Practical model for flexural capacity of corroded PC structures based on interpretable machine learning

This study proposes a machine learning algorithm-based model for flexural capacity of corroded prestressed concrete (PC) structures, which considers the feature parameters including geometric parameters, material parameters and corrosion parameters. First, three black box models are utilized to predict the flexural capacity of corroded PC structures. Then, the sensitivity of feature parameters to the flexural capacity is quantified using the Shapley Additive Explanations (SHAP) additive model, and the model’s key features parameters are extracted. Following this, a practical model is established using Genetic Programming (GP) based on these key feature parameters. The model is verified and compared with the existing models. The results show that the interpretable machine learning model can quantify the influence of the feature parameters on the flexural capacity and obtain the feature importance ranking. The model can fit the calculation formula for the flexural capacity of corroded PC beams by symbolic regression method, which effectively reveal the potential nonlinear relationship between basic feature and flexural capacity. The accuracy and interpretability of the calculation formula are between the existing empirical model and black box models.