CFD-driven symbolic identification of algebraic Reynolds-stress models

A CFD-driven deterministic symbolic identification algorithm for learning explicit algebraic Reynolds-stress models (EARSM) from high-fidelity data is developed building on the frozen-training SpaRTA of [1]. Corrections Reynolds stress tensor and production transported turbulent quantities a baseline linear eddy viscosity model (LEVM) are expressed as functions polynomials selected library candidate functions. The training consists in solving blackbox optimization problem which fitness EARSM evaluated by running RANS simulations. Unlike approach, proposed methodology not restricted to sets full fields available. However, solution high-dimensional expensive function required. Several steps then undertaken reduce associated computational burden. First, sensitivity analysis used identify most influential terms dimensionality search space. Afterwards, Constrained Optimization using Response Surface (CORS) algorithm, approximates black-box cost response surface constructed limited number CFD solves, find optimal parameters. Model discovery cross-validation performed three configurations 2D separated flows channels variable section different show flexibility method. discovered applied prediction an unseen flow with higher geometry. predictions new case shown be only more accurate than LEVM, but also multi-purpose derived purely physical arguments.