Multi-Stage Graph Neural Networks for Data-Driven Prediction of Natural Convection in Enclosed Cavities

Buoyancy-driven heat transfer in closed cavities serves as a canonical testbed for thermal design. High-fidelity CFD modelling yields accurate thermal field solutions, yet its reliance on expert-crafted physics models, fine meshes, and intensive computation limits rapid iteration. Recent developments in data-driven modeling, especially Graph Neural Networks (GNNs), offer new alternatives for learning thermal-fluid behavior directly from simulation data, particularly on irregular mesh structures. However, conventional GNNs often struggle to capture long-range dependencies in high-resolution graph structures. To overcome this limitation, we propose a novel multi-stage GNN architecture that leverages hierarchical pooling and unpooling operations to progressively model global-to-local interactions across multiple spatial scales. We evaluate the proposed model on our newly developed CFD dataset simulating natural convection within rectangular cavities with varying aspect ratios where the bottom wall is isothermal hot, the top wall is isothermal cold, and the two vertical walls are adiabatic. Experimental results demonstrate that the proposed model achieves higher predictive accuracy, improved training efficiency, and reduced long-term error accumulation compared to state-of-the-art (SOTA) GNN baselines. These findings underscore the potential of the proposed multi-stage GNN approach for modeling complex heat transfer in mesh-based fluid dynamics simulations.