Indirect learning and physically guided validation of interatomic potential models | JoVE Visualize

Area of Science:

Computational Materials Science
Machine Learning in Physics
Condensed Matter Physics

Background:

Machine learning (ML) interatomic potentials offer a balance between accuracy and computational speed for material simulations.
Existing methods often require computationally expensive quantum-mechanical calculations for training, limiting dataset size and simulation scale.

Purpose of the Study:

To introduce a novel ML-based approach for training interatomic potentials by leveraging accurate ML models as data generators.
To demonstrate that this method enables the creation of extensive reference datasets without quantum-mechanical computations, thereby improving fast ML potentials.
To apply this technique to simulate complex phenomena in disordered silicon at an unprecedented scale.

Main Methods:

An accurate, computationally intensive ML potential was used to generate reference data (atomic configurations and energies) for training faster ML potentials.
A 'potential-to-potential' training strategy was employed, bypassing the need for secondary quantum-mechanical calculations.
The method was applied to simulate the vitrification and polycrystalline grain formation of disordered silicon using a million-atom system.

Main Results:

The 'potential-to-potential' training significantly improved the accuracy of faster ML potentials, particularly those with less flexible functional forms.
Extensive reference datasets were efficiently generated, facilitating the training of high-quality, computationally inexpensive interatomic potentials.
Large-scale simulations of disordered silicon, including phase transitions and defect formation, were successfully performed.

Conclusions:

This work presents a conceptual advancement in the machine learning of interatomic potentials, offering a scalable route to high-fidelity simulations.
The developed method significantly accelerates the generation of training data, paving the way for faster and more extensive simulations of condensed-phase materials.
This approach provides a powerful tool for exploring material behavior at large scales and under various conditions.