Transferable Dynamic Molecular Charge Assignment Using Deep Neural Networks.

Area of Science:

• Computational Chemistry
• Machine Learning in Science
• Quantum Mechanics

Background:

• Accurate computation of partial atomistic charges is vital for various chemical applications.
• Traditional methods for calculating charges can be computationally expensive.
• Machine learning (ML) presents a promising approach for efficient charge prediction.

Purpose of the Study:

• To develop and validate a machine learning model for predicting transferable, extensible, and conformationally dynamic atomic partial charges.
• To demonstrate the accuracy and efficiency of ML-based charge prediction compared to traditional methods.

Main Methods:

• Utilized the hierarchical interacting particle neural network (HIP-NN) ML model.
• Trained and tested the model on reference density functional theory (DFT) simulations.
• Evaluated performance across various charge partitioning schemes (Hirshfeld, CM5, MSK, NBO) and molecular sizes, including large proteins.

Main Results:

• HIP-NN achieved high accuracy in predicting atomic partial charges across diverse molecules and partitioning schemes.
• Demonstrated remarkable transferability and size extensibility, with errors as low as 0.004e- on the COMP6 benchmark (including proteins larger than training set).
• ML charge predictions are orders of magnitude faster than DFT calculations and accurately reproduced IR spectra from dynamical trajectories.

Conclusions:

• ML, specifically HIP-NN, offers a powerful and efficient pathway for accurate atomic partial charge prediction.
• This approach significantly enhances the feasibility of large-scale simulations while preserving quantum-level accuracy.
• The method has practical applications in areas like IR spectra calculation and force field parametrization.