Lifelong Machine Learning Potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning potentials (MLPs) offer high accuracy with low computational cost.
• Current MLPs require system-specific training and struggle with diverse chemical elements.
• Retraining MLPs on new data is computationally intensive and risks knowledge loss.

Purpose of the Study:

• To develop a continuously adapting machine learning potential (lMLP).
• To enhance MLPs' ability to handle diverse chemical elements and systems.
• To enable autonomous, on-the-fly training of MLPs on new data streams.

Main Methods:

• Introduction of element-embracing atom-centered symmetry functions (eeACSFs) combining structural and elemental information.
• Development of a lifelong machine learning potential (lMLP) using uncertainty quantification.
• Application of continual learning strategies, including the CoRe optimizer, for autonomous training.

Main Results:

• eeACSFs effectively represent diverse chemical elements within MLPs.
• The lMLP demonstrates continuous adaptation and maintains accuracy on new data.
• Continual learning strategies facilitate on-the-fly training and broader applicability.

Conclusions:

• The developed lMLP with eeACSFs overcomes limitations of traditional MLPs.
• Lifelong learning and continual strategies enable robust and adaptable ML potentials for complex chemical systems.
• This approach paves the way for more efficient and versatile computational simulations.