Predicting Mechanical Strength in FDM Printed ABS Parts with In-Process Annealing: A Machine Learning Approach

Predicting Mechanical Strength in FDM Printed ABS Parts with In-Process Annealing: A Machine Learning Approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Predicting Mechanical Strength in FDM Printed ABS Parts with In-Process Annealing: A Machine Learning Approach Tanvir Ahmed Shanto This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7437589/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 Fused deposition modeling (FDM), a popular additive manufacturing (AM) technology, is widely used for extruding thermoplastic filaments. Acrylonitrile butadiene styrene (ABS) is a widely used polymer for the FDM technique due to its cost-effectiveness, strong mechanical properties, incredible durability, and excellent thermal stability, making it suitable for functional parts. Nonetheless, low mechanical strength along the Z-X axis from small-scale production to large-scale production of ABS parts is yet to be overcome. This study uses a patent-pending modified heater block assembly to apply in-process thermal load and a conventional brass nozzle to print ASTM D638 Type IV tensile specimens. The effects of these two nozzle types, print speed, and part spacing, are studied on the mechanical properties of the 3D printed samples, ultimate tensile strength, to be specific. The findings show that nozzle type greatly impacts the ultimate tensile strength, with in-situ annealing outperforming conventional nozzle, while effects of part spacing and print speed are less. Various machine learning models are utilized for regression to enhance the process and forecast tensile stress (Decision Tree, Random Forest, Gradient Boosting, and Support Vector Regression). The highest prediction accuracy was attained by Random Forest, demonstrating its applicability for simulating mechanical properties linked to the FDM process. The results highlight the challenges and opportunities of incorporating machine learning into optimizing the FDM process. Future endeavors will investigate sophisticated modeling methods to improve predictive precision by increasing the dataset and considering more process parameters as predictors. Mechanical Engineering Additive Manufacturing Fused Deposition Modeling Machine Learning Supervised Learning In-situ Annealing Mechanical Strength Ultimate Tensile Strength Figures Figure 1 Figure 2 1. Introduction Additive Manufacturing (AM) has revolutionized contemporary manufacturing by facilitating the creation of intricate, lightweight, and customized components [1]. Fused Deposition Modeling (FDM) is one of the most prevalent additive manufacturing processes, recognized for its cost-effectiveness, user-friendliness, and material versatility [2]. Nowadays, FDM is widely employed in prototyping and manufacturing functional components in sectors like aerospace, automotive, marine, and biomedical engineering [3-6]. Among all the materials used in FDM, acrylonitrile butadiene styrene (ABS) is a popular thermoplastic polymer widely used due to its mechanical and thermal properties, excellent dimensional stability, and low glass transition temperature (Tg) [7]. Obtaining high print quality and optimum mechanical strength in thin and tall structures, large-area additive manufacturing, and batch production continue to pose difficulties. The inferior inter-layer adhesion is relative to intra-layer cohesion, resulting from inadequate molecular diffusion and thermal bonding, which results in shrinkage, warpage, and lower mechanical strength in the z-direction. Maintaining the printing layer temperature above the polymer's glass transition temperature throughout the printing process enhances inter-layer adhesion, allowing polymer chains from adjacent layers to diffuse and entangle effectively [8]. Several techniques have been developed to do that. These include near-infrared (NIR) laser-based pre-deposition heating [9], infrared (IR) preheating [10], cold plasma treatment [11], and modified print head assemblies designed to apply localized heating for in-situ annealing [12]. In-situ annealing during printing has surfaced as a potential method to improve inter-layer adhesion. The optimization of FDM process parameters combined with localized heating for in-situ annealing to enhance the mechanical strength of ABS-printed parts remains an unexplored area of research. Design of Experiments (DOE) based experimental studies offer statistical insights into the effects of process parameters on mechanical strength in FDM printed parts. Machine learning (ML) has emerged as an effective tool in FDM research, providing data-driven insights, predictive modeling, and real-time process optimization to improve print quality, mechanical performance, and efficiency. In this study, standard 3D-printed specimens are manufactured with in-situ annealing using a modified heater block nozzle and without in-situ annealing using a conventional nozzle. The specimens are manufactured by varying process parameters, including not only nozzle type but also print speed and sample spacing. Each specimen is subjected to tensile testing, and the obtained ultimate tensile strength (UTS) serves as the target variable while the process parameters are used as input features to train four machine learning models: Decision Tree, Random Forest, Gradient Boosting Regression Tree, and Support Vector Regression for prediction of ultimate tensile strength (UTS) in unseen data. 2. Related Research Various machine learning models and frameworks have been examined to forecast mechanical characteristics, optimize process parameters, improve part quality, and optimize operations in FDM. Nasrin et al. demonstrated that Linear Regression (LR), Ridge Regression (RR), Gaussian Process Regression (GPR), and K-Nearest Neighbors (KNN) can forecast the tensile characteristics of FDM parts utilizing limited experimental data through the integration of an active learning framework [13]. Butt and Mohaghegh demonstrated that digital twins and machine learning models, such as Random Forests (RF) and Convolutional Neural Networks (CNN), may enhance the surface quality and mechanical characteristics of Poly Lactic Acid (PLA) while minimizing trial-and-error [14]. Kumar et al. employed an LR-based machine learning model and finite element analysis (FEA) to investigate the influence of slicing variables on FDM-printed PLA parts' mechanical and surface characteristics [15]. Bauriedel et al. employed infrared images and machine learning models, such as Support Vector Regression (SVR) and RF, to forecast tensile strength in FFF-printed PLA components, aiming to elucidate the influence of nozzle temperature and material on layer adhesion [16]. Kazemi and Steeves forecasted the elastic modulus distributions of FDM-printed PLA via a neural network trained on FEA-simulated data. The model highlighted the potential of machine learning in topology optimization [17]. Belei et al. utilized regression models informed by a factorial design to optimize tensile strength for carbon fiber-reinforced polyamide composites, identifying critical parameters such as layer height and bed temperature [23]. ML has also enabled real-time quality control and error detection. Goh et al. deployed YOLO-based object detection models within a real-time monitoring system to effectively detect and correct extrusion errors, significantly enhancing FDM print quality [24]. Additionally, Ramkumar et al. demonstrated the exceptional accuracy of RF regression in optimizing flexural properties of ABS-Al composites, highlighting the interplay between infill patterns and material characteristics [25].However, despite these advancements, limited research has focused on applying ML techniques to predict mechanical strength specifically along the Z-axis in FDM-printed ABS parts. Moreover, studies on the influence of inter-part spacing in multi-component print setups remain scarce. The current study addresses these gaps by investigating the combined effect of nozzle types including conventional and in-situ annealing nozzles alongside print speed and component spacing. By comparing multiple ML models, this research highlights their capability to predict mechanical strength, thereby building upon existing research while exploring new avenues for additive manufacturing optimization. 