modest and gradual effects. Other grains export contributes positively but with a shallower slope, suggesting limited marginal gains beyond initial levels of diversification. Population shows a near-linear but relatively flat increase, implying that while larger populations support higher production, the effect is incremental rather than transformative. GDP demonstrates a positive but plateauing trend: wealthier economies generally sustain higher predicted wheat production, but the slope diminishes at very high income levels, reflecting that once basic infrastructure and technology thresholds are met, further gains in GDP contribute less to production growth. The spread of individual ICE curves (Figure 2) further highlights substantial heterogeneity across countries. While the average trends align with economic expectations, country-specific lines deviate in slope and curvature, illustrating how local factors, such as market integration, technological adoption, and policy environments, modulate the marginal effects of each predictor. 2.2 AI and Magnet model comparison Figure 3 compares the absolute and relative national wheat production for 2019 estimated by the MAGNET model, the AI model, and actual observed data. The AI model effectively replicates the absolute production levels across all countries, with predictions closely matching the observed values and preserving the expected rank order of major producers (Figure 4a). For example, in France, the AI prediction (37,604 thousand tonnes) is slightly below the observed value (41,083 thousand tonnes), whereas MAGNET estimates production at 38,990 thousand tonnes. In the United States, the AI model predicts 51,099 thousand tonnes, closer to the observed 52,581 thousand tonnes than MAGNET, which estimates a lower value of 48,998 thousand tonnes. The AI model and MAGNET replicated both the direction and approximate magnitude of production changes for most countries(figure 4b). Larger deviations for certain countries (e.g., Estonia, Lithuania, and Hungary) highlight areas where local conditions or data uncertainty may affect model responsiveness. For instance, in Estonia, the observed increase was +18.7%, while the AI model predicted+59.3% and MAGNET +1.8%. In Lithuania, observed production fell by –1.9%, yet the AI model predicted +35.9% and MAGNET +11.3%. Conversely, in Hungary, the observed production increased by +2.5%, the AI model predicted +17.0% and MAGNET +2.8%. Figure A4 shows outliers with large predicted changes, such as Portugal (+757% predicted vs. +25% observed) and Slovenia (+195% predicted vs. –0.8% observed) illustrating potential data noise or local anomalies where the AI model substantially overestimated production shifts. Overall, these comparisons demonstrated that the AI-based approach replicates historical production scales while providing an alternative, data-driven estimate of interannual variation that differs from the benchmark MAGNET model in several contexts. From a structural standpoint, discrepancies may reflect the machine learning model’s implicit flexibility to capture time-varying elasticities and short-run adjustment costs, unlike the MAGNET model, which is rooted in long-run equilibrium assumptions and fixed behavioral parameters. This suggests that machine learning methods can complement traditional economic simulation models by offering an independent perspective and potentially capturing variability not explicitly modelled in MAGNET, thereby enriching short-term agricultural production analyses. 2.4 AI-Magnet projection under different SSPs Figure 4 shows both modelling approaches capture the expected ranking of scenarios, with SSP3 and SSP5 producing higher outputs than SSP1. However, the magnitude of projections diverges: the MAGNET model projects strong and sustained increases in wheat production across all SSPs, exceeding 500 Mt by 2100 in SSP5. In contrast, the AI-based model predicts a flatter trajectory, stabilizing between 180–270 Mt across scenarios. Even under SSP2 and SSP3, where both models anticipate growth, MAGNET’s projections are consistently 100,000–150,000 thousand tons higher than the AI-based estimates by the end of the century. This divergence reflects the differing foundations of the two approaches. MAGNET embeds forward-looking assumptions about technological progress, policy-driven yield improvements, and long-run equilibrium adjustments, which amplify projected output growth. In contrast, the AI model is constrained by historical patterns in its input variables (harvested area, prices, imports, exports, GDP, population, and interactions with rice and other grains), and thus provides a more conservative outlook that reflects socio-economic and trade dynamics rather than speculative yield gains. From a foresight perspective, the discrepancy between the models underscores the uncertainty inherent in long-term agricultural projections. On one hand, MAGNET illustrates the potential of optimistic pathways driven by technological innovation and global trade integration. On the other, the AI-based model offers a grounded, historically consistent trajectory that highlights structural limits to growth. Figure 5 illustrates country-level projections of wheat production in 2050 under alternative SSP scenarios, comparing outcomes from the MAGNET model with those from the AI-based emulator. In panel (a), both models reproduce the relative ranking of major producers, but the absolute magnitudes diverge. For instance, in the United States, MAGNET projects 55,092 thousand tonnes under SSP1, whereas the AI model estimates only 39,975 thousand tonnes—a shortfall of more than 15,000 tonnes relative to MAGNET. In Canada, MAGNET projects 69,326 thousand tonnes under SSP1, more than double the AI estimate of 25,374 thousand tonnes. In contrast, in France, the two models produce closer results, with MAGNET projecting 35,041 thousand tonnes and the AI model predicting 26,603 thousand tonnes. Germany shows a similar pattern: MAGNET estimates 26,418 thousand tonnes versus the AI model’s 20,154 thousand tonnes. These examples highlight that MAGNET’s forward-looking productivity assumptions consistently inflate projections compared to the more conservative, historically anchored AI estimates. Panel (b) provides further insight by comparing relative changes in 2050 production against the 2017 baseline. In Western and Northern Europe, both models project stagnation or decline, but the AI model often shows sharper contractions. Conversely, in several Southern and Eastern European countries, the AI model amplified positive deviations. The country-level analysis underscores important methodological contrasts between the AI and MAGNET models. MAGNET’s equilibrium structure produces relatively smooth trajectories, dominated by assumed productivity growth and resource reallocation. By contrast, the AI model captures the imprint of historical variability, amplifying signals in countries where production has been volatile or heavily influenced by trade fluctuations. This explains the very large positive deviations observed in Portugal, Slovenia, and Italy, as well as the sharper contractions seen in Northern Europe. From a foresight perspective, these divergences highlight both the strengths and limitations of each approach. The MAGNET model offers consistency grounded in economic theory, but may overstate long-run production capacity by assuming sustained technological and productivity improvements. The AI-based emulator provides a more conservative outlook, shaped by historical evidence, but may exaggerate volatility in smaller markets. 3. Discussion In this study, we analyzed possible short-term and long-term trajectories for global wheat production by combining a structural computable general equilibrium model (MAGNET) with a data-driven AI emulator (XGBoost). Empirically, we find that the AI emulator reproduced historical wheat production with accuracy comparable to MAGNET and closely tracked observed 2019 outcomes. Under long-term SSP scenarios, the AI-based projections were more conservative than MAGNET’s, with larger differences in smaller and more volatile producing countries. By linking scenario-consistent exogenous drivers and internally consistent system responses, we showed how this hybrid approach can replicate key dynamics of a well-established agro-economic model while providing a flexible and computationally efficient tool for rapid scenario exploration. A key feature of our analysis is the focus on wheat production as a critical pillar of global food supply, benchmarked across contrasting SSP pathways. Our findings show that exogenous factors such as population growth and total GDP drive both the extensive (scale) and intensive (per capita) margins of wheat demand, while endogenous responses, including land allocation, price-mediated supply reactions, and trade equilibrium, determine regional production capacity and market outcomes. In this hybrid setting, the CGE framework ensures theoretical consistency and scenario coherence, while the emulator provides responsiveness to historically observed variability. The results also illustrate how the two modelling components play different roles across time horizons. MAGNET provides structurally grounded long-run pathways under SSP narratives, including assumptions on productivity and technical change that support sustained increases in projected output across scenarios. The emulator, trained on historical patterns in the data, produces trajectories that remain anchored in historically learned relationships. Rather than implying that one approach is “right” and the other “wrong,” these differences help clarify uncertainty in long-term agricultural projections and indicate where outcomes are sensitive to productivity pathways that cannot be validated directly from historical relationships alone. At the country level, this hybrid perspective remains important. MAGNET tends to produce smoother trajectories that follow scenario narratives and economy-wide consistency. The emulator can show larger relative movements in smaller or historically volatile producing countries, reflecting sensitivity to patterns in the historical record. This can be informative for identifying where projected outcomes are more sensitive to trade and market fluctuations, while MAGNET provides the broader equilibrium context in which such changes occur. The broader value proposition of the hybrid approach lies in its ability to bridge long-term foresight with rapid stress-testing of short-term shocks. Traditional CGE models are essential for capturing long-run equilibrium pathways under alternative futures, but they are computationally intensive when large ensembles of scenarios are required. In contrast, the emulator offers speed and flexibility and can be used to explore a wide range of counterfactuals, such as trade disruptions, policy interventions, or localized shocks, at low computational cost. This hybrid role provides theoretical grounding on one hand and operational agility on the other, enabling more timely and adaptive policy assessments in food security analysis. The integration of AI into economic modelling is consistent with broader developments in applied policy analysis. The European Commission’s Joint Research Centre (JRC) and IFPRI, for example, use AI to support scenario exploration, data management, and communication of modelling insights, while retaining structural models for theory-consistent interpretation and policy coherence. These examples underscore a shared direction: AI can provide tools for stress-testing and policy communication, while structural economic modelling remains central for internally consistent scenario assessment. Several limitations remain. First, while our approach captures supply-side and trade-driven dynamics, demand-side nutrition metrics, income distribution effects, and explicit household welfare are not endogenised in the AI component. Second, climate change impacts and extremes enter as exogenous inputs derived from scenario narratives rather than being endogenously simulated within the emulator, meaning that feedback loops and adaptive behaviour are not captured within the AI structure. Third, because the emulator is trained on historical socio-economic and trade patterns, it may underestimate the impacts of unprecedented shocks or regime changes outside the range of past experience. Finally, other machine learning algorithms or ensemble approaches could be explored; the focus here was on evaluating whether a single robust algorithm can emulate structural CGE-style outcomes. Future work can build on this proof-of-concept by developing scenario-robust, constraint-aware emulators that are iteratively refined with targeted CGE runs. Incorporating productivity and technology signals more explicitly can strengthen long-horizon emulation where divergence is driven by productivity pathways. Extending the approach beyond wheat to multi-crop systems and food-security indicators could broaden its relevance to the multidimensional challenges of global food security. Overall, the MAGNET-AI framework provides a proof-of-concept for combining structural economic logic with flexible machine learning in global food security research. The study demonstrates how an emulator can complement a structural CGE model by supporting rapid scenario exploration while retaining the role of CGE analysis for internally consistent equilibrium interpretation. In this hybrid setup, MAGNET provides scenario coherence and theory-based linkages, whereas the emulator offers a computationally efficient way to extend exploration to larger ensembles of scenario variants. Taken together, the findings support a practical implication for foresight and policy analysis. AI emulators should not be treated as replacements for CGE models, but as scalable complements that enable rapid screening and stress-testing of “what-if” scenarios while helping to identify sensitive regions and drivers. This can support more timely and adaptive assessments in food security contexts, particularly when decision makers need to explore uncertainty across multiple futures rather than focus on a small set of runs. At the same time, CGE models remain essential for analyzing welfare, trade, and broader equilibrium effects within SSP narratives and for ensuring that scenario results remain internally consistent. 