A Machine Learning-Based Calibration of a 1D Ejector Model from CFD

Abstract

Ejectors are devices that expand a primary flow through a nozzle to entrain and compress a secondary flow without moving parts. They can be modelled in 1D as two streams exchanging momentum. However, the engineering modelling of this exchange is based on closure parameters such as friction coefficients that must be calibrated against experimental or numerical data. This work proposes a general machine learning framework for calibrating engineering models governed by Ordinary Differential Equations (ODEs) and presents its application to the 1D modelling of an ejector. We combine a physics-separated approach with a physicsintegrated approach, with the first acting as an initial guess for the second. The first approach calibrates shear and friction coefficients from their counterpart extracted via post-processing of an axisymmetric CFD simulation. The second consists of calibrating these coefficients from the prediction of physical quantities (pressures, temperatures, and cross-sections), thus making the training process aware of the ODEs driving the forecast.