The use of feed-forward and cascade-forward neural networks to determine swelling potential of clayey soils

Clay soils can exhibit excessive swelling due to changes in water content. The clay swelling threatens the longterm stability of structures and foundations, thus an accurate prediction of clay swelling properties is essential in many geotechnical projects. We present feed-forward and cascade-forward neural network models trained with the Levenberg-Marquardt and Bayesian optimization algorithms to determine the swelling potential of natural and artificial clayey soils. The compiled experimental dataset includes various types of soils covering a wide span of swelling potential, ranging from 0.01 to 168.6%. The activity, water content, dry unit weight, liquid limit, plastic limit, plasticity index, and clay content were considered as the input parameters of the models as they are commonly measured during the experimental testing of soil behaviour. The results show that the feed-forward neural network trained with the Levenberg-Marquardt algorithm is the most accurate model for the prediction task. The performance of the model is satisfactory, showing an acceptable agreement with experimental data. The developed model showed substantial improvements over previous empirical and semi-empirical correlations in determining the swelling potentials of both natural and artificial soils.