Détail des interventions

Olivier Roustant (INSA/IMT)

Titre : An overview of 5 years of research on surrogate modelling in OQUAIDO.

Auteur : Olivier Roustant (INSA-Toulouse).

Résumé : OQUAIDO was a 5 years research project, gathering partners from technological research (BRGM, CEA, IFPEN, IRSN, Safran, Storengy) and academia (Mines Saint-Étienne, École Centrale de Lyon, CNRS, Univ. Grenoble Alpes, Univ. Nice, Univ. Toulouse III). The Chair focused on surrogate modelling and its applications to optimization and uncertainty quantification. The philosophy was to do upstream research guided by case studies. The research outputs include PhD thesis, publications, open source R software and notebooks.

In this talk, I will present an overview of noticeable contributions of OQUAIDO, that address several main scientific limitations: mixed continuous and discrete input variables, monotonicity / inequality constraints, large data sets, large number of inputs / functional inputs, stochastic codes. I will give a couple of illustrations on the case studies, as well as a brief description of the R software. Finally, I will say a word of the future of surrogate modelling through the starting CIROQUO project.

Mini Bio : Olivier Roustant is professor at INSA Toulouse in applied mathematics. He supervised the OQUAIDO research project with N. Durrande and R. Le Riche.

Nathalie Bartoli (ONERA)

Titre : Surrogate-based and adaptive optimization.

Auteur : Nathalie Bartoli (ONERA).

résumé : The objective is to present some surrogate model techniques in order to perform efficient optimization approach to solve complex design systems. The proposed optimization is based on Bayesian optimization (typically Efficient Global Optimization) with adaptive sampling in order to promote a trade-off between exploration and exploitation. Some adapted Gaussian processes are used to handle the number of design variables, combining mixtures of experts to handle non linearity for the objective and/or the constraints and adaptive strategies to handle some equality and/or inequality constraints.

The performance of this approach was evaluated for analytic constrained and unconstrained benchmark problems and more realistic aeronautical applications.

Mini Bio : Nathalie Bartoli received the Eng. Diploma and Ph.D. Degree in applied mathematics form the National Institute of Applied Sciences of Toulouse, France, in 1997 and 2000, respectively. She worked 5 years as a researcher at CERFACS (European Center for Research and Advanced Training in Scientific Computation). Since 2005, she is doing research at ONERA in the applied mathematics domain. Her research interests include surrogate based optimization. She is in charge of some courses at ISAE-SUPAERO in the field of optimization and numerical simulation. In 2012, she joined the MDO and conceptual aircraft design team at ONERA, where she has worked ever since. In December 2019, she defended her academic accreditation (HDR) to supervise research in Applied Mathematics. She worked on multidisciplinary optimization, Bayesian optimization and the mixtures of experts techniques within the framework of national projects or Europeans with a joint supervision of several PhDs on these subjects. She is currently involved in AGILE 4.0 H2020 project (2019-2022) and previously in AGILE (Aircraft 3rd Generation MDO for Innovative Collaboration of Heterogeneous Teams of Experts, 2015-2018).

Société Datadvance

Titre : Surrogate Modeling using pSeven

Auteur : Joan Mas Colomer.

Résumé de la société Datadvance :

Datadvance develops a software platform, for Predictive Modeling and Design Space Exploration (DSE), called pSeven.

Surrogate modeling is at the core of the solution. Other main capabilities include engineering workflow automation, design space exploration including surrogate-based optimization, and uncertainty quantification.

Surrogate models can be trained, without programming, either using existing data loaded directly in the platform, or from data produced by the automated engineering workflow. Model training itself can even be part of the workflow along with the other HiFi models to substitute.

Many approximation techniques (including Gaussian Processes, Neural Networks, and Mixture of Approximators) are available in pSeven.

To overcome the difficulty of choosing and tuning techniques, Datadvance has developed a unique, industry focused, AI technology called SmartSelection, where the user simply specifies hints about the problem and requirements about the expected model (exact fit, smoothness…).

The pSeven platform also provides dedicated tools to assess model quality and model exploration tools to visually explore multi-dimensional models.

