Machine learning for full spatiotemporal acceleration of gas-particle flow simulations

Abstract

Gas-particle flows can be well described by highly resolved simulations (HRS). Limited by the daunting computational cost, HRS at scale remains infeasible. Many methods are proposed to strike tradeoffs between accuracy and computational cost. Here, we design a machine learning (ML) based alternate mode to accelerate HRS without sacrificing accuracy. Spatiotemporal samples (~1.3 × 106) generated from two-fluid model-based HRS of gas-particle flows are trained for developing artificial neural network (ANN) and long short-term memory (LSTM) models. Simple time-series predictions consisting of repeated ANN iterations can accurately predict the flow field in ~10−3 s. Hybrid accelerations combining CFD and the offline trained ML in time scale can approximate to the time-series highly resolved flow fields within 1% error, saving 40% computing time. This work may contribute to a new transformative paradigm that how scientific computing can leverage ML to improve gas-particle simulations without new device requirements.