Seminars

Home Centers & Programs AI and Natural Sciences Seminars

FIELD: AI and Natural Sciences

DATE: May 08 (Thu), 2025

TIME: 14:00 ~ 16:00

PLACE: 7323

SPEAKER: Hyun Woo Kim

HOST: Cho, Kwang Hyun

INSTITUTE: GIST

TITLE: Applying AI to Predict Molecular and Material Properties and Simulate Practical Quantum Molecular Dynamics

ABSTRACT: Computational chemistry aims to calculate molecular properties as accurately as possible and simulate the molecular system within reasonable modeling. To this end, computational chemists usually try to find an approximate solution to the time-dependent or time-independent Schrödinger equation. Recent applications of artificial intelligence (AI) techniques show that AI can reduce the computational effort of solving the Schrödinger equation in the presence of enough data. Specifically, with machine learning (ML), predicting the properties of molecules and materials is intensively studied since it can play a role in the novel design of molecules and materials with target properties in the future. The first step in ML is generating and curating data, including representations of molecules and materials in a machine-readable format. These representations usually come from domain knowledge in chemistry-related fields. Here, I will present two ML methods to explain the essential moiety of molecules and optimize the expression of molecules and materials. At the beginning of my talk, I will shortly review AI and then two algorithms of graph neural networks (GNNs): graph convolutional network (GCN) and graph attention network (GAT). For GAT, I will explain how this algorithm is applied to select essential substructures in organic molecules. Then I will present how to apply neural networks to find data-driven representations, including loss function design. Multiple tests on molecular and materials datasets showed that ML algorithms using the new expression marked improved performance. We also observe that new representations become more valuable in small datasets. Finally, for time-dependent simulations, I will briefly discuss how ML can be used to improve Ehrenfest dynamics.

FILE

List