Rapid seismic damage-state assessment of steel moment frames using machine learning

• Eight machine learning models were used to predict damage-state of steel frames. • Performance of eight machine learning models was examined and compared. • Synthetic minority oversampling technique was used to resolve imbalance of dataset. • Effects of input variables were explored using Shapley additive explanations. • The graphical user interface was created for easy access of practical engineers. The damage state assessment of buildings after an earthquake is an essential and urgent task that typically requires significant manpower and time for the resilience of a city-scale society. This study aims to develop machine learning (ML) models for the rapid seismic damage-state assessment of steel moment frames, which was never tried before to the authors’ knowledge. Eight ML models were examined for this purpose, including K-nearest neighbors, naïve Bayes, decision tree, random forest (RF), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), light gradient boosting, and category boosting. The combination of 468 steel moment frames from the database in DesignSafe cyberinfrastructure and 240 ground motions yielded a total of 112,320 data points. The steel moment frames have a wide variety of geometric configurations (e.g., number of stories from 1 to 19, number of bays from 1 to 5, bay width from 6.1 to 12.19 m), and applied loads (i.e., three cases of dead load and two cases of live load). Nonlinear time history analyses were conducted using OpenSees to produce a comprehensive dataset for the training and testing of the ML models. A reliable procedure to define the damage states of steel moment frames was suggested based on pushover analysis. Damage states of steel moment frames were categorized following the tag definitions (i.e., green, yellow, and red) in ATC-20. Spectral accelerations at five selected periods (1, 2, 3, 4, and 5 s) for the given ground motions and at the first three natural periods of the steel frames were used as input variables for the ML models. From the results, the RF model is suggested for the prediction of the seismic damage states of steel moment frames. The RF model could accurately predict 98% of the assigned tags in the testing dataset. In contrast, the AdaBoost (88%) and naïve Bayes (90%) models displayed the lowest performance. Among the four boosting methods considered, the XGBoost model (97%) exhibited the highest performance. Furthermore, Shapley additive explanations (SHAP) method was used to inspect the importance of input variables on the prediction. It was found that the spectral accelerations at 1 and 2 s strongly influence the prediction, likely because the first natural periods of the considered steel frames fall in the range of 1–2 s. Finally, to provide convenient access to engineers, a graphical user interface based on the developed RF model was created. This study places a pioneering step for the application of machine learning to the rapid damage assessment of building structures.