Uncertainty in Materials Science Property Prediction - The Good, The Bad, and The Uncalibrated

Explore the critical concepts of aleatoric and epistemic uncertainty in machine learning applications for materials science through this comprehensive 51-minute technical presentation by Ashley Dale from the University of Toronto. Learn to distinguish between different types of uncertainty and understand their implications for model trustworthiness in materials property prediction. Master an uncertainty characterization workflow that begins with theoretical principles and progresses to practical approximations, covering how to identify uncertainty in models, estimate uncertainty in predictions for both trained and untrained models, and leverage model uncertainties to identify promising research directions. Examine ALIGNN models for predicting bandgap values from the JARVIS dataset as a practical case study, and work through a leave-one-element-out experiment demonstrating how Random Forest models generalize differently depending on omitted elements. Gain hands-on experience through a scaffolded Jupyter notebook coding exercise designed to improve programming skills while developing intuition about model prediction reliability. Discover state-of-the-art techniques for quantifying uncertainty, understand variance decomposition through conservation assumptions, and explore model calibration methods for regression models. Access the Random Forest Uncertainty tool and related materials to apply these concepts in your own materials science research projects.