Modification of Geotechnical Properties of Soil using Waste Plastic Strips and Fly Ash: An Experimental Study with a Machine-Learning-Based Predictive Framework

Abstract

The solid waste management challenges of plastic and coal fly ash disposal, along with the lack of good construction soil, are common in rapidly urbanizing regions. The influence of shredded waste plastic strips (0-2.5% by dry weight) and Class-F fly ash (5% by dry weight) on soil compaction, strength, and bearing capacity characteristics of soil classified as Clayey Sand (SC) was evaluated. Specific gravity, grain size distribution, Atterberg limits, Proctor compaction, unconfined compression (UCC), and California Bearing Ratio (CBR) tests were performed on eight soil blends with the plastic and fly ash additives. Plastic strips between 0 and 1.5% by dry weight increased the soil maximum dry density (to 1.70 g/cc) and soaked CBR values (to 16.79%) relative to the control soil (1.54 g/cc and 8.76% CBR, respectively). Beyond 1.5% plastic strip inclusion, however, the strength parameters began to decrease as a result of the plastic particles disrupting the soil particles’ natural interlock and increasing the soil particles’ void ratio. Similar results were obtained with Class-F fly ash; its inclusion decreased the optimum moisture content of the soil blends by 1-2 percentage points; no changes in strength characteristics were detected with the addition of 5% Class-F fly ash. Furthermore, within each of the eight soil blends (n=41, 150, and 96 points, respectively), the compaction, UCC, and CBR results were modelled using four artificial intelligence methods: artificial neural networks (ANN), random forests (RF), extreme gradient boosting (XGBoost), and support vector regression (SVR). Each of these models were trained using a leave-one-out cross-validation method that prevents information leakage between data points along the same response curve. For the CBR values, the XGBoost model emerged as the best performing algorithm for predicting California Bearing Ratio values (R2=0.86, RMSE=48.3 kg, MAE=24.2 kg, MAPE=14.3%) outperforming the ANN and SVR models. Furthermore, SHapley Additive exPlanations (SHAP) analyses revealed that the predominant influence on CBR values was the penetration depth of the soil samples, followed by plastic strip content in the soil blend; the contribution of Class-F fly ash to the CBR value was less pronounced and less certain due to the evaluation of only a single dosage level of the additive. The results indicate that the best balance of soil properties and cost occur at approximately 1.5% plastic strips without Class-F fly ash; random forest, gradient boosting, and ANN models have the potential to effectively provide predictions of the response curves with no information leakage between soil data points.