4. Method This study employed an integrated modelling framework combining a global CGE model (MAGNET) with data driven machine learning to assess future wheat production under alternative socioeconomic scenarios. The methodological components and their coupling are described below. 4.1. Economic Theory and Data The design of our machine learning predictive framework is rooted in well-established principles of agricultural economics and international trade theory (details refer to supplementary B), which describe how resource endowments, price signals, and trade structures determine production and market outcomes (Gardner, 1987; Hertel, 1997; Anderson & Martin, 2005). Our primary target variable is wheat production, which is conceptualized as an outcome influenced by both domestic resource allocation and market conditions. Our selection of predictor variables reflects this theoretical structure. From a supply-side perspective, wheat production is a function of available land, which we proxy with harvested area. This represents a key resource allocation decision made by producers. We also include rice and other grain metrics to account for cross-crop competition, reflecting the opportunity cost of land and other inputs. This aligns with the economic principle that farmers allocate land to the most profitable crops, a decision often modeled in CGE frameworks using Constant Elasticity of Transformation (CET) functions. On the demand-side, we include GDP and population as standard macroeconomic drivers. These variables shape the income elasticity and scale effects on consumption, influencing the overall demand for wheat. Prices, both domestic and international, play a crucial role as market-clearing signals, balancing supply and demand. Their inclusion allows the AI model to capture producer and consumer responses to market incentives. Finally, we incorporate trade flows (imports and exports) to capture the extent to which domestic supply shortfalls or surpluses are balanced through international markets. These variables are essential for understanding how national production is integrated into the global system. By using a variable set that aligns with these theoretically grounded causal pathways, we ensure that our data-driven model can be directly compared with the structurally consistent results of the MAGNET CGE framework. These variables reflect the standard structure of general equilibrium models such as MAGNET, capturing the primary supply-side, demand-side, and trade linkages that determine national and global food production outcomes. Data for this study was derived from the MAGNET model database, which integrates multiple global datasets (van Meijl et al., 2020; Woltjer & Kuiper, 2014). The dataset includes annual records for multiple countries and years, with core variables covering harvested area, domestic wheat price, import and export volumes, GDP, population, and cross-crop information for rice and other grains (including harvested area, price, trade flows, and production) (Table1). The key characteristics of the database are: Regions : 141 regions, encompassing 121 individual countries and 20 regional aggregates. This detail allows for both country-specific and broad regional analysis. Sectors : 114 sectors producing 130 commodities, including primary agricultural products as well as important by-products (e.g., oilcake, crop residues). Production Factors : 12 factors of production are included, such as land, five types of labor, capital, natural resources, fossil fuels (coal, oil, gas), and wild fish. Reference Year : The most recent version of the database is based on the year 2017 , serving as the benchmark for all model simulations. MAGNET model database integrates multiple global datasets to ensure accuracy and comprehensive coverage: GTAP : Provides global interrelated social accounting matrices with monetary values, forming the economic backbone of the model. This includes comprehensive bilateral trade data, transport, and protection linkages. FAO : Supplies physical quantities for primary agricultural production and fisheries data. IEA : Contributes detailed energy data in physical quantities. IMAGE : Supplies data on land availability and technical change in the agricultural sector. ILO : Provides essential employment data. The data spans 1973–2019. Earlier years are constructed by combining the underlying FAO, IEA, and other data sources consistent with the MAGNET database structure, ensuring that historical series are harmonized with the 2017 benchmark. After collecting these different sources from within the MAGNET database, we combined them to construct the dataset used in this work. Each observation represents a unique year–country–product combination and includes all present variables on production, harvested area, prices, trade flows, GDP, population, and cross-crop dynamics. Whereas standard CGE and SSP frameworks typically express shocks as percentage changes relative to a baseline, we reconstructed the absolute values of variables by applying these changes to base-year levels. This results in a consistent dataset expressed in actual physical and monetary units, so that the data both follows the structure of CGE models and can be directly used in economic and machine learning analyses. A full overview of descriptive statistics is provided in Appendix Table A1, Figure A1 and Figure A2. Table 2. Key variables, interpretation, and economic role. Variable Interpretation Economic Role Year Calendar year of observation Temporal dimension — allows capturing time trends, policy phases, and temporal shocks in production and trade patterns. EU Country Country within the European Union Spatial dimension — captures country-specific institutional, policy, or agro-ecological differences influencing production, trade, and market behavior. Harvested area Total hectares allocated to wheat cultivation Land input: Core supply-side driver; larger areas typically enable higher production volumes, ceteris paribus. Price Domestic or border price of wheat Market-clearing signal: Reflects equilibrium between supply and demand; higher prices can incentivize increased production and moderate demand. Import Quantity of wheat imported Foreign supply channel: Supplements domestic production shortfalls; buffers domestic market against shocks. Export Quantity of wheat exported Foreign demand channel: Indicates surplus production capacity and international competitiveness; links domestic market to global price signals. GDP Gross Domestic Product (constant prices) Income effect: Higher GDP reflects greater purchasing power and economic capacity, generally associated with higher food demand and investment in agricultural productivity. Population Total number of inhabitants Scale effect on demand: Larger populations increase aggregate consumption needs, driving production and import demand. Wheat_Production Domestic wheat production quantity Key output variable: Endogenous outcome determined by supply-side factors, demand conditions, price signals, and trade flows. Rice_ variables * Metrics for rice (e.g., harvested area, price, trade, production) Substitute crop effect: Competes with wheat for land and other inputs; relative profitability influences allocation decisions. Other grains_ variables * Metrics for other grains Additional substitution or complementarity: Capture how competing or complementary grains influence wheat production through land use trade-offs and market interactions. 4.2 Machine Learning Emulator and model development procedure We applied an XGBoost supervised learning model(Chen & Guestrin, 2016) to assess whether data-driven machine learning models can reproduce selected outputs of an agro-economic equilibrium system under comparable scenario conditions. We selected XGBoost because gradient boosting decision tree algorithms are widely recognized for tabular data, often outperforming deep learning and other machine learning methods in benchmarking studies (Shwartz et al. 2022). The model is designed to learn the functional relationship between a set of macroeconomic, agricultural, and trade-related variables—including GDP, population, harvested area, price levels, import and export volumes, and production metrics for rice and other grains—and wheat production. In this framework, these variables serve as predictors that jointly determine wheat production outcomes. This formulation conceptually parallels the structure of the MAGNET model, in which macroeconomic drivers, input allocation decisions, and market indicators interact to shape production levels. Model development and evaluation followed a structured three-step procedure designed to ensure robustness and meaningful comparison with a benchmark agro-economic model. Historical training and validation: Historical data from 1973 to 2017 were used for machine learning model development. Specifically, we used 1973–2016 as the training set with cross-validation (“Train”), 2017 as hold out test set (“Test”), and 2019 as another independent evaluation set (“2019”). The set “2019” was used for comparison with MAGNET and observed data. A 5-fold cross-validation strategy was employed to assess model performance on the training data, prevent overfitting, and fine-tune hyperparameters. The year 2017 was held out as an independent test set to evaluate the model’s predictive accuracy on unseen historical data. Model evaluation metrics incudes R² (Coefficient of Determination) and RMSE (Root Mean Squared Error). R² measures the proportion of variance in the dependent variable explained by the model. A higher R² indicates better model fit. RMSE quantifies the average magnitude of errors between predicted and observed values. A lower RMSE indicates better model performance. To assess the relative importance of input variables, we employed SHAP (SHapley Additive exPlanations) values. SHAP is a game-theoretic approach that decomposes model predictions into additive contributions from each feature, enabling both global and local interpretability (Lundberg et al, 2020). SHAP summary plots were used to rank influential predictors and visualize how feature values (low vs. high) affect predicted wheat production. To further interpret the model, we applied Partial Dependence Plots (PDPs) and Individual Conditional Expectation (ICE) plots for the most influential predictors (Goldstein et al. 2015). PDPs summarize the marginal effect of a predictor on the outcome by averaging over the distribution of all other features, thereby illustrating how predicted wheat production varies with changes in the predictor within the fitted model. ICE plots complement this by displaying prediction trajectories for individual observations, highlighting heterogeneity across countries and contexts. 4.3 Benchmark comparison between AI emulator and MAGNET To validate the AI emulator’s practical relevance, the AI-based predictions for wheat production were directly compared to the corresponding MAGNET model projections for the year 2019. This comparison also included the actual observed production data for 2019 to assess both approaches against real-world outcomes. This step provided evidence of how well the AI surrogate replicates the structural model’s behavior under contemporary conditions. In other words, this comparative analysis provided insights into how well the AI model captured dynamic production shifts over time, and how its performance stood against MAGNET. 