Master thesis defense by Yifan Liu

Title: A Physics driven approach to solving inverse problem using Neural Networks

Abstract: Recent developments in machine learning made it possible to solve complex prob- lems through a purely statistical approach. One might ask is it possible to combine the power of neural networks and prior information, in this case laws of physics, to make a neural network drastically better? This project explores this idea by imple- menting physics information in the loss function of a fully connected feed forward neural network, in order to generally solve a scaleable inverse problem with out train- ing for each specific case. Despite limitations of computational resource, the final results from physics informed neural network(PINN) is slightly better in goodness of prediction than the non-physics informed neural network(noPINN) when there are sufficient amount of training samples and neural network layers.

Supervisor: Klaus Mosegaard

Censor: Thomas Mejer Hansen