摘要
Modeling cross-scale failure processes remains challenging due to the high computational cost associated with size effects and the need to calibrate macro-scale damage models from micro-scale fracture mechanics. In this work, we propose a model-based neural network (MBNN) that serves as a micro-scale informed constitutive surrogate for macro-scale damage simulations. Within an energy-based homogenization framework, the approach links micro-crack evolution to continuum-scale damage variables while embedding mechanics-consistent constitutive laws into the network architecture, resulting in sparse and physically interpretable parameters. Damage evolution is formulated as a joint model-selection and parameter-identification problem, enabled by a differentiable Gumbel-Softmax mechanism that automatically selects appropriate constitutive and damage behaviors from RVE-level fracture data. Compared with conventional deep neural networks, the proposed framework exhibits improved robustness, interpretability, and extrapolation behavior. The MBNN is validated on multiple cross-scale failure benchmarks, demonstrating favorable accuracy and computational efficiency. Applications to femoral bone damage further illustrate its applicability to biomechanical problems. In addition, an extension of the MBNN to progressive damage modeling is investigated and validated using experimental data, demonstrating its ability to capture gradual degradation, thereby extending the framework beyond ideal brittle fracture.