4.4 Scenario-based forecasting under SSPs The AI emulator was applied to project wheat production for the year 2025-2100 under selected Shared Socioeconomic Pathway (SSP) scenarios. To generate the full set of input variables required by the emulator, we combined exogenous SSP-consistent projections for macroeconomic drivers (e.g., GDP and population) with corresponding scenario-consistent values for market and agricultural system variables (e.g., trade flows, domestic prices, and cross-crop production metrics) derived from MAGNET model outputs. A set of five socio-economic development pathways (SSP1—SSP5) has been developed, serving as a consistent scenario framework. Each SSP is described by a quantification of future developments in population (Samir & Lutz, 2017), urbanization and economic development (Dellink et al., 2017), and by a descriptive storyline to guide model parametrization (O'Neill et al 2017). General characteristics of the SSP storylines, with a focus on food insecurity issues, are summarized in Table A2. Table A3 summarizes the core scenario-specific assumptions implemented in the MAGNET–AI framework, focusing on agriculture, wheat productivity, and trade openness, consistent with the established SSP narratives (O’Neill et al., 2017). These assumptions were quantified by domain experts to ensure internal consistency with the SSP storylines and coherence with the broader socio-economic context. In the simulations, future wheat supply trajectories emerge from the interaction between exogenous macro drivers (GDP and population) and endogenous system responses, including production levels, trade flows, and price formation. SSP1 (Sustainability) assumes inclusive economic growth, low population increase, strong land-use regulation, rapid technological progress in agriculture, and increasing trade openness, supporting sustainable productivity growth and efficient global allocation of wheat. SSP2 (Middle of the Road) reflects a continuation of historical trends, characterised by moderate GDP and population growth, incremental productivity improvements, and largely unchanged land-use and trade policies. SSP3 (Regional Rivalry) represents a fragmented and protectionist world with high population growth, weak economic performance, limited agricultural innovation, and restricted trade, resulting in stagnating productivity and regionally constrained wheat supply. SSP5 (Fossil-fuelled Development) assumes rapid economic growth driven by energy-intensive development, low to medium population growth, fast increases in agricultural productivity through intensive input use and mechanisation, and highly liberalised trade, maximising global wheat output but with limited environmental constraints. Following common practice in agricultural impact studies, we focus on SSP1, SSP2, SSP3, and SSP5, which span a wide range of socioeconomic futures while avoiding redundancy associated with SSP4. Together, these contrasting SSPs provide a coherent set of future pathways within which both the structural MAGNET simulations and the AI-based projections assess the sensitivity of global wheat production to alternative socio-economic and policy conditions. Declarations DATA AVAILABILITY The processed panel dataset (1973–2019) used to train, validate, and test the machine learning emulator, as well as the generated baseline and SSP scenario projection datasets used for comparative analysis, are available from the corresponding author upon request. The final dataset was constructed by aggregating and harmonizing historical records from several underlying databases. The physical quantities, energy data, land availability, and employment data are derived from publicly available sources, including the Food and Agriculture Organization (FAO), Eurostat, the International Energy Agency (IEA), the IMAGE model, and the International Labour Organization (ILO). The monetary and bilateral trade matrices are derived from the proprietary Global Trade Analysis Project (GTAP) database, which requires a valid institutional license for direct access. CODE AVAILABILITY The custom code used to train the XGBoost machine learning emulator, conduct the explainable AI analysis (SHAP, partial dependence, and ICE plots), and generate the comparative SSP scenario projections is available on GitHub at https://github.com/WFSRDataScience/MAGNET_AI. Instructions for installation and execution, along with necessary software dependencies, are provided in the repository's README file. The MAGNET model (Modular Applied GeNeral Equilibrium Tool) is a proprietary Computable General Equilibrium (CGE) model developed and maintained by the authors' institution, Wageningen Social & Economic Research. Due to strict third-party licensing restrictions associated with the underlying GTAP database, the core MAGNET source code cannot be made publicly open-source. However, access to the MAGNET model code for review and verification purposes is available from the corresponding author upon reasonable request, subject to the necessary data licensing conditions. FUNDING SOURCE This project has received funding from the European Union’s HORIZON-CL6-2022 research and Innovation programme under grant agreement N◦101084201. AUTHOR CONTRIBUTIONS Conceptualization: XW, MH, DK, BV, AH, SM; methodology: XW, DK, AH ; formal analysis: XW, ; investigation: XW, ; resources: MH; data extraction and curation: MH, DK, SM; writing—original draft preparation: XW, ; writing—review and editing: XW, MH, DK, BV, AH, SM; visualization: XW, DK; supervision: BV,; project administration: AH, BV, XW ; funding acquisition: BV; MAGNEt CGE projections: SM . All authors have read and agreed to the published version of the manuscript. CONFLICT OF INTEREST The authors declare no conflict of interest. References Carriquiry, M., Dumortier, J., & Elobeid, A. (2022). Trade scenarios compensating for halted wheat and maize exports from Russia and Ukraine increase carbon emissions without easing food insecurity. Nature Food , 3 (10), 847-850. Chen, T., & Guestrin, C. (2016). Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, Dellink, R., Chateau, J., Lanzi, E., & Magné, B. (2017). Long-term economic growth projections in the Shared Socioeconomic Pathways. Global Environmental Change , 42 , 200-214. Erenstein, O., Jaleta, M., Mottaleb, K. A., Sonder, K., Donovan, J., & Braun, H.-J. (2022). Global trends in wheat production, consumption and trade. In Wheat improvement: food security in a changing climate (pp. 47-66). Springer International Publishing Cham. Eyring, V., Collins, W. D., Gentine, P., Barnes, E. A., Barreiro, M., Beucler, T., Bocquet, M., Bretherton, C. S., Christensen, H. M., & Dagon, K. (2024). Pushing the frontiers in climate modelling and analysis with machine learning. Nature climate change , 14 (9), 916-928. Hultgren, A., Carleton, T., Delgado, M., Gergel, D. R., Greenstone, M., Houser, T., Hsiang, S., Jina, A., Kopp, R. E., & Malevich, S. B. (2022). Climate change impacts on global agriculture accounting for adaptation. Availabale at SSRN: http://dx. doi. org/10.2139/ssrn , 4222020 . Lobell, D. B., Burke, M. B., Tebaldi, C., Mastrandrea, M. D., Falcon, W. P., & Naylor, R. L. (2008). Prioritizing climate change adaptation needs for food security in 2030. Science , 319 (5863), 607-610. Mansfield, L. A., Nowack, P. J., Kasoar, M., Everitt, R. G., Collins, W. J., & Voulgarakis, A. (2020). Predicting global patterns of long-term climate change from short-term simulations using machine learning. npj Climate and Atmospheric Science , 3 (1), 44. Moghayer, S., Zurek, M., Muzammil, M., Mason-D’Croz, D., Magrath, J., Tabeau, A., Vervoort, J. M., & Achterbosch, T. (2023). A low-carbon and hunger-free future for Bangladesh: An ex-ante assessment of synergies and trade-offs in different transition pathways. Frontiers in Environmental Science , 10 , 977760. Riahi, K., Van Vuuren, D. P., Kriegler, E., Edmonds, J., O’neill, B. C., Fujimori, S., Bauer, N., Calvin, K., Dellink, R., & Fricko, O. (2017). The Shared Socioeconomic Pathways and their energy, land use, and greenhouse gas emissions implications: An overview. Global Environmental Change , 42 , 153-168. Samir, K., & Lutz, W. (2017). The human core of the shared socioeconomic pathways: Population scenarios by age, sex and level of education for all countries to 2100. Global Environmental Change , 42 , 181-192. Shiferaw, B., Smale, M., Braun, H.-J., Duveiller, E., Reynolds, M., & Muricho, G. (2013). Crops that feed the world 10. Past successes and future challenges to the role played by wheat in global food security. Food Security , 5 (3), 291-317. van der Velden, B. H. (2024). Explainable AI: current status and future potential. European Radiology , 34 (2), 1187-1189. van Meijl, H., Tabeau, A., Stehfest, E., Doelman, J., & Lucas, P. (2020). How food secure are the green, rocky and middle roads: food security effects in different world development paths. Environmental Research Communications , 2 (3), 031002. Wang, X., Liu, C., & van der Fels-Klerx, H. (2022). Regional prediction of multi-mycotoxin contamination of wheat in Europe using machine learning. Food Research International , 159 , 111588. Xuhui, W., Zhao, C., Müller, C., Wang, C., Ciais, P., Janssens, I., Peñuelas, J., Asseng, S., Li, T., & Elliott, J. (2020). Emergent constraint on crop yield response to warmer temperature from field experiments. Nature Sustainability , 3 (11), 908-916. Aragie, E. A., McDonald, S., & Thierfelder, K. (2016). A static applied general equilibrium model: Technical documentation: STAGE_DEV Version 2. IFPRI. (2024a. Machine Learning for Spatially Granular Food Security Mapping (Project update). International Food Policy Research Institute. IFPRI. (2024b). Generative AI for Agriculture (GAIA) Project: Enhancing Advisory Services with LLMs (Project update). International Food Policy Research Institute. IFPRI. (2024c). IFPRI's Complementary Modeling Systems: MIRAGRODEP, IMPACT, RIAPA. International Food Policy Research Institute. Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., ... & Lee, S. I. (2020). From local explanations to global understanding with explainable AI for trees. Nature machine intelligence, 2(1), 56-67. Shwartz-Ziv, R., & Armon, A. (2022). Tabular data: Deep learning is not all you need. Information Fusion, 81, 84-90. Goldstein, A., Kapelner, A., Bleich, J., & Pitkin, E. (2015). Peeking inside the black box: Visualizing statistical learning with plots of individual conditional expectation. journal of Computational and Graphical Statistics, 24(1), 44-65. Additional Declarations There is NO Competing Interest. Supplementary Files Supplementarys.docx 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-9405456","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":625168503,"identity":"73cca32f-bcc5-47c7-a48c-b7fb794adb17","order_by":0,"name":"Xinxin Wang","email":"","orcid":"","institution":"Wageningen Food Safety Research","correspondingAuthor":false,"prefix":"","firstName":"Xinxin","middleName":"","lastName":"Wang","suffix":""},{"id":625168502,"identity":"c80f1bea-8dca-48eb-a491-6c5e948adfad","order_by":1,"name":"Saeed Moghayer","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5ElEQVRIiWNgGAWjYLACHhAhAWbaQEQekKAlDSKSQIKWw4S1yLcfPvjgDcM2Ofno5mcSH/ecT9xw+wDbA3xaDM6kJRvOYbhtbHjnmJnkjGe3EzecS2A3wKuFIcdMmofhduLGGQlAxoHbuRvOMLBJ4HVY//vvvyFa0r9J/zlwjrAWhhs5bMwgLfMlgNYxHDhAWIvBjWfGknMMbhsbSOQUW/YcSK6feYaxjYDDkh9+eFNxW05+RvrGGz8O2BnznWE+JvEBn8MgdgHRAQYWSNQwMDYQ1ACxroGBmbDZo2AUjIJRMCIBAHyZUlsTEy5rAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0001-6497-3360","institution":"Wageningen University \u0026 Research","correspondingAuthor":true,"prefix":"","firstName":"Saeed","middleName":"","lastName":"Moghayer","suffix":""},{"id":625168504,"identity":"0355b017-bf0f-4bc9-9af1-33b6bf6b60cd","order_by":2,"name":"Daan Korporaal","email":"","orcid":"https://orcid.org/0000-0002-8899-4374","institution":"Wageningen Food Safety Research","correspondingAuthor":false,"prefix":"","firstName":"Daan","middleName":"","lastName":"Korporaal","suffix":""},{"id":625168505,"identity":"300146ef-8f26-4d75-8baf-2dee9692f166","order_by":3,"name":"Maryam Hajialibeigi","email":"","orcid":"https://orcid.org/0000-0002-6474-5752","institution":"Wageningen Food Safety Research","correspondingAuthor":false,"prefix":"","firstName":"Maryam","middleName":"","lastName":"Hajialibeigi","suffix":""},{"id":625168506,"identity":"02460a8a-bf38-40b1-9abe-928363ddfe56","order_by":4,"name":"Bas van der Velden","email":"","orcid":"https://orcid.org/0000-0003-3750-2824","institution":"Wageningen Food Safety Research","correspondingAuthor":false,"prefix":"","firstName":"Bas","middleName":"van der","lastName":"Velden","suffix":""},{"id":625168507,"identity":"55eaa467-b471-484c-a483-c97ac502ddb4","order_by":5,"name":"Ali Hürriyetoglu","email":"","orcid":"","institution":"Wageningen University \u0026 Research","correspondingAuthor":false,"prefix":"","firstName":"Ali","middleName":"","lastName":"Hürriyetoglu","suffix":""}],"badges":[],"createdAt":"2026-04-13 14:35:55","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9405456/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9405456/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107452773,"identity":"bd23c4a6-6a69-4e03-81b4-30cbed2d36de","added_by":"auto","created_at":"2026-04-21 15:29:15","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":122523,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP summary plot visualizing the importance and impact of input features on wheat production predictions. The SHAP summary plot ranks the top predictors contributing to the AI model’s estimation of wheat production across countries. Each point represents a SHAP value for an individual observation, coloured by the feature value (low to high). Export volume, other grains production, and harvested area are the dominant predictors, reflecting both direct agronomic drivers and broader cropping system intensity. Socioeconomic variables (e.g., GDP, population) and country-specific indicators (e.g., United States, United Kingdom, Spain) highlight the role of national context in modulating wheat production outcomes. The plot displays the most influential features, demonstrating the model’s reliance on a coherent mix of agronomic, economic, and geographic information.