CatGrass: a feature engineering framework to forecast the seismic response of low-rise RC frames using CatBoost algorithm integrated with grasshopper optimization

Abstract Predicting the maximum drift ratio ( MDR ) of reinforced concrete frames is a critical parameter in the design, evaluation, and rehabilitation of reinforced concrete structures. This study aims to investigate the prediction of the MDR of regular and irregular low-rise reinforced concrete frames using the grasshopper optimization algorithm (GOA) optimized CatBoost algorithm (i.e., CatGrass). For this purpose, a total of 2,300 ground motion records were utilized, and the dataset includes 20 different intensity measures (IMs), spectral acceleration values at the first mode period S a (T 1 ) , as well as the results of nonlinear time history analyses for each record. Also, several feature engineering techniques (advanced ML pipeline) were applied including outlier detection using the multivariate outlier detection and replacement method (i.e., outForest ), missing data imputation with the remove missing values function (i.e., na.omit ), data transformation using the Box-Cox method, and feature selection via the Chi-square method. To evaluate model performance, five statistical evaluation metrics, namely R 2 , RMSE, MSE, MAE, and MAPE were employed. For comparison, the prediction results of the CatGrass models were evaluated against those of the raw CatBoost model (benchmark ML pipeline) and some popular ML algorithms, which do not use any feature engineering techniques. The results demonstrated that the proposed CatGrass model achieved a higher R 2 value (exceeding 0.99) and lower MSE, RMSE, MAE, and MAPE values compared to the raw CatBoost model and the other ML models in predicting the MDR. This study confirms the effectiveness of the proposed novel CatGrass model, establishing it as a computationally efficient, time-saving, and reliable tool for seismic design and assessment of RC frames.