Frequently Asked Questions

Browse benchmarks on the leaderboard, download a starter kit from the showcase, and read the Getting Started guide. Public model upload is not open yet.

A small zip with the design-of-experiments table, a couple of example solved cases (YAML and VTU), reference meshes, and notes on how the simulation is set up. It helps you understand the benchmark. It is not the full training or inference bundle. Full data releases are planned later. See Starter kits for contents.

Published results from Rescale AI Physics tutorial runs: Two-Stream Transformer, MeshGraphNet, Transolver, and DoMINO on the live benchmarks. You can compare architectures, inspect submission scores, and download starter kits from the showcase.

Not publicly. Submit is invite-only for now. Contact us if you want early access.

Rankings combine global scalar accuracy (Cl, Cd, peak stress, Kt, and similar quantities) and nodal field R² on holdout cases. The combined score is the geometric mean of those averages. See the evaluation protocol for details.

No. Holdout ground truth is not released publicly. Starter kits include holdout inputs and reference scalars for orientation, not full field meshes for every holdout case.

Current leaderboard runs used Rescale grossular-1 nodes: 1× NVIDIA A10 GPU, 4 AMD EPYC cores, and 32 GB RAM per node. Training and inference times are shown for context but do not affect rank.

Yes. When public submission opens, any architecture you choose will be welcome unless a specific benchmark says otherwise.

We plan to support that when upload opens. Release of code and weights would remain your choice.

The consortium is in early stages. See the About page for where we are headed on governance. Contact us if you want to get involved.

Yes. We are open to new benchmark ideas as the platform grows. Contact us if you have a proposal.

AIPhysics.org is free to use. Browsing benchmarks, starter kits, and the leaderboard costs nothing. We do not plan to charge researchers a submission fee.