PhysicsNeMo

Type: Technology Tags: CUDA, NVIDIA, GPU, Physics AI, Scientific Computing, Geoscience, Neural Operators, Simulation, PyG, GNN Related: NVIDIA-Modulus, Earth-2, NIM-for-Earth-2-CorrDiff, NIM-for-Earth-2-FourCastNet, NIM-for-DoMINO-Automotive-Aero, PyTorch, PyG, NVIDIA-DGL, NVIDIA-Warp, cuDNN Sources: NVIDIA official documentation, developer.nvidia.com/physicsnemo, https://docs.nvidia.com/physicsnemo/latest/physicsnemo/examples/dgl_to_pyg_migration.html, https://docs.nvidia.com/nim/earth-2/corrdiff/latest/overview.html, https://docs.nvidia.com/nim/earth-2/fourcastnet/latest/overview.html, https://docs.nvidia.com/nim/physicsnemo/domino-automotive-aero/latest/overview.html Last Updated: 2026-04-30

Summary

PhysicsNeMo (formerly part of Modulus) is NVIDIA’s open-source framework specifically designed for large-scale physics AI model training for geoscience, subsurface modeling, and industrial physics applications. It extends the NVIDIA Modulus physics-ML ecosystem with specialized architectures, datasets, and training workflows for Earth science domains including seismic imaging, weather modeling, reservoir simulation, and ocean dynamics. PhysicsNeMo is the framework powering NVIDIA’s Earth-2 climate AI initiative’s large-scale model training.

Detail

Purpose

PhysicsNeMo addresses the need for scalable, production-grade training of very large physics-AI models (billions of parameters) for geoscience and climate applications, where the models must learn from massive observational and simulation datasets while respecting physical constraints. It provides domain-specific model architectures, data loaders, and multi-node distributed training infrastructure for Earth and physics sciences.

Key Features

  • Specialized physics-ML architectures: Spherical Fourier Neural Operator (SFNO), Earth-2 FourCastNet, GraphCast-compatible GNN
  • Distributed training across thousands of GPUs for continental/global-scale models
  • Support for structured (latitude-longitude grids) and unstructured (mesh) geophysical data
  • Data loaders for ERA5, CMIP6, and seismic datasets
  • Physics-constrained loss functions for atmospheric, oceanic, and subsurface domains
  • Built on PyTorch with Megatron-LM-style distribution for large model training
  • PyG / PyTorch Geometric backend support for current and new GNN-based models, with migration guidance from DGL graph objects
  • Integration with NVIDIA Earth-2 inference and visualization platform
  • Deployment adjacency with NIM-for-Earth-2-CorrDiff and NIM-for-Earth-2-FourCastNet for named Earth-2 inference microservices
  • Deployment adjacency with NIM-for-DoMINO-Automotive-Aero for automotive external-aerodynamics surrogate inference.
  • Multi-modal training support (combining observational + simulation data)
  • Checkpoint compatibility with Modulus training pipelines

Use Cases

  • Global weather and climate AI model training (NWP surrogate models)
  • Seismic full waveform inversion (FWI) neural surrogates
  • Subsurface reservoir simulation with neural operators
  • Ocean current and sea surface temperature forecasting
  • Climate downscaling (coarse-to-fine resolution enhancement)
  • Automotive external-aerodynamics surrogate simulation through NIM-for-DoMINO-Automotive-Aero
  • Carbon capture and storage (CCS) subsurface modeling
  • Tsunami and flood early warning systems

Hardware Requirements

  • NVIDIA A100 or H100 GPUs strongly recommended
  • Multi-node clusters required for global-scale model training (1000+ GPUs typical)
  • NVLink for intra-node; InfiniBand NDR for inter-node
  • CUDA 11.8 or higher
  • DGX SuperPOD or equivalent HPC cluster

Language Bindings

  • Python (primary)
  • YAML/Hydra for experiment configuration

Connections

  • NVIDIA-Modulus — PhysicsNeMo is the geoscience-focused evolution of the Modulus framework
  • Earth-2 — Earth-2 climate AI platform uses PhysicsNeMo for training its weather/climate models
  • NIM-for-Earth-2-CorrDiff - deployable CorrDiff NIM for high-resolution weather downscaling.
  • NIM-for-Earth-2-FourCastNet - deployable FourCastNet NIM for global medium-range forecasting.
  • NIM-for-DoMINO-Automotive-Aero - deployable PhysicsNeMo NIM for automotive aerodynamic surrogate prediction.
  • NVIDIA-Warp — Warp differentiable simulation can be used for adjoint-based data assimilation within PhysicsNeMo
  • PyTorch — PhysicsNeMo is built on PyTorch with distributed training extensions
  • PyG — PhysicsNeMo documentation recommends PyG for GNN backends and documents migration from DGL to PyG graph objects
  • NVIDIA-DGL — legacy DGL context for users migrating existing PhysicsNeMo GNN examples and models toward PyG
  • NCCL — multi-node collective communications for distributed physics-ML training

Resources