NVIDIA/physicsnemo

Open-source deep-learning framework for building, training, and fine-tuning deep learning models using state-of-the-art Physics-ML methods

What it solves

NVIDIA PhysicsNeMo provides a standardized, PyTorch-native framework for building and training AI models that simulate physical systems. It addresses the complexity of scientific machine learning by offering reusable components for tasks like fluid dynamics, weather forecasting, structural mechanics, and generative physics, allowing researchers to move from prototypes to scalable production workflows.

How it works

The framework is organized into three layers: a foundation layer for core utilities (mesh handling, distributed training), a building-block layer for specialized AI models (such as FNOs, MeshGraphNets, and diffusion models), and an application layer containing end-to-end training recipes. It supports advanced features like domain parallelism, which allows users to split high-resolution physical samples across multiple GPUs, and it is designed to be composable, meaning you can integrate specific layers or operators into existing PyTorch projects.

Who it’s for

It is designed for scientists, engineers, and AI researchers working in fields like computational fluid dynamics (CFD), climate modeling, geophysics, and manufacturing who need to build high-performance surrogate models for complex physical simulations.

Highlights

  • Scientific Data Support: Native handling for diverse data structures including regular grids, unstructured meshes, graphs, and point clouds.
  • Scalable Training: Built-in support for distributed training and domain parallelism (via ShardTensor) to handle large-scale scientific datasets.
  • Extensive Model Library: Includes implementations of state-of-the-art architectures like FNO, MeshGraphNet, Transolver, and various diffusion-based generative models.
  • End-to-End Recipes: Provides ready-to-use workflows for real-world applications such as aerodynamic design, regional weather forecasting, and structural crash dynamics.

Related

  • Project
    NVIDIA-NeMo/Automodel

    A framework for automating the fine-tuning and pretraining of large-scale AI models with PyTorch-native parallelism and optimized recipes.

  • Dispatch
    Mapping the PyTorch Ecosystem: A Deep Dive into the PyTorch Landscape

    An exploration of the PyTorch Landscape, detailing the vast array of libraries and tools spanning modeling, training, and optimization that define the modern AI ecosystem.

  • Project
    mit-han-lab/torchquantum

    A PyTorch-based framework for simulating quantum computations on classical hardware, enabling the training of parameterized quantum circuits and hybrid classical-quantum models.

  • Project
    fangwei123456/spikingjelly

    SpikingJelly is a PyTorch‑native library for building, training, and deploying spiking neural networks. It provides neuron modules, surrogate‑gradient tools, dataset loaders for neuromorphic data, automatic backend selection (Torch/Triton/native CUDA), mixed‑precision support, and export interfaces to neuromorphic hardware. Installable via PyPI (Python ≥ 3.11, PyTorch ≥ 2.6) with optional CUDA extensions for high performance. Licensed under Apache‑2.0.

  • Project
    pytorch/pytorch

    PyTorch is an open‑source Python library that offers GPU‑accelerated tensor operations and a dynamic, tape‑based autograd system for building and training deep neural networks. It integrates tightly with Python, supports extensions in C/C++ and other languages, and provides fast, low‑overhead execution on CPUs, NVIDIA GPUs, AMD ROCm, and Intel GPUs.