Multi-scale feature fusion network and bi-objective Bayesian optimization for blended-wing-body underwater glider

The shape optimization of blended-wing-body underwater glider (BWBUG) heavily relies on computational fluid dynamics (CFD) simulations, and the complex mapping relationships between design parameters and performance cannot be easily resolved, severely restricting design efficiency. To address these challenges, a shape optimization method based on a multi-scale feature fusion network (MSFFN) and bi-objective Bayesian is proposed for BWBUG. First, an automated parametric modeling and simulation platform is established, and a dataset of one-to-one corresponding design and performance parameters is generated. By exploiting this dataset, a machine learning modal based on MSFFN is established which outperforms various neural network models on multiple metrics with an average MAPE of only 10.25%. On this basis, SHapley Additive exPlanations (SHAP) analysis and parameters sensitivity analysis are further incorporated to reveal the influences of key design parameters on performance. Finally, taking lift-to-drag ratio (L/D) and volume (V) as the optimization objectives, the optimal trade-off solution selected from the Pareto solution set obtained by the bi-objective Bayesian optimization achieves an increase of 16.87% in L/D and 15.10% in V. These findings demonstrate the potential of machine learning for the design and property prediction of BWBUG.