Multi-objective performance optimization and numerical simulation of a cryogenic centrifugal compressor using neural network, active learning, and NSGA-II

Simulating helium flow through a compressor at a few kelvin requires reliable thermophysical data. The computed performance also depends on impeller geometry and off-design operation. This paper studies a pascal scale helium centrifugal compressor impeller with a CFD based multi-objective optimization method that combines a neural network surrogate, active learning, and NSGA-II. The design space included 13 geometric parameters. Isentropic efficiency and total pressure ratio were used as the objectives. An initial Latin hypercube database supplied the first CFD samples. During active learning, Monte Carlo dropout estimated surrogate uncertainty, and expected hypervolume improvement was used to choose additional CFD cases. Sobol analysis identified the effects of individual geometric parameters and their interactions on impeller efficiency. The final CFD validated Pareto front contained 11 non-dominated solutions, and the validated hypervolume increased. A representative design selected with a preference-based composite ranking score extended the sampled mass-flow range of the steady single-passage CFD maps while retaining the high-efficiency region. These results apply to the modeled impeller passage and do not establish the performance or stability limits of a complete compressor stage.