Use of machine learning in determining Gmax from bender element tests

Imperial College London, Civil and Environmental Engineering, Imperial College Rd, South Kensington, London SW7 2BB, UK

^* wenzhang.xu21@imperial.ac.uk, truong.le@imperial.ac.uk

Abstract

The use of bender element is one of the most popular methods of determining shear wave velocity, and hence elastic shear modulus due to its relatively straightforward experimental set-up. While several analysis methods have been proposed, manual interpretation using the first arrival continues to be favoured owing to its simplicity. This paper presents a novel automated program for determining the shear wave velocity and associated maximum shear modulus. The proposed new method involves the use of Convolutional Neural Networks (CNNs) to predict the most probable shear wave velocity using a range of input frequencies as the inputs. Estimates made by the trained CNN are compared to values determined using more traditional interpretation methods (first-arrival, cross-correlation and frequency domain). The program is able to autonomously determining the shear modulus in the three principal orientations (G_vh, G_hv, and G_hh) at a range of stress levels. The shear modulus determined using the range of techniques showed great agreement. Statistical analysis of the determined shear modulus regression of over 0.99 between interpretations made using first arrival and that estimated using the new CNN approach.