Soil stabilization is a technique for improving soil engineering characteristics, such as durability, permeability, plasticity, compressibility, and mechanical strength. It involves combining or mixing different elements with soil to enhance its quality. Soil stabilization is essential for ensuring that soil can effectively support the stresses imposed by structures.
The conventional material used for soil stabilization is Ordinary Portland Cement (OPC), which has been utilized for decades. However, the production of OPC requires significant energy and results in the release of large amounts of carbon dioxide into the atmosphere. To mitigate this environmental impact, an alternative material known as Geopolymer can be used as a substitute for OPC.
Geopolymers are synthetic alkali aluminosilicate materials produced by reacting solid aluminosilicate with a hydroxide-silicate solution or concentrated aqueous alkali hydroxide. They can be created using solid aluminosilicate derived from various industrial waste products, including silicate and/or alumina. Research has shown that geopolymer concrete emits 9% less CO2 compared to conventional concrete.
Several studies have explored different machine learning methods, including artificial neural networks (ANN), support vector machines (SVM), M5 trees, and random forests (RF), to predict the unconfined compressive strength (UCS) of various materials like lime sludge, stabilized pond ash, and geopolymer-treated clayey soil. While these methods have shown promising results, researchers have found that random forests are particularly effective and widely used.
However, there is still a gap in research, as not many studies have applied this approach to predict the UCS of clayey soil that has been stabilized using geopolymer techniques. Typically, to determine essential
soil properties like UCS and how well the soil can be compacted, multiple tests (around four to six) need to be conducted, which can be quite labour-intensive.
To simplify this process, a recent study by Husein Ali Zeini et al., from Al-Awsat Technical University in Najaf, Iraq, published in the Sustainability Journal in January 2023, aimed to use a random forest model to predict compaction parameters, UCS, and other soil properties. This research highlights the potential of machine learning models, like random forests, to improve civil engineering practices by making soil property predictions easier and more efficient.
The study focuses on optimizing the hyperparameters of a Random Forest (RF) model using grid search techniques. In this analysis, a dataset comprising 283 samples was utilized, of which 70% were designated for training and 30% for testing. The input features included various critical parameters, such as the percentage of ground-granulated blast-furnace slag, plasticity index, alkali-to-binder ratios, fly ash percentage, molar concentrations of an alkali solution, and the ratios of silicon to aluminium and sodium to aluminium.
The evaluation metric employed was R-squared, with the RF model achieving a value of 0.9757, indicating a high level of predictive accuracy. Furthermore, the RF model was benchmarked against several previously developed models, including multivariable regression (MLSR), multi-gen genetic programming (MGGP), multiple variable regression (MVR), and logistic regression (LR). The results demonstrated that the RF model significantly outperformed these alternatives.
The evaluation of feature importance was conducted through SHAP analysis, which highlights several key contributors to the prediction of Unconfined Compressive Strength (UCS). Notably, ground-granulated blast-furnace slag percentage (%S), plasticity index (PI), ratios of sodium to aluminum, and molar concentrations of the alkali solution (M) were found to have significant impacts. On the other hand, the percentage of fly ash (FA%), alkali-to-binder ratio (A/B), and silicon-to-aluminum ratio (Si/Al) demonstrated less influence on UCS.
The findings from the SHAP analysis indicate that an increase in the Plasticity Index (PI) correlates with a decrease in UCS. This reduction is attributed to alterations in the ratio of polymer emulsion to bentonite. A higher PI leads to decreased stiffness and peak strength of the soil while enhancing its ductility. In contrast, an increase in the concentration of NaOH, which plays a crucial role in the geopolymerization process, is associated with improved UCS, although this may affect the workability of the mix due to the heightened solubility of aluminosilicate. Furthermore, elevated values of M and A/B have been linked to enhanced UCS in the context of geopolymer-stabilized clayey soil.
In conclusion, this study represents a significant step toward revolutionizing how we understand and predict soil properties. By focusing on the nuances of penetration and compaction tests, and by incorporating a broader range of data that includes the effects of time and material composition, researchers are paving the way for more accurate assessments. These advancements could not only enhance soil analysis but also have far-reaching implications for agriculture, construction, and environmental management. The future of soil science is bright, and the potential for improved predictive capabilities is truly exciting.
Reference
Zeini, H. A., Al-Jeznawi, D., Imran, H., Bernardo, L. F. A., Al-Khafaji, Z., & Ostrowski, K. A. (2023). Random forest algorithm for the strength prediction of geopolymer stabilized clayey soil. Sustainability, 15(2), 1408.