Hyperactive learning for data-driven interatomic potentials.

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Atomistic simulations

Background:

• Interatomic potentials approximate complex quantum mechanical calculations.
• Generating accurate training data is a bottleneck in developing these potentials.
• Existing methods require extensive computational resources for database generation.

Purpose of the Study:

• To introduce hyperactive learning (HAL) for accelerated training database assembly.
• To demonstrate HAL's efficiency in creating data-driven interatomic potentials.
• To validate the accuracy of HAL-generated potentials for materials properties.

Main Methods:

• Developed a HAL framework integrating a biasing term with physical samplers (e.g., molecular dynamics).
• Applied HAL to generate training configurations for AlSi10 alloy and polyethylene glycol (PEG) polymer.
• Utilized the generated data to train Atomic Cluster Expansion (ACE) interatomic potentials.

Main Results:

• HAL significantly reduced the time and data required for training database assembly.
• The HAL-generated ACE potentials accurately predicted macroscopic properties (melting temperature, density) for AlSi10 and PEG.
• Potentials achieved accuracy close to experimental values starting from minimal initial configurations.

Conclusions:

• HAL is an effective strategy for rapid, automated training database generation.
• This approach enables the efficient development of accurate data-driven interatomic potentials.
• HAL has broad applicability in computational materials science and polymer modeling.