Harnessing transfer learning for achieving superior thermal-hydraulic performance in heterogeneous pin-fin arrays

Abstract

In gas turbines, achieving greater fuel efficiency and increased thrust demands higher operating temperatures, which require advanced cooling mechanisms to prevent thermo-mechanical failures. Pin-fin arrays, traditionally designed with uniform circular pin-fins, have played a crucial role in cooling the trailing edge of turbine blades. However, the conventional approach of uniform pin-fin size and spacing fails to fully capitalize on the complex and evolving flow field within the domain, potentially limiting the effectiveness of the cooling system. In this paper, a transfer learning framework, which combines Bayesian optimization with computational fluid dynamics, is used to optimize three complex heterogeneous pin-fin array configurations on a large-scale domain which would otherwise be computationally expensive. This framework leverages a small-scale domain which preserves the flow-thermal behavior, allowing for high-throughput evaluations on a high-dimensional input design space, creating robust surrogate models and enabling efficient optimization. The small-scale optimized design is then transferred to the large-scale domain as a starting point for further optimization, reducing computational costs by up to 60 %. It is found that increasing the heterogeneity of pin-fin arrays leads to increases in heat transfer up to 6.8 % and reductions in pressure drop up to 76.7 % when compared to a traditional circular pin-fin array.