Comparing machine learning and deep learning approaches to predicting the seismic response of slab-column connections

Slab-column joints constitute the most sensitive elements of flat plate structures during seismic forces since they are likely to experience brittle failure. This paper examines the predictive accuracy of an extensive Machine Learning (ML) and Deep Learning (DL) solutions for predicting seismic performance of slab-column connections, such as punching moment (M) and Drift Ratio (dr). The ML models under study are Ridge Regression (RR), Linear Regression (LR), Lasso Regression (Lasso R), Elastic Net (EN), Support Vector Regression (SVR), Gradient Boosting (GB), random Forest (RF), and Extreme Gradient Boosting (XGBoost). The models of DL that were analyzed are Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and a hybrid version that incorporates CNN and LSTM (CNN-LSTM). The analysis shows that in (M) prediction, the performance of GB was the best with a coefficient of determination (R²) of 0.870850, a Root Mean Square Error (RMSE) of 0.378246, and a Mean Absolute Error (MAE) of 0.282686. In the DL models, CNN had the best accuracy using R² of 0.827864, MAE of 0.335228, and RMSE of 0.436680. RF was better than other ML models in dr prediction with an R² of 0.565014, MAE of 0.440262, and RMSE of 0.619478. Once again, CNN performed better than the rest of the DL models, with an R² of 0.470568, MAE of 0.513140, and RMSE of 0.683429.