Reanalysis of Reinforced Concrete Frames via a Three-Layer Machine Learning Framework: Sensitivity-Based Features and Model Interpretability

Structural reanalysis involves repeated evaluation of structural responses under iterative design changes. It is a major computational burden in structural optimization, sensitivity analysis, and health monitoring. The three-layer architecture, which comprises the stiffness, displacement, and force layers, is motivated by the governing structural mechanics relationship F=K·U, which establishes stiffness and displacement as natural intermediate quantities for predicting internal forces. This physics-informed hierarchy reduces dependence on large training datasets while preserving predictive accuracy across all response quantities. The framework predicts member-level stiffness statistics, nodal displacements, and internal forces through three sequential layers: stiffness, displacement, and force. Each layer enriches the feature set of the layer above. Sensitivity-based secondary inputs are derived analytically from closed-form expressions relating cross-sectional dimensions to stiffness and displacement changes. This embeds structural mechanics knowledge directly into the feature engineering process without additional analyses. Member stiffness matrices are recovered as submatrices of the global stiffness matrix, encoding inter-member structural context into each member’s representation. The framework is implemented on a six-floor, three-bay reinforced concrete frame of 42 members. Training uses 1890 data points from 45 finite element iterations. The Random Forest model achieves R² scores of 0.99, 0.98, and 0.91 for axial force, shear force, and bending moment, respectively, on unseen validation data. Once trained on 45 FE iterations, the framework evaluates any number of candidate cross-sectional configurations in a single batch inference pass, enabling a shift from sequential solver-driven reanalysis to model-driven batch optimization. The proposed framework offers a scalable, interpretable, and physics-consistent alternative to both classical reanalysis methods and purely data-driven models, with direct applicability to structural size optimization and structural health monitoring workflows.