Deep Neural Network modeling and analysis of the laboratory compaction parameter of unbound granular materials

Successful field compaction of earth materials depends on accurate laboratory compaction tests. The standardized laboratory compaction test takes time, consumes energy, and can sometimes be affected by instrument/personal errors. There is a need for a quicker and more efficient means of predicting the laboratory maximum dry density (MDD) and optimum moisture content (OMC) of soil. In this study, Artificial-Neural-Network (ANN) along with stepwise multiple linear regression (MLR) was employed to develop deep ANN-models for predicting the MDD and OMC of unbound-granular soil. Factor and correlation analysis were employed for the reduction and selection of predicting variables. The study utilized 1212 soil test results from the Long-Term Pavement Performance (LTPP) database. This dataset consists of reliable and large-volume of laboratory test results for a wide variety of soil-types. Two composite gradation variables (F2 and VF3) were extracted and employed along with Plasticity-Index (PI), Bulk Specific Gravity (BSG) of aggregate, and percent retained on ¾” as predictors. The tested ANN-models performed better with root means square error of 21.36 kg/m³ (R² = 0.99) and 0.33 % (R² = 0.98) for MDD and OMC respectively. Statistical and sensitivity analysis showed that all the selected variables contributed significantly to the performances of the models. The ANN-models were utilized to develop a standalone user-friendly app for the prediction of MDD and OMC. Lower values of F2 and values of VF3 between 60 and 70 signifies gradation with higher MDD and lower OMC. An optimum PI corresponding to higher MDD but not necessarily lower OMC exists.