| How can we design machine learning models that not only predict outcomes but also explain why? Many systems in the natural sciences—molecules, materials, and beyond—exhibit intricate structure–property relationships that are difficult to model analytically. Our group develops interpretable learning frameworks that bridge mathematical modeling, data science, and physical chemistry to uncover the rules governing these relationships.
We explore new ways of representing complex systems—such as graph-based and text-based encodings—that capture hierarchical structure while remaining amenable to rigorous mathematical analysis. These representations, combined with models such as equivariant neural networks and attention-based architectures, allow us to probe how local and global features interact to shape system behavior. The same mathematical tools that describe geometry, symmetry, and topology become powerful means of explaining chemical and physical phenomena. This project focuses on developing and analyzing interpretable machine learning models for structured data, using molecular systems as a rich testbed. Students will explore how graph theory, representation learning, and optimization can be combined to derive transparent, explainable models that generalize beyond chemistry—offering insights relevant to materials science, physics, and other data-driven domains. |
