Source-linked AI summary
Physics and Data Driven Transformer-Mamba Framework for Flow Field
Zhuo Zhang, Shun Zou, Canqun Yang, Xi Yang
TL;DR
Deep-learning CFD methods face challenges with unsteady-flow dependencies, generalization, noise, and physical constraints. TM4FF combines operational-parameter embedding, Mamba feature processing, Transformer-based fusion, and a physics-informed loss; experiments report high accuracy and generalization across varying flow conditions.
Problem
Deep-learning CFD methods face challenges modeling unsteady-flow dependencies, generalizing across conditions, handling noise, and explicitly enforcing physical constraints.
Source: §1 · p. 1Method
TM4FF learns a flow-field operator from the previous velocity field and operational parameters, combining Mamba, Transformer-based feature fusion, and a Fourier-derivative physics-informed loss.
Source: §1 · p. 1, §2.1 · p. 2, §2.4 · p. 4Results
TM4FF achieves high accuracy and generalization across varying flow conditions, with state-of-the-art results across four fluid cases and the lowest error in nearly every metric for both velocity components.
Source: §4 · p. 4, §3.2 · p. 4Takeaways & Limitations
The reported results present TM4FF as a promising alternative to traditional CFD methods across varying flow conditions.
Source: §4 · p. 4
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Deep-learning CFD methods face challenges in modeling unsteady-flow dependencies, generalizing across changing conditions, enforcing physical constraints, and handling noise. TM4FF addresses these challenges with an operator-learning framework that combines operational-parameter embedding, specialized feature processing, attention-based fusion, and a physics-informed loss.
1. INTRODUCTION
- Unsteady-flow modeling, generalization to unseen conditions, explicit physical constraints, and training noise remain challenges for deep-learning CFD methods.The introduction identifies complex temporal dependencies and multi-scale correlations, retraining as PDE parameters or physical properties change, and possible departures from fundamental physical principles.Source: §1 · p. 1
- TM4FF treats fluid flow as a spatiotemporal image sequence and embeds operational parameters to enhance generalization.Its Mamba architecture is used to capture long-range spatiotemporal dependencies.Source: §1 · p. 1
- A physics-informed loss with Fourier derivatives constrains predictions using the Navier-Stokes equations.The introduction reports that comprehensive experiments achieve state-of-the-art performance.Source: §1 · p. 1
- The Residual Wavelet Mamba layer combines a model-driven Haar DWT branch with a data-driven residual branch for noise decoupling.The framework also replaces standard skip connections with Transformer-based attention to fuse multi-scale features.Source: §1 · p. 1
2. METHODOLOGY
TM4FF learns a flow-field operator conditioned on operational parameters and prior velocity, combining wavelet-based feature extraction, BiS feature fusion, and a Fourier-derivative physics loss.
2.1. Overall Framework and Condition Embedding
- Operational parameters and the prior-time velocity field define the input to an operator that predicts the current-time flow field.The framework embeds operational conditions, including boundary conditions and viscosity, into the network.Source: §2.1 · p. 2
2.2. Residual Wavelet Mamba Encoder
- The RWM layer uses Haar wavelet decomposition and a parallel data-driven residual branch, then fuses their features for processing by a VSSM.The wavelet branch separates features into frequency sub-bands to help filter structured noise.Source: §2.2 · pp. 2–3, Figure 3 · p. 3
2.3. Bi-level Skip Connection (BiS)
- BiS replaces standard skip connections with feature fusion that selects relevant regions for attention and combines sparse attention with local context enhancement.The wavelet decomposition and reconstruction of strip features are illustrated in Fig. 4.Source: §2.3 · p. 3, Figure 4 · p. 4, Figure 5 · p. 4
2.4. Physics-Informed Loss Function
- The hybrid loss combines data error with incompressibility and Navier–Stokes residual penalties computed using Fourier derivatives.The incompressibility term enforces a divergence-free condition, while the residual term penalizes deviations from the momentum equation.Source: §2.4 · p. 4
3. EXPERIMENTAL RESULTS
Across four flow cases, TM4FF reports the lowest error in nearly every velocity metric and parameter efficiency versus larger models. Ablations report gains from the proposed components.
3.1. Experimental Setup
- The four CFDBench datasets were split 8:1:1 into training, validation, and test sets with disjoint operating parameters.Models were trained for 100 epochs on an NVIDIA GeForce RTX 3090.Source: §3.1 · p. 4
3.2. Results
- Across Cavity, Tube, Dam, and Cylinder, TM4FF achieves the lowest error in nearly every u- and v-velocity metric while remaining parameter-efficient versus larger models.The comparison includes autoregressive and non-autoregressive baselines.Source: §3.2 · p. 4
3.3. Ablation Study
- Case-parameter embedding, the Mamba encoder, and BiS skip connections progressively improve the baseline; adding RWM boosts all datasets, including reducing Cavity MSE from 0.0226 to 0.0194.The full model achieves the best results in the strategy ablation.Source: §3.3 · p. 4, Table 2 · p. 4, Table 3 · p. 4
4. CONCLUSION
TM4FF embeds operational parameters and combines denoising, attention-based feature fusion, and a physics-informed loss. The paper reports high accuracy and generalization across varying flow conditions.
4. CONCLUSION
- TM4FF enhances CFD prediction by embedding operational parameters in the model.The framework is designed for fluid-flow prediction.Source: §4 · p. 4
- Its innovations are a Residual Wavelet Mamba layer for feature denoising, Transformer attention for feature fusion, and Fourier-derivative loss enforcing Navier-Stokes constraints.Source: §4 · p. 4
- Experiments report high accuracy and generalization across varying flow conditions.Source: §4 · p. 4