TORAX is a differentiable tokamak core transport simulator aimed for fast and accurate forward modelling, pulse-design, trajectory optimization, and controller design workflows. TORAX is written in Python using JAX, with the following motivations:
- Open-source and extensible, aiding with flexible workflow coupling
- JAX provides auto-differentiation capabilities and code compilation for fast runtimes. Differentiability allows for gradient-based nonlinear PDE solvers for fast and accurate modelling, and for sensitivity analysis of simulation results to arbitrary parameter inputs, enabling applications such as trajectory optimization and data-driven parameter identification for semi-empirical models. Auto-differentiability allows for these applications to be easily extended with the addition of new physics models, or new parameter inputs, by avoiding the need to hand-derive Jacobians
- Python-JAX is a natural framework for the coupling of ML-surrogates of physics models
For more comprehensive documentation, see our readthedocs page.
TORAX now has the following physics feature set:
- Coupled PDEs of ion and electron heat transport, electron particle transport, and current diffusion
- Finite-volume-method
- Multiple solver options: linear with Pereverzev-Corrigan terms and the predictor-corrector method, nonlinear with Newton-Raphson, nonlinear with optimization using the jaxopt library
- Ohmic power, ion-electron heat exchange, fusion power, Bremsstrahlung, and bootstrap current with the analytical Sauter model
- Time dependent boundary conditions, sources, geometry.
- Coupling to the QLKNN10D [van de Plassche et al, Phys. Plasmas 2020] QuaLiKiz-neural-network surrogate for physics-based turbulent transport
- General geometry, provided via CHEASE or FBT equilibrium files
- For testing and demonstration purposes, a single CHEASE equilibrium file is available in the data/geo directory. It corresponds to an ITER hybrid scenario equilibrium based on simulations in [Citrin et al, Nucl. Fusion 2010], and was obtained from PINT. A PINT license file is available in data/geo.
- Time dependent geometry is supported by provided a time series of geometry files
Additional heating and current drive sources can be provided by prescribed formulas, user-provided analytical models, or user-provided prescribed data.
Model implementation was verified through direct comparison of simulation outputs to the RAPTOR [Felici et al, Plasma Phys. Control. Fusion 2012] tokamak transport simulator.
This is not an officially supported Google product.
Short term development plans include:
- Extension of and more flexible data structures for prescribed input data
- Implementation of forward sensitivity calculations w.r.t. control inputs and parameters
- More extensive documentation and tutorials
- Visualisation improvements
Longer term desired features include:
- Sawtooth model (Porcelli + reconnection)
- Neoclassical tearing modes (modified Rutherford equation)
- Radiation sinks
- Cyclotron radiation
- Line radiation
- Neoclassical transport + multi-ion transport, with a focus on heavy impurities
- IMAS coupling
- Stationary-state solver
- Momentum transport
Contributions in line with the roadmap are welcome. In particular, TORAX is envisaged as a natural framework for coupling of various ML-surrogates of physics models. These could include surrogates for turbulent transport, neoclassical transport, heat and particle sources, line radiation, pedestal physics, and core-edge integration, MHD, among others.
Install Python 3.10 or greater.
Make sure that tkinter is installed:
sudo apt-get install python3-tk
Install virtualenv (if not already installed):
pip install --upgrade pip
pip install virtualenv
Create a code directory where you will install the virtual env and other TORAX dependencies.
mkdir /path/to/torax_dir && cd "$_"
Where /path/to/torax_dir
should be replaced by a path of your choice.
Create a TORAX virtual env:
python3 -m venv toraxvenv
Activate the virtual env:
source toraxvenv/bin/activate
Download QLKNN dependencies:
git clone https://gitlab.com/qualikiz-group/qlknn-hyper.git
export TORAX_QLKNN_MODEL_PATH="$PWD"/qlknn-hyper
It is recommended to automate the environment variable export. For example, if using bash, run:
echo export TORAX_QLKNN_MODEL_PATH="$PWD"/qlknn-hyper >> ~/.bashrc
The above command only needs to be run once on a given system.
