Answers are grounded in this paper and cite their sources.
Reliable surrogate modeling of femtosecond-laser machined fused silica microchannels is difficult because experimental data are scarce and the responses are strongly nonlinear. We study physics-informed Gaussian process regression (GPR) with deep kernel learning (DKL) on a published orthogonal dataset, using nested leave-one-out cross-validation to separate hyperparameter selection from generalization assessment. Baseline GPR reaches a training R 2 of 0.973 for surface roughness S a , but a generalization R gen 2 of only 0.217, indicating pronounced overfitting. Adding the physics-derived features line energy and pulse density, which reparameterize the process parameters, raises it to 0.695 for GPR; flexible models on the same inputs reach 0.849 ± 0.005 (DKL) and 0.886 ± 0.030 (multilayer perceptron) over five seeds. The nearly linear material removal rate (MRR) is already predicted with R 2 = 0.975 by GPR, and flexible models bring no gain. Model complexity should therefore match the response mechanism. These results apply to the dataset and process window studied; DKL additionally provides the predictive uncertainty used for optimization within the region supported by the data.
No takes yet. Share an insight, caveat, or question.
Sun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: