The goal is to create learning algorithms able to modify their mechanisms according to the domain under analysis and the accumulation of meta-knowledge (knowledge about the behavior of the learning agent in different environments).
In this project, we aim at building predictive models that exploit both data and domain knowledge. Recent work in physics-informed neural networks has set the path to adding the ability to reinforce statistical models with known domain constraints. We aim to find novel ways to incorporate knowledge as we search for data patterns.
Machine learning has already played a significant role in analyzing astrophysical data. Our goal in this project is to generate new learning algorithms that can aid in searching for data patterns in the areas of dark matter and cosmology.
We aim to find automated ways to extract new physical laws and equations from scientific data, mainly in physics and astronomy. The idea is to go beyond search techniques exploring the space of possible new equations. We plan to automate the discovery process by “learning-to-search” solutions that satisfy physical constraints and display properties typical to known proven equations.