TorchANI: A Free and Open Source PyTorch-Based Deep Learning Implementation of the ANI Neural Network Potentials

Area of Science:

• Computational Chemistry
• Machine Learning
• Materials Science

Background:

• Accurate potential energy surfaces are crucial for molecular simulations.
• Existing tools like NeuroChem provide neural network potentials but can be complex.
• There is a need for a more accessible and flexible platform for developing and deploying these models.

Purpose of the Study:

• Introduce TorchANI, a PyTorch-based implementation of the ANAKIN-ME (ANI) deep learning models.
• Provide a lightweight, user-friendly, and modifiable tool for training and inference of molecular properties.
• Enable automatic computation of forces and Hessian matrices using PyTorch's autograd engine.

Main Methods:

• Utilized PyTorch operators for computing atomic environmental vectors and neural networks.
• Leveraged PyTorch's autograd engine for automatic differentiation to calculate forces and Hessian matrices.
• Implemented the ANAKIN-ME (ANI) deep learning architecture within the PyTorch framework.

Main Results:

• Developed TorchANI, a program emphasizing ease of use, readability, and cross-platform compatibility.
• Achieved automatic computation of analytical forces and Hessian matrices, facilitating force training.
• Demonstrated the capability for training and inference of potential energy surfaces and physical properties.

Conclusions:

• TorchANI offers a flexible and accessible alternative to NeuroChem for ANI model development.
• The PyTorch implementation simplifies the process of calculating molecular properties and performing force training.
• TorchANI is open-source, promoting further research and development in deep learning for chemistry.