TorchANI

TorchANI implements a PyTorch-based framework for training and performing inference with ANI (ANAKIN-ME) deep learning models to generate potential energy surfaces and predict physical properties of molecular systems using neural network potentials.

Key Features:

• PyTorch ANI implementation: Provides a PyTorch-based implementation for training and inference of ANI (ANAKIN-ME) deep learning models.
• Potential energy surfaces and physical properties: Generates potential energy surfaces and predicts other physical properties of molecular systems using neural network potentials.
• Atomic Environmental Vectors (AEVs): Implements PyTorch operators to compute atomic environmental vectors (AEVs) and perform operations within atomic neural networks.
• Autograd for forces and Hessians: Leverages PyTorch autograd to compute analytical forces and Hessian matrices automatically.
• Force training: Supports training on forces (force training) using the autograd-enabled computational graph.
• Relation to NeuroChem: Serves as an alternative implementation to the NeuroChem C++/CUDA implementation of ANAKIN-ME.

Scientific Applications:

• Molecular dynamics and PES mapping: Calculation of potential energy surfaces for molecular dynamics and conformational sampling studies.
• Force and Hessian analysis: Computation of analytical forces and Hessian matrices for vibrational analysis and stability assessments.
• Materials modeling: Modeling interatomic interactions and energy landscapes in materials science applications using neural network potentials.
• Drug discovery and molecular interactions: Modeling molecular interactions and energetic properties relevant to drug discovery and computational chemistry.

Methodology:

Implements ANI models in PyTorch, computes atomic environmental vectors (AEVs) with PyTorch operators and atomic neural networks, and uses PyTorch autograd to obtain analytical forces and Hessian matrices while supporting force-inclusive training and inference.