| description abstract | The boundary conditions are critical for analyzing the behavior of the deck slabs in girders. However, the boundary conditions of slabs are hard to predict due to complex influencing factors. Recently, machine learning (ML) has been extensively used in structural analysis, while the lack of interpretability prevents their practical application. This paper innovatively proposes a framework that integrates analytical models with interpretable ML techniques, aiming at explainable prediction of bridge deck slab boundary conditions. The rotational and lateral restraint stiffnesses are each divided into two parts. Analytical models tackle analytically solvable structural behaviors, while ML models address complex behaviors, complemented by interpretive approaches to ensure prediction reliability. Ultimately, by integrating the solutions of these two parts, a prediction model for slab boundary conditions can be established. Taking the reinforced concrete deck slabs in steel–concrete composite box girders as an example, analytical models are established for restraint stiffness under simplified conditions. For restraint stiffness considering spatial effect and material nonlinearity, a validated finite-element model is employed for parametric analysis to construct the data set. Subsequently, three ML models are utilized to predict this part of restraint stiffness, incorporating three interpretability approaches to guarantee model reliability. After a comprehensive assessment of both interpretability and accuracy, the optimal ML models are integrated with the analytical models, creating the interpretable prediction models for boundary conditions of deck slabs in composite box girders. Examples and evidence demonstrate that the combination of theoretical analysis and artificial intelligence can effectively improve the reliability of the entire algorithm when considering the spatial effect of the structure as well as the nonlinearity of the material, providing a new perspective and method for calculating boundary conditions. | |
