pSeven User Conference 2022 | VIRTUAL 

This presentation attempts to demonstrate the contribution of reproducible research in MDO/AI with several recent open source attempts of Prof Morlier’s group. During the last few decades, surrogate modeling has gained in popularity, especially in engineering fields, where they are often used in design analysis and optimization to replace expensive numerical simulations. Coupled to Multidisciplinary Design Optimization (MDO) this process lead to an engineering design acceleration through AI.

The first part of the talk will present some novelties in AI for Engineers such as the Surrogate Modeling Toolbox (SMT developed conjointly between ONERA, ISAE-SUPAERO/ICA, Nasa, and University of Michigan). SMT is a Python package that contains a collection of surrogate modeling methods (Mainly Kriging also known as Gaussian Processes in Machine Learning community), sampling techniques and benchmarks. SMT is different from existing surrogate modeling libraries because of its emphasis on derivatives that can be used directly in MDO processes. It also includes surrogate models or options that are not available elsewhere: KPLS (Kriging with Partial Least Square) for automatic inputs dimension reduction, mixture of experts (with automatic clustering) for simulation codes that include singularities (such as buckling in structural mechanics) and mixed variable design space (continuous, discrete, and categorical). The 2022 version currently available propose new applications such as multi-fidelity approach, Bayesian unconstrained optimization (EGO), fully inter-operable and automated.

The second part will present some MDO applications in Aerospace sciences will be highlighted, including launcher and UAV, HALE optimal design especially in an environmental perspective.

The hot rolling of long steel products is a forming process used to provide the final cross section of rolled bars suited to further manufacturing processes. During this process, the microstructure is refined and the defects generated during the casting of steel such as shrinkage pore are reduced.

The finite element modeling of this process is useful to reduce the trial phase necessary to elaborate an industrial rolling sequence suited to the requirements of customers in terms of product soundness and mechanical properties. To reach this objective, the boundary conditions of the model need to be carefully identified to ensure its ability to correctly model the thermomechanical fields involved in the rolled bloom.

In the current study, pSeven is used to optimize the thermal exchange and the friction coefficients of the rolling model of the Rotoforgia rolling mill of ABS. The simulation model is built using the FORGE software and the optimization procedure is based on the minimization of the difference between simulated and measured rolling forces. The workflow implemented in pSeven is using a surrogate-based optimization approach to explore the defined design space and then optimize the response of the simulation. During the workflow, the setting up of the thermal exchanges and friction coefficients in FORGE and the running of the simulations are fully automated by pSeven. The post-processing phase is automatized as well. During this step, the mean rolling force of each stand of the rolling sequence and other process parameters such as the surface temperature are extracted. Using this approach, a large combination of entry parameters is tested without configuring each simulation case individually. The required storage space is also limited to only one computed case at once for the whole study. The use of computation resources is thus optimized, and the large computation time needed can be executed as background tasks. Using this approach, the performances of the simulation model are improved after 75 runs of the computation.