Abstract:Excited-state molecular dynamics simulations are crucial for understanding processes like photosynthesis, vision, and radiation damage. However, the computational complexity of quantum chemical calculations restricts their scope. Machine learning (ML) offers a solution by delivering high accuracy properties at lower computational costs. We present SpaiNN, an open-source Python software for ML-driven surface hopping nonadiabatic molecular dynamics simulations. SpaiNN combines the invariant and equivariant neural network architectures of SchNetPack with SHARC for surface hopping dynamics. Its modular design allows users to implement and adapt modules easily. We compare rotationally-invariant and equivariant representations in fitting potential energy surfaces of multiple electronic states and properties arising from the interaction of two electronic states. Simulations of the methylenimmonium cation and various alkenes demonstrate the superior performance of equivariant SpaiNN models, improving accuracy, generalization, and efficiency in both training and inference.

What problem does this paper attempt to address?

The paper primarily aims to address the following issues: ### Research Background and Objectives - **Importance of Non-Adiabatic Molecular Dynamics (NAMD)**: Non-adiabatic molecular dynamics simulations are crucial for understanding processes such as photosynthesis, vision, and radiation damage. These simulations help explore photoexcited reactions and can reveal the complex relationships between molecular structure and photochemical properties. - **Limitations of Quantum Chemistry Calculations**: Despite the importance of non-adiabatic molecular dynamics simulations, accurate quantum chemistry calculations are costly, limiting the scale of simulations. This is particularly evident in fully quantum or semiclassical dynamics simulations, which typically can only handle short time scales (a few picoseconds) and small systems (tens of atoms). ### Solution - **Application of Machine Learning**: To overcome the above limitations, researchers use machine learning to accelerate molecular dynamics simulations by predicting potential energy surfaces (PESs) and their related properties, thereby reducing the cost of electronic structure calculations. - **Development of SpaiNN Software**: This paper introduces a new open-source Python software called SpaiNN, which combines invariant and equivariant neural network architectures from SchNetPack with the SHARC surface hopping dynamics program to achieve machine learning-driven non-adiabatic molecular dynamics simulations. ### Specific Contributions - **Software Features**: - Supports rotationally equivariant representations, enhancing prediction accuracy, generalization ability, and efficiency. - Modular design allows users to easily implement and adjust modules. - Provides a direct comparison of two equivariant models (SchNet and PaiNN) to evaluate the advantages of equivariance in fitting multi-electronic state potential energy surfaces and their related properties. - **Performance Demonstration**: - Demonstrates the superior performance of the PaiNN model over the SchNet model in terms of energy, forces, and non-adiabatic couplings through simulations of methylene imine cation and various alkenes. - Shows that the PaiNN model not only improves prediction accuracy but also exhibits higher computational efficiency during training and inference. ### Conclusion In summary, the main objective of this paper is to develop a new machine learning tool—SpaiNN, to enhance the prediction capabilities of non-adiabatic molecular dynamics simulations for multi-electronic state potential energy surfaces and their related properties. By introducing equivariant neural network models, SpaiNN significantly improves prediction accuracy, generalization ability, and computational efficiency. Additionally, through specific case studies, SpaiNN's potential in simulating complex photochemical processes is demonstrated.

SpaiNN: Equivariant Message Passing for Excited-State Nonadiabatic Molecular Dynamics

PiNN: equivariant neural network suite for modelling electrochemical systems

Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation

Neural Message Passing for Quantum Chemistry

Analyzing Atomic Interactions in Molecules as Learned by Neural Networks

Heterogeneous Relational Message Passing Networks for Molecular Dynamics Simulations

Ewald-based Long-Range Message Passing for Molecular Graphs

Combining SchNet and SHARC: The SchNarc Machine Learning Approach for Excited-State Dynamics

Ab initio quantum chemistry with neural-network wavefunctions

Molecular Dynamics with Neural-Network Potentials

Interpolating Nonadiabatic Molecular Dynamics Hamiltonian with Bidirectional Long Short-Term Memory Networks

Enhancing Open Quantum Dynamics Simulations Using Neural Network-Based Non-Markovian Stochastic Schrödinger Equation Method

Machine Learning Nonadiabatic Dynamics: Eliminating Phase Freedom of Nonadiabatic Couplings with the State-Intraction State-Averaged Spin-Restricted Ensemble-Referenced Kohn-Sham Approach

NNP/MM: Accelerating Molecular Dynamics Simulations with Machine Learning Potentials and Molecular Mechanics

NNP/MM: Accelerating molecular dynamics simulations with machine learning potentials and molecular mechanic

Neural Network Potential with Multi-Resolution Approach Enables Accurate Prediction of Reaction Free Energies in Solution

A Simple and Efficient Equivariant Message Passing Neural Network Model for Non-Local Potential Energy Surface

Machine Learning Accelerates Precise Excited-State Potential Energy Surface Calculations on a Quantum Computer

Simplicial Message Passing for Chemical Property Prediction

Learning local equivariant representations for large-scale atomistic dynamics

Long-Short-Range Message-Passing: A Fragmentation-Based Framework to Capture Non-Local Atomistic Interactions