IRT SystemX | "Machine Learning for Physical Simulation" challenge
![IRT SystemX | Challenges](/sites/default/files/2023-10/illustration_airfoil_0.png)
This competition aims to promote the use of ML-based surrogate models to solve physical problems, through a task on a recently published dataset called AirfRANS relating to airfoil design (CFD simulation). The aim is to improve basic solutions for the use case of aircraft wing design by building ML-based surrogate models. The overall objective is to improve the trade-off between the accuracy of the solutions obtained and the associated computational cost, while taking into account off-distribution generalization (OOD) and compliance with certain basic physical constraints.
To achieve this, the competition is based on the performance evaluation framework we recently proposed, called LIPS ("Learning Industrial Physical Systems"). This framework will be used to evaluate the candidate solutions provided by participants against meaningful criteria organized into three categories, namely: ML-related criteria, physical compliance criteria and OOD generalization criteria: ML-related criteria, physical compliance criteria and OOD generalization criteria. For each solution submitted, an overall score will be calculated on the basis of the above criteria in order to rank it.
To facilitate the development of new physics-based ML models, participants can use the NVIDIA MODULUS Framework and receive support from NVIDIA via a dedicated webinar at the start of the competition.
Participants must train and refine their models on their own computers based on the dataset provided. A cluster of 8 NVIDIA RTX A6000 GPUs will be made available by Exaion for participants who do not have their own GPU resources. For the training phase, reference solutions will be made available to participants.
Partners
![Partenaires ML](/sites/default/files/inline-images/Capture%20d%E2%80%99e%CC%81cran%202023-10-18%20a%CC%80%2008.39.21.png)
Organizing team
- Mouadh Yagoubi (IRT SystemX)
- Milad Leyli-Abadi (IRT SystemX)
- David Danan (IRT SystemX)
- Jean-Patrick Brunet (IRT SystemX)
- Maroua Gmati (IRT SystemX)
- Ahmed Mazari (Extrality)
- Florent Bonnet (Extrality, Sorbonne Université)
- Antoine Marot (RTE)
- Jérôme Picault (RTE)
- Asma Farjallah (NVIDIA)
- Marc Schoenauer (Inria)
- Patrick Gallinari (Sorbonne Université, Criteo AI Lab)
The competition is hosted on CodaLab and sponsored by IRT SystemX, Extrality, NVIDIA, Exaion and RTE. It is planned to organize the next iteration of the competition on different physical problems, such as a use case of a power grid (provided by RTE) as a series of challenges.
The competition will be hosted by the Codabench platform. Participants must :
- create an account;
- download a starter kit to prepare their submission;
- upload their trained ML models to the Codabench platform. The platform will then use the LIPS framework to calculate the submission scores. The score will be published on the Codabench competition page, and the participant will also have access to an additional page with detailed metrics.
Who can participate?
Anyone interested in solving physical problems using ML techniques is encouraged to enter this competition. This could be an excellent opportunity to bring together people from the ML and scientific computing communities to exploit the synergies between these two fields.
How to enter the competition?
Instructions:
- Go to the challenge's codabench page: CODABENCH
- Go to the "My submissions" tab
- Click on the "Register" button
For more information, read the full Getting Started guide.
Prices
🏆 1st price: €3K
🥈 2nd price: €2K
🥉 3rd price: €1K
4th price: €500
5th price: €500
Competition phases
The competition will take place in three phases:
- Warm-up phase (4 weeks): participants can familiarize themselves with the material provided and the competition platform, make their first submissions and provide feedback to the organizers. Based on this feedback, the organizers can adjust and improve the contest for the next phase.
- Development phase (9 weeks): participants will develop their solutions and can test their already-trained models on a provided validation dataset. They also have permanent access to the global score corresponding to the submitted solution.
- Final phase (1 week): the organizers prepare the final ranking and official results. A special event will be organized to announce the winners.
Calendar
- Competition kick-off: November 16, 2023
- Warm-up phase: November 16 - December 13, 2023
- Development phase: December 14 - February 15, 2024
- Final phase: February 16 - February 22, 2024
- Announcement of winners: event scheduled between February 22 and 29 (to be confirmed)