Artificial Intelligence at Cerfacs
“Artificial Intelligence” (AI) encompasses many transformative technologies that will affect most fields of information technology in the years to come. However, the precise role that it will play is still very much an open question, field by field. This is even more true in the domain of scientific computing, where decades of developments have yielded high-fidelity reliable techniques. AI as a challenger has a lot to prove, but instead of opposing AI to traditional resolution techniques, many opportunities seem to lie in the synergies that are yet to be proven viable between the two.
At Cerfacs, the HELIOS (High pErformance LearnIng for cOmputational phySics) workgroup concentrates our efforts related to AI in order to coordinate resources, tools and technological surveys. Helios focuses on investigating recent developments in the field of Machine Learning for their potential to revolutionize computational physics, as they have e.g. the field of image processing. Specifically, Deep Learning, with the training of Artificial Neural Networks, is a very promising technique which is actively investigated for several reasons: its capacity to systematically extract information from previously underexploited databases; its ability to integrate complex multiscale patterns in physical models, to a level of complexity never reached in traditional hand-designed approaches; and for compression, generation and parametrization issues regarding high-dimensional data.
OpenData
Databases are freely available on CERFACS’s website. These datasets are meant to be of relevance for physical modeling purposes, but framed as a learning problem in order to apply machine learning techniques to them. They are usually linked to a publication, available on arXiv, and a repository of code to help explore it, on CERFACS’ Gitlab. Please feel free to explore these datasets, and any feedback is welcome.
Fields of focus
Our AI activities mainly focus on the relation between AI and CFD. There are many ways in which data-driven techniques can be hybridized with traditional CFD solvers, some of which are suggested in the following figure:
As the literature and community have not yet matured on this topic, CERFACS is intent on pursuing several of these options in parallel to build insights into the future of hybrid CFD techniques.
Data-driven subgrid-scale models
Several actions related to data-driven sub-grid scale techniques are investigated at CERFACS, notably:
- turbulent combustion models, for which the 2019 Grand Challenge on the Jean-Zay supercomputer gave the opportunity to demonstrate the scaling up of the AVBP-DL methodology on several thousand processors and several hundred GPUs (see the report, p. 16-17). (1 PhD)
- RANS turbulence models, as well as wall models for LES, within the HiFi-Turb project (1 Postdoc)
New resolution techniques for CFD
Actions regarding the future of hybrid CFD solvers including deep learning directly in the resolution loop are centered around:
- using neural networks to find an initial guess for the Poisson equation, enabling very low numbers of iteration from classical iterative solvers. This guarantees both the accuracy of traditional iterative solvers and the speed offered by the good initial guess. (2 PhDs)
- exploring new discretization techniques for numerical schemes based on local flow analysis and dynamic stencil adaptation. (1 PhD)
Generative networks as surrogates
Generative networks, notably GANs, have brought incredible advances to the field of image generation of high quality. Today, these generators are used to learn fast differentiable surrogates for the generation of plausible physical data, both in earth sciences and subsurface modeling. (1 PhD)
Generative techniques also yield fast solutions for surrogate models, enabling rapid exploration of design spaces in aeronautics. (1 PhD with Airbus)
Deep learning for physical time-series forecasting
Time series forecasting is an important result of many physical models, be it for civil security issues, or industrial exploitation forecasting. We focus on including physical constraints in data-driven forecasting (1 PhD with Total), as well as predicting cloud, rain, water flow and floods over time (1 PhD, 1 Postdoc).
