BCAT: A Block Causal Transformer for PDE Foundation Models for Fluid Dynamics

Background: Foundation models have demonstrated remarkable generalization in vision and language, but their accurate use for scientific computing remains challenging. For two-dimensional fluid-dynamics prediction, current transformer-based models are often hindered by limited context and rigid causality information, as they typically map fixed windows to the next step. Objectives: We aim to develop a unified architecture for heterogeneous fluid-dynamics families. The goal is to better capture the complex spatio-temporal dependencies inherent in nonlinear fluid dynamics. Methods: We introduce BCAT, a block-causal transformer for autoregressive prediction of two-dimensional fluid-dynamics solutions. Our approach adapts block-causal attention to next-frame prediction, leveraging previous frames as contextual priors. The model is trained on 69K trajectory sequences from PDEBench, PDEArena, and CFDBench using the Muon optimizer. Results: BCAT achieved an average relative 𝐿 2 error of 1.18% across the six evaluated fluid families, reducing error by 45% relative to a 3.4-times-larger baseline. In ablations, next-frame prediction was 3.5x more accurate than the corresponding nexttoken variant. A five-seed analysis, noisy-input stress test, long-rollout analysis, and physics diagnostics further characterize the model. Conclusions: The block-causal framework captures spatial interactions within frames while enforcing temporal causality across frames. BCAT provides an accurate and computationally efficient model for the two-dimensional fluid-dynamics benchmarks considered here.