RANS predictions of trailing-edge slot flows using heat-flux closures developed with CFD-driven machine learning

Abstract

Accurate prediction of the wall temperature downstream of the trailing-edge slot is crucial to designing turbine\nblades that can withstand the harsh aerothermal environment in a modern gas turbine. Because of their computational\nefficiency, industry relies on low-fidelity tools like RANS for momentum and thermal field calculations, despite their\nknown underprediction of wall temperature. In this paper, a novel framework using a branch of machine learning, geneexpression\nprogramming (GEP) [Zhao et al. 2020, J. Comp. Physics, 411:109413] is used to develop closures for the\nturbulent heat-flux to improve upon this underprediction. In the original use of GEP (“frozen” approach), the turbulent\nheat-flux from a high-fidelity database was used to evaluate the fitness of the candidate closures during the symbolic\nregression, however, the resulting closure had no information of the temperature field during the optimisation process.\nIn this work, the regression process of the GEP instead incorporates RANS calculations to evaluate the fitness of the\ncandidate closures. This allows the inclusion of the temperature field from RANS to advance the iterative regression,\nleading to a more integrated heat-flux closure development, and consequently more accurate and robust models. The\nGEP-based CFD-driven framework is demonstrated on a trailing edge slot configuration with three blowing ratios. Full\na posteriori predictions from the new closures are compared to high-fidelity reference data and both conventional RANS\nclosures and closures obtained from the “frozen” approach.