TorchMD

TorchMD integrates classical empirical potentials with machine learning models to enhance molecular simulations and improve the quality and transferability of potential energy models.

Key Features:

• Mixed Classical and Machine Learning Potentials: Combines traditional force fields with data-driven machine learning potentials to form hybrid potential energy models.
• PyTorch Integration: Implements all force computations—bond, angle, dihedral, Lennard-Jones, and Coulomb interactions—using PyTorch arrays and operations.
• Neural Network Potentials: Supports development and application of neural network–based potentials for learned force fields.
• End-to-end Training: Enables end-to-end training workflows from initial potential learning to final simulation execution.
• Ab initio Potential Learning: Facilitates learning of ab initio potentials from reference data.
• Coarse-grained Model Support: Supports simulation and learning of coarse-grained models.

Scientific Applications:

• Standard Amber Simulations: Validated using standard Amber all-atom simulations to reproduce established molecular dynamics behavior.
• Protein Folding Models: Learns and simulates coarse-grained models for protein folding relevant to structural biology studies.

Methodology:

Classical simulation components (bond, angle, dihedral, Lennard-Jones, Coulomb) are expressed and computed in PyTorch arrays and operations, supporting end-to-end training from potential learning—including ab initio potentials—to simulation execution and coarse-grained model simulation.