Comparison of deep neural network (DNN) and recurrent neural network (RNN) for estimation of california bearing ratio of soils

Abstract

This study applies deep learning models to predict the California Bearing Ratio (CBR) of road subgrade soils. The dataset consists of 214 soil samples, each characterized by ten geotechnical input parameters, including gravel content, coarse sand content, fine sand content, silt–clay content, organic content, liquid limit, plastic limit, plasticity index, optimum water content, and maximum dry density, with CBR as the output variable. Two deep learning models, the Deep Neural Network (DNN) and Recurrent Neural Network (RNN), were developed to model the nonlinear relationship between soil properties and CBR values. The evaluation results indicate that the DNN model outperforms the RNN model on both the training and testing datasets. Specifically, on the testing set, the DNN achieved R² = 0.961, RMSE = 1.791, and MAE = 1.319, whereas the RNN achieved R² = 0.882, RMSE = 3.121, and MAE = 2.059. In addition, supplementary analyses using Taylor diagrams, Partial Dependence Plots (PDP), and Individual Conditional Expectation (ICE) demonstrate that the DNN model not only provides higher prediction accuracy but also exhibits greater stability and better interpretability. These results highlight the strong potential of deep learning models, particularly DNN, for supporting rapid and reliable CBR prediction in road subgrade design and evaluation.