Shallow-neural-network Optimization for Predicting Plasticity Index of Loess with Cone Penetration Test Data

Abstract

Plasticity index is essential for engineering applications, obtaining which would be carried out from situ-fields to the laboratory costly and time-consuming. Cone penetration tests (CPTs), fast, low-cost, reliable and output near-continuous measurement, are widely used in geological and geotechnical engineering, and shallow neural networks can learn and build models of complex nonlinear relationships. This paper presents a methodology of predicting soil plasticity index by CPT using optimized artificial neural networks (SNNs) for reducing laboratory work that represents a significant saving of both time and money. Gathered from fields in Western Henan province in central China, 237 sets of laboratory results and CPT tests divided into 20 groups were used to train, test, and validate the optimization ANN models with single and double hidden layers. A criterion ensuring without underfitting or overfitting is set up by regression coefficient distribution. The optimization covers 12 train functions, four process functions, divide functions and divide models, 2 to 20 neurons selected for two hidden layers. Of the results with double hidden layers, the largest minimum and 2-norm regression coefficients and the least maximum and 2-norm mean square errors are 0.640, 1.318 and 0.775, 1.078 individually, which distinctly larger than the corresponding values in with a single layer, thus indicates improved performances. The influence on the regression values and MSEs is presented.