Applying Hybrid Deep Neural Networks With Manually Derived Layers to Model Compression Strength of Angled CFRP in Aerospace Industry

Abstract This study models the impact of ply orientation and stacking sequence on open-hole compression strength of aerospace-grade composite laminates. Innovating from traditional modeling techniques, this study combines physics-based analytical solutions with data-driven methods to study composite structures. The proposed hybrid model incorporates interply compatibility theory as a manually derived hidden layer in the Multilayer Perceptron architecture, and achieves root mean square error (RMSE) of 20.1 MPa (6.2% average failure strength) and R2 of 0.85 on experimental test records. The hybrid model outperforms either physics-based analytical model or data-driven models when used individually. In addition, it offers enhanced physical interpretation by validating the importance of interply compatibility in laminate failure. The success in modeling non-standard ply angles, not just (0°, ±45°, 90°), and multiple stacking sequence patterns unlocks the potential of designing composite laminates with a much larger design space. The proposed model could facilitate stacking sequence optimization to maximize strength for given laminate compositions.