SchNetPack 2.0: A neural network toolbox for atomistic machine learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Atomistic machine learning (ML) is crucial for simulating molecular behavior.
• Existing toolboxes may lack flexibility for method development and complex applications.
• Efficient data handling and advanced neural network architectures are needed.

Purpose of the Study:

• To introduce SchNetPack version 2.0, a versatile neural network toolbox.
• To enhance capabilities for both atomistic ML method development and application.
• To facilitate complex training tasks, including 3D molecular structure generation.

Main Methods:

• Development of an improved data pipeline for efficient processing.
• Integration of modules for equivariant neural networks.
• Implementation of molecular dynamics simulations using PyTorch.
• Optional integration with PyTorch Lightning and Hydra for flexible configuration.

Main Results:

• SchNetPack 2.0 provides a flexible command-line interface.
• The toolbox is easily extendable with custom code.
• Enables complex training tasks, such as generating 3D molecular structures.
• Supports advanced features like equivariant networks and molecular dynamics.

Conclusions:

• SchNetPack 2.0 is a powerful and flexible tool for atomistic machine learning.
• It addresses key requirements for both research and application in computational chemistry and materials science.
• Facilitates advanced simulations and the generation of molecular structures.