Optimization of an high-pressure turbine vane for pressure gain combustion cycles through convolutional neural network

Rotating Detonation Engine (RDE) represents a pioneering technology in the propulsive and energy production fields, although the coupling with transonic High-Pressure Turbine (HPT) stages is challenging. Several studies introduced the concept of diffusive endwalls in order to ingest high-subsonic flow delivered by the RDE. Machine Learning (ML) gained a prominent role in the optimization procedures because of its fast and accurate prediction, reducing the computational time respect to Computational Fluid Dynamics (CFD). The aim of this study is to optimize the blade and endwall profiles of an HPV, targeting an enhancement of aerodynamics performance. The optimization methodology is based on several algorithms using a Convolutional Neural Network (CNN) as the fitness function, which can estimate the vane performance directly from an image containing blade and endwall profiles. The Design of Experiment (DOE) is performed using Reynolds-Averaged Navier–Stokes (RANS) analysis that takes in account 18 design variables, which represent geometric features of vane profile and diffusive endwalls. The optimized profiles are then tested under pulsating boundary conditions from a Rotating Detonation Combustor (RDC) to understand which configuration can perform adequately under extremely harsh conditions. The CNN coupled with Particle Swarm Optimization (PSO) increases the vane efficiency with respect to nominal configuration and the novel design proved to be affordable also in presence of variable inlet conditions, attenuating pressure oscillations.