A Cross-Validated Reassessment of Regression and Machine-Learning Models for AISI 1045 End Milling

Abstract

Machining-induced residual stress, cutting force, and temperature govern the fatigue life, dimensional stability, and surface integrity of milled components, yet predictive models for these responses are still routinely validated only in-sample, concealing overfitting on small, single-laboratory datasets. This study re-examines a published AISI 1045 end-milling dataset (N = 24, combining one-factor-at-a-time and Taguchi L9 trials) using six regression paradigms: Multiple Linear Regression (MLR), random forest, gradient boosting, Support Vector Regression (SVR), Gaussian process regression (GPR), and a shallow neural network (ANN)—under leave-one-out cross-validation (LOO-CV). The previously reported in-sample R2 of 0.84 (from a smaller n = 9 subset) was substantially higher than the LOO-CV R2 of 0.167 obtained here on the full dataset; although this gap cannot be attributed to cross-validation alone, it shows a substantial in-sample/out-of-sample performance gap. SVR gave the strongest, bootstrap- and nested-CV-confirmed cross-validated residual-stress prediction (R2 = 0.575); its apparent force advantage (R2 = 0.558) was statistically indistinguishable from GPR and did not survive nested tuning, so it is reported cautiously. GPR was narrowly best for temperature (R2 = 0.492); the ANN and SVR underperformed the linear baseline there, though nested tuning traced this largely to a fixed hyperparameter rather than the kernel method itself. Random forest permutation importance identified feed rate as the dominant residual-stress predictor, consistent with the original ANOVA. The contribution is a cross-validated, multi-paradigm reassessment with explicit uncertainty and sensitivity analysis, together with a candidate low-cost screening surrogate for AISI 1045 process planning—not a replacement for XRD or FE—and a broader caution to match model complexity to sample size.