Experimentally grounded machine learning prediction of TIF-derived removal metrics for deterministic optical polishing

Accurate prediction of the tool influence function (TIF) is essential for dwell-time planning and figure correction in deterministic optical polishing. Here we present an experimentally grounded data-driven framework to predict TIF-derived removal depth and removal volume in an orthogonal velocity field (OVT) polishing across three representative optical mirror materials: Al-Si, cordierite, and SiC. The dataset contained 243 experimental entries corresponding to 81 unique processing and material combinations, with each experiment repeated three times and aggregated at the condition level. Linear, tree-based, gradient-based, instance-based, and kernel-based regressors were employed, and parameter effects were interpreted using SHapley Additive exPlanations (SHAP) analysis. Gaussian process regression (GPR) yielded the best generalization, with test R 2 score of 0.9879 for SiC, 0.9176 for cordierite, and 0.9759 for Al-Si. The results indicate that the OVT removal process is smooth but nonlinear in the investigated processing window, and material-dependent behavior is preserved in the TIF metrics, with the largest deviation from linearity observed for cordierite. The proposed framework provides a practical methodology for selecting OVT processing parameters, TIF prediction, and future dwell-time optimization in deterministic optical polishing.