Integrated RSM and machine learning approach for surface topography optimization in SLA 3D-printed functional components

Abstract This study aimed to determine and predict the parameters affecting surface roughness (SR) in stereolithography (SLA) 3D printing processes. Experimentally determined exposure time, lifting speed, and retract speed parameters were used to compare both the traditional Response Surface Method (RSM) and modern Machine Learning (ML) approaches. Due to the small-scale experimental dataset (N = 20), Leave-One-Out Cross Validation (LOOCV) and nested hyperparameter optimization methods were applied to improve the generalizability of the models. In the analyses conducted with the OLS, SVR, LSBoost, and ANN models, the highest prediction accuracy was obtained with the RSM method. However, the OLS and SVR models also showed a strong linear relationship and provided high fit. The findings revealed that retract speed was the most effective parameter on surface roughness. It was concluded that RSM provides high accuracy on small, well-structured datasets, while ML models offer the advantage of flexibility on larger, more complex datasets. The study demonstrates that combining traditional statistical methods with machine learning-based models can be an effective strategy for improving surface quality in 3D printing processes.