Abstract

A new type of section inverse design is presented that offers a 200 times speed up over conventional iterative methods. The method parametrises a blade section in terms of its aerodynamic duty and design style. The duty is defined by the inlet and outlet flow angles, and the Mach and Reynolds numbers at which the blade must operate. The aerodynamic style is represented by a normalised Mach number distribution, parametrised by four control variables. This is found to provide a wide range of desirable loading distributions.

For a given style and duty, the method obtains the section geometry and its performance using a radial basis function network. This network is trained, a priori, on a database of designs produced using a conventional inverse design method. The database covers a range of inlet and exit angles for a range of Mach numbers, at each duty aerofoil families are produced varying the four style parameters. During the database production the radial basis function (RBF) network is sequentially refitted and used to accelerate future designs convergence by providing an increasingly better initial guess.

Errors in prediction accuracy, of the fully trained system, are calculated by evaluating the obtained geometry and comparing its aerodynamic performance with the intent. The average exit flow angle error was calculated to be 0.074° from the intent. The root-mean-squared error between intent and achieved Mach fraction distributions averaged below 0.0275. The loss prediction featured average errors of 0.012% and hence provides a very accurate method of performance estimation.

It is highlighted that not all combinations of aerodynamic duty and design style are compatible. “Infeasible” regions within the design space where the aerodynamic intent is not possible exist. A method is applied to avoid these regions, where section errors are above average, ensuring accurate section generation.

This content is only available via PDF.
You do not currently have access to this content.