Bearing capacity prediction of open caissons in anisotropic clays utilizing a deep neural network coupled with a population based training approach

• Developing a deep neural network (DNN) model for predicting bearing capacity prediction of open caissons. • Novel limit state solutions of open caissons with the investigation of clay anisotropy. • A novel soft-computing approach by combining DNN with a population based training approach (PBT) to optimize machine learning parameters. • Examining DNN-PBT performance by comparing with other machine learning models and also investigating importance score of each input feature. Open caissons are increasingly utilized for underground construction due to the increasing demand for aboveground structures, which employ the principle of submersion using the self-weight of the edge cutting face and the applied bearing pressure to mitigate the vertical soil reaction. This paper examines the bearing capacity factor of the edge cutting face in anisotropic clays, approximated using the finite element limit analysis (FELA) method and considering the average results between the upper and lower bounds. The influence of the adhesion factor at the interface of the cutting edge ( α ), the ratio between the depth of the internal embedment and the embedded width ( H/B ), the ratio between the radius and the embedded width ( R/B ), the anisotropic shear strength ( r e ), and the cutting face angle ( β ) is investigated. The results indicate a significant influence of the anisotropic shear strength on the adhesion factor at the interface of the cutting edge. An increase in r e denotes a decrease in the undrained shear strength obtained from the triaxial compression test, resulting in an increase in the value of N . An increase in α influences β , such that when β <90°, the value of N remains constant when β = 90°. In addition, a highly efficient hybrid model called DNN-PBT was established utilizing a deep neural network (DNN) and a population based training (PBT) approach, specifically for the purpose of accurately predicting the bearing capacity factor of circular open caissons positioned in undrained clay. Both computational and comparative outcomes demonstrate that the proposed DNN-PBT can precisely forecast the bearing capacity, achieving an R 2 value higher than 0.999 and a mean squared error (MSE) <0.007. These findings highlight the accuracy and efficiency of the suggested approach. Furthermore, the sensitivity analysis results demonstrated that the anisotropic shear strength ( r e ) is the most important input variable for estimating the bearing capacity factor of the edge cutting face.