Synergistic Integration of Physical Embedding and Machine Learning Enabling Precise and Reliable Force Field

Area of Science:

• Computational chemistry
• Machine learning in materials science

Background:

• Machine-learning force fields (MLFFs) offer quantum chemical accuracy but struggle with extrapolation and long-range interactions.
• Existing MLFFs face challenges in predicting macroscopic properties and navigating novel chemical spaces.

Purpose of the Study:

• To develop a physically informed neural network (PINN) force field by synergistically integrating physical principles and machine learning.
• To address limitations of current MLFFs in chemical space extrapolation, electrostatic interactions, and macroscopic property prediction.

Main Methods:

• Incorporated physical knowledge into neural network parameters within a PINN framework.
• Employed the novel Tabu-Adam algorithm for efficient global optimization under physical constraints.
• Utilized the AMOEBA+ force field as the physics-based model and trained/tested on the diethylene glycol dimethyl ether (DEGDME) dataset.

Main Results:

• Achieved a precise and noise-robust machine learning force field.
• Demonstrated remarkable generalization and density functional theory (DFT) accuracy in describing molecular interactions.
• Enabled accurate prediction of macroscopic properties, such as diffusion coefficients, with reduced computational cost.

Conclusions:

• The developed PINN force field represents a breakthrough in constructing accurate and robust MLFFs.
• This approach offers a fundamental framework for future development of physically informed machine learning models in chemistry.
• The study highlights the potential of combining physical principles with machine learning for efficient and accurate molecular simulations.