aims-PAX: Parallel Active Exploration Enables Expedited Construction of Machine Learning Force Fields for Molecules...

Area of Science:

• Computational Chemistry and Materials Science
• Machine Learning Applications in Science
• Atomistic Simulations

Background:

• Machine learning force fields (MLFFs) are crucial for advancing atomistic simulations.
• Developing accurate MLFFs requires extensive reference data, robust architectures, and efficient active learning.
• Current methods face challenges in handling complex molecular and material systems.

Purpose of the Study:

• To introduce aims-PAX, an expedited, multitrajectory active learning framework for streamlined MLFF development.
• To provide a modular, high-performance workflow for diverse sampling and scalable training.
• To enable rapid deployment of accurate MLFFs for a wide range of research applications.

Main Methods:

• Developed aims-PAX, a framework coupling diversified sampling with scalable CPU/GPU training.
• Integrated aims-PAX with the FHI-aims *ab initio* code.
• Supported state-of-the-art ML models and dataset generation using general-purpose force fields.

Main Results:

• Demonstrated aims-PAX on challenging systems: flexible peptides, multiple organic molecules, solvated molecules, and CsPbI3 perovskite.
• Achieved up to a 3-order-of-magnitude reduction in required reference calculations.
• Enabled simulation of large solvated molecules (>1000 atoms) and a 10-fold speedup in active learning time.

Conclusions:

• aims-PAX significantly accelerates the development of stable and accurate MLFFs.
• The framework automates the selection of challenging systems for efficient data generation.
• aims-PAX is a versatile platform for next-generation atomistic simulations in academia and industry.