Skip to main content

Physics Informed Neural Networks for Fluid Dynamics

Academic lead
Peter Jimack (Computing)
Amir Khan (Civil Engineering), He Wang (Computing)
Project themes
Underpinning Methods for Fluid Dynamics

This project will investigate the use of deep learning techniques, coupled with more conventional numerical modelling approaches, in order to develop new computational algorithms for fluid flow problems. The goal is to enhance traditional “black box” machine learning with physical knowledge of the flow, as described through conservation laws (such as mass, momentum, etc.)in order to improve the performance of each (either through greater accuracy, reduced computational costs or both). Initially you will develop your understanding of artificial neural networks by creating your own “black-box” network to predict flow features (e.g. the lift and drag coefficients for the flow around a 2-d obstacle) based upon prescribed inputs (e.g. Re, Ma, angle of attack). You will then build upon this, combined with your understanding of fluid dynamics, to incorporate models of the underlying flow physics into your network. This will build upon the work of others to begin with – allowing you to make informed decisions regarding the avenues that you will explore in further detail.