HGATSolver: A Heterogeneous Graph Attention Solver for Fluid-Structure Interaction

AI summary

Overview

Research area: Machine learning for scientific computing, specifically graph neural networks and neural operators for coupled multi-physics PDE simulation.
Technical level: Advanced.
Scope: The paper introduces HGATSolver, a heterogeneous graph attention surrogate model that treats fluid, solid, and interface regions as distinct node and edge types to predict fluid-structure interaction more accurately and stably.

What This Paper Is About

Fluid-structure interaction (FSI) systems involve fluid and solid domains governed by different partial differential equations and coupled at a moving interface, making high-fidelity simulation expensive. Existing learning-based solvers often flatten the system into one homogeneous graph, forcing a single message-passing rule to approximate different physics and ignoring the known fluid-solid decomposition. This paper builds a surrogate solver that explicitly encodes that physical heterogeneity, stabilizes time stepping, and balances learning across domains.

Key Contributions

1. Introduces the first FSI simulation framework based on a heterogeneous graph architecture, with fluid and solid nodes and typed edges for intra-domain and interface interactions, enabling relation-aware message passing.
2. Proposes the Physics-Conditioned Gating Mechanism (PCGM), a learnable adaptive relaxation factor that blends initial and updated node states based on local features and global physics parameters to improve stability.
3. Develops the Inter-domain Gradient-Balancing Loss (IGBL), an uncertainty-based loss weighting scheme that automatically balances fluid and solid prediction objectives without manual tuning.
4. Constructs two challenging FSI benchmarks, FI-Valve and SI-Vessel, and evaluates on the public NS+EW dataset, achieving state-of-the-art accuracy with released code.

Main Findings

• State-of-the-art benchmark accuracy: On FI-Valve, HGATSolver achieves 2.649% fluid and 0.250% solid relative L2 error; on SI-Vessel, 4.569% fluid and 0.652% solid, outperforming U-Net, GCN, GAT, HGAT, GINO, GNOT, Transolver, and AMG.
• Interface-focused gains: Error reductions are most pronounced near fluid-solid interfaces and at elastic-rigid material junctions, where coupled dynamics are strongest.
• Strong few-shot generalization: On NS+EW at Re=400, error is 0.237% with only 5 training samples and 0.055% with 100 samples; at Re=4000, error is 0.270% with 100 samples, leading all baselines.
• PCGM is the largest ablation factor: Removing PCGM increases FI-Valve fluid error from 2.649% to 3.276% and SI-Vessel solid error from 0.652% to 0.851%, confirming its stabilizing role.
• IGBL beats manual loss weighting: Learned uncertainty-based weighting lies below the Pareto front of fixed fluid-to-solid loss ratios, reducing both domain errors simultaneously.
• Other components matter: Time embeddings, learnable aggregation, and physics parameters each contribute measurable accuracy gains in ablations.
• Qualitative error maps: HGATSolver suppresses residuals in valve leaflets, compliant vessel walls, and high-velocity jets better than AMG.

Methodology in Plain English

The researchers convert the FSI mesh into a graph where each node is labeled as fluid or solid. Edges are also typed: fluid-to-fluid, solid-to-solid, fluid-to-solid, and solid-to-fluid. This lets the network use different attention and projection weights for internal domain dynamics versus interface coupling. Node inputs combine past physical states, spatial position, and a time-step embedding. A type-specific encoder and several relation-aware attention layers pass messages only along appropriate edge types; intra-domain and inter-domain messages are aggregated with learnable weights.

After message passing, PCGM computes a gate for each node from its initial encoded features, its graph-updated features, and normalized physics parameters such as density or viscosity. The final node state is a soft interpolation between the initial and updated states, so the model can learn how strongly to trust the graph update in stiff or sensitive regions. For training, IGBL treats fluid and solid predictions as Gaussian distributions with learned variances. The total loss weights each domain’s mean squared error by its inverse variance and adds log-variance terms, so the model automatically increases or decreases the loss weight for each domain as training progresses.

