GNN-Flash: A Hybrid Graph Neural Network Framework for Bypassing Stability Analysis in Two-Phase Flash Calculations with Generalization to Unseen Mixtures

Résumé fourni par la source

Abstract Flash calculations are fundamental for determining phase behavior under specified thermodynamic conditions. Conventional approaches rely on iterative procedures, including phase stability analysis and phase-split calculations, which are computationally demanding due to the nonlinear nature of the governing equations. These steps can account for a substantial portion of the overall computational cost, motivating the need for more efficient methods as large-scale simulations become increasingly complex. This study presents GNN-Flash, a GPU-accelerated hybrid framework that applies Graph Neural Networks (GNNs) to accelerate two-phase flash calculations. GNN-Flash does not replace the thermodynamic solver; instead, it learns a mapping from mixture graphs to high-quality initial states, effectively removing the need for conventional stability analysis, while preserving thermodynamic consistency in flash calculations, resulting in a hybrid approach. Trained on a diverse dataset of real and synthetic mixtures exhibiting vapor–liquid and liquid–liquid equilibria, GNN-Flash accurately predicts phase stability and equilibrium without requiring the conventional stability analysis step. Since GNN-Flash is trained on final flash solutions rather than stability analysis outputs, it produces initial guesses of higher quality than those obtained from conventional stability analysis, enabling more stable and efficient convergence. This significantly reduces computational cost and results in speedups of up to 7× while maintaining high predictive accuracy. In this approach, chemical mixtures are modeled as graphs, with components as nodes and binary interaction parameters as edge weights. Unlike other neural networks, GNN-Flash is not mixture- or composition-specific. Instead, it leverages permutation-invariant graph representations of mixtures, enabling generalization across unseen compositions and mixtures.