A hybrid data-driven architecture for forecasting soil thermal conductivity using optimized gradient boosting frameworks

Soil thermal conductivity (k) plays a vital role in optimizing geotechnical and geo-environmental engineering applications. This property is governed by an interconnected set of physical and environmental variables, including saturation percentage, water content, dry density, void ratio, porosity, initial temperature, and soil type, which creates a clear need for high-fidelity predictive frameworks. To address this, the current study implements a Gradient Boosting Machine (GBM) architecture integrated with four distinct enhancement frameworks: BBO, ES, BPI, and GPO. Predictive pipeline is built upon a compiled dataset of 125 experimental observations, utilizing a 90% training and 10% testing partition to map the seven geological and environmental inputs to measured thermal conductivity values. Overfitting is systematically controlled during the model adjustment phase via a robust k-fold cross-validation scheme. Each hybrid method is rigorously evaluated based on its computational execution speed alongside standard statistical metrics, namely R-squared (R 2 ), MSE, and AARE%. Pearson correlation analysis reveals a highly dominant positive relationship between soil thermal conductivity and saturation levels (correlation coefficient: 0.86), followed by water content (0.66) and dry density (0.42). Conversely, moderate negative associations are observed for porosity (-0.48) and void ratio (-0.43), whereas soil type displays a minor positive correlation (0.12) and initial temperature exerts a nearly negligible negative impact (-0.03). Among the evaluated hybrid configurations, GBM-GPO proves to be the most accurate predictive tool, achieving an impressive R 2 of 0.996311 during training and maintaining a superior test set R 2 of 0.9704336. In terms of algorithmic efficiency, the GBM-ES model exhibits the most rapid execution, requiring only 67.6 s compared to the 111.8 s observed for the top-performing GPO variant. Furthermore, interpretability via SHAP analysis establishes saturation percentage as the most vital parameter dictating the model’s output. This is sequentially followed by water content, void ratio, soil type, dry density, porosity, and finally, initial temperature. Ultimately, these data-driven frameworks serve as highly dependable alternatives for determining soil thermal conductivity, successfully minimizing the necessity for costly, slow, and demanding physical laboratory testing.