Prediction of heel build-up on activated carbon using machine learning

Determining the long-term performance of adsorbents is crucial for the design of air treatment systems. Heel buildup i.e., the accumulation of non-desorbed/ non-desorbable adsorbates and their reaction byproducts, on the surface/pores of the adsorbent is a primary cause of adsorption performance deterioration. However, due to the complexity of heel buildup mechanisms, theoretical models have yet to be developed to map the extent of heel buildup to the adsorption/desorption parameters. In this work, two machine learning (ML) algorithms (XGBoost and neural network (NN)) were applied to predict volatile organic compounds (VOCs) cyclic heel buildup on activated carbons (ACs) by considering the adsorbent characteristics, adsorbate properties and regeneration conditions. The NN algorithm showed better performance in prediction of cyclic heel buildup (R² = 0.94) than XGBoost (R² = 0.81). To analyze interaction between heel buildup and adsorbent characteristics, adsorbate properties, and regeneration conditions, partial dependency plots were generated. The proposed ML-based heel prediction methods can be ultimately used to: (i) optimize adsorption/desorption operating conditions to minimize heel buildup on activated carbon in cyclic adsorption/desorption processes and (ii) quickly screen various adsorbents for efficient adsorption/desorption of a particular family of VOCs by excluding adsorbents prone to high heel formation.