Orthogonal Discrepancy Kernels for Learning with Partial Physics

arXiv:2606.21199v1 Announce Type: cross Abstract: We introduce a semi-parametric framework for nonlinear system identification, which decouples discrepancy functions from physics-based components. Orthogonal Gaussian process regression balances sparse parameter selection (the white box) with discrepancy learning (the black box) to produce interpretable models from incomplete physics.

Source: arXiv cs.LG | 2026-06-23