Machine learning-driven prediction of Gaussian curvature deviation in reverse engineered freeform surfaces

Abstract The present study was undertaken to build a machine learning-based framework for predicting Gaussian Curvature Deviation (GCD) in reverse engineered freeform surfaces. The predictive capabilities of Support Vector Regression (SVR) and Random Forest Regression (RFR) models were explored to overcome the limitations of traditional error assessment methods. The models were trained with experimental GCD data generated by varying Noise Reduction Level (NRL), Triangle Count Percentage (TCP), Smoothing Level (SL), and Sample Percentage (SP). Feature engineering was employed by including higher-order interaction terms to capture nonlinear dependencies inherent in freeform geometries. An exhaustive grid search combined with cross-validation was employed to explore the hyperparameter space and identify the configurations that yielded the best predictive performance for each model. SVR surpassed RFR in terms of predictive accuracy and generalization by achieving a testing R 2 of 0.8801, RMSE of 0.0024, and MAPE of 3.96%. RFR exhibited a higher training R 2 (0.9168) but a lower testing R 2 (0.6323) indicating mild overfitting. Qualitative analyses through scatter plots also confirmed the superior predictive fidelity of SVR. Confirmation experimental results exhibited a deviation of less than 5.3% from the predicted GCD% by SVR model and this further validated its robustness. This research transitions reverse engineering process from a reactive post-process assessment to a predictive, data-driven surface quality control. The outcome of this research can benefit fields like aerospace, biomedical, and precision manufacturing. Future research can focus on embedding deep learning frameworks and expanding datasets to enhance predictive accuracy across even more complex geometrical domains.