DBNets: simulation-based inference of young planets' masses in dusty discs with deep learning

The ubiquitous substructures observed in protoplanetary discs can be signatures of embedded young planets, crucial to fully understanding planet formation and evolution. Current methods to model and characterize these systems are limited in their ability to fully account for the observed complex morphologies or present high computational and time costs. In our recently developed tool DBNets, now publicly available, we implemented a simulation-based-inference pipeline that exploits state-of-the-art deep learning techniques to analyze dust substructures observed in protoplanetary discs and quickly estimate the mass of the allegedly embedded planet that would generate such substructures. The main achievement of our work has been the development of an uncertainty quantification method that, in addition to providing the statistical significance of our results, also allows us to detect out-of-distribution data and provides a rejection criterion to identify unreliable estimates. In this talk, I will delve into DBNets functioning, focusing on how we managed to measure the uncertainties and highlighting why this is crucial in an inference tool. The problem we addressed is characterised by very complex and expensive-to-generate data, with multiple sources of uncertainty. This led to many challenges that we had to address during the development of this tool. I will show our attempts, failures and achievements discussing our ideas, solutions and future plans to take advantage of the rapid development of artificial intelligence techniques in our hunt for young planets.