Determination of the relationship between soil properties and earthquake damage with the aid of neural networks: a case study in Adapazari, Turkey

Abstract

The building stock in the city of Adapazari, Turkey, experienced widespread damage during the 1999 Marmara earthquake. An attempt was made to relate structural damage to the type of subsoil in this study. The Adapazari soil database has been established which contains information on boreholes, cone penetration and laboratory testing since 1996 and is being updated continuously. The database has been organised using a geographical information system software. Several numeric soil profiles across the city were then taken to establish a back propagation neural network model to enable the investigator to estimate probable structural damage by referring to the type of soils at the usual footing embedment depths. Ten cross sections comprising 140 data each were used to form scanlines of 1400 m length. The input for the neural networks were the physical, mechanical and dynamic properties of soils while the resulting damage ratio data formed the target layer. Feedforward, backward spreading networks were employed in modelling. Numeric data for eight cross sections were employed for the learning process, whereas data for two cross sections were used to test the model. The proposed model was found to predict the damage ratios successfully. The general evaluation of the city following the earthquake has shown that the structural damage was minimal in a limited section of the city where the bedrock outcrops. The damage in the flat areas around the outcropping rock covered by lacustrine clays of high and intermediate plasticity was markedly low. However, damage and destruction was obvious in the central parts of the city where liquefaction and cyclic softening cases were abundant.