3. Experimental Setup The study utilizes a modified heater block assembly to enhance FDM inter-layer adhesion and mechanical properties. It employs a conventional brass nozzle (Fig: 1b) and a patent-pending in-process annealing nozzle assembly (Fig: 1a, 1c). The proprietary nozzle assembly features a heater block and aluminum plate mounted on a heated liquefier, maintaining a controlled temperature at the deposition interface. The aluminum plate's thickness and diameter are optimized for thermal efficiency and minimize inertia and weight. The nozzle height (distance between the nozzle tip and the aluminum heater block bottom) is calibrated at 1 mm to optimize heat transfer and polymer chain diffusion. Localized cooling is crucial, with a strategically placed fan providing localized cooling above the heated liquefier zone. This prevents filament softening in the extrusion chamber's upper regions, preventing clogging and print failures. The proprietary nozzle assembly significantly improves inter-layer molecular diffusion, bond strength, and mechanical properties. It also reduces stress concentrations at filament interfaces and common defects, reducing premature mechanical failure and improving printed specimens' ultimate tensile strength and toughness. Experimental tests were conducted using a Creality Ender 3 Pro printer. ASTM D638 Type IV tensile specimens scaled to 150% of their standard size were printed with ABS filament (1.75 mm diameter, Hatchbox brand). To systematically evaluate the impact of the modified nozzle assembly, print speeds of 1200 mm/min, 1800 mm/min, and 2400 mm/min were selected, alongside variable inter-part spacing (7.5 mm, 12.5 mm, 25 mm, 50 mm, and 75 mm). The printed specimens were tested with a Shimadzu Universal Testing Machine (UTM) (Fig: 1d) to evaluate ultimate tensile strength (MPa). This study used nozzle type, print speed, and dogbone spacing as predictors. The average UTS of 3 dogbones of each batch was taken to avoid the repeatability of predictors. The average UTS was used as the output variable. 4. Methodology Two levels of nozzle type, three levels of print speed, and five levels of dogbone spacing were employed to print 30 batches as shown in Table 1. A statistical study was conducted to determine whether there is a statistically significant difference between the two types of nozzles in the outcome of UTS. Next, before applying the Machine Learning models, the dataset was processed. Nozzle type is a categorical feature (1 and 2) by default to reflect its non-ordinal nature. Initially recorded as 1200 mm/min, 1800 mm/min, and 2400 mm/min, the print speed feature was encoded as ordinal categories (1, 2, and 3, respectively) to represent the increasing speed and lessen the possible influence of its very high raw values on model behavior. Meanwhile, the dogbone spacing feature, a continuous numerical variable, was standardized to have a mean of zero and a standard deviation of one. Since this feature had a large range of values, standardization helps to prevent bias or instability during model training, particularly in models like Support Vector Regression that are sensitive to feature scaling. Specifically, decision tree regression, random forest regression, gradient boosting regression tree, and support vector regression were applied to the processed dataset to measure the predictability of the continuous variable, UTS. 80% of the dataset was used for training each model, while the remaining 20% was used for testing. Table 1: Printing Parameters of samples Input Parameters Level 1 Level 2 Level 3 Level 4 Level 5 Nozzle Type Brass Nozzle Modified nozzle with in-situ annealing capability Print speed (mm/min) 1200 1800 2400 Dogbone Spacing (mm) 7.5 12.5 25 50 75 Decision tree (DT) regression, a non-linear regression technique, recursively splits the dataset into subsets based on feature values. It builds a tree-like model where nodes represent decisions based on feature thresholds and leaves reflect the expected result (average of the target values in the subset) . Non-linearity can be effectively handled; it can function well with small datasets and work efficiently with both categorical and continuous variables [18]. Random forest (RF) regression is an ensemble technique that builds multiple decision trees during training and aggregates their outputs (average for regression) to improve performance [19]. RF reduces overfitting by employing bagging, in which every tree is trained on a random subset of the data, or random feature selection, in which each split considers only a random collection of features. Gradient boosting regression tree (GBRT) is an ensemble technique that generates decision trees one after the other, with each tree employing gradient descent to fix the mistakes of the ones before it [20]. It handles non-linearity, minimizes bias, and captures intricate patterns. Support vector machines are the foundation of support vector regression (SVR), which identifies the function that best fits the data while keeping errors within a predetermined range [21,22]. SVR efficiently models both linear and non-linear connections using a variety of kernels. To evaluate the performance of prediction for each model, R-squared, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) have been considered. R-squared calculates the proportion of variance in the response variable explained by the predictor variables. MAE measures the average absolute difference between actual and predicted values, while the square root of the average squared differences between actual and anticipated values is measured by RMSE. MAPE is a regression metric that calculates the average percentage error between actual and predicted values. A higher R-squared, a lower MAE, a lower RMSE, and a lower MAPE are desirable. 