Study on data-driven inverse identification of structural parameters for large-scale structures

Accurate identification of structural parameters from deformation fields remains a pivotal objective in theoretical research and engineering practice. As inverse-problem methodologies evolve, engineering applications now require parameter identification at unprecedented spatial scales and levels of fidelity. Conventional approaches, however, encounter substantial limitations when applied to large-scale structures, primarily because of the high dimensionality, nonlinearity, and heterogeneity inherent in full-field deformation data. Inspired by multi-scale decomposition principles, this study introduces a segmentation–assembly–optimization (SAO) framework that systematically reduces the complexity of large-scale inverse problems. A lightweight convolutional neural network (CNN) is trained to map local displacement fields onto their corresponding structural parameters; these local estimates are subsequently assembled and refined by a mechanics-driven optimization procedure to reconstruct the global parameter distribution. Comprehensive numerical experiments demonstrate that the proposed framework achieves accuracies exceeding 98 % within the region of interest (ROI) for large-scale structures with intricate geometries, and maintains robust reconstruction accuracy (>90 % under 1 % noise), whereas the standalone CNN performance degrades significantly. The SAO framework thereby overcomes scale-dependent constraints and delivers a reliable, data-driven solution for high-resolution structural identification.