RELIABILITY OF DATA-DRIVEN STRATEGIES: CASE STUDIES IN FLUID MECHANICS

Machine learning is rapidly transforming scientific computing; yet its apparent simplicity often masks critical issues such as limited generalizability, lack of guarantees, and strong case dependency. Increasing dataset size or model complexity does not necessarily address these problems and can incur high computational costs. When possible, physical constraints are used to improve robustness.

This talk presents some case studies recently explored by our group, with an emphasis on reliability and performance guarantees.

First, we examine modeling and prediction using neural networks. Specifically, we use Long Short-Term Memory (LSTM) networks to assess how the structure of training data and the role of memory gates affect long-term predictions. Drawing on insights from ergodic theory and curriculum learning, we analyze how dataset design can ensure faithful modeling and open pathways for active learning.

In the second example, we discuss Graph Neural Networks (GNNs) and deploy them for data assimilation, using the Reynolds-Averaged Navier-Stokes (RANS) equations as a baseline. GNNs are particularly promising due to their ability to represent complex multi-connected networks, making them well suited to unstructured meshes in computational fluid mechanics. A supervised learning closure term for the RANS equations is used in combination with direct-adjoint methods and active learning. Our results provide evidence on the extent to which GNN models can be parameterized effectively.

Finally, we turn to flow control, where Reinforcement Learning (RL) is gaining traction over traditional model-reduction approaches. RL requires no prior knowledge of governing equations and learns policies from flow measurements. However, early applications often produce non-intuitive policies, even when simpler solutions exist. As a final example, we present an optimistic policy iteration strategy specifically designed for convective flows.

Speaker

Onofrio Semeraro received his PhD in Mechanical Engineering at KTH-Stockholm (Sweden) in 2013. He served as postdoctoral researcher at Ecole-Polytechnique, Palaiseau (France) and Politecnico of Bari (Italy), and he is currently a CNRS Research Associate since 2017, at the Laboratoire Interdisciplinaire des Sciences du Numérique (LISN) - Universite Paris Saclay, Orsay (France).

His studies focus mainly on control, data assimilation, modelling and data-driven techniques, ranging from system identification to deep learning for fluid mechanics. He is currently PI of an ANR project dedicated to optimal control and Reinforcement Learning for control of fluids, and contributors for projects at the intersection of machine learning, dynamical systems and fluid dynamics.

30.5.2025