Application of machine learning to reduce model uncertainty and enable resource-efficient design of flat slabs

The construction sector has a significant share of global CO₂ emissions due to its high material and resource consumption. Since material and resource consumption can be significantly influenced by the structural design of components, strategies to promote resource-efficient and optimised designs are required providing that the target reliability levels of structures prescribed in building design codes (e.g., DIN EN 1990+NA) are strictly satisfied. Structural components are designed in accordance with building design codes, such as DIN EN 1992-1-1+NA. Since the mechanical behaviour of some components can hardly be analytically described—e.g., punching shear of flat slabs without shear reinforcement—ultimate limit state evaluations partly rely on empirical equations. These equations exhibit significant model uncertainty, meaning that the ratio of actual to predicted load-bearing capacity is subject to statistical scatter. In punching shear verifications, model uncertainty has a great influence on the probability of failure. Previous studies have suggested that the model uncertainty of machine learning models can be lower than that of the empirical design equations specified in international codes. In this work, a machine-learning model (Extreme Gradient Boosting – XGBoost) was trained and utilised to predict the punching shear resistance of flat slabs without punching shear reinforcement. This model was employed in a reliability analysis with decoupled load and resistance side. Monte Carlo simulations were used to determine the probability that the actual load-bearing capacity, determined by the machine-learning model, falls below the design resistance according to DIN EN 1992-1-1+NA. The results indicate that predicting the actual load-bearing capacity with a trained model of lower uncertainty leads to a higher reliability level. This higher reliability can enable reductions in safety margins, for example, by lowering partial safety factors, thereby increasing the potential for resource-efficient designs.