RD-JEPA Enables Few-Trajectory Transfer for Reaction-Diffusion PDEs

RD-JEPA utilizes joint-embedding predictive architecture to pretrain on reaction-diffusion trajectories, allowing adaptation to unseen systems using as few as one trajectory. It outperforms supervised surrogate baselines and architecture-matched models in mean relative discrete l2 and spatial first-difference errors.