Explainable machine learning-based probabilistic prediction in T-shaped steel-concrete composite columns

The modular wall light steel-concrete composite frame serves as a load-bearing structure within a novel ultra-low energy prefabricated building system. The biaxial eccentric bearing capacity of T-shaped composite irregular columns is a key factor for the framework’s performance. An approximate method for calculating the biaxial eccentric bearing capacity is proposed in this study, with the error controlled within 15 %. To enhance accuracy and address uncertainties in material properties and geometry, a machine learning approach was applied to develop a probabilistic prediction model for the bearing capacity of these columns. Various machine learning algorithms were utilized, taking into account factors such as eccentric angle, eccentric distance, specimen length, concrete strength , steel strength, and steel thickness, based on a validated finite element model . Among the models, the artificial neural network (ANN7) exhibited the highest accuracy. Through SHapley Additive exPlanations (SHAP) analysis, the eccentric angle was identified as having the most significant influence on biaxial eccentric bearing capacity. A probabilistic model , incorporating material and geometric uncertainties, was developed using Latin Hypercube Sampling (LHS) and was validated by capturing the probability density of numerical simulation results. A global sensitivity analysis (GSA) indicated that uncertainty in thickness had a substantial effect on the bearing capacity of the T-shaped composite columns. For practical application, interactive software (PEMTC) was created, enabling users to predict the probabilistic biaxial eccentric bearing capacity by inputting nominal design and geometric parameters of the steel material.