5. Results From the experiments, the average UTS obtained using nozzle 1 and 2 are 16.543 MPa and 22.917 MPa, respectively. Using an independent t-test, a statistical analysis was performed to ascertain whether the difference between the two types of nozzles in the outcome of average UTS is statistically significant. The t-statistics were found to be -9.148 with a p-value of 1.319e-09. Since the p-value is very minimal, we can reject the null hypothesis and infer that nozzle type significantly impacts tensile stress. Hence, nozzle 2 produces a significantly higher average UTS than nozzle 1. The distribution of experimentally obtained tensile stress between the two nozzle types is shown in Figure 2a as a box plot. In addition, Figure 2(b-d) shows scatter plots of experimentally obtained tensile stress against nozzle type and two other predictors, dogbone spacing and print speed. Decision tree (DT) yielded an R-squared of 0.623. The model is expected to suffer from overfitting, resulting in poor generalization. The R-squared significantly improves to 0.820 under the random forest (RF) model. With the gradient boosting regression tree (GBRT), an R-squared of 0.760 is achieved. With a sigmoid kernel, an R-squared of 0.711 is obtained in support vector regression (SVR). However, for all the models, the MAPE is less than 10 which suggests good accuracy. The compared performance among various Machine Learning models is summarized in Table 2. Table 2: Comparative performance of ML models Model R-squared MAE RMSE MAPE DT 0.623 1.755 2.239 8.539 RF 0.820 1.312 1.546 6.157 GBRT 0.760 1.246 1.788 5.678 SVR 0.711 1.570 1.960 7.321 RF has achieved the best prediction with an R-squared of 0.820. With RF, the feature importance test is conducted. Nozzle type is the most important predictor, with an importance of 0.713. Figure 2d shows that tensile stress increases as nozzle 1 is switched to nozzle 2. The importance of dogbone spacing and print speed are 0.202 and 0.085, respectively. The time gap between printing successive layers varies with dogbone spacing and print speed. More extended cooling periods reduce polymer chain diffusion and weaken interlayer adhesion. In-situ annealing mitigates these effects, ensuring sublayer temperature remains near the glass transition temperature, Tg. That is why spacing impacts mechanical properties more than print speed. Some strategies should be implemented to improve the R-squared and enhance the model's predictive power for the average UTS. First, feature selection should be broadened by adding more parameters like infill density, layer thickness, nozzle diameter, nozzle temperature, etc. Second, expanding the dataset is crucial since a machine learning prediction model needs adequate data to identify significant trends and perform well. The current dataset's 30 samples restrict the model's capacity to make precise predictions. By adding more tests to the dataset, overfitting will be lessened, and the model's capacity to identify intricate associations between the features and the tensile stress will be improved. Third, after increasing the dataset, along with these models, XGBoost, CATBoost, artificial neural network (ANN), etc., should be applied to see if prediction improves. 6. Conclusions This study explored the impact of specific parameters of the FDM process on the mechanical strength of additively manufactured ABS parts printed in the Z-X direction. By employing a novel in-situ annealing nozzle alongside a conventional brass nozzle, we investigated their effects on the mechanical strength of parts under varying conditions of print speed and part spacing. The experimental results demonstrated that nozzle type significantly influences ultimate tensile strength (UTS), with in-situ annealing improving layer adhesion and mechanical performance compared to conventional nozzle. To predict the response, UTS, we applied multiple regression-based machine learning (ML) models, including Decision Tree, Random Forest, Gradient Boosting, and Support Vector Regression. Among these, Random Forest achieved the highest predictive accuracy, highlighting its suitability for modeling FDM-related mechanical properties. However, the small dataset used in this study is a limitation as ML models require a large dataset to learn and generalize better. Future research will focus on expanding the dataset to improve predictive accuracy, incorporating additional process parameters such as infill density, layer thickness, nozzle diameter, nozzle temperature, etc., and exploring models like XGBoost, CATBoost, and Neural Networks. Future research will also focus on optimizing tensile stress using the Response Surface Methodology. Moreover, a comprehensive full-factorial Design of Experiments (DOE) will be conducted to rigorously analyze factor-level interactions among process parameters. Additionally, future studies will leverage high-resolution Forward-Looking Infrared (FLIR) camera imaging to investigate spatiotemporal thermal gradients during printing. Declarations Data Availability Data will be made available on request. References M. A. Shahriar, M. H. Kobir, S. Rahman, M. Z. Rahman, and B. Saha, “Overview of additive manufacturing and applications of 3D printed composites,” in Comprehensive Materials Processing , Elsevier, 2024, pp. 58–76. doi: 10.1016/b978-0-323-96020-5.00209-0. Campbell, D. Bourell, and I. Gibson, “Additive Manufacturing: Rapid Prototyping comes of age.” [Online]. Available: https://dspace.lboro.ac.uk/ R. Ahmed, R. Saha Niloy, T. Ahmed Shanto, and M. Rahat Mozumder, “APPLICATION OF FUSED FILAMENT FABRICATION IN MARINE SECTOR, FROM RAPID PROTOTYPING TO FINAL PRODUCT,” 2024. [Online]. Available: https://www.researchgate.net/publication/388682275. M. Jawaad Zulqernine, T. Ahmed Shanto, M. Aniruddah Alam, M. Raihan Uddin, and I. Sharmin Dola, “AN INVESTIGATION ON THE APPLICATIONS OF ADDITIVE MANUFACTURING IN THE MARINE INDUSTRY,” 2024. [Online]. Available: https://www.researchgate.net/publication/388830803 S. Mellor, L. Hao, and D. Zhang, “Additive manufacturing: A framework for implementation,” in International Journal of Production Economics , Mar. 2014, pp. 194–201. doi: 10.1016/j.ijpe.2013.07.008. M. A. Sadat, G. Kremer, and K. J. Min, "A real options model for remanufacturing facility installation decisions," Decision Analytics Journal, vol 7, 100222, 2023. S. K. Selvamani et al. , “3D printing: Overview of ABS evolvement,” in AIP Conference Proceedings , American Institute of Physics Inc., Jan. 2019. doi: 10.1063/1.5085984. I. Gibson, D. Rosen, B. Stucker, and M. Khorasani, “Additive Manufacturing Technologies.” A. K. Ravi, A. Deshpande, and K. H. Hsu, “An in-process laser localized pre-deposition heating approach to inter-layer bond strengthening in extrusion based polymer additive manufacturing,” Journal of Manufacturing Processes , vol. 24, pp. 179–185, Oct. 2016, doi: 10.1016/j.jmapro.2016.08.007. V. Kishore et al. , “Infrared preheating to improve interlayer strength of big area additive manufacturing (BAAM) components,” Additive Manufacturing , vol. 14, pp. 7–12, Mar. 2017, doi: 10.1016/j.addma.2016.11.008. C. C. Shih, M. Burnette, D. Staack, J. Wang, and B. L. Tai, “Effects of cold plasma treatment on interlayer bonding strength in FFF process,” Additive Manufacturing , vol. 25, pp. 104–111, Jan. 2019, doi: 10.1016/j.addma.2018.11.005. P. Patel, R. Rane, M. Mrinal, V. Ganesan, R. Taylor, and A. Jain, “Characterization of the effect of in-process annealing using a novel print head assembly on the ultimate tensile strength & toughness of Fused Filament Fabrication (FFF) parts,” Virtual and Physical Prototyping , vol. 17, no. 4, pp. 989–1005, 2022, doi: 10.1080/17452759.2022.2095288. T. Nasrin, M. Pourali, F. Pourkamali-Anaraki, and