Neural networking to model and predict properties of stabilized road base designs.

摘要

Expansive soils have been troublesome for numerous years in the construction of highway projects; however, the utilization of pozzolan additives to stabilize them have been proven successful time and time again. Typically, laboratory testing is needed on each project to determine the type and amount of pozzolan additives. This ends up being cumbersome, time consuming, and costly for state agencies and contractors. This study investigated the use of multilayer perception (MLP) neural networks (NN) to predict the laboratory density and unconfined compressive strength of soil stabilized designs. It was hypothesized that a NN model would have the ability to take multiple data sets from different laboratories to generate a workable prediction model capable of determining the characteristics of any soil mix design.;Common A-6 and A-7 soil types were used in this study to generate a proof of concept NN prediction model. There were 186 data points collected from various state agencies and research using Class C fly ash and lime additives. Independent samples were tested in a control laboratory setting to validate model outcomes. To eliminate possible fault in the neural networking, comparisons were analyzed by using the NN model of the agencies and control data.;Agency NN models using fly ash appeared to have R values within acceptable ranges for NN training, validation, and testing; however, lime had R values trending lower and modeled poorly. When the control samples were compared to the predicted soil properties, the A-6(14) soil blended with fly ash were within ASTM acceptable ranges; however, all other soil types failed to fall within these ranges. Comparisons between agency and control NN models indicated no significant differences between R values.;Results indicated it was possible to create a prediction model using NN that were as accurate and timelier than laboratory testing, but this was not possible using various or multiple data sources. This can be explained due to differences in mineralogy, pozzolan source, operator, and laboratory testing.