PI-JEPA: Label-Free Surrogate Pretraining for Coupled Multiphysics Simulation via Operator-Split Latent Prediction
Researchers have introduced PI-JEPA, a novel pretraining framework for neural operators designed to reduce the need for extensive labeled simulation data in multiphysics simulations. This method leverages unlabeled parameter fields and a physics-informed approach, enabling effective training with significantly fewer completed PDE solves. PI-JEPA demonstrates superior performance compared to existing models like FNO and DeepONet, particularly when fine-tuned with limited labeled runs, thereby lowering the simulation budget required for deploying multiphysics surrogates. AI
IMPACT Reduces simulation costs and accelerates deployment of AI surrogates for complex physics problems.