ABFML: A problem-oriented package for rapidly creating, screening, and optimizing new machine learning force fields

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Developing Machine Learning Force Fields (MLFFs) requires extensive iterative testing and tuning.
• Existing software packages are often limited to single descriptors or models, hindering innovation.
• There is a need for efficient and flexible tools to facilitate MLFF development.

Purpose of the Study:

• To introduce ABFML, a novel PyTorch-based package designed to accelerate MLFF innovation.
• To provide researchers with a rapid, user-friendly tool for constructing, screening, and validating new MLFF models.
• To lower the barriers to entry for developing and applying advanced MLFFs.

Main Methods:

• Development of the ABFML package utilizing the PyTorch framework.
• Implementation of standardized module operations for rapid model construction.
• Integration with graphics processing unit (GPU) environments for accelerated computations.

Main Results:

• ABFML enables swift establishment of MLFF models through standardized operations.
• The platform supports seamless transition to GPU environments for large-scale parallel simulations.
• ABFML significantly reduces the time and effort required for MLFF development compared to traditional methods.

Conclusions:

• ABFML effectively promotes innovation in MLFF development by providing an efficient and accessible platform.
• The package facilitates rapid construction, screening, and validation of novel force field models.
• ABFML is poised to expedite the application of MLFFs across diverse scientific domains.