Physics-informed machine learning

Models are used in almost all scientific disciplines to describe relationships between what we know and what we do not know, but want to learn more about. There are two ways to derive and deduce models - either using theory-based first principles or data-driven approaches.

In this project we will combine these two approaches by developing new data-driven machine learning models which are leveraged with theory-based first principles and also can enable new knowledge discoveries in the physical domains in which they are employed.To realize this we have identified three subprojects.

In the first subproject we will develop flexible probabilistic models that enforces nonlinear constraints.

In the second subproject we will create probabilistic dynamical models based on so-called physics-informed neural networks.

In the third subproject we will build a novel modular neural network model where some modules are specified using physical laws and some are data-driven.The subprojects will be carried out by the applicant and two PhD students.

The models will not target a specific application, but we will nevertheless collaborate with groups from physics, primarily here in Uppsala, to make sure that our research is relevant and for getting access to real data and domain-specific expertise.