Spatial Context or Memory? Sizing the Inputs of Neural Surrogates for Partially Observed Fields
Neural surrogates forecast physical fields from sparse sensors by reading a local window and past history. The study investigates the trade-off between window size (spatial context) and history length (memory) against optimal error. Findings show that for chaotic systems, the best input split shifts toward wider windows as forecasting lead time increases, providing design rules for better surrogate models.