PHysics-based Learning for robUst fluid SIMulation – PHLUSIM

PHLUSIM is an interdisciplinary project that proposes to unlock the potential of recent advances in artificial intelligence for computational fluid dynamics (CFD) and reduced order modeling (ROM) for fluids. Specifically, it seeks to apply innovative techniques emerging in AI to the specificity of their application to fluid modeling and simulation. Classically in CFD, physical laws are encoded in partial differential equations and solved for via numerical schemes, which become intractable at scale and require many internal models. These models historically do not leverage the large volumes of data that are now available from measurements and simulations, and new learning-based strategies can revisit existing CFD and ROM bottlenecks and have the potential to dramatically improve the status quo, notably by relying on innovative context-aware formulations. However, moving to data-driven high-dimensional learning presents new challenges for fluid dynamics experts such as risks of overfitting and generalization difficulties. Conversely, in the machine learning (ML) community, physics-based ML has also emerged as a new research field but has often been limited to simplified academic cases. The complexity of industrial-scale fluid flows introduces new challenges for these methods. Therefore, in both perspectives, this topic presents typical interdisciplinary challenges. To address them, these two distinct but overlapping skill sets are gathered in this consortium of both AI experts (ISIR-Sorbonne Universite) and fluid-modeling experts (CERFACS-Toulouse and d’Alembert-Sorbonne Universite), which will closely collaborate in PHLUSIM to bridge the gap between the disciplines. They will address this by setting-up high-dimensional context-aware learning problems related to fluid simulation and modeling, training state-of-the-art AI algorithms on them, and systematically addressing generalization issues that arise. Four main objectives will guide this work: (1) select context-aware techniques that are compatible with fluid modeling, (2) define the formulations by which they can be used to assist high-fidelity CFD solvers, (3) explore an array of uses they can have for reduced order models, and define quality metrics w.r.t. pre-existing methods; (4) assemble data on a new unseen case and on numerous existing cases to map out the generalization capacity of all the methods explored in the project. The project organization into four work packages directly reflects the four objectives. This project has the potential to significantly impact a wide range of scientific fields, including atmospheric, ocean, and subsurface sciences, as well as engineering systems that rely on fluid-dynamics understanding.