In a recent publication featured in Scientific Reports, researcher Laila Safi explored the complex thermal and physical dynamics of soil freezing temperatures during the environmental remediation of lead-contaminated soils using biochar amendments. Traditional laboratory techniques utilized for observing subzero ground behavior are frequently hindered by high financial costs, extensive time requirements, and restricted experimental scopes, which ultimately make it difficult to forecast how treated soils will react under varying climatic conditions. To overcome these operational challenges, the investigation successfully implemented an advanced data-driven prediction framework that pairs powerful ensemble regression algorithms with a specialized metaheuristic optimization technique known as the catch fish optimization algorithm. By thoroughly analyzing structured datasets extracted from prior empirical research, the study established a robust computational method to estimate freezing points without requiring exhaustive physical testing for every unique combination of soil properties and contaminant levels.

The comprehensive findings revealed exceptional predictive performance when utilizing optimized ensemble models, particularly highlighting the superior accuracy and stability of the optimized random forest regression framework. During the intensive training phase, this optimized model achieved a remarkable coefficient of determination of ninety-five percent and a variance account factor exceeding ninety-five percent, demonstrating an outstanding ability to accurately capture complex nonlinear relationships between soil characteristics and freezing thresholds. Furthermore, independent testing confirmed the model’s robust reliability on unseen data, yielding a coefficient of determination of over ninety-four percent alongside an exceptionally high Willmott’s index of agreement. When benchmarked against conventional unoptimized algorithms and alternative machine learning approaches, the optimized framework significantly reduced prediction errors, successfully lowering the mean squared percentage error by more than thirty-two percent during the evaluation phase. Rigorous statistical validation using Wilcoxon signed-rank tests confirmed that these performance enhancements were robust, non-random, and statistically significant, thereby verifying the clear superiority of the optimization-assisted modeling approach.

To better understand the underlying physical factors driving these accurate predictions, the investigation utilized Shapley additive explanations to evaluate the relative importance of each environmental and chemical parameter. The detailed analysis demonstrated that lead content exerted the primary influence on the predicted soil freezing temperature, followed closely by moisture content and biochar application levels, while the preset ambient temperature played a comparatively minor role. This hierarchical importance aligns closely with established thermodynamic principles of freezing-point depression, wherein heavy metal solutes and water-particle interactions alter the chemical potential of soil pore water. Because dissolved contaminants and varying moisture levels actively change the phase transition dynamics of pore fluids, lower ambient temperatures are required to achieve solid-liquid equilibrium within the contaminated soil matrix.

The practical implications of these scientific findings offer substantial value for geotechnical and environmental engineering projects conducted in cold regions. By providing a rapid computational tool to estimate freezing temperatures across diverse contaminant concentrations and moisture levels, engineers can accurately assess frost heave risks and structural stability for roadways, foundations, and embankments. Practitioners can simulate various biochar application rates and soil amendments digitally before initiating expensive field trials, effectively streamlining remediation strategies and reducing operational expenses. Although the current models serve primarily as reliable decision-support instruments within the evaluated experimental domain, the successful integration of machine learning and soil science marks a vital step toward sustainable environmental management in frigid climates.

Source: Safi, L. (2026). Machine-learning prediction of freezing temperature in Pb-contaminated soils remediated with biochar. Scientific Reports, 16, 77024.