Application of artificial neural network models for mixture design of surplus soils treated with paper sludge ash-based stabilizer

Abstract

Excavation projects for road, highway, and railway construction generate large amounts of surplus soil. Surplus soils frequently exhibit low strength and high fluidity because of their high water content. Hence, stabilizers are used to enhance their properties, including handling and transportability. Paper sludge ash-based stabilizers (PSASs) have recently been developed as sustainable materials for soft surplus soil treatment in Japan. The stabilizer was produced from paper sludge ash discharged from paper mills. Conventionally, cone index tests are conducted in the laboratory to determine the strength of soil after mixing with PSASs. However, the laboratory mix tests require considerable time and effort. Therefore, in this study, a new approach using artificial neural networks (ANNs) was developed for the mixture design of surplus soils treated with four types of PSASs. Two ANN models were developed: one with a backpropagation algorithm (BP-ANN) and the other with a genetic algorithm (GA-ANN). The BP-ANN model is straightforward yet effective in tackling nonlinear challenges, which makes it the preferred choice in this study. However, it sometimes encounters convergence problems and can become trapped in the local minima. To address these concerns, this study considers the GA-ANN algorithm, inspired by biological evolution, as an alternative. These models were developed to predict the amount of PSAS ((APS)200) required to satisfy a cone index of 200 kN/m2. In Japan, a qc of 200 kN/m2 is used as a reference to determine whether the surplus soils generated from road or railroad tunnel constructions can be transported by dump trucks. The (APS)200 was also calculated using an empirical method (hereinafter called Mochizuki’s method). The performance of the developed models was evaluated not only in terms of accuracy, including the coefficient of determination (R2), but also in terms of errors, such as the mean absolute error (MAE) and mean squared error (MSE). A comparison between the (APS)200 obtained by the BP-ANN model and that from the laboratory mix tests showed a coefficient of determination (R2) of 89.70 % for all datasets, which was higher than that obtained using Mochizuki’s method (R2 = 82.68 %); this suggests that the BP-ANN model can predict the required amount of stabilizer with higher accuracy than the prediction equations proposed to date, even with the same input parameters. Moreover, a comparison between the (APS)200 obtained by the GA-ANN model and that obtained from laboratory mix tests showed an R2 of 91.32 % for all the datasets. In addition, the errors represented by MAE and MSE also showed that the GA-ANN had fewer errors than the BP-ANN model, and both ANN models had fewer errors than Mochizuki’s method. In addition, the GA-ANN model converged faster than the BP-ANN model, indicating that the GA-ANN model can predict the required amount of stabilizer with higher accuracy within a shorter period than the BP-ANN model can.