Michael Bauerheim received his PhD in thermoacoustics in 2014 at CERFACS under the supervision of Thierry Poinsot and Laurent Gicquel. His doctoral research focused on low-order modelling, uncertainty quantification, and high-fidelity simulations of both academic and industrial thermoacoustic systems. Then, he extended his research to turbomachinery and cavity noise at École Centrale de Lyon (France) and ETH Zurich (Switzerland). In 2017, he joined ISAE-SUPAERO (Toulouse, France) in the Department of Aerodynamics, Energy and Propulsion.

Building on his longstanding interests in modelling, uncertainty quantification, optimization, and large-scale simulations, his research progressively evolved toward artificial intelligence for fluid mechanics. At ISAE-SUPAERO, he founded the L2F (Learning Fluid Flows) research group, dedicated to the development of machine learning approaches for fluid dynamics. Over the past decade, his work has focused on deep learning and deep reinforcement learning methods for accelerating simulations, optimizing complex systems, and controlling fluid flows.

Day 1: Wednesday, September 9th | 14:00 – 15:30

Room: Opus

Title:  The AI/ML Ecosystem for Engineering Analysis, Design, and Control

Moderator: Prof. Juan Alonso (Stanford University, USA)

Panelists:
Ass. Prof. Michael Bauerheim (ISAE-Supaero, France), Prof. Xiaomo Jiang  (Dalian University of Technology, China), Ass. Prof. Ricardo Vinuesa (University of Michigan, USA)

Abstract:
As data availability (from simulation, test, and operations) becomes more widely available, AI/ML-based methods for the analysis, design, and control of aerospace systems, including power and propulsion systems, are becoming more pervasive. But the pace of development is frenetic and the landscape of physics-informed and agentic AI/ML is changing more rapidly than we can keep up with. In this panel we tackle both the current state and 1-2-year predictions for the capability, accuracy, and cost for such methods. What are the most logical applications in engineering in the near term? What transformational capabilities do we expect in the next 5 years? What are the key bottlenecks to make this technology work in realistic industrial environments?

Day 1: Thursday, September 10th | 14:00 – 14:25

Room: Tchaikovsky

Title:  Beyond better models: Towards engineering-ready AI for turbomachinery

Speaker: Michael Bauerheim

Abstract:
CFD remains the cornerstone of turbomachinery design, yet its computational cost severely limits the exploration of design spaces involving complex geometries and multiple operating conditions. Addressing future challenges in fuel efficiency, emissions, noise reduction, and sustainable aviation will require exploring these design spaces far more extensively than is currently feasible. Over the past few years, deep learning has emerged as a promising alternative through increasingly accurate surrogate models for high-fidelity flow prediction. This talk argues that the field is reaching a turning point. While significant progress has been achieved in neural architectures for isolated components, the next challenge is no longer simply to build more accurate models, but to develop AI systems that can be reliably integrated into industrial design workflows. As an illustration, I will present MARIO, a geometry- and operating-point-aware implicit neural field recently extended to three-dimensional turbomachinery flows. It provides continuous, mesh-agnostic predictions that generalize across blade geometries and operating conditions while preserving the rich flow information required for CFD-like post-processing. These recent developments also pave the way to multiphysics applications, such as rapid aeroacoustic predictions among many others. Nevertheless, these models still represent individual components, raising the fondamental question of how future AI models should capture interactions across multiple components, such as a rotor-stator stage subject to inflow distorsion. Finally, I will argue that future progress will depend less on architectural innovations than on the ecosystem built around these models. Recent advances in uncertainty quantification, reliability and data assimilation illustrate how the community is moving from accurate neural surrogates towards trustworthy engineering AI. This evolution is essential to transform deep learning into a reliable engineering tool for the next-generation turbomachinery design and analysis.