Data selection methods for the construction of neural network-based potentials

• Grant number: 26/00789-3
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: March 01, 2026
• End date: December 31, 2026
• Field of knowledge: Physical Sciences and Mathematics - Physics - Condensed Matter Physics
• Principal Investigator: Luana Sucupira Pedroza
• Grantee: Matheus Tierro de Paula
• Host Institution: Instituto de Física (IF). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Associated research grant: 23/09820-2 - Materials by design: from quantum materials to energy applications, AP.TEM

Abstract

The study of solid/water interfaces (liquid water or ice) is of paramount importance for understanding photocatalytic and electrocatalytic processes, among others, related to the development of more efficient and "clean" energy sources. A realistic description of these interfaces requires highly accurate simulations capable of properly describing both water-water interactions and water-surface interactions, such as those provided by density functional theory (DFT). However, due to their high computational cost, these simulations are limited to small system sizes and short time scales.In recent years, machine-learned interatomic potentials (MLIPs) have emerged as a promising alternative, combining near-DFT accuracy with computational costs comparable to those of classical (parameterized) potentials. Nevertheless, the quality of these potentials strongly depends on the selection of the data used for training-that is, on the choice of configurations sampled in phase space that ensure adequate coverage of the physically relevant regions of the system.The objective of this project is to investigate methodologies for the efficient selection of configurations used in the training stage of neural networks, with a focus on liquid water systems. Different selection approaches, such as random sampling and clustering techniques, will be compared in order to optimize the sampling of the system's configurational space and reduce the need for redundant data. (AU)