Large-Scale Atomic Simulation via Machine Learning Potentials Constructed by Global Potential Energy Surface...

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Quantum mechanics (QM) calculations are vital for understanding chemical problems but are computationally expensive.
• Traditional QM methods face a trade-off between accuracy and speed for large-scale simulations.
• Machine learning (ML) offers a promising alternative for high-speed, high-accuracy atomic simulations.

Purpose of the Study:

• To overview recent advances in ML methodologies for atomic simulations.
• To address challenges in ML model predictivity and robustness.
• To introduce a new class of global neural network (G-NN) potentials for enhanced simulations.

Main Methods:

• Utilizing stochastic surface walking for representative potential energy surface (PES) data generation.
• Employing power-type structure descriptors with feed-forward neural networks (NN).
• Developing global neural network (G-NN) potentials compatible with complex global PES data.

Main Results:

• G-NN potentials demonstrate compatibility with highly complex global PES data.
• The developed methods improve data representation for ML models.
• Successful applications in material and reaction simulations showcase potential for discovery.

Conclusions:

• ML-based atomic simulations, particularly with G-NN potentials, significantly enhance the speed and accuracy of chemical simulations.
• Addressing data generation, model extensity, and data representation is crucial for robust ML applications.
• These advancements facilitate the discovery of novel materials and chemical reactions.