Prediction of phosphate adsorption amount, capacity and kinetics via machine learning: A generally physical-based process and proposed strategy of using descriptive text messages to enrich datasets

Abstract

The use of machine learning to predict phosphate adsorption performance by specific adsorbent holds great promise due to its ability to save time and reveal underlying mechanisms. However, the small size of the dataset, caused by the insufficient detailed information, limits the ability to fully train the model and obtain accurate and generalizable results. To address this issue, we employ a fuzzing strategy that substitutes detailed numeric information with descriptive text messages on the physiochemical properties of adsorbents (such as using “Porous Carbon” to substitute the adsorbent with 100 < SBET < 500 m2⋅g−1 and with C content > 10 wt%). This strategy allows the recovery of the discarded samples that are short in detailed information, which delivers the enrichment of dataset and the accurate prediction of adsorption amount, capacity, and kinetics. The adsorption amount is determined by environmental pressures, while adsorption capacity and kinetics are subject to a synergism of environmental pressures and physicochemical properties of adsorbents. Comprehensive analysis finds that phosphate uptake by adsorbents is generally via physisorption, although chemisorption also plays a role. This study has constructed a framework that requires only the input of common and easily accessible information to quickly predict phosphate adsorption performance by specific adsorbent, which is highly practical, especially in urgent scenarios. Additionally, our strategy provides useful guidelines for recovering discarded samples and enriching the dataset by fuzzing detailed information.