Related Experiment Video
Updated: Oct 3, 2025
14:55
Atomic Scale Structural Studies of Macromolecular Assemblies by Solid-state Nuclear Magnetic Resonance Spectroscopy
Published on: September 17, 2017
15.6K
A nearsighted force-training approach to systematically generate training data for the machine learning of large
Cheng Zeng1, Xi Chen1, Andrew A Peterson1
1School of Engineering, Brown University, Providence, Rhode Island 02912, USA.
The Journal of Chemical Physics
|February 16, 2022
Summary
This study introduces a machine-learning (ML) framework for atomistic simulations. It efficiently generates training data for large systems by focusing on uncertain atoms, improving ML potential robustness.
Area of Science:
- Computational Materials Science
- Atomistic Machine Learning
- Electronic Structure Theory
Background:
- Atomistic machine learning (ML) methods face challenges in generating suitable training data for large atomic systems.
- Current ML approaches often rely on the nearsightedness principle, using local atomic environments to predict energies.
- Ensuring training data relevance and convergence is critical for the accuracy of ML potentials.
Purpose of the Study:
- To develop a systematic framework for creating appropriate training data for large atomic structures in ML simulations.
- To exploit the nearsightedness principle of ML models for efficient training set generation.
- To improve the robustness and accuracy of ML potentials for large-scale atomistic modeling.
Main Methods:
- Developed a framework utilizing per-atom uncertainty estimates to identify critical atoms for training data extraction.
- Extracted small, converged atomic 'chunks' centered around uncertain atoms, ensuring they meet ML cutoff radii and electronic structure convergence criteria.
- Calculated electronic structures for these chunks and incorporated only the central atom's force into the training set to avoid boundary noise.
Main Results:
- Demonstrated the generation of robust ML potentials requiring only single-point calculations on small structures, without direct training on large systems.
- Successfully applied the approach to structure optimization of a 260-atom system.
- Extended the framework to handle clusters containing up to 1415 atoms.
Conclusions:
- The proposed framework effectively addresses the challenge of training data suitability for large atomistic systems in ML.
- By focusing on uncertain atoms and converged local environments, accurate and robust ML potentials can be generated efficiently.
- This method offers a pathway to more reliable and scalable atomistic simulations using ML.
