Séminaire : David Danan et Milad Leyli Abadi

Evaluate hybrid AI/physics simulators with the LIPS framework: Case Study insights

Physic-based simulations are at the core of many real-world critical industrial systems. While classical approaches have proven their accuracy and robustness over the years and come with a rich mathematical foundation, they suffer from several limitations depending on the underlying physics and use cases. For instance, those approaches are known to be computationally expensive and even lead to non-convergence for some feasible cases. Recently, the use of data-driven approaches to learn complex physical phenomena has been considered as a promising approach to address those issues. However, while allegedly much more efficient from a computational point of view, they may not necessarily hold up well regarding physical considerations and accuracy. Thus, there is both a need to design relevant hybrid simulators and propose a standard approach to evaluate those simulators on a same basis. Regarding the former and the experience through various organized competitions, we provide insights into three distinct use-cases from heterogeneous industrial domains, namely: the power grid, the pneumatic and the airfoil. To drive the latter, towards a better real-world applicability, we proposed at IRT SystemX a new benchmark suite "Learning Industrial Physical Simulations"(LIPS) enabling the performance monitoring and continuous improvement of augmented simulators. The need to develop efficient and industrial application-oriented hybrid simulators is also considered through the design and evaluation of several Key Performance Indicators (KPIs).