Application of artificial neural networks (ANNs) to predict pressure meter modulus using geotechnical and geophysical parameters

Abstract

Stress-strain behavior of subgrade material controls the performance of the transmission line foundations subjected to lateral pressure. In the transmission line industry, this concept is also known as the deformation modulus. Lateral shear modulus at intermediate strains (GPMT), an equivalent elastic modulus for lateral loading, is provided by the pressure meter test (PMT). GPMT mimics reaction of transmission line foundations within the elastic range of motion. Pressure meter modulus, EPMT can be calculated from Degradation factor, GPMT/G0. But due to project cost, PMT is rarely performed. Correlations have been developed between pressure meter modulus and SPT_N (blow counts) as well as other index properties but existing correlations have high variability and may result in erroneous foundation design. In this study, an Artificial Neural Networks (ANNs) approach was used to explore the relationships and increase the statistical accuracy of existing correlations. Different geotechnical, geophysical parameters and GPMT/G0 data collected from the literature were used for this research. Some observations had partially missing variables which had been retracted using the QUERY Method developed by one of the authors. Two models were then developed using all the datasets. Model 1 included depth, Standard Penetration Test (SPT), unit weight, P-wave velocity and S-wave velocity of the dataset as inputs and GPMT/G0 as output. For Model 2 depth, P-wave velocity and S-wave velocity were considered as inputs and GPMT/G0 as output. Based on the minimum ASE (Average Square Error) and MARE (Mean Absolute Square Error) and maximum R2(Co-efficient of Determination) values, the ANN structure was chosen for each model. A GUI (Graphical User Interface) was created for each model used to predict GPMT/G0 for different inputs and sensitivity analysis was performed between GPMT/G0 and depth for both models. From the statistical measures of the models, it can be observed that ANN is capable of predicting the pressure meter modulus more accurately than regression because ANNs approach is efficient in exploring nonlinear trends.