Data-driven prognosis of root mean square deviation in reverse-engineered freeform surfaces: A comparative study of support vector, Gaussian process and extreme gradient boosting regressions

Abstract

Freeform surfaces are vital in aerospace, biomedical, and die/mold industries, where dimensional accuracy directly impacts performance and reliability. Reverse engineering workflows are highly sensitive to scan density, noise reduction, and reconstruction parameters, yet current practice relies on trial-and-error tuning, increasing inspection cost and uncertainty. A CNC-milled Al6061 freeform test component was fabricated and scanned under 64 parameter combinations generated using a Taguchi-based orthogonal design. Root Mean Square (RMS) deviation between CAD and scan models was quantified to establish a performance baseline. Three machine learning models, namely, Support Vector Regression (SVR), Gaussian Process Regression (GPR), and eXtreme Gradient Boosting (XGBoost) were developed to predict RMS deviation directly from reverse engineering parameters. Model hyperparameters were optimized via grid search with five-fold cross-validation. GPR achieved the highest predictive accuracy (Testing R 2 =0.9901, RMSE=0.0215, MAPE=4.95%), while SVR demonstrated superior robustness across folds (R 2 range: 0.7717-0.9782). XGBoost captured complex nonlinearities but showed larger fold variability and outlier errors, with MAPE exceeding 20% in select cases. The proposed machine learning methodology enables tolerance-aware, first-time-right scanning of freeform components by replacing heuristic parameter tuning with predictive modelling. GPR is recommended for accuracy-driven applications, while SVR provides a stable alternative for robustness-focused scenarios, significantly advancing industrial freeform metrology planning.