Multimodal machine learning for Cr(Ⅵ) removal and floc settling using image-based floc features and operating parameters.

Electrocoagulation is a promising technique for Cr(Ⅵ) removal from wastewater, but its performance is limited by challenges in adapting to variable conditions, such as pH variations, electrolyte concentrations, and stirring rates, which impact floc formation and settling. To address this, we developed a novel multimodal machine learning framework, integrating image-based floc features, extracted via deep learning (ResNet50), with operating parameters to predict Cr(Ⅵ) removal efficiency and floc settling performance. A classification-regression model, combining deep learning (ResNet50) with Support Vector Classification and Support Vector Regression achieved a Cr(Ⅵ) removal R2 of 0.887, while ResNet50 with Bagging Classifier and Extra Tree achieved an R2 of 0.893 for floc settling. A direct regression model, integrating image features and operating parameters, outperformed the classification approach, with ResNet50-Bagging yielding an R2 of 0.971 for Cr(Ⅵ) removal and ResNet50-Extra Tree achieving an R2 of 0.986 for settling. This pioneering multimodal approach enhances prediction accuracy and process adaptability, offering a robust framework for optimizing electrocoagulation across diverse wastewater conditions.