Source-linked AI summary
Combining Differentiable PDE Solvers and Graph Neural Networks for Fluid Flow Prediction
Filipe de Avila Belbute-Peres, Thomas D. Economon, J. Zico Kolter
TL;DR
Large CFD PDEs are expensive to solve, while purely learned fluid predictions often generalize poorly beyond their training distribution. The paper embeds a differentiable coarse CFD simulator within a graph neural network operating on unstructured meshes. The resulting system is faster than full simulation, more accurate than a purely learned model, and generalizes to parameter settings outside training.
Problem
CFD simulations are computationally expensive, and deep learning approximations may produce poor results on settings well outside their training data.
Method
The method combines a graph convolutional network on unstructured meshes with a differentiable SU2 CFD simulator operating on a learned coarse mesh.
Results
The combined system is faster than full simulation, more accurate than a purely learned model, and generalizes to parameter settings outside the training distribution.
Takeaways & Limitations
Combining coarse fluid simulation with learned components provides complementary benefits for CFD prediction within the demonstrated scope.
Takeaways & Limitations
The approach involves a trade-off between the cost and generalization ability of full CFD and those of GCN predictions.
Abstract
from arXiv · showhide
Solving large complex partial differential equations (PDEs), such as those that arise in computational fluid dynamics (CFD), is a computationally expensive process. This has motivated the use of deep learning approaches to approximate the PDE solutions, yet the simulation results predicted from these approaches typically do not generalize well to truly novel scenarios. In this work, we develop a hybrid (graph) neural network that combines a traditional graph convolutional network with an embedded differentiable fluid dynamics simulator inside the network itself. By combining an actual CFD simulator (run on a much coarser resolution representation of the problem) with the graph network, we show that we can both generalize well to new situations and benefit from the substantial speedup of neural network CFD predictions, while also substantially outperforming the coarse CFD simulation alone.
1. Introduction
Deep learning offers faster PDE and CFD prediction, but learned models can fail outside their training domain. The paper therefore combines graph neural networks with a differentiable coarse CFD simulator to improve speed, physical fidelity, and generalization.
- CFD simulations solve nonlinear Navier–Stokes PDEs numerically and can require days to weeks on supercomputing infrastructure.The runtime bottleneck motivates faster surrogate models for applications such as aerodynamic design iteration.
- Deep learning models for fluid dynamics can quickly produce poor results on settings well outside their training data.
- The proposed hybrid system combines a graph convolutional network with an industry-grade CFD simulator operating on a much coarser mesh.The graph network operates directly on the non-uniform meshes used in typical CFD tasks.
- The embedded differentiable simulator can provide gradients through adjoint methods, allowing its outputs and coarse mesh to be integrated into network learning.
- The combined approach performs substantially better than coarse CFD alone, generalizes better to novel situations than a pure graph-network approach, and remains faster than full-resolution CFD.
2. Related Work
Prior work applies learning to meshes, CFD fields, physical-model components, and differentiable fluid simulation, but differs in representation, physical fidelity, or integration scope. This work combines graph operators with a full differentiable CFD simulation and optimizes the mesh for a downstream task.
- Earlier CFD learning work often directly predicts physical processes with deep networks, including convolutional models operating on small image representations of airfoil flow fields.
- Other approaches embed learned function approximators for selected equation terms, whereas this work embeds a full physical simulation into a deep learning system.
- Graphics-oriented differentiable fluid methods target realistic animations or control, rather than accurately modeling physical processes for CFD applications.
- Adjoint-based differentiation enables fluid simulations to serve as differentiable modules, using established formulations and the SU2 open-source CFD suite.
- Prior mesh-based graph models and positional graph methods generally use mesh modifications for pooling or learn dynamics without a PDE solver.
- The paper claims its method is the first to modify a mesh through a differentiable simulator to optimize mesh functionality for a downstream task.
3. Methodology
CFD-GCN combines a differentiable SU2 simulator on a learned coarse mesh with graph convolutions on a fine mesh. The coarse simulation is upsampled and integrated into fine-mesh graph processing to produce flow-field predictions.
- The CFD-GCN Architecture: CFD-GCN operates over fine and coarse graphs, using the coarse mesh as input to the CFD simulator and the fine mesh for prediction processing.The coarse mesh preserves the fine mesh structure while reducing node count by almost 20x for faster simulation.
- Graph structure and network input: The model represents the CFD domain as a triangular mesh-derived graph with node coordinates, triangular elements, and boundary points.Graph edges are inferred directly from mesh elements, allowing the network to operate on the unstructured meshes used by traditional CFD software.
- The SU2 Fluid Simulator: The differentiable SU2 simulator takes the coarse mesh, angle of attack, and Mach number, then outputs coarse-node velocity and pressure fields.The coarse mesh acts as trainable simulator parameters whose configuration is adjusted for integration with the rest of the network.
- Upsampling and graph convolutions: The coarse outputs are upsampled with successive squared distance-weighted k-nearest-neighbor interpolation before graph-convolution processing.The default interpolation uses k = 3, and the upsampled values are concatenated into an intermediate fine-graph representation.
