Programme

O.19 -- Oral, VCTP-47

Date: Tuesday, 2 August 2022

Time: 16:50 - 17:15

Construction molecular potential using machine learning

Duong D. Hoang-Trong (1), Khang Tran (2), Quan-Hao Truong (1), Doan-An Trieu (1), Ngoc-Loan Phan (1)*, Van-Hoang Le (1)

(1) Ho Chi Minh City University of Education; (2) New Jersey Institute of Technology

Theoretical simulation of physical and theoretical phenomena is an essential way to understand their nature and examine experimental observation. The prerequisite task for theoretical simulation is constructing a potential model for molecules, and then solving its Time-Independent Schrödinger Equation (TISE) to get the expected values that are needed to be consistent with experiments. Various methods have been developed for constructing potential model – including the Single Active Electron model with a soft-Coulomb pseudo-potential governed by a set of parameters [1], [2]. For complex molecules, finding a large number of parameters that need to satisfy many outcome conditions by traditional statistical learning faces to challenge due to the limit of computational resources. Besides, solving the TISE is also a costly and time-consuming process. Concurrently, the emergence of machine learning with advanced artificial models capabilities to overcome these difficulties. In this report, we apply machine learning to solving two-fold goals: The first one is re-constructing the soft-Coulomb pseudo-potential of HCN molecule with multiple parameters by optimizing in a way that the energies and permanent dipoles of different molecular orbitals converge to the experimental values. For doing this task, we apply two machine learning algorithms - Neural Network (NN) and Light Gradient Boosting Machine (LGBM) - to predict potential parameters, then combine both predictions to make the final predictions with another algorithm - Random Forest. Similarly, the second one is to solve the inverse problem. We use NN to build surrogate models which solve the TISE to determine rapidly the energies and permanent dipoles based on potential parameters. Both works have fruitful results. The Mean Absolute Percent Error (MAPE) of most paraments in the first task is less than 0.5%. And the Symmetry Mean Absolute Percent Error (sMAPE) of most energies value in the second task is less than 1%. However, the sMAPE of the permanent dipoles is high for some states – this is still in the improvement process. [1] M. Peters, T. T. Nguyen-Dang, E. Charron, A. Keller, and O. Atabek, “Laser-induced electron diffraction: A tool for molecular orbital imaging,” Phys. Rev. A, vol. 85, no. 5, p. 53417, May 2012, doi: 10.1103/PhysRevA.85.053417. [2] C. I. Blaga et al., “Imaging ultrafast molecular dynamics with laser-induced electron diffraction,” Nature, vol. 483, no. 7388, pp. 194–197, 2012, doi: 10.1038/nature10820.

Presenter: Hoàng Trọng Đại Dương