Interpretable Machine Learning Based Prediction of Bearing Capacity of Rectangular RC Columns

Reinforced concrete (RC) columns are the main stress components in reinforced concrete structures; accurately obtaining the yield load and ultimate load of RC columns under low cyclic load can guarantee the safety of reinforced concrete structures and improve structural stability and reliability. Firstly, 160 sets of test data of RC columns were collected. Five machine learning algorithms, namely, Random Forest, AdaBoost, GBRT, XGBoost, and CatBoost, were used to establish the yield load and ultimate load prediction models, and the Bayesian optimisation algorithm was used to optimise the hyper-parameter combinations of the prediction models, to obtain the five prediction models under the optimal hyper-parameter combinations. Then, different evaluation indexes are used to compare the prediction results of the optimised prediction models horizontally, and the empirical formula is used to calculate the ultimate bearing capacity of the test set sample data; the calculated results are compared with the prediction results of the optimal model. Finally, the SHAP method gives the global and local interpretations of the optimal prediction models. The results show that among the five machine learning prediction models, the CatBoost algorithm-based prediction model outperforms the other prediction models and empirical formulas in four evaluation indexes, R2, MAE, MAPE, and MSE, in yield and ultimate load prediction, and R2 reaches 0.9347 for the yield load prediction model and 0.9382 for the ultimate load prediction model, the two prediction models have higher overall prediction accuracy. Among all the essential characteristics, the shear-to-span ratio, section height, axial compression ratio, and section width have a more significant influence on the predictions of yield and ultimate loads; the results of the study can provide a reference for the calculation of yield and ultimate loads of rectangular RC columns under low-cycle reversed loading.