Download and install the TORAX codebase via http:
git clone https://github.com/google-deepmind/torax.git
or ssh (ensure that you have the appropriate SSH key uploaded to github).
git clone [email protected]:google-deepmind/torax.git
Enter the TORAX directory and pip install the dependencies.
cd torax; pip install -e .
If you want to install with the dev dependencies (useful for running pytest
and installing pyink
for lint checking), then run with the [dev]
:
cd torax; pip install -e .[dev]
Optional: Install additional GPU support for JAX if your machine has a GPU: https://jax.readthedocs.io/en/latest/installation.html#supported-platforms
The following command will run TORAX using the default configuration file
examples/basic_config.py
.
python3 run_simulation_main.py \
--config='torax.examples.basic_config' --log_progress
To run more involved, ITER-inspired simulations, run:
python3 run_simulation_main.py \
--config='torax.examples.iterhybrid_rampup' --log_progress
and
python3 run_simulation_main.py \
--config='torax.examples.iterhybrid_predictor_corrector' --log_progress
Additional configuration is provided through flags which append the above run command, and environment variables:
Path to the QuaLiKiz-neural-network parameters. Note: if installation instructions above were followed, this may already be set.
$ export TORAX_QLKNN_MODEL_PATH="<myqlknnmodelpath>"
Path to the geometry file directory. This prefixes the path and filename provided in the geometry_file
geometry constructor argument in the run config file. If not set, TORAX_GEOMETRY_DIR
defaults to the relative path torax/data/third_party/geo
.
$ export TORAX_GEOMETRY_DIR="<mygeodir>"
If true, error checking is enabled in internal routines. Used for debugging. Default is false since it is incompatible with the persistent compilation cache.
$ export TORAX_ERRORS_ENABLED=<True/False>
If false, JAX does not compile internal TORAX functions. Used for debugging. Default is true.
$ export TORAX_COMPILATION_ENABLED=<True/False>
The following implements the JAX persistent cache and will cause jax to store compiled programs to the filesystem, reducing recompilation time in some cases:
$ export JAX_COMPILATION_CACHE_DIR=<path of your choice, such as ~/jax_cache>
$ export JAX_PERSISTENT_CACHE_MIN_ENTRY_SIZE_BYTES=-1
$ export JAX_PERSISTENT_CACHE_MIN_COMPILE_TIME_SECS=0.0
Output simulation time, dt, and number of stepper iterations (dt backtracking with nonlinear solver) carried out at each timestep.
python3 run_simulation_main.py \
--config='torax.examples.iterhybrid_predictor_corrector' \
--log_progress
Live plotting of simulation state and derived quantities.
python3 run_simulation_main.py \
--config='torax.examples.iterhybrid_predictor_corrector' \
--plot_progress
Combination of the above.
python3 run_simulation_main.py \
--config='torax.examples.iterhybrid_predictor_corrector' \
--log_progress --plot_progress
Once complete, the time history of a simulation state and derived quantities is written to state_history.nc
. The output path is written to stdout.
To take advantage of the in-memory (non-persistent) cache, the process does not end upon simulation termination. It is possible to modify the runtime_params, toggle the log_progress
and plot_progress
flags, and rerun the simulation. Only the following modifications will then trigger a recompilation:
- Grid resolution
- Evolved variables (equations being solved)
- Changing internal functions used, e.g. transport model, or time_step_calculator
You can get out of the Python virtual env by deactivating it:
deactivate
Under construction
A TORAX paper is available on arXiv. Cite this paper to cite TORAX:
@article{torax2024arxiv,
title={{TORAX: A Fast and Differentiable Tokamak Transport Simulator in JAX}},
author={Citrin, Jonathan and Goodfellow, Ian and Raju, Akhil and Chen, Jeremy and Degrave, Jonas and Donner, Craig and Felici, Federico and Hamel, Philippe and Huber, Andrea and Nikulin, Dmitry and Pfau, David and Tracey, Brendan, and Riedmiller, Martin and Kohli, Pushmeet},
journal={arXiv preprint arXiv:2406.06718},
year={2024}
}