\u003c/p\u003e","description":"","filename":"Picture1.png","url":"https://assets-eu.researchsquare.com/files/rs-9405456/v1/8fd543983da8b33f564df700.png"},{"id":107487648,"identity":"2fe3223b-16cc-41ae-a57e-225f1a938635","added_by":"auto","created_at":"2026-04-22 02:42:30","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":421827,"visible":true,"origin":"","legend":"\u003cp\u003ePartial dependence and individual conditional expectation (ICE) plots for the top predictors of national wheat production. a, The figure presents partial dependence (dashed lines) and ICE curves (light blue lines) for key predictors identified by the model, including: Export volume, Other grains production, Harvested area, Other grains export, Population, and GDP. Partial dependence plots represent the average change in predicted wheat production as a function of a given predictor, computed by averaging model predictions over the distribution of the remaining variables. ICE curves display the corresponding prediction trajectories for individual countries, thereby revealing cross-country heterogeneity in the model response.\u003c/p\u003e","description":"","filename":"Picture2.png","url":"https://assets-eu.researchsquare.com/files/rs-9405456/v1/2e5590ac5f0f438f90d95e84.png"},{"id":107705409,"identity":"cbee92f4-aeeb-447e-9524-44dac510525a","added_by":"auto","created_at":"2026-04-24 09:12:24","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":202470,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of absolute wheat production and relative changes in 2019 estimates from MAGNET, the AI model, and observations.\u003cstrong\u003e \u003c/strong\u003ea, Country-level absolute wheat production in 2019 as estimated by the MAGNET model, the AI-based prediction, and the observed data.\u003cstrong\u003e \u003c/strong\u003eb, Relative changes in 2019 wheat production compared to the 2017 baseline for each source. Overall, the AI model reproduces the magnitude and ranking of national wheat production more closely aligned with observations than MAGNET, and captures both the direction and scale of year-on-year changes across diverse production contexts.\u003c/p\u003e","description":"","filename":"Picture3.png","url":"https://assets-eu.researchsquare.com/files/rs-9405456/v1/605574549f871059ec3ed411.png"},{"id":107452775,"identity":"d5ad0b07-64e2-4fde-8c6d-ce88f29537eb","added_by":"auto","created_at":"2026-04-21 15:29:15","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":218439,"visible":true,"origin":"","legend":"\u003cp\u003eTotal wheat production under alternative SSP scenarios (2025–2100): comparison between MAGNET and AI-based projections. The figure shows aggregated wheat production (thousand tonnes) from 2025 to 2100 under four Shared Socioeconomic Pathways (SSP1, SSP2, SSP3, SSP5). Dashed lines represent MAGNET model projections; solid lines represent AI-based predictions.\u003c/p\u003e","description":"","filename":"Picture4.png","url":"https://assets-eu.researchsquare.com/files/rs-9405456/v1/0a11926df0b00d6ea041e852.png"},{"id":107452777,"identity":"46aa9922-0237-41df-abc7-c7774e45ac64","added_by":"auto","created_at":"2026-04-21 15:29:15","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":203568,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of MAGNET and AI-based wheat production projections under alternative SSP scenarios for 2050. (a) Absolute wheat production levels (thousand tonnes) by country, comparing MAGNET model projections (blue bars) with AI-based predictions (orange bars). (b) Relative changes in wheat production compared to the 2017 baseline, showing percentage deviations for each scenario and country.\u003c/p\u003e","description":"","filename":"Picture5.png","url":"https://assets-eu.researchsquare.com/files/rs-9405456/v1/ea06edbb65c50bbd24f68ab2.png"},{"id":107711332,"identity":"df65ef8c-6c84-4145-bfa0-25121214db18","added_by":"auto","created_at":"2026-04-24 09:45:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1431102,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9405456/v1/79a46274-3ebb-424d-b5a7-d5a6218b8ed0.pdf"},{"id":107488912,"identity":"3a93de76-ecc2-47ca-bedd-d790c86615cb","added_by":"auto","created_at":"2026-04-22 02:46:09","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":255499,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarys.docx","url":"https://assets-eu.researchsquare.com/files/rs-9405456/v1/b65bdf14c366fc47fa0792c0.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Foresight Scenario Modelling for Global Wheat Production Under Shared Socioeconomic Pathways: Testing AI Emulators of CGE Models","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eEnsuring global food security under growing socio-economic pressures and escalating climate risks is one of the most pressing challenges (Lobell et al., 2008; Xuhui et al., 2020). Wheat, as one of the world’s principal staple crops\u0026nbsp;(Shiferaw et al., 2013), provides a substantial share of daily caloric intake for billions of people worldwide (Carriquiry et al., 2022; Erenstein et al., 2022; Lobell et al., 2008). Projections of future wheat production are therefore central to informing robust policy decisions, designing sustainable food systems, and mitigating risks associated with climate change and market shocks\u0026nbsp;(Hultgren et al., 2022; Lobell et al., 2008).\u003c/p\u003e\n\u003cp\u003eTraditionally, projections of wheat production rely on well-established agro-economic models, including partial and computable general equilibrium (CGE) frameworks such as MAGNET(Modular Applied GeNeral Equilibrium Tool) (Moghayer et al., 2023; van Meijl et al., 2020). These structural models explicitly embed production functions, supply-demand linkages, trade balances, and policy feedbacks, allowing researchers and policymakers to explore how macro drivers—such as Gross domestic product (GDP) \u0026nbsp;growth, population dynamics, and policy interventions—translate into production outcomes and trade flows under various scenarios, including the widely used Shared Socioeconomic Pathways (SSPs)(O’Neill et al. 2017).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWhile robust and theoretically consistent, these models are computationally intensive and time-consuming to run, especially when exploring large ensembles of policy or climate scenarios to quantify uncertainty and stress-test resilience. As noted by van Meijl et al. (2020), running MAGNET scenarios under diverse SSPs requires full equilibrium recalibration, which is computationally intensive when testing alternative futures. Similarly, Antle et al. (2017) emphasize that multi-scenario CGE applications, particularly under climate uncertainty, require substantial computational effort due to the need for repeated model convergence and calibration. In addition, CGE models like MAGNET are built on large nonlinear systems calibrated to social accounting matrices and solved under equilibrium constraints. Although they capture key nonlinear economic mechanisms, their functional forms and parameterizations may not fully reflect the complex, high-dimensional nonlinear relationships observed in empirical data.\u003c/p\u003e\n\u003cp\u003eAdvances in artificial intelligence methods such as machine learning offer a promising complementary pathway to accelerate foresight and scenario exploration in global food systems. Machine learning is particularly well-suited for learning highly nonlinear relationships from data, capturing interactions that are difficult to encode in structural models, and generating large numbers of simulations at negligible computational cost. While there has been recent interest in using machine learning \u0026nbsp;to emulate biophysical crop models or climate impact modules (Eyring et al., 2024; Mansfield et al., 2020; Wang et al., 2022), there remains limited evidence on whether machine learning \u0026nbsp;can accurately replicate the core behavioral logic of equilibrium agro-economic systems for staple crops under realistic socio-economic drivers.\u003c/p\u003e\n\u003cp\u003eIn this study, we investigated whether machine learning approach can serve as a robust surrogate for a structural CGE-style model of global wheat production. Using a panel dataset that captures the nonlinear relationships among essential economic drivers, such as GDP, population, harvested area, prices, and trade flows, and cross-crop substitution effects, we used machine learning \u0026nbsp;to emulate and reproduce wheat production outcomes consistent with theoretical supply, demand, and trade relationships. Beyond benchmarking predictive performance under historical and short-term conditions, we evaluated the emulator’s behaviour under Shared Socioeconomic Pathway (SSP) scenarios, comparing AI-based projections with those generated by the MAGNET CGE model to assess robustness across contrasting long-term development trajectories. The emulator’s performance, robustness under pseudo-unseen scenarios, and interpretability are further examined using explainable artificial intelligence tools (van der Velden, 2024).\u003c/p\u003e\n\u003cp\u003eThe aim of this study was to assess the extent to which data-driven AI emulators can replicate the behaviour and policy-relevant outcomes of structurally grounded CGE models in global wheat production scenarios under alternative socio-economic and climate futures. In particular, the study aimed to (i) assess the emulator’s ability to reproduce historical and near-term CGE-based wheat production outcomes, (ii) examine divergences between AI-based and CGE-based projections under alternative SSP scenarios, and (iii) demonstrate how explainable AI tools can be used to interpret emulator behaviour in a manner that is consistent with established economic reasoning. By pursuing these aims, the study clarified when and how AI-based emulators can complement structural economic models in short term and long term SSP-based food security foresight, while delineating the boundaries of their applicability for policy-relevant analysis\u003c/p\u003e"},{"header":"2. Results","content":"\u003cp\u003eThis section reports the outcomes of the MAGNET and AI emulator, focusing on changes in wheat production across scenarios. Results are presented at both global and regional scales.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.\u003c/strong\u003e\u003cstrong\u003e1 AI Model performance:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe AI model achieved a root mean square error (RMSE) between actual and predicted wheat production of 1,301 thousand tons and a high coefficient of determination (R\u0026sup2; = 0.99) under five-fold cross-validation, indicating strong generalization within the training dataset (Table 1). On the test set, the RMSE was 1603 with R\u003csup\u003e2\u003c/sup\u003e = 0.98, and on the 2019 data, RMSE was 1,301 with R\u0026sup2; = 0.99. This closely matches the predictive accuracy of the MAGNET model for the 2019 (RMSE = 1,012; R\u0026sup2; = 0.99, Table 2).