A. M. Peterson, “Active learning for prediction of tensile properties for material extrusion additive manufacturing,” Scientific Reports , vol. 13, no. 1, Dec. 2023, doi: 10.1038/s41598-023-38527-6. J. Butt and V. Mohaghegh, “Combining Digital Twin and Machine Learning for the Fused Filament Fabrication Process,” Metals , vol. 13, no. 1, Jan. 2023, doi: 10.3390/met13010024. A. Kumar, K. S. Boparai, J. S. Chohan, and R. Singh, “Investigations and predictions for mechanical and surface properties of FFF prints using DOE, ML and FEA,” Advances in Materials and Processing Technologies , vol. 10, no. 3, pp. 1767–1783, 2024, doi: 10.1080/2374068X.2023.2201089. N. Bauriedel, R. Q. Albuquerque, J. Utz, N. Geis, and H. Ruckdäschel, “Monitoring of fused filament fabrication (FFF): An infrared imaging and machine learning approach,” Journal of Polymer Science , Dec. 2024, doi: 10.1002/pol.20240586. Z. Kazemi and C. A. Steeves, “Machine learning for characterizing uncertain elastic properties of fused filament fabricated materials for topology optimization applications,” Aug. 2024, [Online]. Available: http://arxiv.org/abs/2408.05850. A. Raza, A. Haider, and W. Haider, “A Consolidated Approach towards Application of Machine Learning Principles in Additive Manufacturing,” Electro/Information Technology , pp. 363–368, May 2021, doi: 10.1109/EIT51626.2021.9491833. A. Deka and J. Hall, “Predictive modeling of mechanical properties for fused deposition modeling parts: A focus on processing and environmental parameters,” Manufacturing letters , Aug. 2023, doi: 10.1016/j.mfglet.2023.08.053. G. Barrionuevo, J. Ramos‐Grez, and F. Montero, “Machine Learning Regressors in Forecasting Mechanical Properties in Advanced Manufacturing Processes,” Springer International Publishing, 2024, pp. 279–292. doi: 10.1007/978-3-031-52255-0_20. H. Kaneko, “Support vector regression that takes into consideration the importance of explanatory variables,” Journal of Chemometrics , vol. 35, no. 4, Apr. 2021, doi: 10.1002/CEM.3327. R. Saha Niloy, R. Ahmed, M. Shariful Islam, A. Jahin, and M. Rahat Mozumder, “Machine Learning-Based Resistance Prediction of AMECRC Hull,” pp. 2024–108, 2024, doi: 10.5281/zenodo.14213316. Belei, C., Joeressen, J., & Amancio-Filho, S. T. (2022). Fused-Filament Fabrication of Short Carbon Fiber-Reinforced Polyamide: Parameter Optimization for Improved Performance under Uniaxial Tensile Loading. Polymers, 14(7). https://doi.org/10.3390/polym14071292 Goh, G. D., Hamzah, N. M. bin, & Yeong, W. Y. (2023). Anomaly Detection in Fused Filament Fabrication Using Machine Learning. 3D Printing and Additive Manufacturing, 10(3), 428–437. https://doi.org/10.1089/3dp.2021.0231 Ranjan, N., Kumar, R., Kumar, R., Kaur, R., & Singh, S. (2023). Investigation of Fused Filament Fabrication-Based Manufacturing of ABS-Al Composite Structures: Prediction by Machine Learning and Optimization. Journal of Materials Engineering and Performance, 32(10), 4555–4574. https://doi.org/10.1007/s11665-022-07431-x Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7437589","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":504401772,"identity":"6fd6b30f-0f34-43a3-abc6-8d571040346b","order_by":0,"name":"Tanvir Ahmed Shanto","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5ElEQVRIiWNgGAWjYHAC9g8SBhJyyCIGeNXzMDCwMVhU2BgjqyZCS8WZtMQGorXYS6Q/e3Cz7XD62vYzxh8Y2/7IM7A3b5PAawvPGXPDmW2Hc7edyTGTYGwzMGzgOVaGXwt7D4O0JEjLgbQ0BqAWxgYJoF68WpjZH0j/BTrM7PyzZKDDDOwb5N8Q0MLeYCYhcSYtwexG8gGQwxIbJHgIaDlzxthAosLGcNuNx8ckEs4ZJ7fxpBVb4NPCPiP94QNgVMqbnU9s/vChTM62n/3wxhv4tKCCBAZQNI2CUTAKRsEooBgAAJVdRl1yYWMNAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0009-0006-8243-5690","institution":"","correspondingAuthor":true,"prefix":"","firstName":"Tanvir","middleName":"Ahmed","lastName":"Shanto","suffix":""}],"badges":[],"createdAt":"2025-08-22 22:29:26","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7437589/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7437589/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":89885445,"identity":"509e5293-c9b1-4870-8ebe-ce71e07fb56d","added_by":"auto","created_at":"2025-08-26 06:26:01","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":199019,"visible":true,"origin":"","legend":"\u003cp\u003eModified printhead assembly (a), printing with regular brass nozzle (b) and modified-printhead (c), and tensile testing (d)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7437589/v1/356e35fd554d1aaeacb11f35.png"},{"id":89885444,"identity":"e394dd13-2412-4cda-80fd-2dd0e1b079ee","added_by":"auto","created_at":"2025-08-26 06:26:01","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":45400,"visible":true,"origin":"","legend":"\u003cp\u003eBoxplot of maximum tensile stress for different nozzle types (a), and Scatter plot of maximum tensile stress against different predictors (b-d)\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7437589/v1/b9f15c22842ae9db79767cb8.png"},{"id":89885834,"identity":"12a4e138-f902-4e98-94e3-9840e51dad6d","added_by":"auto","created_at":"2025-08-26 06:34:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":601952,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7437589/v1/9e96b213-2a62-41e6-92fc-5eaa5781a0a7.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003ePredicting Mechanical Strength in FDM Printed ABS Parts with In-Process Annealing: A Machine Learning Approach\u003c/p\u003e","fulltext":[{"header":"1.\tIntroduction","content":"\u003cp\u003eAdditive Manufacturing (AM) has revolutionized contemporary manufacturing by facilitating the creation of intricate, lightweight, and customized components [1]. Fused Deposition Modeling (FDM) is one of the most prevalent additive manufacturing processes, recognized for its cost-effectiveness, user-friendliness, and material versatility [2]. Nowadays, FDM is widely employed in prototyping and manufacturing functional components in sectors like aerospace, automotive, marine, and biomedical engineering [3-6]. Among all the materials used in FDM, acrylonitrile butadiene styrene (ABS) is a popular thermoplastic polymer widely used due to its mechanical and thermal properties, excellent dimensional stability, and low glass transition temperature (Tg) [7].