Differentiable simulation and optimization of particle-fluid flows using graph neural networks

Abstract

Particle-fluid two-phase flow phenomena are prevalent in various engineering applications, including resource management and environmental hazard mitigation. Accurate modeling and control of these flows are crucial for optimizing system performance and preventing catastrophic events. In this study, we propose a novel approach that integrates the Discrete Element Method (DEM) and Smoothed Particle Hydrodynamics (SPH) to simulate complex particle-fluid interactions. A Graph Neural Networks (GNNs) framework is subsequently trained to capture these two-phase dynamics. The developed approach is validated through real-world experiments, demonstrating its effectiveness in practical scenarios. To optimize flow control structures, such as baffles or barriers, we develop an inverse design framework powered by the differentiable capability of GNNs. This framework efficiently explores the design space to minimize destructive flow behaviors and control the processes. The results show that this approach provides a powerful tool for designing resilient systems capable of optimizing the two-phase flows in various engineering contexts.