Adaptive Batch Sampling Strategy for the Hundreds-Dimensional Aerodynamic Optimization of the Centrifugal Impeller

Abstract

In recent years, in order to achieve higher performance, the number of design variables used in the aerodynamic optimization of turbomachinery (based on surrogate models) has increased significantly. However, the increase in the design dimensions brings the problems of sparse distribution of training samples and numerous inaccurate local optima to the surrogate model, making it difficult to obtain an accurate Pareto front in high-dimensional aerodynamic optimization. To solve this, adaptive batch sampling strategy (ABSS) is proposed in this paper. By searching numerous local optimum points with large prediction errors in the surrogate model, ABSS can provide batches of valuable samples for each iterative update of the model. Compared with the current model update strategy based on the best point, which requires hundreds or thousands of iterative model updates, ABSS only needs to go through a few model updates to make the predicted Pareto front close to the real one. ABSS not only greatly shortens the whole optimization time, but also makes it easier to jump out of the local optima for the optimization. Based on this, a 114-dimensional aerodynamic optimization of the full three-dimensional centrifugal impeller (including arbitrary blade surfaces, independent splitter surfaces, and non-axisymmetric hub surfaces) is carried out. The results show that, compared with the baseline impeller, the design point isentropic stage efficiency of the optimal impeller is increased by 2.1%, the design point pressure ratio is well controlled to 1.84, and the choke margin is increased by 9.2%. This research breaks through the largest number of variables in the current centrifugal impeller optimization, proposes the optimization methods of the full-3D centrifugal impeller, and also provides a reference for the future high-dimensional aerodynamic optimization of turbomachinery.