Incremental machine learning-based accelerator for computational fluid dynamics simulations

Abstract

The simulation of physicochemical processes with computational methods is key for engineering design, with applications in a variety of industries, ranging from pharmaceuticals to aerodynamics. Despite its importance and widespread use, significant challenges related to the accuracy and computational complexity of these simulations remain prominent. These systems are governed by non-linear transport equations with physical and chemical processes occurring at different spatiotemporal scales in complex geometries. This leads to problems which are computationally expensive and often infeasible to solve. As such, reducing the computational complexity of multiphysics problems without compromising on accuracy is a central goal in the engineering community. Recently, machine learning has proven to be a promising direction towards this goal. The availability of data from both multiphysics experiments and simulations have led to high-performing neural networks capable of accelerating traditional methods for solving multiphysics problems. Despite these hopeful results, there still exists a gap between machine learning and its optimal application in a realistic engineering design process. This work aims to bridge that gap through two main approaches. The first approach is by developing a framework which hosts neural network training and existing computational multiphysics software in a unified framework. The second approach is to incrementally determine optimal neural network parameters by running computational multiphysics problems and neural network training in parallel. This has shown to reduce data collection and training time while increasing the speedup of multiphysics simulations over increments.