This repository functions as submission portal for the competition hosted in conjunction with the Workshop on Geometry-grounded Representation Learning and Generative Modeling (GRaM) at ICLR 2026. For description of the challenge refer to the competition website (link above).

Deadline is on April 22, 2026 (AoE).

In order to participate in the competition, your team has to create a valid submission in the form of a pull request to this repository. The requirements for a valid submission are listed in the following. You can mimic our implementation of a basic MLP for reference.

Create a class implementation of your model (fully contained) in the directory models/<model name>/ that can be instantiated without arguments, i.e.,

model = ModelName()

and is callable (e.g., via model.forward) by the signature

def __call__(
    t: torch.Tensor,
    pos: torch.Tensor,
    idcs_airfoil: list[torch.Tensor],
    velocity_in: torch.Tensor
) -> torch.Tensor:
    ...
    return velocity_out

with tensor dimensions

t: (batch size, 10)
pos: (batch size, 100k, 3)
velocity_in: (batch size, 5, 100k, 3)
velocity_out: (batch size, 5, 100k, 3)

where elements of the list idcs_airfoil are variable-length tensors indexing pos, i.e., take values in [0, 100k).

Feel free to use a different backend than PyTorch (JAX, etc.) but please match the typing in the signature above. It is fine if your model depends on external libraries (xFormers, your own, etc.) as long as they are easy to install.

Provide model weights along with your pull request, either by uploading to the directory models/<model name>/ or via download link if their file size would be too large. Your model must load the weights during construction.

Create an import entry in models/__init__.py that imports you model's constructor from models/<model name>/.

Optional: provide a Markdown file under models/<model name>/ detailing your training process and other important information to reproduce your approach.

Each team may create one submission. We will monitor submissions on a rolling basis and notify participants once their submission is valid.