To guarantee safety while maximizing performance, REGALE partner Electricité de France (EDF) executes safety analysis on a regular basis in all its electricity production infrastructure, such as nuclear, hydroelectric or wind power plants, using a variety of physics simulations. In the last years machine learning has become more and more interesting for scientists, promising faster and more detailed simulations.
Alejandro Ribés is principal scientist at the Industrial AI Laboratory SINCLAIR from EDF Lab Paris-Saclay and works on the question, how to use Artificial Intelligence techniques for numerical simulations. We wanted to know more about his research and asked him some questions.
Can you tell us a little bit more about your work?
Alejandro: I work at the Research and Development area of EDF, a French electric utility company, which is one of the world’s largest producer of electricity. My expertise is in applied mathematics and computer science. In the last six years I’ve been working closely with scientists and engineers that perform simulations for EDF. My role is to provide them with tools to facilitate their simulations. Usually, we use our own software on our own supercomputers at EDF. Over the past years, we have developed a lot of opensource software that is used across industries.
What project do you currently work on?
Alejandro: My research focuses for a few years now on machine learning techniques. My goal is to find out, how we can use these techniques at EDF to improve our simulations – make them more detailed and faster.
We summarized our latest research results in two papers:
L. Meyer, M. Schouler, R-A. Caulk, A. Ribes, B. Raffin. High Throughput Training of Deep Surrogates from Large Ensemble Runs. In Proceedings of The International Conference for High Performance Computing, Networking, Storage, and Analysis, Denver, Colorado, USA, November 2023 (SC’23). DOI: https://dl.acm.org/doi/10.1145/3581784.3607083
L. Meyer, M. Schouler, R-A. Caulk, A. Ribes, B. Raffin. Training Deep Surrogate Models with Large Scale Online Learning, ICML (International Conference on Machine Learning), Honolulu, Hawaii, USA, July 23-29, 2023.
They got accepted in two major conferences last year – The International Conference on Machine Learning (MCML) and Supercomputing Conference (SC). That’s a great confirmation of our work and we are very happy about that. In these papers, we discuss how to train at scale deep learning algorithms, which are a viable alternative for obtaining fast solutions for Partial Differential Equations (PDEs). These equations play important roles in the mathematical description of the world’s physical phenomena. We propose an opensource online training framework for deep surrogate models, which is an extension of our open source project Melissa.
How is this connected to the REGALE project?
Alejandro: The REGALE project is one of the R&D projects EDF is involved in. In REGALE we created a database of numerical simulations, in this case floodings, that can have negative effects on the operation of hydroelectric power plants. We simulated a lot of floodings with different parameters. These large simulations produce a lot of data that you normally need to write down. That’s the problem of big data. We are speaking of terabytes or even petabytes of data produced, thus we have the problem of saving and storing this huge amount of data. Our approach to tackling this issue is the Melissa framework.
Melissa is an on-line data processing solution, developed by EDF and the University Grenoble Alpes (UGA), coupling member execution with on-the-fly statistics computing. The data goes to Melissa via the network and Melissa does the computation on the fly without saving the data. It is used in three pilots in the REGALE project.
Before REGALE we had Melissa for sensitivity analysis. We used our participation in the REGALE project to upgrade Melissa for training neural networks. For Melissa, participating in this project has been a big step forward and we are really satisfied with the results.
In fact, Melissa is currently one of the few solutions on the world for the problem of this kind of big data.
Why is this work important?
Alejandro: Writing down data is a very energy-intense process. But we believe that working with data on the fly through Melissa will soon be more efficient than the old way of storing data and this is a huge opportunity to reduce energy consumption and cost. The same is true for training models for artificial intelligence. Normally, if you train neural networks you have a lot of data to store and you repeat the training very often. For REGALE we have adapted Melissa to do this training on the fly and at the same time, the training is better quality because we don’t repeat trainings, all simulations are different.
That’s a change of paradigm and a very exciting development. In fact, Melissa is currently one of the few solutions on the world for the problem of this kind of big data.
Links
Scientific Paper: https://openreview.net/forum?id=WT70GgYdLI
Melissa: https://github.com/melissa-sa/