Predicting the UCS of polyhydroxyalkanoate and xanthan gum treated sandy soil using gradient boosting algorithms

• Published in: Journal of cleaner production Vol. 489; p. 144672
• Main Authors: Abbas Jaffar, Syed Taseer, Iqbal, Mudassir, Bao, Xiaohua, Jalal, Fazal E., Chen, Xiangsheng
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.01.2025
• Subjects: Biopolymer; CatBoost; cement; compression strength; data collection; granite; Granite residual soil; greenhouses; LightGBM; polyhydroxyalkanoates; prediction; sandy soils; Sustainable materials; UCS; xanthan gum; Sustainable materials; UCS; Granite residual soil; LightGBM; CatBoost; Biopolymer
• ISSN: 0959-6526
• DOI: 10.1016/j.jclepro.2025.144672

In geotechnical engineering, it has been reported that bio-based materials reduce environmental pollutants such as greenhouse gases and heavy metals. However, due to short study periods and inadequate engineering performance verification, bio-treated soil techniques are less reliable than conventional materials (such as cement and lime). Therefore, in this study, two sustainable materials namely Polyhydroxyalkanoate (PHA) and Xanthan gum (XG) biopolymers are utilized to treat the granite residual soils (GRS). For this purpose, an elaborate experimental program was designed to collect an extensive experimental dataset, and the capability of recently developed and powerful algorithms such as Categorical Boosting (CatBoost) and Light Gradient Boosting Machine (LightGBM) was evaluated to predict the unconfined compressive strength (UCS) of biopolymer-treated GRS. The Shapley Additive exPlanations technique has been applied to study the feature significance of the selected variables in the dataset. Experimental results showed that the UCS of biopolymer-treated GRS increased from 521 kPa to 1123 kPa at 90 days of curing. Moreover, prediction results showed that both gradient-boosting algorithms performed well in predicting the UCS of the GRS. However, CatBoost outperformed the LightGBM in terms of performance metrics, explaining approximately 99% of the variability in both training (R2 = 0.99) and testing phases (R2 = 0.99), and thus achieving the lowest mean absolute error (MAE) of 2.11 for training, and 6 for testing data points, respectively. In addition, the curing period has been the most significant feature influencing the UCS property, followed by biopolymer content. This study validates the efficacy of the suggested CatBoost and LightGBM models, and it is recommended that they be employed before laboratory testing and field application to save time and cost.