The heat transfer coefficient and pressure losses are among the main parameters to be evaluated in gas turbine cooling network design. Due to the complexity of these estimates, correlation-based computations are typically used as a result of time-consuming and expensive experimental activities. One of the main problems that the industry has to face is that these correlations, based on non-dimensional experimental data, produce reliable results in a range of validity typically different from that encountered in gas turbine applications.
This paper will present preliminary results of an innovative procedure based on CFD analyses and Artificial Neural Networks, able to extend correlation predictions out of their range of validity, without any additional experimental data. Well-known test cases were replicated by building corresponding CAD geometries which were discretized by means of appropriate meshes, resulting from grid-independence studies. CFD analyses, based on the RANS approach, were performed to overlay the computations of the Nusselt number obtained from experimental activities. A preliminary comparison among turbulence models was carried out to find one leading to a good agreement with the experimental data. Then, an optimization method, based on Evolutionary Algorithms, was applied to the CFD analyses in order to find the best set of constant values for the chosen turbulence model, leading to the most accurate prediction of the experimental dataset. The resulting ad hoc CFD model was adopted in order to analyse test case configurations characterized by parameters within and external to the correlation validity field, building a sufficiently wide feeding database. A feed-forward multi-layer neural network was selected among network architectures typically used in engineering applications for prediction analyses. ANNs were chosen because they enable the solution of these complex nonlinear problems by using simple computational operations. The selected Artificial Neural Network was trained by a back-propagation procedure on the CFD results regarding Nusselt number. The validation of the resulting ANN was performed comparing its outputs with experimental data external to the correlation range of validity, which had not been used in the training session. Good agreement has been found. Results are presented and discussed.