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e AI Model Performance Metrics\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 24px;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 27px;\"\u003e\n \u003cp\u003eDataset\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003eRMSE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 24px;\"\u003e\n \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 24px;\"\u003e\n \u003cp\u003eAI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 27px;\"\u003e\n \u003cp\u003eTrain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e989\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 24px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 27px;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e1603\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 24px;\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 27px;\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e1301\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 24px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 24px;\"\u003e\n \u003cp\u003eMAGNET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 27px;\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e1012\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 24px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eExplainable AI showed that export volume, production of other grains, and total harvested area were the top three predictors (Figure 1). These variables capture both direct and indirect production dynamics: higher export capacity and larger multi-crop production areas signal more robust agricultural systems that likely support greater wheat yields. The magnitude of the SHAP values for export indicates that high export volumes are consistently associated with substantially higher predicted wheat production, aligning with the economic principle that surplus-oriented producers scale output beyond domestic needs. Similarly, harvested area exhibits strong positive SHAP contributions, reflecting the role of land allocation in sustaining wheat productivity. Socioeconomic variables also exert measurable impacts: \u003cem\u003ePopulation\u003c/em\u003e and \u003cem\u003eGDP\u003c/em\u003e show consistently positive SHAP values, suggesting that countries with larger populations and stronger economies tend to have higher predicted wheat production, likely due to their capacity to mobilize agricultural inputs, infrastructure, and technology. \u0026nbsp;Notably, specific country indicators (e.g., the United States of America, the United Kingdom, Spain) also appear among the top features, indicating persistent regional elements\u0026mdash;such as institutional, climatic, or policy environments. This unobserved heterogeneity acts similarly to the region-specific productivity parameters or fixed effects commonly used in panel data econometrics and CGE calibration.\u003c/p\u003e\n\u003cp\u003eCollectively, the SHAP analysis confirms that the AI model relies on a coherent blend of agronomic, economic, and geographic determinants to accurately estimate wheat production across diverse national contexts.\u003c/p\u003e\n\u003cp\u003ePartial dependence and ICE analyses were conducted to interpret how the most important predictors shape the AI model\u0026rsquo;s estimation of national wheat production (Figure 2). The partial dependence plots (dashed lines) represent the average model-implied response of wheat production to a given predictor, holding the joint distribution of other variables constant. The ICE curves (light blue lines) show the corresponding response for individual countries, thereby revealing cross-country heterogeneity around the average trend. \u0026nbsp;For the top three predictors (Export, Other grains production, and Harvested area), \u0026nbsp;the plots indicate clear, positive relationships with distinct curve shapes. The partial dependence plot for Export shows a sharp, near-linear increase in predicted production as export volume rises, particularly up to ~15,000 units(1000 tone per unit), after which the curve begins to level off. This pattern suggests a strong positive association between export capacity and production, but also a saturation effect where additional exports yield diminishing marginal gains. For Other grains production, the relationship is initially steep but flattens beyond ~50,000\u0026ndash;100,000 units(1000 tone per unit), indicating that diversification into other cereals supports wheat output up to a point, after which production synergies stabilize. Similarly, Harvested area displays a steadily increasing but concave relationship, where gains in predicted production become less pronounced once cultivated area surpasses ~5,000\u0026ndash;7,000 units (1000 ha per unit), reflecting land constraints and diminishing returns to expansion.\u003c/p\u003e\n\u003cp\u003eFor the next three predictors (Other grains export, Population, and GDP), the partial dependence plots reveal more modest and gradual effects. Other grains export contributes positively but with a shallower slope, suggesting limited marginal gains beyond initial levels of diversification. Population shows a near-linear but relatively flat increase, implying that while larger populations support higher production, the effect is incremental rather than transformative. GDP demonstrates a positive but plateauing trend: wealthier economies generally sustain higher predicted wheat production, but the slope diminishes at very high income levels, reflecting that once basic infrastructure and technology thresholds are met, further gains in GDP contribute less to production growth.\u003c/p\u003e\n\u003cp\u003eThe spread of individual ICE curves (Figure 2) further highlights substantial heterogeneity across countries. While the average trends align with economic expectations, country-specific lines deviate in slope and curvature, illustrating how local factors, such as market integration, technological adoption, and policy environments, modulate the marginal effects of each predictor.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2 AI and Magnet model comparison\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 3 compares the absolute and relative national wheat production for 2019 estimated by the MAGNET model, the AI model, and actual observed data. The AI model effectively replicates the absolute production levels across all countries, with predictions closely matching the observed values and preserving the expected rank order of major producers (Figure 4a). For example, in France, the AI prediction (37,604 thousand tonnes) is slightly below the observed value (41,083 thousand tonnes), whereas MAGNET estimates production at 38,990 thousand tonnes. In the United States, the AI model predicts 51,099 thousand tonnes, closer to the observed 52,581 thousand tonnes than MAGNET, which estimates a lower value of 48,998 thousand tonnes. The AI model and MAGNET replicated both the direction and approximate magnitude of production changes for most countries(figure 4b). Larger deviations for certain countries (e.g., Estonia, Lithuania, and Hungary) highlight areas where local conditions or data uncertainty may affect model responsiveness. For instance, in Estonia, the observed increase was +18.7%, while the AI model predicted+59.3% and MAGNET \u0026nbsp;+1.8%. In Lithuania, observed production fell by \u0026ndash;1.9%, yet the AI model predicted +35.9% and MAGNET +11.3%. Conversely, in Hungary, the observed production increased by +2.5%, the AI model predicted +17.0% and \u0026nbsp;MAGNET +2.8%. Figure A4 shows outliers with large predicted changes, such as \u0026nbsp;Portugal (+757% predicted vs. +25% observed) and Slovenia (+195% predicted vs. \u0026ndash;0.8% observed) illustrating \u0026nbsp;potential data noise or local anomalies where the AI model substantially overestimated production shifts.\u003c/p\u003e\n\u003cp\u003eOverall, these comparisons demonstrated that the AI-based approach replicates historical production scales while providing an alternative, data-driven estimate of interannual variation that differs from the benchmark MAGNET model in several contexts. From a structural standpoint, discrepancies may reflect the machine learning model\u0026rsquo;s implicit flexibility to capture time-varying elasticities and short-run adjustment costs, unlike the MAGNET model, which is rooted in long-run equilibrium assumptions and fixed behavioral parameters. This suggests that machine learning methods can complement traditional economic simulation models by offering an independent perspective and potentially capturing variability not explicitly modelled in MAGNET, thereby enriching short-term agricultural production analyses.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.4 AI-Magnet projection under different SSPs\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 4 shows both modelling approaches capture the expected ranking of scenarios, with SSP3 and SSP5 producing higher outputs than SSP1. However, the magnitude of projections diverges: the MAGNET model projects strong and sustained increases in wheat production across all SSPs, exceeding 500 Mt by 2100 in SSP5. In contrast, the AI-based model predicts a flatter trajectory, stabilizing between 180\u0026ndash;270 Mt across scenarios. Even under SSP2 and SSP3, where both models anticipate growth, MAGNET\u0026rsquo;s projections are consistently 100,000\u0026ndash;150,000 thousand tons higher than the AI-based estimates by the end of the century.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis divergence reflects the differing foundations of the two approaches. MAGNET embeds forward-looking assumptions about technological progress, policy-driven yield improvements, and long-run equilibrium adjustments, which amplify projected output growth. In contrast, the AI model is constrained by historical patterns in its input variables (harvested area, prices, imports, exports, GDP, population, and interactions with rice and other grains), and thus provides a more conservative outlook that reflects socio-economic and trade dynamics rather than speculative yield gains.\u003c/p\u003e\n\u003cp\u003eFrom a foresight perspective, the discrepancy between the models underscores the uncertainty inherent in long-term agricultural projections. On one hand, MAGNET illustrates the potential of optimistic pathways driven by technological innovation and global trade integration. On the other, the AI-based model offers a grounded, historically consistent trajectory that highlights structural limits to growth.\u003c/p\u003e\n\u003cp\u003eFigure 5 illustrates country-level projections of wheat production in 2050 under alternative SSP scenarios, comparing outcomes from the MAGNET model with those from the AI-based emulator. In panel (a), both models reproduce the relative ranking of major producers, but the absolute magnitudes diverge. For instance, in the United States, MAGNET projects 55,092 thousand tonnes under SSP1, whereas the AI model estimates only 39,975 thousand tonnes\u0026mdash;a shortfall of more than 15,000 tonnes relative to MAGNET. In Canada, MAGNET projects 69,326 thousand tonnes under SSP1, more than double the AI estimate of 25,374 thousand tonnes. In contrast, in France, the two models produce closer results, with MAGNET projecting 35,041 thousand tonnes and the AI model predicting 26,603 thousand tonnes. Germany shows a similar pattern: MAGNET estimates 26,418 thousand tonnes versus the AI model\u0026rsquo;s 20,154 thousand tonnes. These examples highlight that MAGNET\u0026rsquo;s forward-looking productivity assumptions consistently inflate projections compared to the more conservative, historically anchored AI estimates.\u003c/p\u003e\n\u003cp\u003ePanel (b) provides further insight by comparing relative changes in 2050 production against the 2017 baseline. In Western and Northern Europe, both models project stagnation or decline, but the AI model often shows sharper contractions. Conversely, in several Southern and Eastern European countries, the AI model amplified positive deviations.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe country-level analysis underscores important methodological contrasts between the AI and MAGNET models. MAGNET\u0026rsquo;s equilibrium structure produces relatively smooth trajectories, dominated by assumed productivity growth and resource reallocation. By contrast, the AI model captures the imprint of historical variability, amplifying signals in countries where production has been volatile or heavily influenced by trade fluctuations. This explains the very large positive deviations observed in Portugal, Slovenia, and Italy, as well as the sharper contractions seen in Northern Europe.