FeNNol: An efficient and flexible library for building force-field-enhanced neural network potentials

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Neural network interatomic potentials (NNPs) offer accurate molecular modeling, but can be computationally expensive.
• Advances in ML and hybrid models have expanded the possibilities for NNPs.
• Combining ML with physics-based force fields presents a promising avenue for efficient simulations.

Purpose of the Study:

• Introduce FeNNol, a novel library for developing and deploying force-field-enhanced neural network potentials.
• Provide a flexible and modular framework for constructing hybrid ML-force field models.
• Enhance the performance and accessibility of NNPs for molecular simulations.

Main Methods:

• Developed FeNNol, a Python library for building, training, and running hybrid NNPs.
• Integrated state-of-the-art embeddings with ML-parameterized physical interactions.
• Utilized Jax for automatic differentiation and just-in-time compilation for accelerated NNP evaluation.

Main Results:

• FeNNol enables flexible construction of hybrid ML-force field models without complex programming.
• Achieved near-parity simulation speeds with traditional force fields (e.g., AMOEBA) for the ANI-2x model on GPUs.
• Demonstrated significant performance gains in NNP evaluation through Jax integration.

Conclusions:

• FeNNol facilitates the creation and application of advanced hybrid NNP architectures.
• The library bridges the performance gap between ML potentials and standard force fields.
• FeNNol is expected to accelerate research across various molecular simulation problems.