Soil contamination by heavy metals (HMs) poses significant environmental and public health risks due to their persistence and tendency to bioaccumulate. BiocharBiochar is a carbon-rich material created from biomass decomposition in low-oxygen conditions. It has important applications in environmental remediation, soil improvement, agriculture, carbon sequestration, energy storage, and sustainable materials, promoting efficiency and reducing waste in various contexts while addressing climate change challenges. More, a cost-effective and environmentally sustainable material, has gained recognition for its effectiveness in immobilizing HMs. However, achieving consistent HM removal efficiency remains challenging due to the inherent variability of biochar and its interactions with complex environmental factors. Addressing this, Mohammad Sadegh Barkhordari and Chongchong Qi, in Ecotoxicology and Environmental Safety, have introduced an innovative machine learning (ML) framework utilizing deep forest (DF) algorithms to predict and optimize HM removal efficiency using biochar.

The new framework tackles key challenges in data management by employing data imputation for missing information and data augmentation to overcome limitations posed by small datasets. This robust approach enhances the model’s reliability and generalization capabilities. The findings indicate that the DF model outperforms conventional ML methods, achieving a coefficient of determination (R2) of 0.88 on the testing dataset. Furthermore, data augmentation specifically boosted the DF model’s accuracy, with the augmented training dataset yielding a mean squared error (MSE) of 0.012 and a mean absolute error (MAE) of 0.075 during training. This demonstrates the value of data augmentation in achieving robust performance in ML models.

Probabilistic reliability analysis, a crucial component of this research, offers valuable insights into the likelihood of achieving various levels of remediation efficiency (RE). For lower RE thresholds, such as 20–30%, the model predicts a high probability (over 80%) of substantial HM removal, confirming biochar’s effectiveness in mitigating contamination. However, as target RE thresholds rise to moderate levels (50–70%), the probability drops significantly to below 5%, indicating the increasing difficulty of achieving higher remediation efficiencies. At very high RE thresholds (80-100%), the likelihood of near-complete removal diminishes significantly. This exceedance probability curve is vital for environmental engineers and soil scientists, helping them determine the performance ranges where biochar is most effective and assess the probability of reaching specific performance targets.

The study also delves into the interpretability of the DF model, revealing the most influential input parameters for predicting HM removal efficiency. Biochar dosage emerged as the most critical factor, aligning with expectations that dosage directly affects the availability of active sites for HM adsorption. An increase in biochar concentration expands the total surface area and increases functional groups like hydroxyl, carboxyl, and carbonyl, which promote metal adsorption. Organic carbon (OC) and sand content in the soil also play significant roles, influencing the adsorption capacity and mobility of heavy metals. Higher OC concentrations can increase metal binding due to chelating properties, particularly in sandy soils.

To bridge the gap between advanced ML models and practical applications, an accessible and intuitive web-based application has been developed. This tool allows engineers to input relevant parameters and receive immediate predictive outputs, facilitating data-driven decision-making in real-world environmental remediation scenarios. This user-friendly platform provides a cost-effective, quick, and efficient solution for computing predictions, reducing the need for extensive computational resources or specialized expertise.

Despite these advancements, the researchers acknowledge limitations. The web application’s scalability for high concurrent users (e.g., 1000 simultaneous users) needs further testing to ensure stability and efficiency. Additionally, while the study primarily relies on experimental datasets, a comparison between the model’s predicted outcomes and actual field data is crucial for improving its practical applicability. Future research should include field trials to validate predictions against real-world observations, enhancing the model’s accuracy, robustness, and relevance to environmental management. Further refinement of predictive models could also incorporate additional biochar properties like functional group composition and surface chemistry, and explore hybrid remediation approaches combining biochar with other technologies.

Source: Barkhordari, M. S., & Qi, C. (2025). Integrating machine learning and reliability analysis: A novel approach to predicting heavy metal removal efficiency using biochar. Ecotoxicology and Environmental Safety, 299, 118381.