\u003c/p\u003e\n\u003cp\u003eFrom a foresight perspective, these divergences highlight both the strengths and limitations of each approach. The MAGNET model offers consistency grounded in economic theory, but may overstate long-run production capacity by assuming sustained technological and productivity improvements. The AI-based emulator provides a more conservative outlook, shaped by historical evidence, but may exaggerate volatility in smaller markets.\u0026nbsp;\u003c/p\u003e"},{"header":"3. Discussion","content":"\u003cp\u003eIn this study, we analyzed possible short-term and long-term trajectories for global wheat production by combining a structural computable general equilibrium model (MAGNET) with a data-driven AI emulator (XGBoost). Empirically, we find that the AI emulator reproduced historical wheat production with accuracy comparable to MAGNET and closely tracked observed 2019 outcomes. Under long-term SSP scenarios, the AI-based projections were more conservative than MAGNET’s, with larger differences in smaller and more volatile producing countries. By linking scenario-consistent exogenous drivers and internally consistent system responses, we showed how this hybrid approach can replicate key dynamics of a well-established agro-economic model while providing a flexible and computationally efficient tool for rapid scenario exploration.\u003c/p\u003e\n\u003cp\u003eA key feature of our analysis is the focus on wheat production as a critical pillar of global food supply, benchmarked across contrasting SSP pathways. Our findings show that exogenous factors such as population growth and total GDP drive both the extensive (scale) and intensive (per capita) margins of wheat demand, while endogenous responses, including land allocation, price-mediated supply reactions, and trade equilibrium, determine regional production capacity and market outcomes. In this hybrid setting, the CGE framework ensures theoretical consistency and scenario coherence, while the emulator provides responsiveness to historically observed variability.\u003c/p\u003e\n\u003cp\u003eThe results also illustrate how the two modelling components play different roles across time horizons. MAGNET provides structurally grounded long-run pathways under SSP narratives, including assumptions on productivity and technical change that support sustained increases in projected output across scenarios. The emulator, trained on historical patterns in the data, produces trajectories that remain anchored in historically learned relationships. Rather than implying that one approach is “right” and the other “wrong,” these differences help clarify uncertainty in long-term agricultural projections and indicate where outcomes are sensitive to productivity pathways that cannot be validated directly from historical relationships alone.\u003c/p\u003e\n\u003cp\u003eAt the country level, this hybrid perspective remains important. MAGNET tends to produce smoother trajectories that follow scenario narratives and economy-wide consistency. The emulator can show larger relative movements in smaller or historically volatile producing countries, reflecting sensitivity to patterns in the historical record. This can be informative for identifying where projected outcomes are more sensitive to trade and market fluctuations, while MAGNET provides the broader equilibrium context in which such changes occur.\u003c/p\u003e\n\u003cp\u003eThe broader value proposition of the hybrid approach lies in its ability to bridge long-term foresight with rapid stress-testing of short-term shocks. Traditional CGE models are essential for capturing long-run equilibrium pathways under alternative futures, but they are computationally intensive when large ensembles of scenarios are required. In contrast, the emulator offers speed and flexibility and can be used to explore a wide range of counterfactuals, such as trade disruptions, policy interventions, or localized shocks, at low computational cost. This hybrid role provides theoretical grounding on one hand and operational agility on the other, enabling more timely and adaptive policy assessments in food security analysis.\u003c/p\u003e\n\u003cp\u003eThe integration of AI into economic modelling is consistent with broader developments in applied policy analysis. The European Commission’s Joint Research Centre (JRC) and IFPRI, for example, use AI to support scenario exploration, data management, and communication of modelling insights, while retaining structural models for theory-consistent interpretation and policy coherence. These examples underscore a shared direction: AI can provide tools for stress-testing and policy communication, while structural economic modelling remains central for internally consistent scenario assessment.\u003c/p\u003e\n\u003cp\u003eSeveral limitations remain. First, while our approach captures supply-side and trade-driven dynamics, demand-side nutrition metrics, income distribution effects, and explicit household welfare are not endogenised in the AI component. Second, climate change impacts and extremes enter as exogenous inputs derived from scenario narratives rather than being endogenously simulated within the emulator, meaning that feedback loops and adaptive behaviour are not captured within the AI structure. Third, because the emulator is trained on historical socio-economic and trade patterns, it may underestimate the impacts of unprecedented shocks or regime changes outside the range of past experience. Finally, other machine learning algorithms or ensemble approaches could be explored; the focus here was on evaluating whether a single robust algorithm can emulate structural CGE-style outcomes.\u003c/p\u003e\n\u003cp\u003eFuture work can build on this proof-of-concept by developing scenario-robust, constraint-aware emulators that are iteratively refined with targeted CGE runs. Incorporating productivity and technology signals more explicitly can strengthen long-horizon emulation where divergence is driven by productivity pathways. Extending the approach beyond wheat to multi-crop systems and food-security indicators could broaden its relevance to the multidimensional challenges of global food security.\u003c/p\u003e\n\u003cp\u003eOverall, the MAGNET-AI framework provides a proof-of-concept for combining structural economic logic with flexible machine learning in global food security research. The study demonstrates how an emulator can complement a structural CGE model by supporting rapid scenario exploration while retaining the role of CGE analysis for internally consistent equilibrium interpretation. In this hybrid setup, MAGNET provides scenario coherence and theory-based linkages, whereas the emulator offers a computationally efficient way to extend exploration to larger ensembles of scenario variants.\u003c/p\u003e\n\u003cp\u003eTaken together, the findings support a practical implication for foresight and policy analysis. AI emulators should not be treated as replacements for CGE models, but as scalable complements that enable rapid screening and stress-testing of “what-if” scenarios while helping to identify sensitive regions and drivers. This can support more timely and adaptive assessments in food security contexts, particularly when decision makers need to explore uncertainty across multiple futures rather than focus on a small set of runs. At the same time, CGE models remain essential for analyzing welfare, trade, and broader equilibrium effects within SSP narratives and for ensuring that scenario results remain internally consistent.\u0026nbsp;\u003c/p\u003e"},{"header":"4. Method","content":"\u003cp\u003eThis study employed an integrated modelling framework combining a global CGE model (MAGNET) with data driven machine learning to assess future wheat production under alternative socioeconomic scenarios. The methodological components and their coupling are described below.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1. \u0026nbsp;Economic Theory and Data\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe design of our machine learning predictive framework is rooted in well-established principles of agricultural economics and international trade theory (details refer to supplementary B), which describe how resource endowments, price signals, and trade structures determine production and market outcomes (Gardner, 1987; Hertel, 1997; Anderson \u0026amp; Martin, 2005). Our primary target variable is wheat production, which is conceptualized as an outcome influenced by both domestic resource allocation and market conditions.\u003c/p\u003e\n\u003cp\u003eOur selection of predictor variables reflects this theoretical structure. From a supply-side perspective, wheat production is a function of available land, which we proxy with harvested area. This represents a key resource allocation decision made by producers. We also include rice and other grain metrics to account for cross-crop competition, reflecting the opportunity cost of land and other inputs. This aligns with the economic principle that farmers allocate land to the most profitable crops, a decision often modeled in CGE frameworks using Constant Elasticity of Transformation (CET) functions.\u003c/p\u003e\n\u003cp\u003eOn the demand-side, we include GDP and population as standard macroeconomic drivers. These variables shape the income elasticity and scale effects on consumption, influencing the overall demand for wheat. Prices, both domestic and international, play a crucial role as market-clearing signals, balancing supply and demand. Their inclusion allows the AI model to capture producer and consumer responses to market incentives.\u003c/p\u003e\n\u003cp\u003eFinally, we incorporate trade flows (imports and exports) to capture the extent to which domestic supply shortfalls or surpluses are balanced through international markets. These variables are essential for understanding how national production is integrated into the global system. By using a variable set that aligns with these theoretically grounded causal pathways, we ensure that our data-driven model can be directly compared with the structurally consistent results of the MAGNET CGE framework.\u003c/p\u003e\n\u003cp\u003eThese variables reflect the standard structure of general equilibrium models such as MAGNET, capturing the primary supply-side, demand-side, and trade linkages that determine national and global food production outcomes. Data for this study was derived from the MAGNET model database, which integrates multiple global datasets (van Meijl et al., 2020; Woltjer \u0026amp; Kuiper, 2014). The dataset includes annual records for multiple countries and years, with core variables covering harvested area, domestic wheat price, import and export volumes, GDP, population, and cross-crop information for rice and other grains (including harvested area, price, trade flows, and production) (Table1).\u003c/p\u003e\n\u003cp\u003eThe key characteristics of the database are:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eRegions\u003c/strong\u003e: 141 regions, encompassing 121 individual countries and 20 regional aggregates. This detail allows for both country-specific and broad regional analysis.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eSectors\u003c/strong\u003e: 114 sectors producing 130 commodities, including primary agricultural products as well as important by-products (e.g., oilcake, crop residues).\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eProduction Factors\u003c/strong\u003e: 12 factors of production are included, such as land, five types of labor, capital, natural resources, fossil fuels (coal, oil, gas), and wild fish.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eReference Year\u003c/strong\u003e: The most recent version of the database is based on the year \u003cstrong\u003e2017\u003c/strong\u003e, serving as the benchmark for all model simulations.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eMAGNET model database integrates multiple global datasets to ensure accuracy and comprehensive coverage:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eGTAP\u003c/strong\u003e: Provides global interrelated social accounting matrices with monetary values, forming the economic backbone of the model. This includes comprehensive bilateral trade data, transport, and protection linkages.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eFAO\u003c/strong\u003e: Supplies physical quantities for primary agricultural production and fisheries data.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eIEA\u003c/strong\u003e: Contributes detailed energy data in physical quantities.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eIMAGE\u003c/strong\u003e: Supplies data on land availability and technical change in the agricultural sector.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eILO\u003c/strong\u003e: Provides essential employment data.