- Graph convolutions: The fine-mesh graph receives physical parameters and signed-distance features, applies graph convolutions around the injected coarse result, and generates the final prediction.The architecture uses the GCN formulation of Kipf and Welling, with ReLU nonlinearities between convolutional layers.
- Training the CFD-GCN: The complete system is differentiable through the SU2 layer, enabling joint optimization of GCN weights and coarse-mesh node positions.Projected gradient updates prevent mesh elements from degenerating; with this correction, mesh optimization learns smoothly rather than diverging.
4. Experiments
Experiments evaluate CFD-GCN on interpolation and stronger generalization tasks using coarse and fine meshes, comparing it with ablated and pure GCN baselines. The hybrid model improves on coarse simulation, generalizes better to unseen shocks, and predicts faster than full fine-mesh CFD.
- Experimental setup: Experiments use a 6648-node fine mesh and a 354-node coarse mesh for the NACA0012 airfoil, with steady compressible inviscid Navier–Stokes simulations solved by SU2.The coarse mesh is triangularized for CFD-GCN use, while fine-mesh simulations provide converged ground truth.
- Interpolation: The interpolation task samples training and test pairs from different AoA and Mach-number combinations with similar parameter ranges and qualitative flow behaviors.This evaluates interpolation to unseen but similar conditions, though the authors describe it as a weak learning test.
- Interpolation: The CFD-GCN outperforms the upsampled coarse-mesh baseline but performs worse than the pure GCN on interpolation because the similar test distribution rewards stronger training-set overfitting.This indicates the graph-processing stages improve coarse predictions, while the pure GCN benefits from distribution similarity.
- Generalization: The generalization task places higher-Mach cases, where shocks are frequent, exclusively in the test set, creating qualitatively and quantitatively different behavior from training.The split therefore tests prediction on unseen shock behavior rather than interpolation within a similar distribution.
- Generalization: The CFD-GCN generalizes better than the pure GCN to unseen shocks, produces predictions closer to ground truth, and outperforms the frozen-mesh baseline.The model also improves beyond simply upsampling the coarse simulation, with learned mesh optimization contributing to generalization.
- Runtime: Approximately 85% of CFD-GCN batch-prediction time is spent in CFD simulation, while training takes approximately 19 hours versus 1.3 hours for the pure GCN baseline.Despite this cost, downsampling to 354 nodes makes prediction much faster than running full fine-mesh CFD.
5. Conclusion
The system integrates a differentiable CFD simulator into a graph-based deep learning model operating on unstructured meshes. Combining coarse fluid simulation with learned components yields predictions that are faster than full simulation, more accurate than purely learned models, and capable of generalizing beyond training settings.
- The system integrates a differentiable CFD simulator as a module within a larger deep learning system operating directly on unstructured meshes.Graph operators enable the model to work with the mesh representation used by realistic CFD solvers.
- Combining coarse-mesh fluid simulation with learned model components produces predictions faster than full simulation and more accurate than a purely learned model.
- The model generalizes to parameter settings outside the training distribution.
A. Multi-Airfoil Experiment
The multi-airfoil experiment tests generalization by training on two airfoils and evaluating on a previously unseen airfoil. The CFD-GCN generalizes better to this test set, while the upsampled coarse-mesh baseline has RMSE 0.34.
- A. Multi-Airfoil Experiment: The CFD-GCN generalizes better to the test set containing the previously unseen airfoil than the baselines.
- A. Multi-Airfoil Experiment: 0.34 RMSE is reported for the upsampled coarse mesh baseline, whose results were too high to display in the same plot.
B. Mesh Optimization
Mesh optimization changes the coarse mesh most strongly around the airfoil, where loss gradients are large. Regions farther from the wing remain mostly unchanged because they affect prediction less strongly.
- B. Mesh Optimization: Mesh changes during optimization are greatest around the airfoil, where the loss gradients are large.
- B. Mesh Optimization: Regions farther from the wing are mostly unaltered because they do not strongly affect prediction.
C. Interpolation Experiment
The interpolation experiment presents CFD-GCN predictions and ground truth for velocity and pressure fields. The supplementary figure covers fields omitted from the main interpolation-task discussion.
- C. Interpolation Experiment: The figure presents fields omitted from the main text for the interpolation task.
- C. Interpolation Experiment: Figure 13 compares CFD-GCN predictions with ground truth for the y component of velocity and pressure output fields.The figure uses the same test sample as Figure 6.
D. Generalization Experiment
The generalization experiment presents predictions for a test sample with a large shock, including CFD-GCN, GCN, and upsampled coarse-mesh baseline outputs.
- D. Generalization Experiment: The accompanying text states that Figures 14 and 15 present predictions and ground truth for fields omitted from the main text, while Figure 16 presents the coarse-mesh baseline.
- D. Generalization Experiment: Figure 16 shows the upsampled coarse-mesh baseline prediction for the same large-shock generalization sample.The figure presents the x and y velocity components and pressure output fields.