By NVIDIALearning Objectives
- Understand what NVIDIA PhysicsNeMo is and what a physics-machine-learning framework provides
- Evaluate how AI surrogate models accelerate engineering simulation on GPUs
- Assess who builds with PhysicsNeMo and where it fits versus packaged tools
What Is NVIDIA PhysicsNeMo?
NVIDIA PhysicsNeMo is an open-source deep-learning framework for building AI models that approximate physics simulations. Where a packaged tool gives an engineer a finished application, PhysicsNeMo is a toolkit for developers and researchers to train their own surrogate models — machine-learning models that predict the results of fluid-dynamics, structural, or thermal simulations far faster than solving the equations directly.
The framework was previously known as Modulus (and before that, SimNet); NVIDIA renamed it PhysicsNeMo in 2025. It is released under a permissive open-source license, so anyone can use and extend it, and it is built to run efficiently on NVIDIA graphics processors.
💡Key Concept
Physics-informed machine learning: Training neural networks to respect the laws of physics — for example, by penalizing predictions that violate the governing equations. The result is a model that generalizes from limited data and predicts physical behavior quickly, bridging traditional simulation and modern deep learning.
How AI Accelerates Simulation
PhysicsNeMo provides the building blocks for several families of AI simulation models: physics-informed neural networks that embed governing equations directly, neural operators such as Fourier neural operators that learn mappings between whole input and output fields, and graph neural networks like MeshGraphNet that operate on simulation meshes. With these, a team can train a model on prior simulations and then predict new cases in a fraction of the time a solver would take.
That speed unlocks workflows that are impractical with conventional solving — real-time digital twins, rapid design sweeps, and interactive what-if analysis. NVIDIA positions PhysicsNeMo as the AI engine behind its broader push into computer-aided-engineering acceleration and Omniverse digital twins, where trained surrogates feed live, visual models of physical systems.
Open Source and the Wider Ecosystem
Because PhysicsNeMo is open source, it is used both directly by research and engineering teams and as a foundation that commercial simulation vendors build on. NVIDIA has shown large acceleration figures for computer-aided engineering when surrogates run on its latest hardware, and partners across the simulation industry increasingly connect their tools to this kind of AI acceleration. For engineers, it represents the open, developer-facing layer of the AI-simulation shift.
Who Uses NVIDIA PhysicsNeMo?
PhysicsNeMo is aimed at developers, researchers, and engineering teams with the skills to train models — computational scientists, simulation specialists, and AI engineers in automotive, aerospace, energy, electronics, and climate. It suits organizations that want to build custom AI surrogates rather than buy a packaged application, and that have access to GPU compute and simulation data.
Company Details
| Detail | Info |
|---|---|
| Product | NVIDIA PhysicsNeMo — open-source physics machine-learning framework |
| Developer | NVIDIA (founded 1993, Santa Clara, California) |
| History | Formerly Modulus (and SimNet); renamed PhysicsNeMo in 2025 |
| Model types | Physics-informed networks, neural operators, and graph neural networks |
| License | Open source (permissive), GPU-accelerated |
| Use cases | Surrogate models, design sweeps, and real-time digital twins |
| Target users | Developers, researchers, and simulation teams with GPU compute |
| Website | developer.nvidia.com/physicsnemo |
Strengths
- Open source — free to use and extend, with an active developer community
- Broad model toolkit — physics-informed networks, neural operators, and graph networks in one framework
- GPU-accelerated — built to run efficiently on NVIDIA hardware for large speed-ups
- Enables real-time twins — fast surrogates unlock interactive and live simulation workflows
- Foundation for others — used directly and as a base that commercial tools build on
Limitations and Considerations
- For builders, not end users — it is a framework requiring machine-learning and coding skill, not a packaged app
- Compute and data needed — training accurate surrogates takes GPU resources and simulation data
- Approximation by design — surrogates accelerate exploration; exact solvers still verify final results
- Hardware-aligned — best performance assumes NVIDIA GPUs
Pricing
NVIDIA PhysicsNeMo is open-source software available at no cost under a permissive license. The practical cost lies in the GPU compute used to train and run models, whether on local hardware or in the cloud. Enterprise support and related NVIDIA platforms are available separately.
Key Takeaways
- NVIDIA PhysicsNeMo is an open-source framework for building AI surrogate models of engineering simulations
- Formerly called Modulus, it supports physics-informed neural networks, neural operators, and graph neural networks
- Trained surrogates predict fluid, structural, and thermal behavior far faster than solvers, enabling design sweeps and real-time digital twins
- It is the developer-facing, open layer of the AI-simulation shift, used directly and as a foundation for commercial tools
