Advanced machine learning prediction of the unconfined compressive strength of geopolymer cement reconstituted granular sand for road and liner construction applications

The construction of flexible pavement foundations (subgrade and subbase) and landfill liners has been the cause of serious research in recent decades. In this research work, machine learning predictions have been proposed for the evaluation of the unconfined compressive strength (UCS) of geopolymer cement (GPC) reconstituted granular sand with 149 entries revised from 283 entries due to redundancy. The relations between the measured and predicted values, impact of each input on the UCS values and the best fitting of each developed model, and the comparison of the accuracies of the developed models using Taylor charts were conducted. The results show that the outliers beyond the envelops of the ± 25% are more with the GP model with a parametric line fit of y= 0.965x, R2of 0.925, MAE of 1.15 MPa, RMSE of 1.51 MPa and MSE of 2.27 MPa, while the EPR has a parametric line fit of y= 0.986x, R2of 0.944, MAE of 0.98 MPa, RMSE of 1.33 MPa and MSE of 1.78 MPa and lastly the decisive model, the ANN shows a parametric line fit of y= 0.998x, R2of 0.992, MAE of 0.39 MPa, RMSE of 0.52 MPa and MSE of 0.27 MPa. The impact of the Na/Al and the binder on the UCS model showed remarkable agreement with the models especially the ANN and GP models depicted on the impact curves. The Taylor chart agrees with the error indices and performance of the ANN as the decisive model even though the ANN can only be applied in an intelligent interface since it did not produce a closed-form parametric equation. the performance accuracies of the three models are compared with other previous works that have used the same global database to predict the UCS, which deployed the random forest (RF) with R2of 0.985, MGGP with R2of 0.924, MLSR with R2of 0.788, MVR with R2of 0.817 and the ANN with R2of 0.989. These previous works reported models based on the 283-entry database without recourse to the redundant entries in the database. In the present research work, the redundant entries were eliminated to improve the speed and performance of the models. Also, revising the original database deployed previously showed that the maximum UCS value is about 30 MPa, but assume that the measuring precision is about 1% (0.3 MPa), that means a measuring error more than 30% for the samples with UCS less than 1.0 MPa, and that is why these samples were excluded from the study due to redundancy effects to have total of 149 active entries. This procedure has given advantage to the present models, which perform better than the previous models. This agrees with previous research results as binder and Na/Al ratio theoretically increases the unconfined compressive strength of the granular sand beyond 200 and 150 MPa required for subgrade and sub-base in pavement foundations and landfill liner, respectively, and also improves the permeability coefficient to below 1E−9 m/s required for an efficient landfill construction and performance. Understandably, the liquid limit of granular sands plays a role in its performance in soaked conditions.