Data-driven shell model for progressive damage analysis of composite aerostructures

Modeling progressive damage in laminated composite aerostructures is computationally demanding due to the explicit definitions of unidirectional lamina required to apply failure criteria and the numerous iterations needed to capture nonlinear behavior, especially for large-scale models. To address this challenge, the primary novelty of this paper lies in the development of an explicitly differentiable data-driven section-level shell constitutive model aimed at accelerating such analysis. In the offline phase, mesoscale 3D solid unit cells subjected to rigorous shell deformation boundary conditions are utilized. This strategy extracts 3D laminate's nonlinear responses at the mesoscale and homogenizes them into a macroscopic shell representation. Subsequently, artificial neural networks (ANNs) are employed to construct a regression model for the generalized stress–strain relationship and a classification model for damage states. The training efficiency is significantly enhanced by a Boosted Region Adaptive Sampling (BRAS) strategy, which autonomously densifies sampling data exclusively in highly nonlinear regimes. Crucially, to guarantee implicit nonlinear convergence during online simulations, the analytical Jacobian matrix is derived from the ANN via the chain rule. Implemented into a user subroutine, this framework allows instantaneous section-level updates. Validation on a representative omega-section stiffened laminated plate, through a like-for-like comparison with established 3D finite element reference models, demonstrates that the proposed framework retains predictive fidelity while substantially reducing computational costs.