Jan Van den Berghe
PhD student
Ir. at VUB in 2020

Main project: Reconciling engineering models of ejectors with experiments and CFD using physics-informed machine learning
Funding: UCL grant of 1 year
Supervisor(s): Yann Bartosiewicz

Although Computational Fluid Dynamics (CFD) have proven sufficiently accurate to analyse the complex flow fields in ejectors, their computational cost remains too high to drive the design and optimization phases at system scale operation.

In that case, Lumped Parametric Models (LPM) are vastly preferable because of their lower computational cost. LPM models are based on integral balances and usually 0D formulations from isentropic gas dynamics. However, the lumping of complex physical phenomena such as turbulent mixing, oblique shock patterns and shock-boundary layer interaction into LPMs requires several closure parameters such as isentropic efficiencies. Moreover, LPMs are designed to predict global quantities such as the entrainment ratio or efficiency but are unaware of the stream-wise evolution of local parameters such as velocity, pressure, or Mach number, which only CFD can access. Finally, none of the LPMs presented in the literature consider the problem of modelling ejectors in transient conditions, which can be of primary importance at system scale operation.

This thesis aims to bridge the gap between classic LPMs and CFD approaches by introducing a new family of self-calibrating 1D models that use physics informed machine learning. In particular, the proposed models will combine 1D and unsteady gas dynamics with closure parameters that depend on local variables and properties. The functional relation linking closure models and flow parameters will be encoded in the form of Artificial Neural Networks (ANNs), and their calibration (i.e., the training of the ANN) will be automatic and online, i.e., while data is progressively collected. The model and the automated calibration procedure will be tested on experimental and numerical data. Hence the machine learning techniques will be used here to help the physical model to be closed, i.e., where our knowledge of the governing equations reaches its limit to derive universal relations for exchange of mass, momentum and energy in complex situations.

IMMC main research direction(s):
Fluid mechanics

machine learning

Research group(s): TFL
Collaborations: In collaboration with the von Karman Institute for Fluid Dynamics (VKI). The co-supervisor is Prof. Miguel Mendez.