Accelerating Flash Calculations for Compositional Reservoir and Flow Assurance Simulations Using Graph Neural Networks

Résumé fourni par la source

Abstract We present a novel machine learning framework based on Graph Neural Networks (GNNs) for accelerating phase stability analysis in flash calculations, a key bottleneck in compositional simulations used in reservoir and flow assurance applications. The model represents multicomponent mixtures as graphs, where nodes correspond to components with their thermodynamic properties and edges encode binary interaction parameters. This formulation enables the network to generalize across mixtures with varying size and composition, eliminating the need for case-specific training data. Trained on a diverse dataset of synthetic and real mixtures with up to 50 components, the GNN accurately replicates the behavior of the traditional Tangent Plane Distance (TPD) criterion and replaces the iterative stability analysis phase without modifying the underlying thermodynamic model. Benchmark tests on unseen mixtures demonstrate an average 3$\times$ speedup compared to conventional solvers, while preserving phase identification accuracy. The framework is fully implemented on GPUs, supporting large-scale parallel inference and enabling up to one million flash calculations per second for mixtures with more than 20 components. This acceleration significantly reduces the computational cost in compositional PVT simulations, enhancing the feasibility of detailed modeling in enhanced oil recovery (EOR), carbon capture and storage (CCUS), and transient flow scenarios such as severe slugging. By integrating physics-aware learning with scalable deployment, the proposed method offers a robust and generalizable solution for next-generation thermodynamic solvers.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.