Abstract
The stacking sequence of composite laminates directly determines the internal stress and strain distributions and thereby affects fatigue life. To examine differences in prediction performance among layup categories that may be obscured by overall metrics, this study develops an XGBoost-based framework for category-specific evaluation and further investigates local stacking sequence screening under fixed conditions. In addition to basic material, geometric, and loading information, we define a multilevel classification of laminate layups in the datasets. Under the 1st level classification of layup, laminates are classified as symmetric, antisymmetric, or nonsymmetric according to their symmetry with respect to the laminate mid plane, and under the 2nd level classification of layup, they are further subdivided into unidirectional, angle ply, and other types according to their angular composition. Symmetric angle ply (S-AP) and antisymmetric angle ply (AS-AP) laminates, which exhibit different forward prediction performance, are selected for inverse design case studies. Under fixed baseline conditions, angle composition, total ply number, and layup category, exhaustive enumeration and paired model-ensemble evaluation are used to identify candidate sequences with consistent predicted fatigue life improvements. Results from 10 repeated experiments show that introducing layup category information produces only limited improvement in overall performance, whereas clear prediction differences exist among categories. Under the 2nd-level classification configuration, the mean R2 values for S-AP and AS-AP are 0.885 and 0.780, respectively, differing by 0.105. Local inverse design is subsequently performed for these two categories with different forward prediction performance. The recommended S-AP and AS-AP sequences exhibit predicted fatigue life improvements of 24.90% and 14.83%, respectively, relative to their corresponding baseline predictions, with both sequences yielding longer predicted fatigue lives than the original layups. In these two cases, the category with better forward prediction performance also exhibits a larger predicted improvement and greater consistency across models. These results demonstrate the value of layup classification for analyzing category-dependent differences in prediction performance and the computational feasibility of local layup screening.