Physics-aware machine learning

Physics-aware machine learning

Large scale physical simulations are important for understanding nature and for developing new technologies. However, simulations are limited by computational power. Our idea is to do physical simulations where the physics only have to be solved in a few points, and then have a trained neural network fill in the simulation in the remaining space, in a physical correct way. We will focus on magnetic fields, where we will train a neural network to interpolate and extrapolate in a physically correct way by training it not only on magnetic field simulations but also defining losses defined on Maxwells equations, ensuring that these are fulfilled. This technique can then be transferred to additional physics subsequently, by defining appropriate losses for these. In terms of physical simulations, the project will be beyond state-of-the-art in several ways. Currently simulations are interpolated using a variety of techniques, linear, cubic and spline interpolation but none of these produce fields that obey Maxwells equations. In computer science the project will be beyond state-of-the-art as we will determine the ideal machine approach among two different methods - generative adversarial networks and deep generative models.