Mini Bio :

Joan Mas Colomer graduated in Aerospace Engineering in a double degree program at ISAE-Supaéro and the Polytechnic University of Catalonia in 2014. In 2018, he obtained a PhD in Aerospace Engineering from ISAE-Supaéro for a thesis on multidisciplinary optimization applied to structures and aerodynamics carried out at ONERA. At the beginning of 2020, he joined Datadvance as an application engineer for the pSeven platform.

Société ADAGOS

Auteur : M. Masmoudi. Société ADAGOS.

Titre : Apprendre les structure cachées dans les données.

Résumé : L'apprentissage profond va de pair avec “big data”, or le “big data” est loin d'être une réalité industrielle. Pour cette raison, nous proposons une approche parcimonieuse, qui réduit d'une manière considérable la quantité de données nécessaire à l'apprentissage, tout en faisant de meilleures prédictions.

En effet, apprendre, c'est ajuster le poids des connexions neuronales. Moins on a de connexions neuronales et moins on a besoin de données pour les ajuster.

En plus, la parcimonie de nos réseaux neuronaux oblige le processus d'apprentissage à aller chercher les structures cachées dans les données pour un apprentissage plus intelligent et une meilleure prédiction.

La construction de réseaux neuronaux de taille minimale est basée sur la théorie du gradient topologique.

Nous présenterons quelques modèles dynamiques, où tout écart à la parcimonie se traduit par une dégradation de la prédiction.

Mini Bio : Mohamed Masmoudi est professeur à l'Institut de Mathématiques de Toulouse et fondateur de la société ADAGOS. Il a consacré sa carrière à mettre les mathématiques au service de la société.

Société Monolith AI

Titre : a platform to accelerate product development

Auteur : Dr Joël Henry.

Résumé de la société : Monolith developed a platform using AI and historical data to accelerate product development. The platform is a no-code environment for engineers and was tailored to work with simulation and test data. Multiple tools are available in the platform, including intelligent and interactive data exploration, data preparation, modelling as well as the use of these models in production. The platform contains off-the-shelf models with as Neural network, Gaussian processes and Random forests, but also patent-pending AI models for 3D that were developed in-house. These models can be used to learn the relationship between a 3D design (e.g. CAD file, mesh, …) and its performance, whether it’s structural, aero, … (e.g. drag coefficient, maximum strength, …). Finally, the models can be deployed and used (i) to make new instantaneous predictions (communicating uncertainty), (ii) to understand better the designs (explainability) and (iii) to provide recommendations of new designs (optimisation). These capabilities are demonstrated in a few cases studies on both tabular and 3D data.

Mini Bio : Dr Joël Henry is Principal Engineer at Monolith. In 2018, he completed a PhD in bio-inspired multi-scale structural composites at Imperial College London. He published work in Physical Modelling and analytical virtual testing framework and has previously worked on improving test and simulation methods in the aerospace industry. In 2018, he joined Monolith AI, where he works with the majority of their clients to create lasting value for them.

Michael Bauerheim

Titre : Reinforcement Learning applied CFD.

Auteur : Michael Bauerheim.

Résumé : Deep Reinforcement Learning (DRL) has shown outstanding results in many fields, from video games to robotics. The objective is to control a system evolving in an environment by discovering an optimal sequence of actions to maximize a desired cummulative cost function. To do so, the DRL algorithm can interact directly with the system, trying actions, and learning its long-term reward using a deep neural network. Such a strategy allows model-free optimization of complex systems, which is highly relevant in Computational Fluid Dynamics (CFD) to find for instance optimal kinematics and flight trajectories in perturbed aerology, or to control flow instabilities. Compared with classical DRL applications where the response to a system's action is fast, CFD requires a significant time to produce such output, thus limiting the number of samples available for the learning task. Since DRL applied to CFD is still rare, a little is known on how tractable these DRL algorithms are in this context. Consequently, this presentation will highlight the potential of DRL methods and its difficulties (e.g. sample efficency, robustness etc.) on several applications of increasing complexity involving CFD, from a simple aerodynamic toy-problem, towards turbulent CFD problems.

Mini Bio : Michael Bauerheim obtained his PhD at Cerfacs on theoretical and numerical thermoacoustics in 2014. He joined the departement of aerodynamics and propulsion (DAEP) at ISAE-SUPAERO in 2017, working on acoustics, bio-inspired aerodynamic optimization, and data-driven methods. His group investigates deep learning and deep reinforcement learning methods applied to fluid mechanics.