Research

Machine Learning Chemical Reactions

Machine learning methods, especially deep learning, have significantly expanded a chemist's toolbox, enabling the construction of quantitatively predictive models directly from data. These models make it possible to explore the gigantic chemical space to make chemical discoveries.



Our group focus on the chemical reaction space. We develop novel graph neural networks that are able to represent any chemical reactions with bond alterations and apply them to model reaction properties such as reaction energy, reaction type, and activation energy. These models have been used to investigate the reaction pathways in battery electrolytes.



---

Accelerated Atomistic Molecular Simulations

Molecular simulations are a powerful computational technique for exploring material behavior and properties based on an understanding of the physics of bonding at the atomic scale. At the core of any molecular simulation lies a description of the interactions between atoms that produces the forces governing the atomic motion. In classical molecular simulations, such interactions are modeled via interatomic potentials, which make it possible to simulate a large number of atoms for extended periods of time.



We develop both physics-based and machine learning interatomic potentials and apply them to study the energetic, structural, thermal, and mechanical properties of materials and devices for renewable energy applications. We also develop model analysis frameworks (e.g. uncertainty quantification methods), aiming at making classical molecular simulations more reliable, reproducible, and accessible.



---

High-throughput Materials Discovery

Quantum chemical theory and computations are powerful tools for understanding and designing materials. Conventional approaches that manually perform the calculations are difficult to manage complex materials science workflows and difficult to fully utilize modern high-performance computing resources.



Building on top of the software stack that powers the Materials Project, we develop high-throughput materials discovery recipes for density functional theory (DFT) and apply them to discover new materials with unique mechanical and photonic properties with applications in solid-state batteries and thermal-photonic devices.



---