Abstract:Quantum mechanics/molecular mechanics (QM/MM) molecular dynamics (MD) simulations have been developed to simulate molecular systems, where an explicit description of changes in the electronic structure is necessary. However, QM/MM MD simulations are computationally expensive compared to fully classical simulations as all valence electrons are treated explicitly and a self-consistent field (SCF) procedure is required. Recently, approaches have been proposed to replace the QM description with machine learned (ML) models. However, condensed-phase systems pose a challenge for these approaches due to long-range interactions. Here, we establish a workflow, which incorporates the MM environment as an element type in a high-dimensional neural network potential (HDNNP). The fitted HDNNP describes the potential-energy surface of the QM particles with an electrostatic embedding scheme. Thus, the MM particles feel a force from the polarized QM particles. To achieve chemical accuracy, we find that even simple systems require models with a strong gradient regularization, a large number of data points, and a substantial number of parameters. To address this issue, we extend our approach to a delta-learning scheme, where the ML model learns the difference between a reference method (DFT) and a cheaper semi-empirical method (DFTB). We show that such a scheme reaches the accuracy of the DFT reference method, while requiring significantly less parameters. Furthermore, the delta-learning scheme is capable of correctly incorporating long-range interactions within a cutoff of 1.4 nm. It is validated by performing MD simulations of retinoic acid in water and the interaction between S-adenoslymethioniat with cytosine in water. The presented results indicate that delta-learning is a promising approach for (QM)ML/MM MD simulations of condensed-phase systems.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the high computational cost of quantum mechanics/molecular mechanics (QM/MM) molecular dynamics (MD) simulations in condensed - phase systems and the difficulty in handling long - range interactions. Specifically, QM/MM MD simulations are more expensive than fully classical simulations because all valence electrons need to be explicitly processed and a self - consistent field (SCF) process is required. In addition, although machine - learning (ML) models can replace the QM description to reduce costs, they face challenges in handling long - range interactions in condensed - phase systems. To solve these problems, the authors propose a new workflow that incorporates the MM environment as an element type into the high - dimensional neural - network potential (HDNNP). Through this method, they solve the following key problems: 1. **Computational cost**: By using a machine - learning model to replace the QM description, the amount of computation is reduced. 2. **Long - range interactions**: By introducing a charge - embedding scheme, MM particles can feel the force of polarized QM particles, thus correctly handling long - range interactions. 3. **Chemical accuracy**: To achieve chemical accuracy, the authors find that even for simple systems, strong - gradient regularization and training with a large number of data points and parameters are required for the model. 4. **Δ - learning scheme**: Through the Δ - learning scheme, the ML model learns the differences between the reference method (such as DFT) and the cheaper semi - empirical method (such as DFTB), thus achieving the accuracy of DFT while reducing the number of parameters. Finally, the authors verify the effectiveness of this method in simulating the behavior of retinoic acid in water and the interaction between S - adenosylmethionine and cytosine in water, indicating that Δ - learning is a promising method for (QM)ML/MM MD simulations in condensed - phase systems. ### Formula summary 1. **Electron energy under the Born - Oppenheimer approximation**: \[ \hat{H}_{\text{QM}} \psi_{\vec{R}}(\vec{r}) = E_{\text{QM}}(\vec{R}) \psi_{\vec{R}}(\vec{r}) \] where $\psi_{\vec{R}}(\vec{r})$ is the electron wave function, $\vec{r}$ is the electron coordinate, and $\vec{R}$ is the nuclear coordinate. 2. **Electronic Hamiltonian operator**: \[ \hat{H}_{\text{QM}} = -\frac{1}{2} \sum_{i = 1}^{N_{\text{el}}} \nabla_i^2+\sum_{i < j}^{N_{\text{el}}} \frac{1}{|\vec{r}_i - \vec{r}_j|}-\sum_{i = 1}^{N_{\text{el}}} \sum_{j = 1}^{N_{\text{QM}}} \frac{Z_j}{|\vec{r}_i - \vec{R}_j|}+\sum_{i < j}^{N_{\text{QM}}} \frac{Z_i Z_j}{|\vec{R}_i - \vec{R}_j|} \] 3. **Energy of the classical force field**: \[ E_{\text{MM}}(\vec{R}) = E_{\text{bond}}(\vec{R}) + E_{\text{angle}}(\vec{R}) + E_{\text{dihedral}}(\vec{R}) + E_{\text{el}}(\vec{R}) + E_{\text{vdW}}(\vec{R}) \] 4. **Total QM/MM energy (addition scheme)**: \[ E_{\text{QM/MM}}(\vec{R}) = E_{\text{QM}}(\vec{R}_{\text{QM}}) + E_{\text{el}}^{\text{QM - MM}}(\vec{R}) + E_{\text{vdW,SR}}^{\text{QM - MM}}(\vec{R}) \]

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

Graph Convolutional Neural Networks for (QM)ML/MM Molecular Dynamics Simulations

Learning the Quantum Centroid Force Correction in Molecular Systems: A Localized Approach

Machine learning for molecular simulation

Molecular Dynamics with On-the-Fly Machine Learning of Quantum-Mechanical Forces

Electrostatic Embedding Machine Learning for Ground and Excited State Molecular Dynamics of Solvated Molecules

Molecular Dynamics with Neural-Network Potentials

Towards Exact Molecular Dynamics Simulations with Machine-Learned Force Fields

Accurate Machine Learned Quantum-Mechanical Force Fields for Biomolecular Simulations

Machine Learning for Molecular Dynamics on Long Timescales

Solvation Free Energy Calculations with Quantum Mechanics/Molecular Mechanics and Machine Learning Models.

Accelerated Quantum Mechanics/Molecular Mechanics Simulations Via Neural Networks Incorporated with Mechanical Embedding Scheme.

Machine learning accelerates quantum mechanics predictions of molecular crystals

Machine learning model for non-equilibrium structures and energies of simple molecules

Machine Learning for Quantum Mechanical Properties of Atoms in Molecules

Assessment of Embedding Schemes in a Hybrid Machine Learning/Classical Potentials (ML/MM) Approach

Machine Learning in Molecular Dynamics Simulations of Biomolecular Systems

Machine learning force fields for molecular liquids: Ethylene Carbonate/Ethyl Methyl Carbonate binary solvent

Machine learning heralding a new development phase in molecular dynamics simulations

Breaking the size limitation of nonadiabatic molecular dynamics in condensed matter systems with local descriptor machine learning

Machine Learning for Parameter Auto-tuning in Molecular Dynamics Simulations: Efficient Dynamics of Ions near Polarizable Nanoparticles