This paper proposes an efficient data augmentation strategy for partial differential equation (PDE) modeling. Neural PDEs, which utilize neural networks, are used instead of conventional numerical PDE solvers. Conventional methods, however, require a large amount of data to train neural PDEs using the solution trajectories obtained through long-term integration of the PDE solver. In this paper, we present a sample-efficient data augmentation strategy that generates neural PDE training data from a computer model by space-filling sampling of local "stencil" states. This method removes much of the spatiotemporal redundancy present in trajectory data and oversamples states that are rarely visited but help neural PDEs generalize across the state space. Experiments using synthetic data demonstrate the effectiveness of the proposed method, demonstrating superior performance to conventional methods.
