Prediction of Compaction Characteristics of Soils from Index Test's Results

This paper presents some attempts at prediction of compaction characteristics of soils using the results of the index tests. A data bank, including 728 compaction tests, was compiled. Each case includes the results of soil type, grain size distribution, Atterberg limits (W-L and W-P) and specific gravity of soil particles, as well as the compaction characteristics, maximum dry density and optimum moisture content were calculated under different levels of compaction energy. Using artificial neural networks (ANNs) and multi-linear regression (MLR), the applicability of basic information about soils to estimate the compaction characteristics was evaluated. A sensitivity analysis accomplished on the results of ANN method, demonstrated that fine content has the most pronounced effect on the accuracy of compaction characteristics prediction. Using a trial and error approach and combining the different individual variables, the efficiency of multi-linear regression models were improved. However, the comparisons showed that ANN models are more effective in capturing the correlation among compaction characteristics of soils and their index properties, while the ANN shortcomings, due to their black box nature, make MLR models more useful in prompt estimations.