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe data spans 1973–2019. Earlier years are constructed by combining the underlying FAO, IEA, and other data sources consistent with the MAGNET database structure, ensuring that historical series are harmonized with the 2017 benchmark. After collecting these different sources from within the MAGNET database, we combined them to construct the dataset used in this work. Each observation represents a unique year–country–product combination and includes all present variables on production, harvested area, prices, trade flows, GDP, population, and cross-crop dynamics. Whereas standard CGE and SSP frameworks typically express shocks as percentage changes relative to a baseline, we reconstructed the absolute values of variables by applying these changes to base-year levels. This results in a consistent dataset expressed in actual physical and monetary units, so that the data both follows the structure of CGE models and can be directly used in economic and machine learning analyses. A full overview of descriptive statistics is provided in Appendix Table A1, Figure A1 and Figure A2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e2.\u003c/strong\u003e\u0026nbsp; Key variables, interpretation, and economic role.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"99%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eVariable\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eInterpretation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eEconomic Role\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eYear\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCalendar year of observation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTemporal dimension — allows capturing time trends, policy phases, and temporal shocks in production and trade patterns.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eEU Country\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCountry within the European Union\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSpatial dimension — captures country-specific institutional, policy, or agro-ecological differences influencing production, trade, and market behavior.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eHarvested area\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTotal hectares allocated to wheat cultivation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eLand input:\u003c/strong\u003e Core supply-side driver; larger areas typically enable higher production volumes, ceteris paribus.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePrice\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDomestic or border price of wheat\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eMarket-clearing signal:\u003c/strong\u003e Reflects equilibrium between supply and demand; higher prices can incentivize increased production and moderate demand.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eImport\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eQuantity of wheat imported\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eForeign supply channel:\u003c/strong\u003e Supplements domestic production shortfalls; buffers domestic market against shocks.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eExport\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eQuantity of wheat exported\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eForeign demand channel:\u003c/strong\u003e Indicates surplus production capacity and international competitiveness; links domestic market to global price signals.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eGDP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eGross Domestic Product (constant prices)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eIncome effect:\u003c/strong\u003e Higher GDP reflects greater purchasing power and economic capacity, generally associated with higher food demand and investment in agricultural productivity.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePopulation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTotal number of inhabitants\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eScale effect on demand:\u003c/strong\u003e Larger populations increase aggregate consumption needs, driving production and import demand.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eWheat_Production\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDomestic wheat production quantity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eKey output variable:\u003c/strong\u003e Endogenous outcome determined by supply-side factors, demand conditions, price signals, and trade flows.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003eRice_ variables\u003c/em\u003e*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMetrics for rice (e.g., harvested area, price, trade, production)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eSubstitute crop effect:\u003c/strong\u003e Competes with wheat for land and other inputs; relative profitability influences allocation decisions.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003eOther grains_ variables\u003c/em\u003e*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eMetrics for other grains\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAdditional substitution or complementarity:\u003c/strong\u003e Capture how competing or complementary grains influence wheat production through land use trade-offs and market interactions.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e4.2\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eMachine Learning Emulator\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;and model development procedure\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe applied an XGBoost supervised learning model(Chen \u0026amp; Guestrin, 2016) to assess whether data-driven machine learning models can reproduce selected outputs of an agro-economic equilibrium system under comparable scenario conditions. We selected XGBoost because gradient boosting decision tree algorithms are widely recognized for tabular data, often outperforming deep learning and other machine learning methods in benchmarking studies (Shwartz et al. 2022). The model is designed to learn the functional relationship between a set of macroeconomic, agricultural, and trade-related variables—including GDP, population, harvested area, price levels, import and export volumes, and production metrics for rice and other grains—and wheat production. In this framework, these variables serve as predictors that jointly determine wheat production outcomes. This formulation conceptually parallels the structure of the MAGNET model, in which macroeconomic drivers, input allocation decisions, and market indicators interact to shape production levels.\u003c/p\u003e\n\u003cp\u003eModel development and evaluation followed a structured three-step procedure designed to ensure robustness and meaningful comparison with a benchmark agro-economic model.\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eHistorical training and validation:\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eHistorical data from 1973 to 2017 were used for machine learning model development. Specifically, we used 1973–2016 as the training set with cross-validation (“Train”), 2017 as hold out test set (“Test”), and 2019 as another independent evaluation set (“2019”). The set “2019” was used for comparison with MAGNET and observed data. A 5-fold cross-validation strategy was employed to assess model performance on the training data, prevent overfitting, and fine-tune hyperparameters. The year 2017 was held out as an independent test set to evaluate the model’s predictive accuracy on unseen historical data.\u003c/p\u003e\n\u003cp\u003eModel evaluation metrics incudes R² (Coefficient of Determination) and RMSE (Root Mean Squared Error). R² measures the proportion of variance in the dependent variable explained by the model. A higher R² indicates better model fit. RMSE quantifies the average magnitude of errors between predicted and observed values. A lower RMSE indicates better model performance. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo assess the relative importance of input variables, we employed SHAP (SHapley Additive exPlanations) values. SHAP is a game-theoretic approach that decomposes model predictions into additive contributions from each feature, enabling both global and local interpretability (Lundberg et al, 2020). SHAP summary plots were used to rank influential predictors and visualize how feature values (low vs. high) affect predicted wheat production.\u003c/p\u003e\n\u003cp\u003eTo further interpret the model, we applied Partial Dependence Plots (PDPs) and Individual Conditional Expectation (ICE) plots for the most influential predictors (Goldstein et al. 2015). PDPs summarize the marginal effect of a predictor on the outcome by averaging over the distribution of all other features, thereby illustrating how predicted wheat production varies with changes in the predictor within the fitted model. ICE plots complement this by displaying prediction trajectories for individual observations, highlighting heterogeneity across countries and contexts.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Benchmark comparison between AI emulator and MAGNET\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo validate the AI emulator’s practical relevance, the AI-based predictions for wheat production were directly compared to the corresponding MAGNET model projections for the year 2019. This comparison also included the actual observed production data for 2019 to assess both approaches against real-world outcomes. This step provided evidence of how well the AI surrogate replicates the structural model’s behavior under contemporary conditions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn other words, this comparative analysis provided insights into how well the AI model captured dynamic production shifts over time, and how its performance stood against MAGNET.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4 Scenario-based forecasting under SSPs\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe AI emulator was applied to project wheat production for the year 2025-2100 under selected Shared Socioeconomic Pathway (SSP) scenarios. To generate the full set of input variables required by the emulator, we combined exogenous SSP-consistent projections for macroeconomic drivers (e.g., GDP and population) with corresponding scenario-consistent values for market and agricultural system variables (e.g., trade flows, domestic prices, and cross-crop production metrics) derived from MAGNET model outputs.\u003c/p\u003e\n\u003cp\u003eA set of five socio-economic development pathways (SSP1—SSP5) has been developed, serving as a consistent scenario framework. Each SSP is described by a quantification of future developments in population (Samir \u0026amp; Lutz, 2017), urbanization and economic development (Dellink et al., 2017), and by a descriptive storyline to guide model parametrization (O'Neill et al 2017). General characteristics of the SSP storylines, with a focus on food insecurity issues, are summarized in Table A2.\u003c/p\u003e\n\u003cp\u003eTable A3 summarizes the core scenario-specific assumptions implemented in the MAGNET–AI framework, focusing on agriculture, wheat productivity, and trade openness, consistent with the established SSP narratives (O’Neill et al., 2017). These assumptions were quantified by domain experts to ensure internal consistency with the SSP storylines and coherence with the broader socio-economic context. In the simulations, future wheat supply trajectories emerge from the interaction between exogenous macro drivers (GDP and population) and endogenous system responses, including production levels, trade flows, and price formation.