Source: Published in the July 2025 Issue of Modern Casting by American Foundry Society
Abstract:
Multiphase Computational Fluid Dynamics (CFD) simulations are essential for analyzing industrial systems involving gas–solid flows. Consequently, high-fidelity CFD simulations are computationally intensive for design exploration and optimization and are generally impractical for real-time decision-making in large-scale systems.
This work presents a Graph Neural Network (GNN)–based Reduced Order Modeling (ROM) framework that achieves approximately 80–105× speedup for transient and time-averaged field predictions, respectively, and up to three orders of magnitude speedup for isolated snapshot queries, while preserving predictive accuracy. The approach is demonstrated on two representative multiphase systems: a fluidized bed gasifier and proppant transport in hydraulic fractures.
The proposed framework is first evaluated on the proppant transport problem, where experimental measurements enable direct validation of ROM predictions. An Eulerian GNN-ROM is trained on CFD-generated Eulerian fields and validated against experimental measurements at 300, 630, and 1140 seconds, exhibiting strong spatial accuracy, with MAE consistently below 0.02 m.
The framework is then applied to a fluidized bed gasifier to assess its ability to generalize to unseen operating conditions using high-fidelity CFD as reference. GNN-ROMs are trained using Barracuda Virtual Reactor CFD data across multiple operating conditions defined by varying bottom air flow rates (0.691, 0.8, 0.9, and 1.0 kg/s). The model is evaluated at an intermediate, unseen flow rate (0.85 kg/s) and accurately predicts both time-averaged and transient fields for key quantities such as CH₄, CO₂, pressure, and particle volume fraction. Inference time is reduced from hours to under 10 minutes, with time-averaged relative errors below 7% and transient predictions preserving dominant spatial trends.
A key innovation of this work is the GNN architecture, which uniquely combines spatial graph convolutions on CFD meshes with explicit time-conditioning via Fourier feature embeddings. This design enables high-resolution, temporally flexible predictions at arbitrary time instances without the constraints of iterative time-stepping, offering unprecedented flexibility and computational efficiency for complex multiphase flow systems.
This demonstrates the potential for generalization to unseen conditions and highlights GNN-ROMs as transformative tools for multiphase flow modeling, enabling real-time process control, rapid design optimization, and digital twin applications in energy and subsurface engineering.