Key Features

• Modular Transformer Architecture: Built on building blocks optimized for physical systems.
• Hardware Agnostic: Seamless execution across CPU, MPS (Apple Silicon), and NVIDIA GPUs.
• Industrial Grade: Designed for high-fidelity industrial simulations and engineering verticals.
• Ready for Scale: Built-in support for Multi-GPU and SLURM cluster environments.

Table of contents

• Installation
  • Pre-requisites
  • Working with the source code
    • How to clean up and do a fresh installation
  • Working with pre-built packages
• Quickstart
• Recipes
• Performance Benchmarks
• Contributing
  • Guidelines
  • Third-party contributors
  • Configuring IDEs
• Supported systems
• Licensing
• Endorsements
• Citing

Installation

It is possible to use the framework either from source or from the pre-built packages.

Pre-requisites

• install uv as the package manager on your system
• clone the repo into your desired folder, if you wish to work with the code: git clone https://github.com/Emmi-AI/noether.git

Working with pre-built packages

Installable package is available via pip and can be installed as:

pip install emmiai-noether

or if you use uv:

uv pip add emmiai-noether

Working with the source code

If you prefer to work with the source code directly without installing a prebuilt package.

Create a fresh virtual environment and synchronize the core dependencies after cloning the repo:

uv venv && source .venv/bin/activate
uv sync

Note: Initial installation may take several minutes. Duration depends on your hardware and network speed.

Validate your installation by simply running the tests (if something fails with module import errors it means that the installation was incomplete):

pytest -q tests/

if the tests are passed (warnings are okay to be logged) then you're all set and ready to go!

Install via GitHub

You can install Noether directly from GitHub as well:

uv add "emmiai-noether @ git+https://github.com/Emmi-AI/noether.git"

This way you will stay always up-to-date with the recent updates and don't wait for the built packages being released.

How to clean up and do a fresh installation

You might be in a situation when your venv won't be configured as intended anymore, to fix this:

• Deactivate existing environment in your terminal by running: deactivate
• Remove existing .venv (optionally add uv.lock): rm -rf .venv uv.lock
• [Optional] Clean uv cache: uv cache clean
• Create a new venv and activate it: uv venv && source .venv/bin/activate
• [Optional] If deleted, generate a new uv.lock file: uv lock
• [Optional] If contributor: pre-commit install

Quickstart

Scaffold a New Project

You can get started with scaffolding a project directly with uvx:

uvx --from emmiai-noether noether-init my_project
cd my_project
uv run noether-train --hp my_project/configs/base_experiment.yaml

Or if you have emmiai-noether already installed:

uv run noether-init my_project
cd my_project
uv run noether-train --hp my_project/configs/base_experiment.yaml

See the scaffolding tutorial for all options and the generated project structure.

Recipes

End-to-end training and evaluation examples for specific engineering tasks.

Featured: AB-UPT Showcase on DrivAerML

Train AB-UPT on the DrivAerML aerodynamics benchmark with a preset-based CLI. Ready-to-use model sizes for CPU, GPU, and Apple Silicon.

• Walkthrough: AB-UPT Showcase docs
• Source: recipes/aero_cfd/showcase/
• Dataset: EmmiAI/DrivAerML_subsampled_10x

cd recipes/aero_cfd/
export PYTHONPATH=$(git -C ../.. rev-parse --show-toplevel)/recipes:$PYTHONPATH

python -m showcase.cli train \
  --dataset-root /path/to/drivaerml \
  --output-path /path/to/outputs \
  --model-size scaled \
  --accelerator gpu \
  --precision float16

All recipes

Browse the full list in the Noether Recipe Zoo.

Performance Benchmarks

Training time for the AB-UPT model on the ShapeNet-Car dataset across hardware configurations.

[!NOTE] All benchmarks were conducted using FP32 precision to establish a baseline for raw computational performance.

Contributing

Guidelines

We follow these standards:

• Use typed coding in Python.
• Write documentation to new features and modules:
  • In case of larger modules make sure to update the documentation that is not autogenerated under the docs/.
  • For smaller features writing a clear API documentation is enough and required.
• Before committing your changes:
  • Run tests via pytest -q tests/.
  • Ensure that pre-commit hooks are not disabled and are runnable at every commit. We are using ruff as a linter and formatter as well as mypy for type checking. Their configuration is defined in the project's root pyproject.toml.
• Creating pull requests (PRs) is a mandatory step for any incoming changes that will end up on the main branch.
  • For a PR to be merged at least one core maintainer must give their approval.
  • All test must be green

Pre-commit Hooks

To install pre-commit execute:

pre-commit install

To run the pre-commit configuration on all files, you can use:

pre-commit run --all-files

To run the pre-commit configuration on specific files use:

pre-commit run --files /your/file/path1.py /your/file/path2.py

Third-party contributors

In case of bugs use a corresponding template to create an issue.

In case of feature requests you can submit a PR with clear description of the proposed feature. In that case it must follow the guidelines, or file a feature request as an issue. In that case, we will consider adding it to our backlog.

Configuring IDEs

Pycharm

• Mark src/ directory as Sources Root (right mouse button click on the folder -> Mark Directory as)
• Settings -> Editor -> Code Style -> Python -> Tabs and Indents -> change Continuation indent from 8 to 4.
• Settings -> Editor -> Code Style -> Python -> Spaces -> Around Operators -> Power operator (**)

Supported systems

Worth noting that we work with macOS and Linux environments thus in case of any issues on Windows, at this time, you have to find workarounds yourself.

Licensing

[!NOTE] TL;DR: Research & development ✅| Production deployment ❌ (without commercial license)

The Noether Framework is licensed under a Non-Production License (based on Mistral AI's MNPL). This means you're free to use, modify, and research with the framework, but commercial/production use requires a separate commercial license from Emmi AI.

We're committed to open AI innovation while sustainably growing our business. For commercial licensing, contact us at partner@emmi.ai .

Read the full license here.

Endorsed by research groups from

Citing

If you use Noether in your research or industrial applications, please cite this repository. A formal BibTeX entry for our forthcoming ArXiv publication will be provided here shortly.

@misc{noether2026,
  author = { Bleeker, Maurits AND Hennerbichler, Markus AND Kuksa, Pavel },
  title = {Noether: A PyTorch-based Framework for Engineering AI},
  year = {2026},
  publisher = {GitHub},
  note = {Equal contribution},
  url = {https://github.com/Emmi-AI/noether}
}