CAREER: Minimize ab initio Tasks in Dynamics Simulations of Chemical Reactions with Active Machine Learning

University Of Hawaii, Honolulu

Investigators

Abstract

With support from the Chemical Theory, Models and Computational Methods (CTMC) program in the Division of Chemistry and the Established Program to Stimulate Competitive Research (EPSCoR), Rui Sun of University of Hawaii, Mānoa will work to develop a novel machine learning algorithm to accelerate simulations of chemical reactions. Since a chemical reaction can take place at a very fast rate and on a very small scale, sometimes too fast and small for equipment to directly measure, computer simulations, which follow the motions of atoms by solving equations of motion, play an essential role in seeking a thorough understanding of the nature of a chemical reaction. However, such simulations are computationally very demanding (e.g., require a large number of computers to run for a long period of time), thus severely limiting the scope of their applications. Rui Sun and his research group are developing a novel machine learning algorithm that utilizes the information gathered along the study of the chemical reaction to speed up simulations potentially by an order of magnitude or more. This algorithm, along with a specifically designed data storage and fetch system, will be open-source and implemented with state-of-the-art computational chemistry software. Simulations boosted by this machine learning algorithm have the potential to achieve unprecedented efficiency and accuracy, and thus to push the boundary of our knowledge on chemical reactions, perhaps the central element of the field of chemistry. By introducing computation as a different means for problem solving, Rui Sun will also develop educational programs to enhance the learning experience of students at the University of Hawaii, which hosts the largest population of Pacific islander students in America. Rui Sun is developing an active machine learning protocol with the aim of increasing the efficiency of ab initio molecular dynamics simulations of chemical reactions by at least one order of magnitude. This is to be achieved by replacing 90+% of the ab initio energy gradient calculations with a specifically designed machine learning algorithm, interpolating moving ridge regression (IMRR). IMRR is trained on data fetched from an indexed library containing all the ab initio energy gradients calculated in the previous simulations and updated on the fly as each trajectory progresses. Rui Sun and his research group will also develop an optimal molecular descriptor to efficiently identify ab initio training data that yields the smallest error in IMRR-predicted energy gradients. Each IMRR will provide a risk factor, indicating its probability of reproducing the ab initio energy gradient and maintaining a well-behaved trajectory. A high-risk factor will be used as a rejection criterion to refer back to ab initio calculation in order to protect the integrity of the simulation. Due to the expected dramatic boost in efficiency, the proposed active machine learning protocol has the potential to push the capability of AIMD to an unprecedented level and to set a new standard for dynamics simulation of chemical reactions. Rui Sun will also develop computation modules to support current chemistry labs as well as the very first computational course for non-CS (Computer Science) STEM (Science, Technology, Engineering and Mathematics) majors at the University of Hawaii at Mānoa. 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.