ScholarWorks@성균관대학교

초록

Large-scale simulations of atomic systems play a vital role in fields, including chemistry, materials science, and more. However, they often struggle with efficiency due to high computational costs. To overcome this, machine learning force fields (MLFFs) have emerged as a promising alternative, providing accuracy comparable to density functional theory (DFT) while lowering costs. Nonetheless, variations in potential energy based on atomic environments make it challenging to apply kernel-based models across different compounds. To address this issue, we present an MLFF using the Robust Bayesian Committee Machine (RBCM) for various compounds containing carbon, nitrogen, and hydrogen (C-N-H). The MLFF is trained with first-principles calculations and molecular dynamics simulations of various C-N-H molecules. Testing these models with longer amine structures and two Diels-Alder reactions shows excellent agreement with DFT results, demonstrating that machine learning models accurately predict potential energy surfaces of organic molecules and offer opportunity for studying a wide range of C-N-H compounds.

키워드