Fig 1.
Correlation plot of input variables and output variables of cohesive soils stabilized with geopolymer.
Table 1.
Statistical description of all dataset.
Table 2.
Statistical description of training dataset.
Table 3.
Statistical description of testing dataset.
Fig 2.
K-fold cross validation diagram.
Fig 3.
Methodology of this study.
Fig 4.
Influence of hyperparameter value on fitness function R2 value: (a) Number of estimations vs Max depth, (b) Max features vs Min samples split for RF model; (c) Neuron number of hidden 1 vs Neuron number of hidden 2 with solver LBFGS, (d) Neuron number of hidden 1 vs Neuron number of hidden 2 with solver ADAM for ANN model.
Fig 5.
Influence of hyperparameter value on fitness function R2 value for: (a) Number of estimations vs Learning rate and (b) Max depth vs Min split loss for XGB model; (c) Number of estimations vs Max depth, (d) Max features vs Min samples split for GB model.
Table 4.
Optimal hyperparameters of ML models.
Fig 6.
Experimental UCS of geopolymer vs predicted UCS based on (a) RF model, (b) GB model, (c) XGB model and (d) ANN model.
Table 5.
Performance value including R2, RMSE and MAE of ML models.
Fig 7.
Taylor diagram for comparing UCS predicted by ML models and experimental UCS of geopolymer in (a) training dataset and (b) Testing dataset.
Fig 8.
Global interpretation of two best ML model-based Shapley additive explanations.
Table 6.
New performance values value including R2, RMSE and MAE of XGB and ANN models.
Table 7.
Comparison of performance values and database between this study and other investigation of literature.
Fig 9.
Global interpretation of ANN model-based Shapley additive explanations using the new database.
Fig 10.
Local interpretation of ANN model-based Shapley additive explanations using the new database.
Fig 11.
Individual interpretation of ANN model-based Shapley additive explanations using the new database for three specific cases.