Fluid-DiT: Graph-Free Diffusion Transformers for Fluid Flow Simulations Learning

2026-08-07

AI Summary

High-fidelity fluid simulation is computationally expensive, and existing graph-based diffusion models are limited by handcrafted graph structures, restricted receptive fields, and difficulties in multiscale modeling. This work proposes a graph-free diffusion Transformer framework that eliminates explicit graph construction and leverages global attention mechanisms in a latent space to decouple geometric fidelity from distribution learning, enabling efficient modeling and sampling of the equilibrium distribution of chaotic flow fields. The method substantially reduces high-frequency artifacts and improves sampling efficiency. On benchmarks including laminar wakes, elliptic flows, and three-dimensional turbulent airfoils, it outperforms graph-based models in both sample quality and distributional accuracy, achieving higher R² correlation and lower Wasserstein distance, while generalizing effectively to unseen Reynolds numbers and geometric configurations.

Abstract

Simulating complex fluid flows requires capturing full equilibrium distributions rather than just mean trajectories, yet high-fidelity solvers remain computationally prohibitive. Recent advances, such as Diffusion Graph Networks (DGNs), have combined diffusion models with graph neural networks to sample equilibrium states directly from unstructured meshes, enabling distributional accuracy even from short simulations. However, graph-based diffusion approaches suffer from hand-crafted architectural constraints, limited receptive fields in message passing, and costly multi-scale designs, which restrict scalability to larger and more complex domains. We propose Fluid-DiT, a Graph-Free Diffusion Transformer that replaces graph message passing with attention-based denoising, eliminating explicit graph design while preserving the ability to model distributions of chaotic flows. Our framework introduces a latent-space formulation that disentangles geometric fidelity from distributional learning, reducing high-frequency artifacts and accelerating sampling. By leveraging the transformer's global receptive field, Fluid-DiT naturally captures both local flow structures and long-range correlations without requiring hierarchical graph coarsening. On canonical benchmarks including laminar cylinder wakes, ellipse-flow systems, and turbulent 3D wing experiments, Fluid-DiT consistently outperforms graph-based diffusion baselines in both sample quality and distributional accuracy, achieving higher $R^2$ correlations and lower Wasserstein distances. Moreover, it generalizes robustly from short, incomplete trajectories to unseen Reynolds numbers and geometries, demonstrating strong scalability.

Problem

• fluid flow simulation
• diffusion models
• graph neural networks
• scalability
• distributional accuracy

Innovation

• Diffusion Transformer
• Graph-Free Modeling
• Fluid Flow Simulation
• Latent-Space Disentanglement
• Attention-Based Denoising