Related Experiment Video
Updated: May 21, 2025
Computation of Atmospheric Concentrations of Molecular Clusters from ab initio Thermochemistry
Published on: April 8, 2020
Fine-Tuned Global Neural Network Potentials for Global Potential Energy Surface Exploration at High Accuracy
Xin-Tian Xie1, Tong Guan1, Zheng-Xin Yang1
1State Key Laboratory of Porous Materials for Separation and Conversion, Collaborative Innovation Center of Chemistry for Energy Material, Shanghai Key Laboratory of Molecular Catalysis and Innovative Materials, Key Laboratory of Computational Physical Science, Department of Chemistry, Fudan University, Shanghai 200433, China.
AtomFT is a new machine learning potential (MLP) architecture that accurately predicts potential energy surfaces (PES) for materials science. This method achieves high accuracy for complex systems, enabling better predictions of material properties.
Area of Science:
- Computational materials science
- Chemical physics
- Machine learning applications
Background:
- Machine learning potentials (MLPs) are valuable for exploring global potential energy surfaces (PES).
- Achieving high accuracy (<1 meV/atom) on complex PES data is a significant challenge for determining material properties.
- Existing methods struggle with the diversity and complexity of global PES datasets.
Purpose of the Study:
- To develop a lightweight, fine-tunable MLP architecture (AtomFT) for accurate global PES exploration.
- To enable simultaneous global exploration and high-accuracy description of target system PES.
- To demonstrate efficient training and inference on common CPU platforms.
Main Methods:
- Developed AtomFT, a lightweight fine-tuning MLP architecture.
- Utilized a pretrained many-body function corrected global neural network (MBNN) potential as a basis.
- Iteratively updated atomic features from the MBNN model to generate fine-tuning energy contributions.
- Implemented AtomFT on CPU platforms for efficient computation.
Main Results:
- AtomFT achieved high efficiency in both training and inference.
- Demonstrated high performance in challenging PES problems, including oxides, molecular reactions, and crystals.
- Significantly enhanced PES prediction accuracy to below 1 meV/atom across all tested systems.
Conclusions:
- AtomFT offers an efficient and accurate solution for global PES exploration and prediction.
- The method successfully addresses the challenge of high-accuracy PES description for complex systems.
- AtomFT has broad applicability in computational materials science for predicting thermodynamics and kinetics properties.
