Application of Artificial Intelligence for Prediction of Swelling Potential of Clay-Rich Soils

Susceptibility of fine grained soils to swelling and shrinkage problems is crucial for safe design of infrastructure, construction and maintenance. However, quantification of soil response and measurement of their geotechnical properties is time taking, expensive and involves destructive tests. Therefore, dependable forecasting models are necessary that calculate swell percentage from results of quick, inexpensive and non-destructive tests. In this paper, a three-layer feedforward neural network (ANN-TFN) was applied in order to envisage swell percentage of fine grained soils and the results were compared with that of multiple regression (MR). The parameters considered as input were activity, clay, liquid limit, plastic limit, plasticity index and fines while swell percentage was used as output. The best ANN-TFN model demonstrated root mean square errors (RMSE) of 1.529, sum of squares errors (SSE) of 369.3, and coefficient of correlation (R²) of 0.80. MR model displayed 1.756 (RMSE), 487.2 (SSE), and 0.508 (R²). The maximum R² values obtained by simple regression was 0.5. Overall, the established three-layer feedforward neural network models (ANN-TFN 1-6) showed significantly higher prediction performances than either multiple regression or simple regression models. Moreover, the use of Levenberg–Marquardt as training parameter and tan sigmoid as transfer function were noted to be more appropriate for good prediction performance in this problem. Hence, the result of the present study concludes that practice of the ANN-TFN model to determine swell percentage of fine grained soil is a promising approach for increasing the confidence of making accurate decisions during the soil engineering works.