A Hierarchy of Fragment-based Quantum Chemical Models Incorporating Machine Learning for Applications in Nanoscale Systems

点击翻译按钮获取中文摘要

Krishnan Raghavachari of Indiana University is supported by an award from the Chemical Theory, Models and Computational Methods program in the Division of Chemistry to develop a set of quantum chemical computational methods incorporating machine learning for broad applications in nanoscale systems. While many accurate methods have previously been developed in quantum chemistry by different groups, their applicability thus far has been limited to small molecules due to their associated prohibitive computational cost. Attaining computational efficiency along with accuracy represents the most fundamental obstacle for quantum chemistry today. The new methods that are being proposed by Raghavachari aim to fill this need to treat medium-sized and large molecules accurately, providing systematic well-tested models to the study of nanoscale systems. The methods will combine ideas based on molecular fragmentation, systematic error-correction, and state-of-the-art machine learning to achieve high accuracy in conjunction with computational efficiency, with the aim of providing new tools to solve challenging problems involving intermediate-sized to large molecular systems and materials. Computational nanoscience, as a rapidly expanding field, is attracting student interest, and these projects are expected to provide an excellent training platform for the next generation of researchers in computational chemistry. In order to accomplish goals of this project, Dr. Raghavachari and coworkers will build on two different lines of research that have been developed in the group. In the first approach, they will develop a stepping-stone model based on Connectivity-based Hierarchy (CBH) to provide systematic error corrections to density functional theory (DFT) to result in accuracy comparable to coupled cluster calculations. This will be done using a two- or three-layer model where more accurate calculations are carried out on small fragments to correct for the DFT errors and achieve chemical accuracy. In the second approach, Raghavachari will develop a general computational framework that unifies the advantages of connectivity-based fragmentation with graph network-based machine learning to attain sub-kcal accuracy (“chemical accuracy”) in the calculated energies. Raghavachari has proposed that node embeddings based on molecular fragments will outperform most molecular fingerprints used traditionally in most machine learning applications. The newly developed methods have the potential to provide unprecedented accuracy for the treatment of complex problems involving nanoscale systems. The resulting computational tools will be developed in a platform-independent manner and should work with multiple quantum chemical packages, and will be made freely available for use by other research groups.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.