NVIDIA PhysicsNeMo, Formerly Modulus
Find your way from older Modulus tutorials to the current PhysicsNeMo framework, then choose a reproducible first experiment.
Last reviewed: September 29, 2026github.com ↗What happened to NVIDIA Modulus?
NVIDIA Modulus is now NVIDIA PhysicsNeMo. The former Modulus repository address redirects to NVIDIA/physicsnemo. Start with the current repository and documentation when an older tutorial mentions Modulus.
PhysicsNeMo is an open-source PyTorch framework for physics machine learning. Its models and examples support work such as learning surrogate models from simulation data and combining physical constraints with training.
The name change is not a guarantee of code compatibility. Match the tutorial, package version and example configuration. Check the current documentation and release notes before translating old imports or reusing a saved model.
Install the version your example needs
NVIDIA documents package and container installation routes. Use the official installation instructions to choose the Python, PyTorch, CUDA and optional dependencies for your environment.
- Choose an example first. Read its hardware, data and dependency requirements before installing anything.
- Use an isolated environment. Record the package versions or container tag. For older Modulus work, keep the original environment separate while you migrate.
- Check the compute requirements. A small example and a large aerodynamic model can have very different memory needs. Ask your university or lab about GPU access before planning a large training run.
- Run the example unchanged. Confirm that it loads the data and completes its documented checks before modifying the physics or model.
A common source of confusion is mixing an old tutorial with a current installation. Use documentation for the installed release; changing an import name alone may not fix changed APIs or configuration files.
A first physics-ML experiment
Start with a small problem from NVIDIA's example collection. Darcy flow or a documented fluid example can make it easier to inspect inputs, predictions and errors before attempting an aircraft model.
- Define the prediction. State the input field, the output field and the physical quantity you want to predict.
- Keep evaluation data separate. Describe what the model has not seen during training.
- Reproduce a baseline. Save the configuration, code version, data source and random seed.
- Measure error and cost. Plot reference and predicted fields, an error map, and training/inference time on named hardware.
- Change one condition. Test a new parameter or boundary condition and report where accuracy deteriorates.
Deliverable: a short reproducibility report with plots and an explanation of where the surrogate is useful. Report training cost as well as inference speed; a fast prediction alone does not establish an overall speedup.
Connect it to aerospace work
NVIDIA's repository includes examples for aerodynamics, weather and structural problems. Treat these as starting points for learning and evaluation. A learned model needs validation for the geometry, operating conditions and intended use you care about.
For an aerospace portfolio, make the engineering question explicit: which design decision could the model help with, what reference result did you compare against, and where would you still need a conventional solver or experiment?
- Explore the aerospace engineering pathway.
- Compare the learning workflow in our DeepXDE guide.
- Find a project at your current level and connect your evidence to an internship application.