Prediction of Tunnelling-Induced Settlement Trough by Artificial Neural Networks

Tunnelling-induced settlement is usually estimated based on field data. However, the data are representative of the local study area only depending on such parameters as geology setting and tunnel geometry. Moreover, the number of training data samples is also limited. In this study, surrogate models are developed to account for the variation of the tunnel parameters, so that they are representative of many types of conditions. The data is generated with numerical simulations by employing the Hardening Soil Model and considering various stress reduction factors. Exploiting their pattern recognition capabilities, single and multi-output artificial neural networks are trained to predict the maximum settlement and the trough width. The networks employ only 10 features and return very accurate predictions with a coefficient of determination generally higher than 90%. The network architecture, activation functions and weight initialisers are optimised by grid search. The relative importance of the various features is also studied. A computer script is provided to predict the settlement and trough width with custom input data based on the trained networks.