Label-Free Finite-Volume-Residual Training of Attention Graph Neural Networks for Coupled Thermo-Fluid Fields

2026-07-22 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Scientific Machine Learning · Depth: Expert, quick

Summary

A novel label-free training method for attention graph neural networks (GNNs) is proposed for predicting three-dimensional (3D) thermo-fluid fields. This approach minimizes finite-volume method (FVM) residuals of governing equations directly on the mesh, eliminating the substantial computational and storage costs typically associated with generating labeled training data from numerical solvers. The trained GNN surrogates were evaluated across four scenarios against computational fluid dynamics (CFD) references and a data-supervised baseline. For two steady-state benchmarks, the FVM-loss model achieved an all-field normalized root-mean-square error (nRMSE) of 2.3-2.8%, demonstrating strong agreement with CFD, including buoyancy-energy coupling. In two parametric transient cases, the FVM-loss model surpassed the supervised baseline in accuracy while entirely avoiding data generation expenses, indicating its practicality for reducing model development costs.

Key takeaway

For AI Scientists developing neural surrogates for complex physical systems, this label-free training method offers a significant advantage. You can now train attention GNNs for 3D thermo-fluid fields by minimizing FVM residuals, completely bypassing the expensive and time-consuming generation of labeled CFD data. This approach not only reduces your model development costs but also achieves high accuracy, even outperforming supervised baselines in transient scenarios. Consider integrating FVM-loss into your scientific machine learning workflows to accelerate model deployment.

Key insights

Training GNNs with FVM residuals eliminates labeled data needs for thermo-fluid field prediction, reducing costs.

Principles

• FVM residuals provide a practical training signal.
• Label-free training reduces data generation costs.
• Attention GNNs can model complex thermo-fluid coupling.

Method

An attention graph neural network is trained by directly minimizing finite-volume method (FVM) residuals of governing equations evaluated on the mesh, bypassing labeled data generation.

In practice

• Apply FVM-loss to GNNs for scientific ML.
• Reduce data generation for 3D thermo-fluid models.
• Improve accuracy in parametric transient simulations.

Topics

• Graph Neural Networks
• Scientific Machine Learning
• Finite Volume Method
• Thermo-Fluid Dynamics
• Neural Surrogates
• Label-Free Training

Best for: AI Scientist, Research Scientist

Related on AIssential

• Advances in Scientific Machine Learning for Coupled Fluid Flow and Transport — SciML offers efficient surrogate models for complex, computationally expensive coupled fluid flow and transport systems.
• GENERIC-FNO: Embedding Energy Conservation and Entropy Production into Fourier Neural Operators — GENERIC-FNO embeds full nonequilibrium thermodynamic structure into neural operators, ensuring exact energy conservation and entropy production.
• CoFINN: Conservation Flux Informed Neural Networks for Physics Problems Governed by Conservation Laws — CoFINN integrates finite-volume conservation physics into CNN training for accurate, physically consistent compressible flow predictions.
• Neural Network-Based Parametric Model Reduction for Predicting Turbulent Flow for Different Vehicle Geometries — A VAE-enhanced neural network model reduces computational cost for high-fidelity turbulent flow predictions around diverse vehicle geometries.
• AI-powered design comes to high-speed racing — AI surrogate models, like GIST, can drastically accelerate complex physics simulations while maintaining high accuracy.
• A Review of Pseudo-Labeling for Computer Vision — Pseudo-labeling extends beyond semi-supervised learning to enhance self-supervised and unsupervised methods.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.