Automatic Feature Selection for Atom-Centered Neural Network Potentials Using a Gradient Boosting Decision Algorithm
- 19 November 2024
- journal article
- research article
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 20 (23) , 10564-10573
Abstract
Atom-centered neural network (ANN) potentials have shown high accuracy and computational efficiency in modeling atomic systems. A crucial step in developing reliable ANN potentials is the proper selection of atom-centered symmetry functions (ACSFs), also known as atomic features, to describe atomic environments. Inappropriate selection of ACSFs can lead to poor-quality ANN potentials. Here, we propose a gradient boosting decision tree (GBDT)-based framework for the automatic selection of optimal ACSFs. This framework takes uniformly distributed sets of ACSFs as input and evaluates their relative importance. The ACSFs with high average importance scores are selected and used to train an ANN potential. We applied this method to the Ge system, resulting in an ANN potential with root-mean-square errors (RMSE) of 10.2 meV/atom for energy and 84.8 meV/& Aring; for force predictions, utilizing only 18 ACSFs to achieve a balance between accuracy and computational efficiency. The framework is validated using the grid searching method, demonstrating that ACSFs selected with our framework are in the optimal region. Furthermore, we also compared our method with commonly used feature selection algorithms. The results show that our algorithm outperforms the others in terms of effectiveness and accuracy. This study highlights the significance of the ACSF parameter effect on the ANN performance and presents a promising method for automatic ACSF selection, facilitating the development of machine learning potentials.Keywords
Funding Information
- National Key Research and Development Program of China (2022YFA1503102)
- Shenzhen Municipal Fundamental Research Program (JCYJ20210324115809026)
- Department of Education of Guangdong Province (2021KCXTD012)
- National Natural Science Foundation of China (92270103)
This publication has 54 references indexed in Scilit:
- Charge-Optimized Electrostatic Interaction Atom-Centered Neural Network AlgorithmJournal of Chemical Theory and Computation, 2024
- Atom-centered machine-learning force field packageComputer Physics Communications, 2023
- Physics-Inspired Structural Representations for Molecules and MaterialsChemical Reviews, 2021
- Four Generations of High-Dimensional Neural Network PotentialsChemical Reviews, 2021
- Machine Learning Force FieldsChemical Reviews, 2021
- DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamicsComputer Physics Communications, 2018
- Amp: A modular approach to machine learning in atomistic simulationsComputer Physics Communications, 2016
- Molecular graph convolutions: moving beyond fingerprintsJournal of Computer-Aided Molecular Design, 2016
- An implementation of artificial neural-network potentials for atomistic materials simulations: Performance for TiO2Computational Materials Science, 2016
- Gaussian approximation potentials: A brief tutorial introductionInternational Journal of Quantum Chemistry, 2015