Date/Time Date(s) - 06/03/20262:30 pm - 3:30 pm
Categories
Prof. Oleksandr Voznyy
University of Toronto
The current zeitgeist in materials informatics operates on a fundamentally flawed premise: that feeding enough structural data into a Neural Network will magically yield quantum mechanical truths. It does not. Standard AI in materials science is not learning physics; it is merely interpolating existing trends.
In this talk, I will delineate the strict boundaries of what machine learning can realistically achieve, and where it catastrophically fails. We will begin with the “easy” local-minima successes: Machine Learning Interatomic Potentials (MLIPs) for solid-state electrolytes, where local atomic environment mapping is sufficient to predict ion diffusion.
However, we quickly hit the wall when moving to systems governed by non-local effects. In catalysis, standard MLIPs completely collapse due to their inability to describe long-range charge transfer. I will outline the necessity of 4th-generation potentials and discuss our recent work on electronegativity training as a remedy.
Finally, I will present our recent pivot toward a fundamentally different architecture: moving away from data-fitting and toward law-discovery. By embedding low-rank constraints (via Singular Value Decomposition and related methods), we force the algorithm to compress the data into underlying physical laws. This shift from interpolation to true extrapolation represents the most optimistic path forward for discovering the next generation of energy materials.
Prof. Oleksandr (Alex) Voznyy is an Assistant Professor in Clean Energy at the University of Toronto Scarborough. A semiconductor physicist by training (Ph.D., Chernivtsi National University, Ukraine), his background spans theoretical modeling of laser-assisted quantum well intermixing (Sherbrooke), many-body problems and Auger processes in quantum dots (NRC Canada), and the synthesis of nanomaterials for photovoltaics (Sargent Group). Today, his research group works at the intersection of atomistic simulations, high-throughput automated synthesis, and machine learning. His lab focuses on strictly theory-guided discovery of novel materials for Li-ion batteries, hydrogen storage, and CO2 reduction—ruthlessly separating the physical signal from the algorithmic hype.
In-Person: ABB 102
Online: Echo360