postdoc, Lille

Subject: Thanks to the ever-growing resources available at supercomputing centers, High Performance Computing (HPC) analyses of flow configurations including several complex concurring aspects are becoming an established reality. Thus, the development of reliable numerical strategies capable of providing an accurate representation of multi-physics problems is a timely central challenge in Computational Fluid Dynamics (CFD). The accurate prediction of numerous flow features of unstationary flows, such as aerodynamic forces, is driven by the precise representation of localized near-wall dynamics. In the last decades, several numerical strategies have been proposed to handle these two problematic aspects. Among these, the Immersed Boundary Method (IBM) has emerged as one of the most popular approaches. Among the state-of-the-art proposals reported in the literature, the IBM developed by the team shows favorable features of accuracy and efficiency. The main open challenge with IBM is the representation of wall turbulence, which is a governing aspect in most engineering cases. The wall resolution required increases with the Reynolds number as finer coherent structures are observed, leading to a rapid rise in computational resources. Therefore, a strategy based on online Data Assimilation combined with data stream learning is proposed to train a new generation of IBM methods, with the aim to obtain high precision with limited computational resources. More precisely, the data-driven strategies will aim to identify and optimize new IBM formulations capable to represent complex features of the flow, such as separation and re-attachment of the boundary layers.
The research group has recently developed a software able to perform online sequential data assimilation exploiting a reduced-order technique referred to as multigrid Ensemble Kalman Filter (MGEnKF). Following some works recently proposed in the literature, data stream learning will be integrated within the code to derive new paradigms for data-augmented IBM.
Objectives: the work of the candidate will aim for the development of the following tasks:
1. Implementation of data streaming machine learning techniques within the C++ software developed by the research group, aiming to reconstruct an augmented IBM formalism. The AI learning procedure will be fed by online data generated by the MGEnKF.
2. The performance of the code will be assessed via the analysis of progressively more complex scale-resolved turbulent flows. The test cases investigated include the turbulent channel flow and, in case of success, complex test cases such as a radial pump will be investigated. For every test case considered, high fidelity DNS / experimental data are already available from previous analyses of the research group.
Host Institution: Arts et Métiers - LMFL. The candidate will be based in Lille (France).
Scientific Leader: M. Meldi (marcello.meldi@ensam.eu)
Candidate’s profile: the candidate must have strong competences in machine learning techniques, ideally for applications with streaming data. A PhD degree in this area of expertise is required. In addition, skills in the numerical simulation of turbulent flows and / or IBM tools and / or OpenFOAM are welcome, but they are not mandatory.
Duration of the contract and start date: the proposed contract is for a duration of 18 months with a scheduled start date for January 2023. Some flexibility for the start date can be granted depending on the availability of the candidate.