\u003c/p\u003e\n\u003cp\u003eSSP1 (Sustainability) assumes inclusive economic growth, low population increase, strong land-use regulation, rapid technological progress in agriculture, and increasing trade openness, supporting sustainable productivity growth and efficient global allocation of wheat. SSP2 (Middle of the Road) reflects a continuation of historical trends, characterised by moderate GDP and population growth, incremental productivity improvements, and largely unchanged land-use and trade policies. SSP3 (Regional Rivalry) represents a fragmented and protectionist world with high population growth, weak economic performance, limited agricultural innovation, and restricted trade, resulting in stagnating productivity and regionally constrained wheat supply. SSP5 (Fossil-fuelled Development) assumes rapid economic growth driven by energy-intensive development, low to medium population growth, fast increases in agricultural productivity through intensive input use and mechanisation, and highly liberalised trade, maximising global wheat output but with limited environmental constraints. Following common practice in agricultural impact studies, we focus on SSP1, SSP2, SSP3, and SSP5, which span a wide range of socioeconomic futures while avoiding redundancy associated with SSP4. Together, these contrasting SSPs provide a coherent set of future pathways within which both the structural MAGNET simulations and the AI-based projections assess the sensitivity of global wheat production to alternative socio-economic and policy conditions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eDATA AVAILABILITY\u003c/p\u003e\n\u003cp\u003eThe processed panel dataset (1973–2019) used to train, validate, and test the machine learning emulator, as well as the generated baseline and SSP scenario projection datasets used for comparative analysis, are available from the corresponding author upon request. The final dataset was constructed by aggregating and harmonizing historical records from several underlying databases. The physical quantities, energy data, land availability, and employment data are derived from publicly available sources, including the Food and Agriculture Organization (FAO), Eurostat, the International Energy Agency (IEA), the IMAGE model, and the International Labour Organization (ILO). The monetary and bilateral trade matrices are derived from the proprietary Global Trade Analysis Project (GTAP) database, which requires a valid institutional license for direct access.\u003c/p\u003e\n\u003cp\u003eCODE AVAILABILITY\u003c/p\u003e\n\u003cp\u003eThe custom code used to train the XGBoost machine learning emulator, conduct the explainable AI analysis (SHAP, partial dependence, and ICE plots), and generate the comparative SSP scenario projections is available on GitHub at\u0026nbsp;https://github.com/WFSRDataScience/MAGNET_AI. Instructions for installation and execution, along with necessary software dependencies, are provided in the repository's README file.\u003c/p\u003e\n\u003cp\u003eThe MAGNET model (Modular Applied GeNeral Equilibrium Tool) is a proprietary Computable General Equilibrium (CGE) model developed and maintained by the authors' institution, Wageningen Social \u0026amp; Economic Research. Due to strict third-party licensing restrictions associated with the underlying GTAP database, the core MAGNET source code cannot be made publicly open-source. However, access to the MAGNET model code for review and verification purposes is available from the corresponding author upon reasonable request, subject to the necessary data licensing conditions.\u003c/p\u003e\n\u003cp\u003eFUNDING SOURCE\u003c/p\u003e\n\u003cp\u003eThis project has received funding from the European Union’s HORIZON-CL6-2022 research and Innovation programme under grant agreement N◦101084201.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAUTHOR CONTRIBUTIONS\u003c/p\u003e\n\u003cp\u003eConceptualization: XW, MH, DK, BV, AH, SM; methodology: XW, DK, AH ; formal analysis: XW, ; investigation: XW, ; resources: MH; data extraction and curation: MH, DK, SM; writing—original draft preparation: XW, ; writing—review and editing: \u0026nbsp; XW, MH, DK, BV, AH, SM; visualization: XW, DK; supervision: BV,; project administration: AH, BV, XW ; funding acquisition: BV; MAGNEt CGE projections: SM . All authors have read and agreed to the published version of the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCONFLICT OF INTEREST\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflict of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCarriquiry, M., Dumortier, J., \u0026amp; Elobeid, A. (2022). Trade scenarios compensating for halted wheat and maize exports from Russia and Ukraine increase carbon emissions without easing food insecurity. \u003cem\u003eNature Food\u003c/em\u003e,\u003cem\u003e 3\u003c/em\u003e(10), 847-850.\u003c/li\u003e\n\u003cli\u003eChen, T., \u0026amp; Guestrin, C. (2016). Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining,\u003c/li\u003e\n\u003cli\u003eDellink, R., Chateau, J., Lanzi, E., \u0026amp; Magn\u0026eacute;, B. (2017). Long-term economic growth projections in the Shared Socioeconomic Pathways. \u003cem\u003eGlobal Environmental Change\u003c/em\u003e,\u003cem\u003e 42\u003c/em\u003e, 200-214.\u003c/li\u003e\n\u003cli\u003eErenstein, O., Jaleta, M., Mottaleb, K. A., Sonder, K., Donovan, J., \u0026amp; Braun, H.-J. (2022). Global trends in wheat production, consumption and trade. In \u003cem\u003eWheat improvement: food security in a changing climate\u003c/em\u003e (pp. 47-66). Springer International Publishing Cham.\u003c/li\u003e\n\u003cli\u003eEyring, V., Collins, W. D., Gentine, P., Barnes, E. A., Barreiro, M., Beucler, T., Bocquet, M., Bretherton, C. S., Christensen, H. M., \u0026amp; Dagon, K. (2024). Pushing the frontiers in climate modelling and analysis with machine learning. \u003cem\u003eNature climate change\u003c/em\u003e,\u003cem\u003e 14\u003c/em\u003e(9), 916-928.\u003c/li\u003e\n\u003cli\u003eHultgren, A., Carleton, T., Delgado, M., Gergel, D. R., Greenstone, M., Houser, T., Hsiang, S., Jina, A., Kopp, R. E., \u0026amp; Malevich, S. B. (2022). Climate change impacts on global agriculture accounting for adaptation. \u003cem\u003eAvailabale at SSRN: http://dx. doi. org/10.2139/ssrn\u003c/em\u003e,\u003cem\u003e 4222020\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eLobell, D. B., Burke, M. B., Tebaldi, C., Mastrandrea, M. D., Falcon, W. P., \u0026amp; Naylor, R. L. (2008). Prioritizing climate change adaptation needs for food security in 2030. \u003cem\u003eScience\u003c/em\u003e,\u003cem\u003e 319\u003c/em\u003e(5863), 607-610.\u003c/li\u003e\n\u003cli\u003eMansfield, L. A., Nowack, P. J., Kasoar, M., Everitt, R. G., Collins, W. J., \u0026amp; Voulgarakis, A. (2020). Predicting global patterns of long-term climate change from short-term simulations using machine learning. \u003cem\u003enpj Climate and Atmospheric Science\u003c/em\u003e,\u003cem\u003e 3\u003c/em\u003e(1), 44.\u003c/li\u003e\n\u003cli\u003eMoghayer, S., Zurek, M., Muzammil, M., Mason-D\u0026rsquo;Croz, D., Magrath, J., Tabeau, A., Vervoort, J. M., \u0026amp; Achterbosch, T. (2023). A low-carbon and hunger-free future for Bangladesh: An ex-ante assessment of synergies and trade-offs in different transition pathways. \u003cem\u003eFrontiers in Environmental Science\u003c/em\u003e,\u003cem\u003e 10\u003c/em\u003e, 977760.\u003c/li\u003e\n\u003cli\u003eRiahi, K., Van Vuuren, D. P., Kriegler, E., Edmonds, J., O\u0026rsquo;neill, B. C., Fujimori, S., Bauer, N., Calvin, K., Dellink, R., \u0026amp; Fricko, O. (2017). The Shared Socioeconomic Pathways and their energy, land use, and greenhouse gas emissions implications: An overview. \u003cem\u003eGlobal Environmental Change\u003c/em\u003e,\u003cem\u003e 42\u003c/em\u003e, 153-168.\u003c/li\u003e\n\u003cli\u003eSamir, K., \u0026amp; Lutz, W. (2017). The human core of the shared socioeconomic pathways: Population scenarios by age, sex and level of education for all countries to 2100. \u003cem\u003eGlobal Environmental Change\u003c/em\u003e,\u003cem\u003e 42\u003c/em\u003e, 181-192.\u003c/li\u003e\n\u003cli\u003eShiferaw, B., Smale, M., Braun, H.-J., Duveiller, E., Reynolds, M., \u0026amp; Muricho, G. (2013). Crops that feed the world 10. Past successes and future challenges to the role played by wheat in global food security. \u003cem\u003eFood Security\u003c/em\u003e,\u003cem\u003e 5\u003c/em\u003e(3), 291-317.\u003c/li\u003e\n\u003cli\u003evan der Velden, B. H. (2024). Explainable AI: current status and future potential. \u003cem\u003eEuropean Radiology\u003c/em\u003e,\u003cem\u003e 34\u003c/em\u003e(2), 1187-1189.\u003c/li\u003e\n\u003cli\u003evan Meijl, H., Tabeau, A., Stehfest, E., Doelman, J., \u0026amp; Lucas, P. (2020). How food secure are the green, rocky and middle roads: food security effects in different world development paths. \u003cem\u003eEnvironmental Research Communications\u003c/em\u003e,\u003cem\u003e 2\u003c/em\u003e(3), 031002.\u003c/li\u003e\n\u003cli\u003eWang, X., Liu, C., \u0026amp; van der Fels-Klerx, H. (2022). Regional prediction of multi-mycotoxin contamination of wheat in Europe using machine learning. \u003cem\u003eFood Research International\u003c/em\u003e,\u003cem\u003e 159\u003c/em\u003e, 111588.\u003c/li\u003e\n\u003cli\u003eXuhui, W., Zhao, C., M\u0026uuml;ller, C., Wang, C., Ciais, P., Janssens, I., Pe\u0026ntilde;uelas, J., Asseng, S., Li, T., \u0026amp; Elliott, J. (2020). Emergent constraint on crop yield response to warmer temperature from field experiments. \u003cem\u003eNature Sustainability\u003c/em\u003e,\u003cem\u003e 3\u003c/em\u003e(11), 908-916.\u003c/li\u003e\n\u003cli\u003eAragie, E. A., McDonald, S., \u0026amp; Thierfelder, K. (2016).\u0026nbsp;A static applied general equilibrium model: Technical documentation: STAGE_DEV Version 2.\u003c/li\u003e\n\u003cli\u003eIFPRI. (2024a.\u0026nbsp;Machine Learning for Spatially Granular Food Security Mapping\u0026nbsp;(Project update). International Food Policy Research Institute.\u003c/li\u003e\n\u003cli\u003eIFPRI. (2024b).\u0026nbsp;Generative AI for Agriculture (GAIA) Project: Enhancing Advisory Services with LLMs\u0026nbsp;(Project update). International Food Policy Research Institute.\u003c/li\u003e\n\u003cli\u003eIFPRI. (2024c).\u0026nbsp;IFPRI's Complementary Modeling Systems: MIRAGRODEP, IMPACT, RIAPA. International Food Policy Research Institute.\u003c/li\u003e\n\u003cli\u003eLundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., ... \u0026amp; Lee, S. I. (2020). From local explanations to global understanding with explainable AI for trees. Nature machine intelligence, 2(1), 56-67.\u003c/li\u003e\n\u003cli\u003eShwartz-Ziv, R., \u0026amp; Armon, A. (2022). Tabular data: Deep learning is not all you need. Information Fusion, 81, 84-90.\u003c/li\u003e\n\u003cli\u003eGoldstein, A., Kapelner, A., Bleich, J., \u0026amp; Pitkin, E. (2015). Peeking inside the black box: Visualizing statistical learning with plots of individual conditional expectation. journal of Computational and Graphical Statistics, 24(1), 44-65.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":true,"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":"Computable General Equilibrium (CGE), Artificial Intelligence, Emulation, Hybrid Modelling, Shared Socioeconomic Pathways (SSPs), Economic Foresight","lastPublishedDoi":"10.21203/rs.3.rs-9405456/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9405456/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"While Computable General Equilibrium (CGE) models such as MAGNET are very useful to informing robust policy decisions, their high computational demands often hinder the exploration of large ensembles of future scenarios. Although artificial intelligence (AI) has been applied to biophysical crop modelling, there is limited evidence regarding its capacity to accurately replicate the core behavioral logic of economic equilibrium models. To address this methodological gap, this study evaluates whether a data-driven AI emulator can serve as a robust surrogate for a structural CGE model, using global wheat production as an illustration application case. We trained the emulator using a historical panel dataset derived from MAGNET, incorporating a set of key macroeconomic drivers, and benchmarked its performance against MAGNET projections under diverse Shared Socioeconomic Pathways (SSPs). The results indicate that the AI emulator reproduces historical production patterns with high accuracy and closely replicates the MAGNET model’s behavior in near term scenario projections. However, distinct divergences emerge in long-term projections, where the AI model follows historical constraints while the CGE framework embeds forward-looking assumptions such as technical progress. Ultimately, this research demonstrates that hybrid CGE-AI frameworks can successfully complement structural economic logic with computational agility, offering a scalable methodology for rapid scenario stress-testing.","manuscriptTitle":"Foresight Scenario Modelling for Global Wheat Production Under Shared Socioeconomic Pathways: Testing AI Emulators of CGE Models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-21 15:29:08","doi":"10.21203/rs.3.rs-9405456/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":"5c10be34-1b40-4fe1-8795-463d9700c666","owner":[],"postedDate":"April 21st, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":66544288,"name":"Scientific community and society/Agriculture"},{"id":66544289,"name":"Scientific community and society/Social sciences"}],"tags":[],"updatedAt":"2026-04-22T22:56:42+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-21 15:29:08","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9405456","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9405456","identity":"rs-9405456","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}