Predicting the post-fire load-bearing capacity of slab-column structures based on a physics-embedded deep learning algorithm

While reinforced concrete slab-column structures have been widely adopted in practical buildings, this type of structure is vulnerable to brittle punching shear failure under fire. Predicting the post-fire structural capacity is of great importance to accelerate the post-fire retrofitting or replacement. However, the post-fire performance of slab-column structures is influenced by multiple factors and their interactions. This study proposes a physics-embedded deep learning algorithm to predict the load-bearing capacity of slab-column connections after fire. First, 108 scenarios considering the key factors of concrete compressive strength, slab thickness, column width and fire duration were simulated using a validated sequentially coupled thermomechanical simulation approach to form training datasets. Then, a deep learning algorithm was established to predict the post-fire load-bearing capacity of slab-column connections. The theoretical model to calculate the load-bearing capacity based on rigid-plastic virtual work theory was explicitly embedded into the loss function of the proposed model. Results show that the physics-embedded model performed comparably to the purely data-driven with an overall predicted R ² value of 0.97 on the training datasets, and the physics-embedded model yielded lower relative errors in five cases including four of the five extrapolation cases, showing a modest advantage in the selected extrapolation cases. Shapley Additive Explanations and correlation analyses showed that the average slab temperature has the strongest negative correlation of −0.8768 with bearing capacity, and the proposed model prioritized slab thickness over column width as a positive contributing factor. This study provides an efficient and reliable method to predict the post-fire structural performance of the slab-column structures.