SpaiNN: equivariant message passing for excited-state nonadiabatic molecular dynamics

Area of Science:

• Computational chemistry
• Machine learning
• Chemical dynamics

Background:

• Excited-state molecular dynamics (ESMD) simulations are vital for studying photochemical processes but are computationally intensive.
• Quantum chemical calculations limit the scope and speed of traditional ESMD.
• Machine learning (ML) presents a promising approach to reduce computational costs while maintaining high accuracy.

Purpose of the Study:

• To introduce SpaiNN, an open-source Python software package for ML-driven surface hopping nonadiabatic molecular dynamics (NAMD).
• To integrate advanced ML architectures with established NAMD codes.
• To evaluate the performance of different ML representations for fitting potential energy surfaces and properties relevant to NAMD.

Main Methods:

• Developed SpaiNN by combining SchNetPack's invariant and equivariant neural networks with SHARC for surface hopping dynamics.
• Implemented a modular design for easy adaptation and extension.
• Compared rotationally-invariant and equivariant ML representations for fitting potential energy surfaces and inter-state properties.
• Performed NAMD simulations using the developed models for the methyleneimmonium cation and various alkenes.

Main Results:

• Equivariant SpaiNN models demonstrated superior performance compared to invariant models.
• Equivariant models showed improved accuracy and generalization in fitting potential energy surfaces and properties.
• Significant improvements in both training and inference efficiency were observed with equivariant SpaiNN models.
• Successful application to NAMD simulations of relevant chemical systems.

Conclusions:

• SpaiNN provides an efficient and accurate platform for ML-driven NAMD simulations.
• Equivariant neural network representations offer significant advantages for modeling excited-state dynamics.
• The software facilitates the study of complex photochemical processes previously limited by computational cost.