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

PhysicsNeMo is a deep-learning framework designed to accelerate the development of AI models for science and engineering (AI4Science). It solves the challenge of combining physics knowledge with data to create high-fidelity, real-time predictions for complex physical systems, reducing the time required to develop simulations in domains like computational fluid dynamics (CFD), structural mechanics, and electromagnetics.

How it works

Built on top of PyTorch, the framework provides a modular stack of components for building, training, and deploying physics-informed machine learning models. It includes:

  • Model Zoo: A collection of optimized architectures including Neural Operators (FNO, DeepONet), Graph Neural Networks (GNNs), Diffusion models, and Physics-Informed Neural Networks (PINNs).
  • Data Pipelines: Scalable datapipes specifically tuned for scientific data structures such as meshes and point clouds.
  • Distributed Computing: A sub-module based on torch.distributed for scaling training from single GPUs to multi-node clusters.
  • Symbolic PDE Utilities: Tools to define equations via SymPy and compute physics-informed losses using automatic spatial derivatives.

Who it’s for

It is intended for SciML researchers, climate scientists, and engineering domain experts who need to train large-scale physics-ML models with optimal performance on NVIDIA GPUs.

Highlights

  • Seamless PyTorch Integration: Works as a complementary tool to PyTorch with a familiar Python interface.
  • GPU Optimization: Highly optimized for NVIDIA GPUs to maximize training speed and scalability.
  • Extensible Design: Supports ONNX for flexible deployment and provides APIs for defining custom physics models and constraints.
  • Domain-Specific Packages: Includes specialized sub-modules for CFD and weather/climate research (Earth-2 Studio).