Surrogate neural network and multi-objective direct algorithm for the optimization of a Swiss-roll type Recuperator

Micro-turbines are promising high power-density engines for distributed power generation. In this paper, an innovative optimization procedure is proposed to design a Swiss-roll type recuperator that recovers the exhaust heat of a micro gas turbine. There are several design parameters to be optimized for the recuperator, including the number of turns, channel width, plate thickness, and mass flow rate. The complex interconnections of the parameters make it difficult to analyze the process and select an adequate design with the highest effectiveness and lowest pressure drop. In order to reduce the number of numerical analysis required in the two-objective optimization process, a neural network was trained to serve as a surrogate model for the analysis and a multi-objective DIRECT (DIviding RECTangle) algorithm, named as MO-DIRECT, is proposed. After merely 5 iterations of MO-DIRECT search, we were able to fond a min-max solution with prediction error lower than 4%. In the search process, only 24 numerical simulations are required to achieve the result with totally 2,313 steps conducted in the MO-DIRECT search, rather than 35,343 simulations for an exhaustive search.