Machine learning approach in the quantitative evaluation of the seismic behaviour for 3D reinforced concrete frame structures

Abstract

The implementation of performance-based seismic design (PBSD) in the design and optimization of the structural system of buildings becomes paramount in view of recent seismic design codes. The large number of simulations and analyses to be conducted for specific types of structures and specific seismic areas are prerequisites of the PBSD methodology. The application of machine learning (ML) techniques proved effective for the development of prediction models that have the potential to significantly minimize the amount of time required for structural damage assessment. Artificial neural networks (ANNs) demonstrated the ability to generalize by accurately predicting output parameters for unseen input parameters not included in the training dataset. In this research, an ANN was used for predicting structural damage parameters corresponding to 3D reinforced concrete frame structure subjected to seismic scenarios and for evaluating the influence of various seismic actions on the structures. A total of 243 3D-reinforced concrete models were generated and subjected to 14 seismic scenarios. Thus, 3402 input-output data sets were obtained and were used for ANN training (80 %) and validation (20 %). A total of 10 input parameters were considered to influence the seismic behaviour and damage levels in the RC frame structures. Out of the 6 input structural parameters, the number of stories, the span width and the width of the column cross-section have the highest impact on the seismic damage of reinforced frame structures. From the 4 input parameters characterizing the seismic motion, the peak ground velocity (PGV) and peak ground acceleration (PGA), were found to be the most important seismic parameters which influenced the damage of the analysed structures. The performance of the ANN was compared against two other machine learning algorithms commonly used in civil engineering applications: Random Forest (RF) and Extreme Gradient Boosting (XGBoost). While these algorithms performed marginally better than ANN in the training and validation stages, they did not manage to be accurate in the testing phase when using newly generated data sets. The considered output parameters were: fundamental period of vibration for the non-damaged and damaged states, final softening index, interstory drift ratio, maximum displacements and maximum absolute accelerations. The ANN was able to accurately predict all output parameters (correlation coefficient larger than 0.85) with the exception of final softening index that may be influence by more complex phenomena that are beyond the scope of this paper. The proposed ANN-based prediction model proves to be a fast and reliable tool for quickly assessing the damage state of 3D reinforced concrete frame structures subjected to different seismic scenarios. It can be further enhanced and extended to include other parameters not considered at this stage of research as well as being included in stacked ML algorithms.