Union Adaptive Tokenization for Efficient Spatiotemporal Transformer Modeling of Multiscale Reacting Flows

Chemically reacting flows prevalent in combustion systems exhibit highly localized and dynamically evolving reaction zones, where steep gradients in thermochemical fields and reaction rates encode critical information about the underlying flow dynamics. Uniform tokenization commonly used in PDE transformer models allocates the representation capacity uniformly across the domain, without explicitly accounting for this spatial heterogeneity. Although adaptive tokenization addresses this issue, existing approaches either operate on single frames or, when handling multiple frames, recompute the token layout onto the full grid at each time step. In this work, we present a union adaptive tokenization (UAT) approach, where a single adaptive tokenization structure is constructed over all context frames and frozen across them, so that each token refers to the same spatial region at every time step within the context window. We demonstrate this approach on a 2D lifted hydrogen jet flame, evaluating multiple mesh refinement sensor configurations while holding the transformer architecture, loss, and training budget fixed. The best-performing adaptive configuration nearly matches the uniform-grid baseline across all evaluated fields while consuming 3.9× fewer tokens. The results show that UAT can improve the computational efficiency of transformer-based surrogate and foundation modeling for multiscale reacting flows while providing a temporally consistent spatiotemporal representation of the localized combustion dynamics.