Rapid damage state classification for underground box tunnels using machine learning

Abstract

This study develops and compares the performance of eight machine learning (ML) models to rapidly predict the seismic damage state of underground box tunnels. Nonlinear time history analyses of 24 soil-tunnel configurations subject to 85 ground motions were performed to generate the dataset for the ML models. The aspect ratio, buried depth, flexibility ratio, and 23 ground motion intensity measures (IMs) are employed as input variables of ML models. The output variables are four damage states, namely ‘none’, ‘minor’, ‘moderate’, and ‘extensive’. Among the eight ML models, LightGBM is found to yield the most favorable prediction of the damage states, resulting in an accuracy of 91%. The effects of earthquake IMs were also examined. Results show that the spectral acceleration () and spectral displacement () at the fundamental period of the site (T 1) have the strongest correlation with the damage prediction. Finally, the effect of reducing the input variables to two groups (i.e. combinations of soil-tunnel configuration parameters with top five and top ten ranked IMs) on the model prediction capability was investigated. Accordingly, ), acceleration spectrum intensity, spectral velocity, and velocity spectrum intensity were identified as the key parameters representing the ground-motion characteristics needed for the predictive model.