PINN-Phase: Physics-informed neural integration for accurate, generalizable and scalable multiphase-field evolution
Polycrystal phase-field simulations are repeated for many related initial conditions, calling for integrators trained once and reused without retraining. We introduce PINN-Phase, a physics-informed neural time integrator that advances the multiphase field autonomously from its initial condition, enforcing phase bounds and unit sum at every step. Its hybrid realization learns from the governing residual on its own predicted states, without post-initial-condition reference states in the loss or checkpoint selection. A fixed 25-grain model meets all predefined criteria on seven of eight prospectively fixed unseen microstructures and both atypical stress cases, with 0.94-3.71% terminal label disagreement. At native 128^3, a model trained on four microstructures keeps 98.79% label agreement and exact survivor identities on a training case over 24,000 steps, 5.86 times its represented training horizon; five of six unseen microstructures keep the exact terminal survivor set. A complementary phase-local realization, with 2,160 learned parameters independent of the phase count, trained once at 48^3 with eight grains using one-step physical-map targets on its own states, meets all acceptance criteria with unchanged weights on unseen 64^3, 96^3 and 128^3 microstructures with up to 38 grains at fixed grid spacing. At 128^3 with 150 grains and comparable mean initial grain volume, it reproduces all 368 grain extinctions over 13,500 steps with at most 0.88% label disagreement and meets the per-grain volume criterion up to t = 80. These results demonstrate structurally admissible long-horizon neural integration transferring zero-shot across initial conditions within fixed benchmark families and, in phase-local form, across domain sizes and grain counts.