KU Leuven Centre for mathematical Plasma-Astrophysics Seminar
Title: Combining physics and neural networks: an overview
Speaker: Brecht Laperre from CmPA
Abstract
For the past 15 years, machine learning and subsequently, artificial neural networks, has seen a rise in popularity and applicability. While initially most of its success has been in the field of computer vision, for the past 10 years scientist have been looking how the techniques and advantages of deep learning can be applied to physical applications. The first applications of scientific machine learning leverage the capability of neural networks to interfere physics from (observational) data, resulting in data-driven predictive modelling. However, the volume of useful experimental data for this type of problem is limited. Recently a new learning philosophy called `physics informed neural networks' (or PINNs) has been developed that can seamlessly integrate data and mathematical operators, such as partial differential equations, in the design of the deep learning framework. This has opened the pathway for new research that seeks to improve or replace existing numerical methods with new artificial models.
In this seminar, I will provide an overview of how the field of scientific informed machine learning has developed, starting with an introduction behind artificial neural networks, and how their techniques are leveraged to create data-driven models bound by physical laws. Currently, three `bias'-methods exists to enforce a priori physics in artificial neural networks:
Observational bias, which assumes model learns the physics from the provided data;
Inductive bias, which enforces physical restraints through the network's architecture;
Learning bias, which enforces mathematical constraints in the network's loss function.
All three methods are explained and examples are shown from recent publications. The successes of the techniques are demonstrated, together with their limitations.
We also touch on a new technique named `differential physics' that combines numerical modelling with neural networks with the same kind of philosophy.
The seminars are in hybrid mode, you can follow in person in room 200B 02.16 or online at the (permanent) link:
https://eu.bbcollab.com/guest/7406a5ec00dc4ec6948200f9c769d454
