Physics and Data Driven Transformer-Mamba Framework for Flow Field
In the authors' words
While deep learning accelerates expensive partial differential equation solving in computational fluid dynamics (CFD), existing methods like PINNs and FNOs often struggle with generalization, noise robustness, and physical consistency. We introduce the Transformer-Mamba for Flow Field (TM4FF) framework, a physics-constrained operator learning model with three key innovations: a Residual Wavelet Mamba (RWM) layer for feature denoising, a Transformer-based attention mechanism for enhanced feature fusion, and a physics-informed loss using Fourier derivatives to enforce the Navier-Stokes equations. Experiments on four CFD datasets show TM4FF achieves high accuracy and robust generalization across varying flow conditions.
Appeared: Friday, September 25. arXiv. Preprint with a published version.
DOI: 10.1109/ICASSP55912.2026.11460626
Published version: 10.1109/ICASSP55912.2026.11460626
Authors' comment: Corrected version of our ICASSP'26 paper: Corrected seven Dam/MISSFormer metrics in Table 1 that were mistakenly estimated using the MSE(u) ratio. The conclusions remain unchanged