Machine Learning-Assisted Prediction of Mercury Removal Efficiency of Carbon-Based Adsorbents

Abstract

Adsorbent injection is the most promising technology for solving anthropogenic mercury (mainly Hg⁰) emission from stationary sources. Carbon-based adsorbents have strong potential for Hg⁰ removal due to their high specific surface area and abundant functional groups. However, traditional experimental methods focus on a single adsorbent under specific mercury removal conditions, making it difficult to obtain universal influencing laws and optimal preparation methods for the adsorbents. This study used machine learning (ML) to predict Max. Hg⁰ removal efficiency based on the experimental data including adsorbent parameters and removal conditions published over the past 25 years. It shows that the gradient boosting decision tree (GBDT) model has the best prediction effect (test R² = 0.87). The Brunauer–Emmett–Teller (BET) surface area and Cl are important factors affecting the Max. Hg⁰ removal efficiency, especially within a certain range. By adjusting the BET surface area and the halogen (Cl, Br, and I) ratio, the Max. Hg⁰ removal efficiencies of carbon-based adsorbents can be improved from 85 to 98.4, 90.7, and 88.6%, respectively. The maximum error between the experimental and predicted values is within 10%, proving the accuracy of the ML model prediction. The finding has important guiding significance for the design and development of high-performance mercury removal adsorbents.