\u003c/p\u003e\n\u003cp\u003eObtaining high print quality and optimum mechanical strength in thin and tall structures, large-area additive manufacturing, and batch production continue to pose difficulties. The inferior inter-layer adhesion is relative to intra-layer cohesion, resulting from inadequate molecular diffusion and thermal bonding, which results in shrinkage, warpage, and lower mechanical strength in the z-direction. Maintaining the printing layer temperature above the polymer's glass transition temperature throughout the printing process enhances inter-layer adhesion, allowing polymer chains from adjacent layers to diffuse and entangle effectively [8]. Several techniques have been developed to do that. These include near-infrared (NIR) laser-based pre-deposition heating [9], infrared (IR) preheating [10], cold plasma treatment [11], and modified print head assemblies designed to apply localized heating for in-situ annealing [12]. In-situ annealing during printing has surfaced as a potential method to improve inter-layer adhesion. The optimization of FDM process parameters combined with localized heating for in-situ annealing to enhance the mechanical strength of ABS-printed parts remains an unexplored area of research. Design of Experiments (DOE) based experimental studies offer statistical insights into the effects of process parameters on mechanical strength in FDM printed parts. Machine learning (ML) has emerged as an effective tool in FDM research, providing data-driven insights, predictive modeling, and real-time process optimization to improve print quality, mechanical performance, and efficiency. In this study, standard 3D-printed specimens are manufactured with in-situ annealing using a modified heater block nozzle and without in-situ annealing using a conventional nozzle. The specimens are manufactured by varying process parameters, including not only nozzle type but also print speed and sample spacing. Each specimen is subjected to tensile testing, and the obtained ultimate tensile strength (UTS) serves as the target variable while the process parameters are used as input features to train four machine learning models: Decision Tree, Random Forest, Gradient Boosting Regression Tree, and Support Vector Regression for prediction of ultimate tensile strength (UTS) in unseen data.\u003c/p\u003e"},{"header":"2.\tRelated Research","content":"\u003cp\u003eVarious machine learning models and frameworks have been examined to forecast mechanical characteristics, optimize process parameters, improve part quality, and optimize operations in FDM. Nasrin et al. demonstrated that Linear Regression (LR), Ridge Regression (RR), Gaussian Process Regression (GPR), and K-Nearest Neighbors (KNN) can forecast the tensile characteristics of FDM parts utilizing limited experimental data through the integration of an active learning framework [13]. Butt and Mohaghegh demonstrated that digital twins and machine learning models, such as Random Forests (RF) and Convolutional Neural Networks (CNN), may enhance the surface quality and mechanical characteristics of Poly Lactic Acid (PLA) while minimizing trial-and-error [14]. Kumar et al. employed an LR-based machine learning model and finite element analysis (FEA) to investigate the influence of slicing variables on FDM-printed PLA parts\u0026apos; mechanical and surface characteristics [15]. Bauriedel et al. employed infrared images and machine learning models, such as Support Vector Regression (SVR) and RF, to forecast tensile strength in FFF-printed PLA components, aiming to elucidate the influence of nozzle temperature and material on layer adhesion [16]. Kazemi and Steeves forecasted the elastic modulus distributions of FDM-printed PLA via a neural network trained on FEA-simulated data. The model highlighted the potential of machine learning in topology optimization [17]. Belei et al. utilized regression models informed by a factorial design to optimize tensile strength for carbon fiber-reinforced polyamide composites, identifying critical parameters such as layer height and bed temperature [23]. ML has also enabled real-time quality control and error detection. Goh et al. deployed YOLO-based object detection models within a real-time monitoring system to effectively detect and correct extrusion errors, significantly enhancing FDM print quality [24]. Additionally, Ramkumar et al. demonstrated the exceptional accuracy of RF regression in optimizing flexural properties of ABS-Al composites, highlighting the interplay between infill patterns and material characteristics [25].However, despite these advancements, limited research has focused on applying ML techniques to predict mechanical strength specifically along the Z-axis in FDM-printed ABS parts. Moreover, studies on the influence of inter-part spacing in multi-component print setups remain scarce. The current study addresses these gaps by investigating the combined effect of nozzle types including conventional and in-situ annealing nozzles alongside print speed and component spacing. By comparing multiple ML models, this research highlights their capability to predict mechanical strength, thereby building upon existing research while exploring new avenues for additive manufacturing optimization.\u003c/p\u003e"},{"header":"3.\tExperimental Setup","content":"\u003cp\u003eThe study utilizes a modified heater block assembly to enhance FDM inter-layer adhesion and mechanical properties. It employs a conventional brass nozzle (Fig: 1b) and a patent-pending in-process annealing nozzle assembly (Fig: 1a, 1c). The proprietary nozzle assembly features a heater block and aluminum plate mounted on a heated liquefier, maintaining a controlled temperature at the deposition interface. The aluminum plate\u0026apos;s thickness and diameter are optimized for thermal efficiency and minimize inertia and weight. The nozzle height (distance between the nozzle tip and the aluminum heater block bottom) is calibrated at 1 mm to optimize heat transfer and polymer chain diffusion. Localized cooling is crucial, with a strategically placed fan providing localized cooling above the heated liquefier zone. This prevents filament softening in the extrusion chamber\u0026apos;s upper regions, preventing clogging and print failures. The proprietary nozzle assembly significantly improves inter-layer molecular diffusion, bond strength, and mechanical properties. It also reduces stress concentrations at filament interfaces and common defects, reducing premature mechanical failure and improving printed specimens\u0026apos; ultimate tensile strength and toughness.\u003c/p\u003e\n\u003cp\u003eExperimental tests were conducted using a Creality Ender 3 Pro printer. ASTM D638 Type IV tensile specimens scaled to 150% of their standard size were printed with ABS filament (1.75 mm diameter, Hatchbox brand). To systematically evaluate the impact of the modified nozzle assembly, print speeds of 1200 mm/min, 1800 mm/min, and 2400 mm/min were selected, alongside variable inter-part spacing (7.5 mm, 12.5 mm, 25 mm, 50 mm, and 75 mm). The printed specimens were tested with a Shimadzu Universal Testing Machine (UTM) (Fig: 1d) to evaluate ultimate tensile strength (MPa). This study used nozzle type, print speed, and dogbone spacing as predictors. The average UTS of 3 dogbones of each batch was taken to avoid the repeatability of predictors. The average UTS was used as the output variable.\u003c/p\u003e"},{"header":"4. Methodology","content":"\u003cp\u003eTwo levels of nozzle type, three levels of print speed, and five levels of dogbone spacing were employed to print 30 batches as shown in Table 1. A statistical study was conducted to determine whether there is a statistically significant difference between the two types of nozzles in the outcome of UTS. Next, before applying the Machine Learning models, the dataset was processed. Nozzle type is a categorical feature (1 and 2) by default to reflect its non-ordinal nature. Initially recorded as 1200 mm/min, 1800 mm/min, and 2400 mm/min, the print speed feature was encoded as ordinal categories (1, 2, and 3, respectively) to represent the increasing speed and lessen the possible influence of its very high raw values on model behavior. Meanwhile, the dogbone spacing feature, a continuous numerical variable, was standardized to have a mean of zero and a standard deviation of one. Since this feature had a large range of values, standardization helps to prevent bias or instability during model training, particularly in models like Support Vector Regression that are sensitive to feature scaling. Specifically, decision tree regression, random forest regression, gradient boosting regression tree, and support vector regression were applied to the processed dataset to measure the predictability of the continuous variable, UTS. 80% of the dataset was used for training each model, while the remaining 20% was used for testing.