Autoencoded Chemical Feature Interaction Machine Learning Method Boosting Performance of Piezoelectric Catalytic Process

Piezoelectric catalytic process can reduce energy consumption in water treatment processes. However, the design of high-performance piezoelectric materials and the search for operating parameters are still challenging tasks. This study explored a modified machine learning (ML) technology, autoencoded chemical feature interaction machine learning (AutoCFI-ML), by employing the autoencoder and factorization machine for designing piezoelectric materials and optimizing operating parameters to improve the performance of the piezoelectric catalytic process. This method improved the performance of regression-based ML methods through carefully designed chemical features boost. The extreme gradient boosting (XGBoost) model was considered the optimal model with R2 = 0.88 and RMSE = 1.02. The catalyst composition, initial pH value, catalyst/pollutant dosage ratio, and ultrasound power were identified as relatively important features among the 34 features. When targeting RhB or phenol as the typical pollutant, the errors between the prediction results of the trained AutoCFI-XGBoost model and the experimental results in the reverse experiment were both less than 10%. This work provides novel insights and improvement strategies by ML technique to design piezoelectric catalysts and optimize operating parameters for enhancing the performance of piezoelectric catalytic process, improving the application potential of the piezoelectric catalytic process.