Experiments use two new benchmarks: FI-Valve, which models flow-induced deformation of cardiovascular valve leaflets under pulsatile inflow at transitional-to-high Reynolds numbers, and SI-Vessel, which models structure-induced flow variation in compliant vessels with rigid and elastic wall segments. The public NS+EW dataset simulates flow past a fixed cylinder with an elastic strap. All models are trained in PyTorch with AdamW, batch size 16, a temporal window of 10, cosine learning rate scheduling, and 500 epochs on two NVIDIA RTX 5090 GPUs. Evaluation uses mean relative L2 error.

Why This Matters

This work shows that embedding known physical structure—separate fluid and solid domains plus their interface—into a graph neural network improves both accuracy and stability for coupled multi-physics surrogate modeling. It offers a template for other multi-domain PDE problems where a homogeneous model would have to rediscover domain boundaries from data.

Real-world applications:

• Cardiovascular hemodynamics: predicting blood flow through heart valves and compliant vessels for surgical planning and device design.
• Aerospace and automotive design: aeroelastic simulation of wings, panels, and vehicle structures under fluid loads.
• Biomedical devices: modeling flexible implants, stents, or soft actuators interacting with surrounding flow.
• Digital twins and real-time monitoring: fast surrogate predictions for offshore structures, wind turbines, or energy systems where fluid and structure interact.

Industry relevance: The method targets the CFD/FSI simulation bottleneck. Faster, stable learned surrogates could reduce reliance on expensive numerical solvers in design loops, support uncertainty-aware multi-domain training without manual loss tuning, and enable real-time or many-query scenarios such as optimization, control, and medical planning.

Future Directions

• Extend the heterogeneous graph formulation to three-dimensional, larger-scale FSI problems with more complex geometries and turbulent flows.
• Incorporate additional physics domains, such as thermal, acoustic, or electromagnetic coupling, into the same typed-graph framework.
• Improve long-horizon rollout stability and conservation properties, possibly by combining learned updates with traditional numerical solvers or physical constraints.
• Reduce data requirements further and test transfer learning across geometries, materials, and Reynolds numbers.
• Provide theoretical analysis of PCGM and IGBL, including convergence and uncertainty calibration in coupled multi-domain training.
• Broaden benchmarking to standard public FSI datasets and real experimental measurements.

Target Audience

The paper is best suited for advanced readers in machine learning for scientific computing, computational mechanics, CFD/FSI, and graph neural networks. It also benefits biomedical, aerospace, and mechanical engineers interested in surrogate models, plus graduate students and researchers working on neural operators, multi-physics simulation, and physics-informed machine learning.

Authors’ abstract

Fluid-structure interaction (FSI) systems involve distinct physical domains, fluid and solid, governed by different partial differential equations and coupled at a dynamic interface. While learning-based solvers offer a promising alternative to costly numerical simulations, existing methods struggle to capture the heterogeneous dynamics of FSI within a unified framework. This challenge is further exacerbated by inconsistencies in response across domains due to interface coupling and by disparities in learning difficulty across fluid and solid regions, leading to instability during prediction. To address these challenges, we propose the Heterogeneous Graph Attention Solver (HGATSolver). HGATSolver encodes the system as a heterogeneous graph, embedding physical structure directly into the model via distinct node and edge types for fluid, solid, and interface regions. This enables specialized message-passing mechanisms tailored to each physical domain. To stabilize explicit time stepping, we introduce a novel physics-conditioned gating mechanism that serves as a learnable, adaptive relaxation factor. Furthermore, an Inter-domain Gradient-Balancing Loss dynamically balances the optimization objectives across domains based on predictive uncertainty. Extensive experiments on two constructed FSI benchmarks and a public dataset demonstrate that HGATSolver achieves state-of-the-art performance, establishing an effective framework for surrogate modeling of coupled multi-physics systems.