Uncertainty-aware AI-assisted process selection for quality-efficiency-depth trade-offs in femtosecond-laser thinning of 4H-SiC wafers

Artificial intelligence (AI)-assisted process selection for femtosecond-laser thinning of 4H-SiC wafers requires balancing thinning depth, surface quality, and material removal efficiency across distinct process regions. A full-factorial dataset was established on the Si face by varying laser power, hatch spacing, and scanning speed. Thinning depth T and areal roughness Sa were measured after post-treatment, and scan-based material removal rate (MRR) was derived from T and programmed scan coverage. Higher areal energy generally increased thinning depth and surface roughness, while the maximum MRR and maximum depth occurred under different process conditions. Six machine-learning regression pipelines were evaluated using 20 repetitions of fivefold cross-validation together with leave-one-power-level-out, leave-one-hatch-spacing-level-out, and leave-one-scanning-speed-level-out validation. Gaussian process regression (GPR) achieved mean repeated-cross-validation R2 values of 0.978 for T, 0.952 for Sa, and 0.904 for propagated MRR. Structured validation showed strong GPR performance for power and hatch-spacing withholding, while ExtraTrees showed stronger generalization to withheld scanning-speed levels. GPR was retained for uncertainty-aware screening because it combined competitive predictive performance with local predictive uncertainty estimates required for conservative bound construction. Empirical rescaling increased the 90% marginal coverage from 0.868 to 0.902 for T and from 0.883 to 0.909 for Sa. Conservative screening identified surface-quality-prioritized, quality-efficiency-balanced, and depth-efficiency-balanced process conditions. Independent validation across seven predefined conditions comprised 21 processing runs, all of which satisfied the frozen conservative bounds. The reported condition-level values showed relative errors of 2.05%–4.68% for T, 2.05%–5.23% for Sa, and 2.05%–4.68% for propagated MRR. The proposed framework provides application-oriented process choices for femtosecond-laser thinning within the investigated process domain.