Hybrid gradient boosting with meta-heuristic algorithms prediction of unconfined compressive strength of stabilized soil based on initial soil properties, mix design and effective compaction

Abstract

In the investigation, hybrid Gradient Boosting (GB) with four optimization techniques named Simulated Annealing (SA), Random Restart Hill Climbing (RRHC), Particle Swarm Optimization (PSO) and Evolution Strategy Optimization (ESO) are used to develop machine learning models for predicting unconfined compressive strength (UCS) of stabilized soils based on initial soil properties, mix design and effective compaction. The results show that the hybrid model GB_PSO can successfully predict the UCS of stabilized soils with high performance R2 = 0.9655, RMSE = 0.1633 MPa for testing part. The feature importance analyses are performed with using GB_PSO based Permutation Importance of Sklearn library and Shapley Additive Explanations (SHAP value), which reveal that maximum dry density (MDD), linear shrinkage (LS), and Gravel content are the most important characteristics in the UCS prediction performance of the GB_PSO model. The partial dependence plot (PDP) analysis consisting of PDP 1D and PDP 2D show respectively the impact way of each feature on the UCS and the interactions between initial soil properties, mix design and effective compaction in the UCS improvement of stabilized soil. The PDP analysis show that (i) the LS is the most important criterion in determining the success of improving the UCS of stabilized soil, (ii) the predominance of effective compaction over other conditions on the improvement of UCS. The present investigation can be considered to be new insights for better understanding process of stabilized soil and UCS improvement.