Physically informed input design for neural friction models in mechanical simulations

Abstract

Friction models are essential to accurate mechanical simulation, but classical analytical laws are often difficult to calibrate and can induce numerical stiffness. Data-driven surrogate models offer an attractive alternative, yet require appropriate input selection. This work demonstrates that kinematic-only inputs are insufficient to uniquely predict friction force, especially in the static and stick-slip regimes, and proposes an improved input-output design for neural-network-based friction surrogates. A feed-forward neural network with temporal input windows was trained to perform one-step-ahead prediction. Training data were generated using four classical friction models: smoothed Coulomb, Brown–McPhee, Dahl, and LuGre. To assess generalization, the trained models were evaluated on a Rabinowicz-type spring-mass-on-a-treadmill system and a low-velocity reversal benchmark without additional training or fine-tuning. Using the combined histories of relative velocity and force, the network reproduced all four friction models with high accuracy, achieving R-squared scores between 0.993 and 0.9999 and capturing both sliding and stick-slip behaviors. Conversely, velocity-only networks failed to replicate bristle-type models, confirming the non-uniqueness of purely kinematic mapping. These results provide a practical recipe for designing friction surrogates that generalize across different mechanical configurations.