Prediction of Subgrade Strength of Black Cotton Soils Stabilized with Cement Waste Dust and Marble Powder Using Artificial Neural Network

Abstract

Highway flexible pavement should be capable of carrying traffic axle loads down to the subgrade soil. In some circumstances, marginal materials may be found along proposed routes, which extend to a depth well below the subgrade level. For instance, Black Cotton (BC) soils are highly expansive clay soil grayish to blackish, resulting in deformation and premature failure of road pavements. Ethiopia's large areas of highland and lowlands, specifically Addis Ababa city, BC soil coverage is abundant. Some stabilizing agents have been used with these types of soils in different parts of the world, including Ethiopia, but the problems persist. California Bearing Ratio (CBR) test is widely used as an index test to assess the strength characteristic of stabilized subgrade. Still, the test is laborious, costly, time-consuming, and complex, subject to erroneous results. Hence, this study focused on investigating the effect of CWD and MP on local soil properties and developing a model for predicting the CBR value of stabilized material from simple identification tests, i.e. (LL, PL, PI, OMC, and MDD). Laboratory tests were conducted on natural BC soils and stabilized material, then developed a prediction model using the ANN approach to estimate CBR value. The results revealed that CWD and MP could effectively stabilize BC soils using index properties and CBR values as evaluation criteria. The OAC indicated 8% and 18% for CWD and MP, respectively. On the other hand, Multilayer Perceptrons (MLPs) ANN model was utilized to simulate CBR values of modified black clay, and it performed satisfactorily. The mean absolute error (MAE), root mean square error (RMSE), and R2 -value were used as yardsticks and criteria. In the process of neural network development, NN 5-8-1 and NN 5-2-1 respectively for CBR of a CWD- and MP modified BC soils that gave the lowest MSE value were used in the hidden layer of the network's architecture and performed satisfactorily. Hence, the performance of the simulated network was very good, having R-values of 0.9645 and 0.9457 for the CBR of CWD- and MP-modified BC soils, respectively. These values met the criteria conventionally recommended of 0.8 for strong correlation conditions. In addition, a strong correlation was observed between the experimental CBR of a CWD and MP modified black clay soils CBR values as obtained by laboratory tests and the predicted values using ANN.