TorchFF: A high-performance GPU-accelerated differentiable force field library

Area of Science:

• Computational chemistry
• Molecular modeling
• Scientific computing

Background:

• Molecular dynamics (MD) simulations and force field development increasingly use automatic differentiation.
• Standard frameworks face performance bottlenecks in MD simulations due to small-kernel operations and memory limitations.

Purpose of the Study:

• To introduce TorchFF, a high-performance, differentiable library for molecular modeling.
• To address performance bottlenecks in PyTorch for MD simulations.

Main Methods:

• Developed TorchFF, extending PyTorch with customized CUDA operators.
• Implemented performance-critical routines (bonded interactions, electrostatics, Ewald summation, neighbor lists) as backend-optimized primitives.

Main Results:

• TorchFF bridges the gap between Python's flexibility and compiled MD engine speeds.
• Achieved high performance for molecular modeling tasks within the PyTorch ecosystem.

Conclusions:

• TorchFF offers a solution for efficient and differentiable molecular dynamics simulations.
• Enables faster property prediction and parameter optimization in molecular modeling.