Elucidating the black-box nature of data-driven models in the adsorption of reactive red M-2BE on activated carbon and multi-walled carbon nanotubes through SHapley Additive exPlanations

• Published in: Adsorption : journal of the International Adsorption Society 2024-06, Vol.30 (5), p.457-471
• Main Authors: Gasparetto, Henrique, Lima, Éder Claudio, Machado, Fernando Machado, Dotto, Guilherme Luiz, Salau, Nina Paula Gonçalves
• Format: Article
• Language: English
• Subjects: Activated carbon; Adsorption; Algorithms; Aqueous solutions; Artificial neural networks; Chemistry; Chemistry and Materials Science; Engineering Thermodynamics; Fuzzy logic; Heat and Mass Transfer; Industrial Chemistry/Chemical Engineering; Machine learning; Multi wall carbon nanotubes; Special Issue: EBA14 - 14th Brazilian Meeting on Adsorption; Support vector machines; Surfaces and Interfaces; Thin Films
• ISSN: 0929-5607; 1572-8757
• DOI: 10.1007/s10450-023-00420-z

The removal of reactive red M-2BE dye textile from aqueous solution was performed using multi-walled carbon nanotubes (MWCN) and powdered activated carbon (PAC). Kinetic adsorption modeling has been performed using machine learning (ML) algorithms of artificial neural networks, adaptive-neuro fuzzy inference system (ANFIS), random forest, gradient boosting, and support vector machine. Although ML models are more accurate, they often fail to interpret the reasoning behind predictions. Therefore, the SHapley Additive exPlanations (SHAP) were used to understand the effect of each feature on the adsorption capacity. The ANFIS has presented the best statistical metrics with R = 0.9993 , R M S E = 0.0214 , and S A E = 7.1172 . A higher adsorption capacity was observed for MWCN compared to PAC; while the first peaked at 300 mg L −1 , the second approached 230 mg L −1 . Temperature was found to have the smallest contribution in describing adsorption capacity. This novel application of ML with SHAP can provide important insights for adsorption researchers.