Prediction of compaction parameters of compacted soil using LSSVM, LSTM, LSBoostRF, and ANN

The present research introduces a robust approach for predicting the maximum dry density (MDD) and optimum moisture content (OMC) of compacted soil by comparing models based on least-square support vector machine (LSSVM), long short-term memory (LSTM), least-square boost random forest (LSBoostRF), and artificial neural network (ANN) approaches. For this purpose, the database of 243 soil samples (190 training + 53 testing) has been used to train and test the developed models. For the measurement of the training and testing performance, the root-mean-square error (RMSE), mean absolute error (MAE), correlation coefficient ( R ), variance accounted for (VAF), performance index (PI), a20 index, index of agreement (IOA), and index of scatter (IOS), statistical parameters have been utilized. The performance comparison demonstrates that models based on the LSSVM approach have gained the highest performance in predicting MDD (RMSE = 0.0171 g/cc, R = 0.9920, MAE = 0.0125 g/cc, VAF = 98.41, PI = 1.95, a20-index = 100.00, IOA = 0.9446, and IOS = 0.0093) and OMC (RMSE = 0.6070%, R = 0.9936, MAE = 0.5294%, VAF = 98.49, PI = 1.37, a20-index = 100.00, IOA = 0.9357, and IOS = 0.0450) of compacted soil than LSTM, LSBoostRF, and ANN models. This study also reports that the Adam-optimized LSTM model performs better than the RMSProp LSTM model in predicting the compaction parameters of soil. Also, the number of leaves affects the performance of LSBoostRF models. The performance comparison of ANN models depicts that selecting suitable hyperparameters such as backpropagation algorithm, number of hidden layers, and neurons is necessary to achieve better prediction. However, Levenberg–Marquardt algorithm-based ANN model (single hidden layer interconnected with 5/10 neurons) has predicted compaction parameters better than other ANN models. The score analysis has been performed to confirm that models based on the LSSVM approach are the best architecture model for predicting MDD and OMC. The Monte Carlo sensitivity analysis demonstrates that fine content, sand content, liquid limit, and plasticity index are the most influencing parameters in predicting the compaction parameters of soil.