\u003c/p\u003e\n\u003cp\u003eTable 1: Printing Parameters of samples\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 169px;\"\u003e\n \u003cp\u003eInput Parameters\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 65px;\"\u003e\n \u003cp\u003eLevel 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 186px;\"\u003e\n \u003cp\u003eLevel 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003eLevel 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eLevel 4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003eLevel 5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 169px;\"\u003e\n \u003cp\u003eNozzle Type\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 65px;\"\u003e\n \u003cp\u003eBrass Nozzle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 186px;\"\u003e\n \u003cp\u003eModified nozzle with in-situ annealing capability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 169px;\"\u003e\n \u003cp\u003ePrint speed (mm/min)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 65px;\"\u003e\n \u003cp\u003e1200\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 186px;\"\u003e\n \u003cp\u003e1800\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e2400\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 169px;\"\u003e\n \u003cp\u003eDogbone Spacing (mm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 65px;\"\u003e\n \u003cp\u003e7.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 186px;\"\u003e\n \u003cp\u003e12.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eDecision tree (DT) regression, a non-linear regression technique, recursively splits the dataset into subsets based on feature values. It builds a tree-like model where nodes represent decisions based on feature thresholds and leaves reflect the expected result (average of the target values in the subset) . Non-linearity can be effectively handled; it can function well with small datasets and work efficiently with both categorical and continuous variables [18]. Random forest (RF) regression is an ensemble technique that builds multiple decision trees during training and aggregates their outputs (average for regression) to improve performance [19]. RF reduces overfitting by employing bagging, in which every tree is trained on a random subset of the data, or random feature selection, in which each split considers only a random collection of features. Gradient boosting regression tree (GBRT) is an ensemble technique that generates decision trees one after the other, with each tree employing gradient descent to fix the mistakes of the ones before it [20]. It handles non-linearity, minimizes bias, and captures intricate patterns. Support vector machines are the foundation of support vector regression (SVR), which identifies the function that best fits the data while keeping errors within a predetermined range [21,22]. SVR efficiently models both linear and non-linear connections using a variety of kernels. To evaluate the performance of prediction for each model, R-squared, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) have been considered. R-squared calculates the proportion of variance in the response variable explained by the predictor variables. MAE measures the average absolute difference between actual and predicted values, while the square root of the average squared differences between actual and anticipated values is measured by RMSE. MAPE is a regression metric that calculates the average percentage error between actual and predicted values. A higher R-squared, a lower MAE, a lower RMSE, and a lower MAPE are desirable.\u003c/p\u003e"},{"header":"5.\tResults","content":"\u003cp\u003eFrom the experiments, the average UTS obtained using nozzle 1 and 2 are 16.543 MPa and 22.917 MPa, respectively. Using an independent t-test, a statistical analysis was performed to ascertain whether the difference between the two types of nozzles in the outcome of average UTS is statistically significant. The t-statistics were found to be -9.148 with a p-value of 1.319e-09. Since the p-value is very minimal, we can reject the null hypothesis and infer that nozzle type significantly impacts tensile stress. Hence, nozzle 2 produces a significantly higher average UTS than nozzle 1. The distribution of experimentally obtained tensile stress between the two nozzle types is shown in Figure 2a as a box plot. In addition, Figure 2(b-d) shows scatter plots of experimentally obtained tensile stress against nozzle type and two other predictors, dogbone spacing and print speed.\u003c/p\u003e\n\u003cp\u003eDecision tree (DT) yielded an R-squared of 0.623. The model is expected to suffer from overfitting, resulting in poor generalization. The R-squared significantly improves to 0.820 under the random forest (RF) model. With the gradient boosting regression tree (GBRT), an R-squared of 0.760 is achieved. With a sigmoid kernel, an R-squared of 0.711 is obtained in support vector regression (SVR). However, for all the models, the MAPE is less than 10 which suggests good accuracy. The compared performance among various Machine Learning models is summarized in Table 2.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Table 2: Comparative performance of ML models\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 139px;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 109px;\"\u003e\n \u003cp\u003eR-squared\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 70px;\"\u003e\n \u003cp\u003eMAE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 71px;\"\u003e\n \u003cp\u003eRMSE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003eMAPE\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 139px;\"\u003e\n \u003cp\u003eDT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 109px;\"\u003e\n \u003cp\u003e0.623\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 70px;\"\u003e\n \u003cp\u003e1.755\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 71px;\"\u003e\n \u003cp\u003e2.239\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e8.539\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 139px;\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 109px;\"\u003e\n \u003cp\u003e0.820\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 70px;\"\u003e\n \u003cp\u003e1.312\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 71px;\"\u003e\n \u003cp\u003e1.546\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e6.157\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 139px;\"\u003e\n \u003cp\u003eGBRT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 109px;\"\u003e\n \u003cp\u003e0.760\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 70px;\"\u003e\n \u003cp\u003e1.246\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 71px;\"\u003e\n \u003cp\u003e1.788\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e5.678\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 139px;\"\u003e\n \u003cp\u003eSVR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 109px;\"\u003e\n \u003cp\u003e0.711\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 70px;\"\u003e\n \u003cp\u003e1.570\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 71px;\"\u003e\n \u003cp\u003e1.960\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e7.321\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eRF has achieved the best prediction with an R-squared of 0.820. With RF, the feature importance test is conducted. Nozzle type is the most important predictor, with an importance of 0.713. Figure 2d shows that tensile stress increases as nozzle 1 is switched to nozzle 2. The importance of dogbone spacing and print speed are 0.202 and 0.085, respectively. The time gap between printing successive layers varies with dogbone spacing and print speed. More extended cooling periods reduce polymer chain diffusion and weaken interlayer adhesion. In-situ annealing mitigates these effects, ensuring sublayer temperature remains near the glass transition temperature, Tg. That is why spacing impacts mechanical properties more than print speed.\u003c/p\u003e\n\u003cp\u003eSome strategies should be implemented to improve the R-squared and enhance the model\u0026apos;s predictive power for the average UTS. First, feature selection should be broadened by adding more parameters like infill density, layer thickness, nozzle diameter, nozzle temperature, etc. Second, expanding the dataset is crucial since a machine learning prediction model needs adequate data to identify significant trends and perform well. The current dataset\u0026apos;s 30 samples restrict the model\u0026apos;s capacity to make precise predictions. By adding more tests to the dataset, overfitting will be lessened, and the model\u0026apos;s capacity to identify intricate associations between the features and the tensile stress will be improved. Third, after increasing the dataset, along with these models, XGBoost, CATBoost, artificial neural network (ANN), etc., should be applied to see if prediction improves.\u003c/p\u003e"},{"header":"6. Conclusions","content":"\u003cp\u003eThis study explored the impact of specific parameters of the FDM process on the mechanical strength of additively manufactured ABS parts printed in the Z-X direction. By employing a novel in-situ annealing nozzle alongside a conventional brass nozzle, we investigated their effects on the mechanical strength of parts under varying conditions of print speed and part spacing. The experimental results demonstrated that nozzle type significantly influences ultimate tensile strength (UTS), with in-situ annealing improving layer adhesion and mechanical performance compared to conventional nozzle. To predict the response, UTS, we applied multiple regression-based machine learning (ML) models, including Decision Tree, Random Forest, Gradient Boosting, and Support Vector Regression. Among these, Random Forest achieved the highest predictive accuracy, highlighting its suitability for modeling FDM-related mechanical properties. However, the small dataset used in this study is a limitation as ML models require a large dataset to learn and generalize better. Future research will focus on expanding the dataset to improve predictive accuracy, incorporating additional process parameters such as infill density, layer thickness, nozzle diameter, nozzle temperature, etc., and exploring models like XGBoost, CATBoost, and Neural Networks. Future research will also focus on optimizing tensile stress using the Response Surface Methodology. Moreover, a comprehensive full-factorial Design of Experiments (DOE) will be conducted to rigorously analyze factor-level interactions among process parameters. Additionally, future studies will leverage high-resolution Forward-Looking Infrared (FLIR) camera imaging to investigate spatiotemporal thermal gradients during printing.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData will be made available on request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eM. A. Shahriar, M. H. Kobir, S. Rahman, M. Z. Rahman, and B. Saha, \u0026ldquo;Overview of additive manufacturing and applications of 3D printed composites,\u0026rdquo; in \u003cem\u003eComprehensive Materials Processing\u003c/em\u003e, Elsevier, 2024, pp. 58\u0026ndash;76. doi: 10.1016/b978-0-323-96020-5.00209-0.\u003c/li\u003e\n \u003cli\u003eCampbell, D. Bourell, and I. Gibson, \u0026ldquo;Additive Manufacturing: Rapid Prototyping comes of age.\u0026rdquo; [Online]. Available: https://dspace.lboro.ac.uk/\u003c/li\u003e\n \u003cli\u003eR. Ahmed, R. Saha Niloy, T. Ahmed Shanto, and M. Rahat Mozumder, \u0026ldquo;APPLICATION OF FUSED FILAMENT FABRICATION IN MARINE SECTOR, FROM RAPID PROTOTYPING TO FINAL PRODUCT,\u0026rdquo; 2024. [Online]. Available: https://www.researchgate.net/publication/388682275.\u003c/li\u003e\n \u003cli\u003eM. Jawaad Zulqernine, T. Ahmed Shanto, M. Aniruddah Alam, M. Raihan Uddin, and I. Sharmin Dola, \u0026ldquo;AN INVESTIGATION ON THE APPLICATIONS OF ADDITIVE MANUFACTURING IN THE MARINE INDUSTRY,\u0026rdquo; 2024. [Online]. Available: https://www.researchgate.net/publication/388830803\u003c/li\u003e\n \u003cli\u003eS. Mellor, L. Hao, and D. Zhang, \u0026ldquo;Additive manufacturing: A framework for implementation,\u0026rdquo; in \u003cem\u003eInternational Journal of Production Economics\u003c/em\u003e, Mar. 2014, pp. 194\u0026ndash;201. doi: 10.1016/j.ijpe.2013.07.008.\u003c/li\u003e\n \u003cli\u003eM. A. Sadat, G. Kremer, and K. J. Min, \u0026quot;A real options model for remanufacturing facility installation decisions,\u0026quot; Decision Analytics Journal, vol 7, 100222, 2023.\u003c/li\u003e\n \u003cli\u003eS. K. Selvamani \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;3D printing: Overview of ABS evolvement,\u0026rdquo; in \u003cem\u003eAIP Conference Proceedings\u003c/em\u003e, American Institute of Physics Inc., Jan. 2019. doi: 10.1063/1.5085984.\u003c/li\u003e\n \u003cli\u003eI. Gibson, D. Rosen, B. Stucker, and M. Khorasani, \u0026ldquo;Additive Manufacturing Technologies.\u0026rdquo;\u003c/li\u003e\n \u003cli\u003eA. K. Ravi, A. Deshpande, and K. H. Hsu, \u0026ldquo;An in-process laser localized pre-deposition heating approach to inter-layer bond strengthening in extrusion based polymer additive manufacturing,\u0026rdquo; \u003cem\u003eJournal of Manufacturing Processes\u003c/em\u003e, vol. 24, pp. 179\u0026ndash;185, Oct. 2016, doi: 10.1016/j.jmapro.2016.08.007.\u003c/li\u003e\n \u003cli\u003eV. Kishore \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Infrared preheating to improve interlayer strength of big area additive manufacturing (BAAM) components,\u0026rdquo; \u003cem\u003eAdditive Manufacturing\u003c/em\u003e, vol. 14, pp. 7\u0026ndash;12, Mar. 2017, doi: 10.1016/j.addma.2016.11.008.\u003c/li\u003e\n \u003cli\u003eC. C. Shih, M. Burnette, D. Staack, J. Wang, and B. L. Tai, \u0026ldquo;Effects of cold plasma treatment on interlayer bonding strength in FFF process,\u0026rdquo; \u003cem\u003eAdditive Manufacturing\u003c/em\u003e, vol. 25, pp. 104\u0026ndash;111, Jan. 2019, doi: 10.1016/j.addma.2018.11.005.