Machine Learning in QM/MM Molecular Dynamics Simulations of Condensed-Phase Systems

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Quantum mechanics/molecular mechanics (QM/MM) molecular dynamics (MD) simulations are crucial for systems requiring electronic structure changes but are computationally expensive.
• Existing machine learning (ML) models struggle with long-range interactions in condensed-phase systems.

Purpose of the Study:

• To develop a computationally efficient workflow for QM/MM MD simulations of condensed-phase systems.
• To incorporate the molecular mechanics (MM) environment into a high-dimensional neural network potential (HDNNP).
• To improve the accuracy and efficiency of ML-based QM/MM MD simulations.

Main Methods:

• Incorporating the MM environment as an element type in a high-dimensional neural network potential (HDNNP) with electrostatic embedding.
• Developing a Δ-learning scheme where the ML model learns the difference between density functional theory (DFT) and density functional tight binding (DFTB).
• Validating the approach using MD simulations of retinoic acid in water and S-adenosylmethionine/cytosine interaction in water.

Main Results:

• The Δ-learning scheme achieves DFT accuracy with significantly fewer parameters than direct ML models.
• The approach successfully incorporates long-range interactions up to 1.4 nm.
• MD simulations of retinoic acid and S-adenosylmethionine/cytosine demonstrated the method's validity.

Conclusions:

• Δ-learning is a promising approach for accurate and efficient QM/MM MD simulations in condensed-phase systems.
• This method overcomes the computational cost limitations of traditional QM/MM MD.
• The workflow effectively handles long-range interactions crucial for condensed-phase simulations.