Accelerating Materials Discovery with Application-Focused Machine Learning.

24^th October 2025, 15:00
Felix Therrien
Mila Institute, Montreal, Canada

Abstract

Machine learning is often seen as a solution to accelerate the discovery of functional materials that are central to several challenges pertaining to sustainability. However, many ML models are disconnected from the reality of making and testing such materials leading to models that perform well in theory, but are not used in practice. This presentation will discuss our work on building application-focused generative and predictive models for catalyst and battery materials discovery. First, I will discuss our physics based uncertainty aware model to predict the performance of gas diffusion electrodes for CO2 reduction and drive an automated laboratory with bayesian optimization. Then, I will present our catalyst generation and discovery platform and the steps we are taking to identify realistic materials. Finally, I will talk about our work on making accurate and affordable models for predicting ionic conductivity in solid state electrolytes, including our efforts in assembling the OBELiX dataset.