A displacement-decomposed neural model for analyzing the low-velocity impact response of sandwich structures

Low-velocity impact response of sandwich structures is governed by global bending, local indentation, and core deformation mechanisms that are related to different architectures and support conditions. Existing analytical and numerical approaches are often restricted to specific structural configurations or require computationally intensive, high-fidelity simulations. To address these limitations, a displacement-decomposed neural model (DDNM) is proposed to analyze the low-velocity impact response of sandwich structures. The model represents the predicted impactor displacement response through non-negative learned displacement contributions associated with bending/support compliance, local indentation/contact deformation, and core deformation. Mechanics-informed descriptor routing is used to associate structural and impact descriptors with the corresponding displacement branches during the response assembly. A source-audited experimental database containing 171 impact cases covering foam-core, honeycomb/cellular, corrugated, and hybrid sandwich configurations was compiled from published studies. The model was evaluated using repeated random interpolation and complete-source grouped validation to distinguish the interpolation capability from the source-level transferability. Compared with an unconstrained multilayer perceptron trained under identical preprocessing and validation conditions, the DDNM reduced the complete-source grouped-validation RMSE from 13.60 to 12.38 mm and the MAE from 11.11 to 9.98 mm. Tree-based regressors nevertheless achieved lower prediction errors, and the DDNM is therefore not positioned as a state-of-the-art accuracy model. Its contribution is instead the mechanics-structured representation of maximum impactor displacement through non-negative bending/support, indentation/contact, and core-deformation branches, which provides regime-wise displacement-contribution indices while retaining quantitative predictive capability within the heterogeneous literature-derived database.