Transfer learning in bayesian optimization for aircraft design

Bayesian optimization (BO) is widely employed for expensive blackbox functions, yet it suffers from the cold start problem when data are scarce. While transfer learning offers a solution by leveraging source data to accelerate target optimization, existing constrained BO transfer-learning approaches face limitations when dealing with heterogeneous design variables and complex constraint structures, leading to poor generalization. To address these issues, we propose a novel methodology integrating an ensemble of surrogate models within a constrained BO framework. Specifically, we employ a partial-least-squares dimension reduction algorithm to manage design space heterogeneity and a meta data surrogate selection method to handle constraint heterogeneity. The proposed approach is validated on numerical benchmark problems and a realistic aircraft conceptual design optimization problem. Results demonstrate significant improvement in convergence during early optimization iterations compared to standard BO, alongside enhanced prediction accuracy for both objective and constraint-surrogate models.