Physics-informed graph surrogate for spatial layout optimization of structural components considering collapse safety

Optimizing the spatial layout of discrete structural components requires iterative non-linear time–history analysis (NLTHA) within a vast combinatorial design space to evaluate seismic performance indicators, such as the maximum inter-story drift ratio and collapse margin ratio (CMR). However, this process often has prohibitive computational costs. To support seismic performance-oriented component layout optimization, this paper proposes a multi-fidelity physics-informed surrogate framework. A hierarchical multi-fidelity physics-informed neural network is used as the main predictor and utilizes low-cost physical information, such as linear elastic analysis and non-linear static pushover analysis, to provide structural priors for predicting high-fidelity NLTHA indicators. A mechanics-anchored variational graph autoencoder is employed to map discrete layout topologies into a searchable continuous latent space, thereby enhancing cross-configuration generalization. A structural response field adaptive sampling strategy is also introduced to adapt the expensive NLTHA budget to performance-controlled and high-information-gain regions, thereby improving data efficiency. The framework is validated through a numerical case study in which layout variants are derived from a 10-story reinforced concrete frame–shear wall prototype tested at E-Defense, with all surrogate training data generated from parametric OpenSees analyses. The framework can output optimized layout schemes, as confirmed by independent incremental dynamic analysis under a limited high-fidelity verification budget. This study provides a data-efficient surrogate modeling framework for the spatial layout optimization of structural components.