The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Designing New Energy Materials with Machine Learning with Rafael Gomez-Bombarelli - #558
Feb 7, 2022
Rafael Gomez-Bombarelli, an MIT assistant professor in material science, dives into the fusion of machine learning and atomistic simulations for energy materials. He discusses virtual screening and inverse design techniques, sharing insights on their unique challenges. The conversation highlights generative models and the crucial role of training data in simulations. Rafael also explains how simulation results inform modeling efforts and the significance of hyperparameter optimization in making predictive models more effective for material design.
AI Snips
Chapters
Transcript
Episode notes
Zeolite Discovery
- Gomez-Bombarelli's team used virtual screening to discover a new zeolite material.
- This material shows promise for cleaning diesel exhaust fumes.
Permutation Invariance
- Generative models for materials design face a permutation invariance challenge.
- Atom indexing doesn't matter physically, but algorithms struggle with this.
Generative Model Evolution
- Early generative models for molecules adapted text-based models using string representations.
- Researchers now explore graph-based models to address limitations of string representations.