Resilient Modulus Prediction from Regression and Machine Learning Algorithms

Laboratory determination of resilient modulus (MR) is time consuming and requires expensive equipment and extensive expertise. This has led researchers to develop empirical correlations for MR based on stress state, index, and engineering properties of the soil. In the current study, subgrade soils were collected from 10 road construction projects covering a range of soils (A-1 to A-7-6; untreated to chemically treated) found across Indiana. These soils were tested in the laboratory, and attempts were made to establish correlations between select soil properties such as Atterberg limits, percentage of fines, compaction parameters, and MR. A review of existing models as well as model verification was conducted for Indiana soils. For untreated specimens, the results showed a weak dependence on deviatoric stress and confining stress, while for treated specimens, it was observed that the model parameters increased with treatment, consistent with an increase in MR. It was also found that machine learning tools, specifically the gradient boost regressor model, provide better predictions than traditional regression analyses.