Physics-Informed Neural Network with Decoupled Pathways for Multi-Material Crystal Plasticity Prediction

We present a Hybrid Physics-Informed Neural Network (CP-PINN) with decoupled LSTM pathways for predicting stress tensors, slip system activity, and critical resolved shear stress (CRSS) evolution across three crystal systems: FCC copper, BCC iron, and HCP titanium. The architecture uses gradient-isolated pathways with orientation-dependent elastic stress priors via Hooke's law. Trained on synthetic CP-FEM trajectories spanning monotonic, cyclic, and biaxial loading paths, the model achieves CRSS R-squared = 0.988, stress R-squared = 0.970, and perfect CRSS monotonicity on the test set. An ablation study shows multi-material encoding improves stress R-squared by 19.4% over single-material baselines. Post-hoc temperature-scaled MC-Dropout provides calibrated uncertainty estimates with 96.2% coverage at 2-sigma. The framework establishes a computationally efficient surrogate for multi-material crystal plasticity that preserves key physical constraints while enabling uncertainty-aware predictions.