Precision-driven intelligent framework for thin-walled aerospace structures: Dynamic neutral-layer migration and springback compensation

Dimensional deviations arising from the sheet metal forming of aerospace thin-walled components directly reduce assembly accuracy. There are two main sources of such deviations: inaccurate positioning of the neutral layer during unfolding, and unreliable and uncontrollable springback prediction in the bending process. This study proposes an accuracy-driven intelligent prediction framework for aerospace thin-walled structures, which integrates the Radial Basis Function Neural Network (RBFNN) and the Genetic Algorithm-optimized Categorical Boosting (GA-CatBoost) model. Specifically, RBFNN is applied to calculate the neutral-layer coefficient during sheet metal forming, while the GA-CatBoost model is adopted to predict the springback angle in bending. Based on an experimentally validated finite element (FE) model, this study establishes a multi-parameter FE database covering material properties, geometric parameters, and process conditions. Dominant influencing factors are identified through Pareto analysis. The RBFNN achieved a test-set R² value of 0.9786 for neutral-layer coefficient prediction and reduced the blank-development time by 56.69–68.16% for two representative components. The GA-CatBoost model achieved a test-set R² value of 0.9901 for springback-angle prediction. Compared with the unoptimized CatBoost model, the test-set R² increased from 0.9023 to 0.9901. SHAP-based analysis was further employed to quantify feature contributions and identify nonlinear interactions among the governing variables. Experimental results confirm the effectiveness of the integrated system in forming compensation. For the tested parts, the maximum dimensional and angular deviations were 0.80 mm and 0.50°, respectively. These values were within the applicable ISO 2768–1-c tolerance limits for the relevant nominal dimension ranges and support precision control in flexible manufacturing of aerospace thin-walled components.