Efficient removal of Ni, Al, and As from water using Micrococcus yunnanensis: batch experiments and explainable machine learning.

In this study, a bacterial strain of Micrococcus yunnanensis was isolated from the discharge water of an organized industrial zone wastewater treatment plant and used in the bioremoval of Al, Ni, and As. The effects of different experimental conditions such as initial metal concentration, pH, temperature, contact time, and biomass concentration on metal bioremoval were evaluated, and peak removal conditions were identified. The highest removal rates were 80, 68, and 27.33% for Al, Ni, and As, respectively. The isotherm and kinetic results for all metals showed that they fit the non-linear Langmuir and pseudo-second-order models, respectively. Ten representative machine-learning algorithms from linear, tree-based, ensemble, kernel, instance-based, and neural-network approaches were comparatively evaluated using a combined multi-metal dataset. The Extra Trees model was found to exhibit the best generalization performance (CV R2 = 0.948; test R2 = 0.941; RMSE = 5.98; and MAE = 4.553). To improve the interpretability of the results, SHAP analysis was performed, and the most important parameters were identified as metal type, initial metal concentration, and contact time, respectively. Of the two different training strategies followed in our study, models trained with a complete metal dataset were shown to be more successful than models trained with subsets containing each metal separately. This study contributes to the growing application of explainable machine learning in bioremoval research by integrating an environmentally isolated M. yunnanensis strain with a unified multi-metal predictive framework for Ni, Al, and As bioremoval.