NVIDIA

physicsnemo: Build and Train Physics AI Models

NVIDIA PhysicsNeMo is a PyTorch framework for building and training machine-learning models for physics and engineering. It combines reusable model components with end-to-end recipes for scientific data and workloads.

View on GitHub Compare
Stars
3.3k
Forks
803
Stars, last 30 days
+86

Our take

Health score

68/100

How it is scored
Activity
72
Community
87
Issues
54
Pull requests
58

+86 stars in the last 30 days

  • Merges more pull requests than 81% of the projects we track
  • More contributors than 77% of the projects we track

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

3,447

Forks

845

Change since Nov 16, 2026: +3 stars·+1 forks

Overview

PhysicsNeMo is a PyTorch framework for scientific machine learning, physics AI, and AI for science and engineering. It provides reusable models and supporting components alongside runnable training recipes, helping researchers and engineers build learned approximations for physical systems without assembling every workflow from scratch.

It is most relevant when data has scientific structure such as grids, meshes, point clouds, or graphs, and the goal is to train or fine-tune models for simulation, forecasting, inverse problems, or design. It is a framework and recipe collection, not a general-purpose turnkey simulator.

Key Features

  • Model families for regular grids, meshes, point clouds, and graph-based representations.
  • Components for data loading and preprocessing, metrics, optimization, and model lifecycle management.
  • Recipes covering areas such as aerodynamics, weather, structural mechanics, geophysics, healthcare, and manufacturing.
  • Diffusion tools for generative and inverse-physics workflows.
  • Distributed training support, including DDP, FSDP2, and ShardTensor domain parallelism.
  • Mesh and neural-network utilities for geometry, numerical operations, and differentiable workflows.
  • Deployment helpers for ONNX export and runtime.

Use Cases

  • Engineering and CFD teams can train surrogate models to predict flow or surface fields when repeated simulation is costly and suitable simulation data is available.
  • Weather and climate researchers can adapt forecasting recipes to regional or global field prediction tasks.
  • Scientific ML researchers can compare model families on structured grids, unstructured meshes, graphs, or point clouds.
  • Geophysics and inverse-problem practitioners can explore generative approaches such as diffusion-based full-waveform inversion.
  • Teams working with large scientific samples can consider domain parallelism when a single high-resolution sample must be split across GPUs.

Getting Started

Install the CUDA 13 package with:

pip install "nvidia-physicsnemo[cu13]"

See the README for other installation options, supported dependencies, examples, and setup guidance.

Considerations

  • The framework targets scientific and engineering ML workflows, so users should expect to bring domain data and adapt recipes to their task.
  • GPU support is central to the project, and CUDA-specific package options are available. The README notes that the CUDA 12 backend is Linux-only.
  • APIs under physicsnemo.experimental may change between releases; consult the documentation and changelog before relying on them in long-lived projects.
  • The project provides model and training infrastructure, not a substitute for validating predictions against appropriate physical simulations or measurements.

Found this useful?

Share it with someone who would like physicsnemo.