This design supplies physics-plausible stepping references while letting learned corrections receive gradients throughout training, improving convergence and robustness to modeling errors and ...