\u003c/li\u003e\n \u003cli\u003eP. Patel, R. Rane, M. Mrinal, V. Ganesan, R. Taylor, and A. Jain, \u0026ldquo;Characterization of the effect of in-process annealing using a novel print head assembly on the ultimate tensile strength \u0026amp; toughness of Fused Filament Fabrication (FFF) parts,\u0026rdquo; \u003cem\u003eVirtual and Physical Prototyping\u003c/em\u003e, vol. 17, no. 4, pp. 989\u0026ndash;1005, 2022, doi: 10.1080/17452759.2022.2095288.\u003c/li\u003e\n \u003cli\u003eT. Nasrin, M. Pourali, F. Pourkamali-Anaraki, and A. M. Peterson, \u0026ldquo;Active learning for prediction of tensile properties for material extrusion additive manufacturing,\u0026rdquo; \u003cem\u003eScientific Reports\u003c/em\u003e, vol. 13, no. 1, Dec. 2023, doi: 10.1038/s41598-023-38527-6.\u003c/li\u003e\n \u003cli\u003eJ. Butt and V. Mohaghegh, \u0026ldquo;Combining Digital Twin and Machine Learning for the Fused Filament Fabrication Process,\u0026rdquo; \u003cem\u003eMetals\u003c/em\u003e, vol. 13, no. 1, Jan. 2023, doi: 10.3390/met13010024.\u003c/li\u003e\n \u003cli\u003eA. Kumar, K. S. Boparai, J. S. Chohan, and R. Singh, \u0026ldquo;Investigations and predictions for mechanical and surface properties of FFF prints using DOE, ML and FEA,\u0026rdquo; \u003cem\u003eAdvances in Materials and Processing Technologies\u003c/em\u003e, vol. 10, no. 3, pp. 1767\u0026ndash;1783, 2024, doi: 10.1080/2374068X.2023.2201089.\u003c/li\u003e\n \u003cli\u003eN. Bauriedel, R. Q. Albuquerque, J. Utz, N. Geis, and H. Ruckd\u0026auml;schel, \u0026ldquo;Monitoring of fused filament fabrication (FFF): An infrared imaging and machine learning approach,\u0026rdquo; \u003cem\u003eJournal of Polymer Science\u003c/em\u003e, Dec. 2024, doi: 10.1002/pol.20240586.\u003c/li\u003e\n \u003cli\u003eZ. Kazemi and C. A. Steeves, \u0026ldquo;Machine learning for characterizing uncertain elastic properties of fused filament fabricated materials for topology optimization applications,\u0026rdquo; Aug. 2024, [Online]. Available: http://arxiv.org/abs/2408.05850.\u003c/li\u003e\n \u003cli\u003eA. Raza, A. Haider, and W. Haider, \u0026ldquo;A Consolidated Approach towards Application of Machine Learning Principles in Additive Manufacturing,\u0026rdquo; \u003cem\u003eElectro/Information Technology\u003c/em\u003e, pp. 363\u0026ndash;368, May 2021, doi: 10.1109/EIT51626.2021.9491833.\u003c/li\u003e\n \u003cli\u003eA. Deka and J. Hall, \u0026ldquo;Predictive modeling of mechanical properties for fused deposition modeling parts: A focus on processing and environmental parameters,\u0026rdquo; \u003cem\u003eManufacturing letters\u003c/em\u003e, Aug. 2023, doi: 10.1016/j.mfglet.2023.08.053.\u003c/li\u003e\n \u003cli\u003eG. Barrionuevo, J. Ramos‐Grez, and F. Montero, \u0026ldquo;Machine Learning Regressors in Forecasting Mechanical Properties in Advanced Manufacturing Processes,\u0026rdquo; Springer International Publishing, 2024, pp. 279\u0026ndash;292. doi: 10.1007/978-3-031-52255-0_20.\u003c/li\u003e\n \u003cli\u003eH. Kaneko, \u0026ldquo;Support vector regression that takes into consideration the importance of explanatory variables,\u0026rdquo; \u003cem\u003eJournal of Chemometrics\u003c/em\u003e, vol. 35, no. 4, Apr. 2021, doi: 10.1002/CEM.3327.\u003c/li\u003e\n \u003cli\u003eR. Saha Niloy, R. Ahmed, M. Shariful Islam, A. Jahin, and M. Rahat Mozumder, \u0026ldquo;Machine Learning-Based Resistance Prediction of AMECRC Hull,\u0026rdquo; pp. 2024\u0026ndash;108, 2024, doi: 10.5281/zenodo.14213316.\u003c/li\u003e\n \u003cli\u003eBelei, C., Joeressen, J., \u0026amp; Amancio-Filho, S. T. (2022). Fused-Filament Fabrication of Short Carbon Fiber-Reinforced Polyamide: Parameter Optimization for Improved Performance under Uniaxial Tensile Loading. Polymers, 14(7). https://doi.org/10.3390/polym14071292\u003c/li\u003e\n \u003cli\u003eGoh, G. D., Hamzah, N. M. bin, \u0026amp; Yeong, W. Y. (2023). Anomaly Detection in Fused Filament Fabrication Using Machine Learning. 3D Printing and Additive Manufacturing, 10(3), 428\u0026ndash;437. https://doi.org/10.1089/3dp.2021.0231\u003c/li\u003e\n \u003cli\u003eRanjan, N., Kumar, R., Kumar, R., Kaur, R., \u0026amp; Singh, S. (2023). Investigation of Fused Filament Fabrication-Based Manufacturing of ABS-Al Composite Structures: Prediction by Machine Learning and Optimization. Journal of Materials Engineering and Performance, 32(10), 4555\u0026ndash;4574. https://doi.org/10.1007/s11665-022-07431-x\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"The University of Texas at Arlington","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Additive Manufacturing, Fused Deposition Modeling, Machine Learning, Supervised Learning, In-situ Annealing, Mechanical Strength, Ultimate Tensile Strength","lastPublishedDoi":"10.21203/rs.3.rs-7437589/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7437589/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eFused deposition modeling (FDM), a popular additive manufacturing (AM) technology, is widely used for extruding thermoplastic filaments. Acrylonitrile butadiene styrene (ABS) is a widely used polymer for the FDM technique due to its cost-effectiveness, strong mechanical properties, incredible durability, and excellent thermal stability, making it suitable for functional parts. Nonetheless, low mechanical strength along the Z-X axis from small-scale production to large-scale production of ABS parts is yet to be overcome. This study uses a patent-pending modified heater block assembly to apply in-process thermal load and a conventional brass nozzle to print ASTM D638 Type IV tensile specimens. The effects of these two nozzle types, print speed, and part spacing, are studied on the mechanical properties of the 3D printed samples, ultimate tensile strength, to be specific. The findings show that nozzle type greatly impacts the ultimate tensile strength, with in-situ annealing outperforming conventional nozzle, while effects of part spacing and print speed are less. Various machine learning models are utilized for regression to enhance the process and forecast tensile stress (Decision Tree, Random Forest, Gradient Boosting, and Support Vector Regression). The highest prediction accuracy was attained by Random Forest, demonstrating its applicability for simulating mechanical properties linked to the FDM process. The results highlight the challenges and opportunities of incorporating machine learning into optimizing the FDM process. Future endeavors will investigate sophisticated modeling methods to improve predictive precision by increasing the dataset and considering more process parameters as predictors.\u003c/p\u003e","manuscriptTitle":"Predicting Mechanical Strength in FDM Printed ABS Parts with In-Process Annealing: A Machine Learning Approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-26 06:25:57","doi":"10.21203/rs.3.rs-7437589/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":"fa231b1c-5da7-4ba5-bf7d-e37cbf82b036","owner":[],"postedDate":"August 26th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":53592089,"name":"Mechanical Engineering"}],"tags":[],"updatedAt":"2025-08-26T06:25:57+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-26 06:25:57","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7437589","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7437589","identity":"rs-7437589","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}