Quickstart

AutoSim uses Hydra configurations to instantiate simulators and generate train, validation, and test splits. These configurations are stored as YAML files in the src/autosim/configs directory.

You can list available simulator configurations with:

uv run autosim list

All of these represent different systems of equations which can be simulated; you can see their docstrings for more information.

Generate a dataset

The following command will generate a dataset for the spatiotemporal/advection_diffusion simulator, with:

• a 16x16 grid (simulator.n=16),

• a total simulation time of 0.2 (simulator.T=0.2)

• a time step of 0.1 (simulator.dt=0.1) (i.e., there will be three time steps),

• a single training example (dataset.n_train=1),

• a single validation example (dataset.n_valid=1), and

• a single test example (dataset.n_test=1).

uv run autosim \
  simulator=spatiotemporal/advection_diffusion \
  simulator.n=16 simulator.T=0.2 simulator.dt=0.1 \
  dataset.n_train=1 dataset.n_valid=1 dataset.n_test=1

You can change the arguments to generate a dataset with different parameters.

By default, the generated dataset is saved to the directory

outputs/${now:%Y-%m-%d}/${hydra:runtime.choices.simulator}_${shortuuid:7}

so, for example, the above command might generate a dataset in

outputs/2026-12-31/spatiotemporal/advection_diffusion_1a2b3c4

If desired, you can change the output path by supplying an extra argument:

uv run autosim [...] ++dataset.output_dir=my_custom_path

What’s in a dataset?

The generated dataset has the following structure:

.
├── cli.log
├── examples
│   └── train
│       └── batch_0.mp4
├── resolved_config.yaml
├── stats.yml
├── test
│   └── data.pt
├── train
│   └── data.pt
└── valid
    └── data.pt

The data.pt files are PyTorch-serialized dictionaries. The simulation trajectories are stored as tensors under the data key, alongside the constant_scalars and constant_fields entries. The examples directory contains a video of the training example, which can be used to visualize the simulation.

cli.log is just the output of the uv run autosim command.

resolved_config.yaml contains a copy of the configuration used to generate the dataset. This is useful for inspecting the parameters used to generate the dataset, and for reproducing the dataset later.

stats.yml contains normalisation statistics which are used during subsequent model training.