This big question drives our research on machine learning for atomic and molecular systems. Our philosophy is simple: use what we already know to help machines discover what we do not yet know. Machine learning is a computational technique where the machine learns patterns from an initial dataset of inputs without being specifically programmed for a task. It does so by generating a set of algorithms that find correlations between a given data set (training set) and then uses this learned pattern to predict unknown outputs from a given input (test set). Rather than treating machine learning as a black box, we combine it with physical intuition. Our understanding of atomic and molecular physics guides the development of meaningful features and representations that capture the essential physics of a problem. By combining data, physical insight, and machine learning, we aim to build models that are not only accurate, but also capable of revealing new trends and connections across chemical systems.
From Atoms to Molecules
One of the central goals of our group is to understand how molecular properties emerge from the properties of their constituent atoms. We develop machine-learning models that predict molecular properties using atomic information together with physically motivated descriptors. To support this effort, we have developed and continuously maintain extensive databases of spectroscopic constants, dipole moments, and other properties of diatomic molecules. These data provide the foundation for training and testing our models. Our long-term goal is to develop machine-learning models that can reproduce the accuracy of high-level quantum-chemistry calculations at a fraction of the computational cost. Such models could ultimately serve as fast surrogates for quantum chemistry, allowing us to explore the properties of thousands of molecular systems that would otherwise be computationally expensive to study.
Learning Chemical Reactions
Chemistry is ultimately about transformation: atoms and molecules collide, interact, and rearrange to form new substances. Predicting the outcome and rate of an arbitrary chemical reaction remains one of the grand challenges of chemical physics. Our approach is bottom-up. Rather than immediately increasing the size of the molecules, we increase the complexity of the dynamics. We begin with bimolecular reactions and then move to termolecular reactions, where three particles participate simultaneously in the collision dynamics. Termolecular reactions represent one of the major research thrusts of our group and provide a particularly demanding test for machine learning. By learning relationships between molecular properties, interaction forces, collision dynamics, and reaction rates, we aim to develop models capable of predicting chemical reactivity across increasingly broad classes of systems.
