Fast Prediction Methods for Fluid Simulation Results Using Deep Neural Networks

Computational fluid dynamics is an important application in the field of computational science and high-performance computing, and requires a large number of computations to be performed with high accuracy. In recent years, relying on the fast inference provided by deep learning, there has been a lot of research and development on fast prediction of numerical simulation results using a data-driven approach. We have been working on developing a method that uses deep learning to predict the results obtained from steady and unsteady fluid flow calculations, with the aim of developing a fast surrogate model that predicts fluid flow simulations. For steady flow prediction, we predict the flow velocity over the entire region based on the boundary conditions of the computational domain and a signed distance function representing the objects. To predict simulation results over a computational domain larger than the training data, we propose a prediction method that combines deep learning inference on a decomposed subdomain and boundary exchange between these subdomains. For unsteady flow prediction, we attempt to improve the prediction accuracy by improving the loss function. In this talk, we will present these approaches and results.