aiqm/torchani

TorchANI 2.0 is an open-source library that supports training, development, and research of ANI-style neural network interatomic potentials. It was originally developed and is currently maintained by the Roitberg group.

TorchANI 2.0 – Neural‑Network Interatomic Potentials in PyTorch

TorchANI is an open‑source Python library built on top of PyTorch that lets researchers train, evaluate, and use ANI‑style neural network interatomic potentials (NN‑IPs). These are machine‑learning models that predict the potential energy and forces of a molecule directly from its atomic coordinates, enabling fast, accurate molecular dynamics (MD) and hybrid ML/MM simulations.

What the library does

  • Model library – provides ready‑to‑use pretrained ANI models and a clean API to define new ones.
  • Training utilities – tools for fitting NN‑IPs to quantum‑chemical reference data.
  • Fast inference – C++/CUDA extensions compute the ANI descriptors and run the networks on GPUs for near‑real‑time MD.
  • Command‑line interface – the ani executable offers quick utilities (e.g., energy/force evaluation, building extensions).
  • Integration points – a separate TorchANI‑Amber package lets you plug the potentials into Amber’s classical MD engine for ML/MM simulations.

Who might use it

  • Computational chemists or materials scientists who need high‑performance, ML‑based force fields for simulations.
  • Researchers developing new NN‑IP architectures and wanting a PyTorch‑native baseline.
  • Anyone looking to run ML‑driven molecular dynamics on GPUs without writing low‑level CUDA code.

Getting started (from the README)

  1. Set up a Python environment (conda/mamba or a virtualenv). The library is distributed on PyPI, but the authors recommend installing with pip even inside a conda env because the conda package is unmaintained.
  2. Install a compatible PyTorch (e.g., pip install torch==2.13 --index-url https://download.pytorch.org/whl/cu130 for CUDA 13.0). TorchANI 2 requires PyTorch ≥ 2.0.
  3. Install TorchANI: pip install torchani.
  4. Build the optional C++/CUDA extensions for descriptor computation and fast inference:
    ani build-extensions          # builds for all detected GPU SMs
# or limit to specific SMs, e.g.:
ani build-extensions --sm 8.0 --sm 8.9
  5. (Optional) Clone the repo and create the development environment via the provided environment.yaml if you need to build docs, run tests, or modify the source.

Platform notes

  • GPU support: Works on NVIDIA GPUs with CUDA; strongly recommended for any non‑trivial workload. AMD/ROCm is not tested.
  • macOS: No CUDA; the library is untested with Apple’s MPS backend.
  • Conda package: Available for developers; can be built locally using the recipe directory.

Documentation & help

  • Full user guide and API reference: https://aiqm.github.io/torchani/
  • Migration guide for moving from TorchANI 1.x to 2.0.
  • Bug reports and feature requests are welcomed via GitHub issues.

Citing TorchANI

If you use the library in published work, cite the two papers listed in the README (the original TorchANI paper and the TorchANI 2.0 extension paper).

TL;DR: TorchANI 2.0 is a PyTorch‑based toolkit for building and running neural‑network potentials for molecular simulations, offering GPU‑accelerated inference, training utilities, and integration with classical MD packages like Amber. It is a genuine AI/ML software project aimed at the computational chemistry community.

Related

  • Project
    torchmd/torchmd-net

    A PyTorch-based library providing state-of-the-art neural network potentials for molecular simulations and the tools to train them.

  • Project
    pytorch/rl

    TorchRL is a PyTorch‑native, modular reinforcement‑learning library that provides a unified TensorDict data model and interchangeable components (environments, collectors, replay buffers, loss modules, trainers, etc.) for building, scaling, and experimenting with RL, multi‑agent, model‑based, and LLM post‑training pipelines.

  • Project
    meta-pytorch/torchtune

    torchtune is a PyTorch‑based library that provides ready‑to‑run recipes and configs for fine‑tuning, RL‑HF, distillation, quantization‑aware training, and inference of many modern LLMs (Llama, Gemma, Qwen, Phi, etc.). It abstracts distributed training, memory‑saving tricks, and device support behind a simple CLI (`tune`) and YAML files, integrating with Hugging Face, torch‑ao, FSDP2, and logging tools. Though development stopped in 2025, it remains a practical, open‑source toolkit for LLM experimentation.

  • Project
    torchgeo/torchgeo

    TorchGeo is a PyTorch library that makes working with satellite and aerial imagery easy. It offers ready‑made geospatial datasets, smart samplers for patch extraction, multi‑spectral pre‑trained models, and tight integration with PyTorch Lightning, letting researchers train and evaluate remote‑sensing models with just a few lines of code.

  • Project
    AIFrontierLab/TorchUMM

    TorchUMM is a unified, config‑driven toolkit for running, evaluating, and fine‑tuning a wide range of multimodal models (text‑to‑image, image understanding, editing). It supplies adapters for 14 back‑bones, supports 10+ benchmarks, offers post‑training methods, and can run locally, on AMD HPC clusters, or via Modal cloud. The library standardises inference/evaluation, making fair, reproducible comparisons easy.