Toward Improving Multiple Time Step QM/MM Simulations with Δ-Machine Learning.

Area of Science:

• Computational Chemistry
• Method Development
• Machine Learning in Chemistry

Background:

• Semiempirical quantum mechanics/molecular mechanics (QM/MM) methods offer speed for condensed-phase simulations but often lack accuracy.
• Multiple time step (MTS) methods enhance QM/MM accuracy by using higher-level calculations periodically.
• Standard semiempirical methods show limitations in MTS due to insufficient similarity with high-level methods.

Purpose of the Study:

• To investigate the limitations of standard semiempirical methods in MTS QM/MM simulations.
• To explore the application of delta-machine learning (Δ-ML) for enhancing MTS QM/MM efficiency.
• To provide guidance for Δ-ML based MTS simulations in chemical research.

Main Methods:

• Assessed limitations of AM1 semiempirical method for MTS QM/MM in a condensed-phase proton transfer reaction.
• Trained neural network potentials and Δ-learning corrections for the AM1 method in gas-phase reactions.
• Performed gas-phase MTS simulations using Δ-ML corrected methods and compared results to high-level DFT (B3LYP).

Main Results:

• Standard semiempirical methods severely limit MTS outer time steps (e.g., to 4 for AM1).
• Δ-learning corrections significantly outperform ML potentials and improve accuracy and transferability with sufficient training data.
• Δ-ML enabled MTS simulations achieved near-exact results at an outer integration frequency of 25 and acceptable error at 30.

Conclusions:

• Δ-machine learning is a promising approach to enhance the efficiency and accuracy of MTS QM/MM simulations.
• The accuracy and transferability of Δ-corrections are highly dependent on the amount and quality of training data.
• This work validates Δ-learning for MTS and paves the way for its application in complex condensed-phase chemical reactions.