Spatial Interpolation of SPT with Artificial Neural Network

Abstract

In large infrastructure projects, initial geotechnical investigation is conducted at large spacing (~ 100m to 250m), in which SPT is the common test performed while dynamic tests are limited in number. The preliminary planning and design of the buildings are performed based on this information. Hence, estimate of dynamic properties of soil (say, shear wave velocity) at building locations becomes necessary. This can be performed by estimation of SPT at building locations, by interpolation from borehole locations, and thereafter using correlation expressions for estimating shear wave velocity at building location. Interpolation of SPT has been handled earlier in literature with statistical and geospatial techniques. In this article, an artificial intelligence technique, namely, artificial neural network (ANN) is explored for addressing this problem. ANN allows multiple degrees of freedom to data and optimizes weights and biases of the network to yield the best possible estimates of the desired output, in this case, the SPT at intermediate locations. ANN is known to be robust in handling data with noise and thus would be suitable for this application. Five neighbouring points were found suitable for efficient and accurate spatial interpolation of SPT using ANN with two to three neurons in one hidden layer. The performance was very good (correlation higher than 0.9 and errors lower than 2) and better than the geo-statistical approaches reported in literature (correlation lower than 0.9 and errors higher than 6). Within the limits of the study, the number of degrees of freedom (varying from 9 to 37) of the ANN did not affect its generalization capability.