Taking On the "Curse of Dimensionality" in Chemical Kinetics: Complex Chemical Reaction Prediction Using Manifold Learning

University Of California-Irvine, Irvine CA

Investigators

Abstract

With support from the Chemical Theory, Models and Computational Methods program in the Division of Chemistry, Dr. Dmitrij Rappoport of the University of California, Irvine aims to make computational discovery of new chemical reactions and optimization of known reactions faster and cheaper. While the demand for efficient and economical ways of making chemical compounds—pharmaceuticals, organic light-emitting diode (OLED) materials, and many more-is only increasing, systematic search for these synthetic methods remains a challenge. One major obstacle is fundamental: the number of possible pathways that a chemical reaction may take increases exponentially with the number of atoms, making it impossible to test them all even with the most powerful computers. In order to tackle this “curse of dimensionality”, Rappoport will develop methods to recognize low-dimensional structures in the abundance of possibilities, removing information from computational models that is unrelated to chemical reactions. Taking advantage of machine learning methods of dimensionality reduction to distill enormous data sets into compact representations, this research will enable modeling and discovery of complex chemical reactions for green chemistry and heavy metal-free catalytic processes. With its emphasis on data science and machine learning techniques to explore chemical reactions, this work will introduce the next generation of physical scientists to the tools and techniques of data science and helps to improve their data literacy. Under this CTMC award, Dmitrij Rappoport will develop methods for constructing low-dimensional coordinate sub-manifolds from potential energy surfaces of chemical reactions using non-linear dimensionality reduction and discretization techniques. This new set of computational tools is designed to complement the existing semilocal transition state search methods and to explicitly address the problem of high dimensionality of potential energy surfaces. Non-linear dimensionality reduction techniques separate reactive and nonreactive degrees of freedom and thus create computational models of chemical reaction mechanisms that are computationally efficient to sample and lend themselves to definitions of similarity between reaction mechanisms in terms of changes in bonding. These low-dimensional computational models will be used by Rappoport to make machine learning–based predictions of chemical reactivities that avoid the fundamental limitations of methods that operate on the full high-dimensional potential energy surfaces. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.