VAIL
September 30, 2026

Top 20 in Track 2 of the NeurIPS 2026 RealPDE Competition

Our team finished in the top 20 of Track 2 (Long-Term Test-Time Adaptation) at the NeurIPS 2026 RealPDE Competition with a final score of 81.82, and has been invited to the final evaluation as a reserve team.

We are proud to share that our team, competing as probability, finished in the top 20 of Track 2 (Long-Term Test-Time Adaptation) at the NeurIPS 2026 RealPDE Competition, with a final score of 81.82. Following the development phase, the organizers have invited us to the Track 2 final evaluation as a reserve team.

The team is lab members Taher Merchant, Aarush Mahajan and Rishi Garg, advised by Seifedine Kadry. (With thanks to Abhinav Joshi and Siddharth Bharti for their help and contributions in Track 1.)


About the competition

RealPDE tests whether machine learning models can predict real fluid flow, measured in a wind tunnel, rather than only the clean output of a simulation.

RealPDE is part of the NeurIPS 2026 competition track. It pairs wind-tunnel measurements (PIV) with matching simulations (CFD) of flow around a NACA4418 airfoil across several Reynolds numbers and angles of attack. The competition had a $21,000 prize pool across its two tracks, and Track 2 alone drew more than 350 registered participants and over 1,600 submissions.

The two tracks:

  • Track 1 (Sim2Real) asks a model trained on simulation to predict the real measured flow.
  • Track 2 (Long-Term Test-Time Adaptation) asks a model to forecast the flow far ahead while it adapts during the test, within a time budget per step.

Each submission received five scores: relative L2 error, turbulent kinetic energy (TKE), mean velocity profile error (MVPE), runtime, and a safe prediction score (SPS) that rewards narrow uncertainty bands that still contain the true flow. The leaderboard combined them into one final score with a formula the organizers did not publish, and each team could submit once per day.


Our results

Track 2: Long-Term Test-Time Adaptation

We finished in the top 20 of the Track 2 development leaderboard with a final score of 81.82.

Our subscores:

  • Relative L2: 94.73
  • TKE: 79.84
  • MVPE: 94.35
  • Time: 92.76
  • SPS: 43.44

RealPDE Track 2 development leaderboard, top 20, with team probability highlighted at rank 18

The Track 2 development leaderboard at the close of the development phase.

Track 2 detailed results: rel_l2 94.73, TKE 79.84, MVPE 94.35, time 92.76, SPS 43.44, final 81.82

Detailed Track 2 results for our final submission.

Track 1: Sim2Real

In Track 1 our final score was 81.28, with subscores of 94.55 (relative L2), 78.87 (TKE), 93.65 (MVPE), 92.94 (time) and 42.21 (SPS).

RealPDE Track 1 development leaderboard showing team probability with a final score of 81.2832

Our entry on the Track 1 development leaderboard.

Track 1 detailed results: rel_l2 94.55, TKE 78.87, MVPE 93.65, time 92.94, SPS 42.21, final 81.28

Detailed Track 1 results for our final submission.


What comes next

In the final evaluation, the organizers retrain each invited team’s method from scratch on the public training data and score it on a private test set. We are preparing our Track 2 training package now. Final results will be announced on 10 November 2026, ahead of the competition’s NeurIPS session on 6 December.

We wrote about the experience and what we learned in a blog post.


Thank you to the RealPDE organizers for a demanding and carefully run competition, to Seifedine Kadry for his guidance throughout, and to Abhinav Joshi and Siddharth Bharti for their help and contributions in Track 1.