A feature analysis-based physical experience-driven neural network for predicting VOCs adsorption performance of activated carbon

Purification of industrial gaseous emissions containing volatile organic compounds (VOCs) by adsorption of activated carbon (AC) has been identified as an established method, which is creating a strong demand for rapid predicting VOCs adsorption performance of adsorbents. However, experimental data acquisition is costly and time-consuming. Meanwhile, traditional machine learning methods fail to effectively consider physical knowledge. This study proposes a physical experience -driven neural network (PEDNN) based on feature analysis for predicting VOCs adsorption performance of AC with small sample size. 33 data points were collected at first and extended to 83 using the synthetic minority over-sampling technique (SMOTE). Then the feature correlation and importance analysis results were exploited as physical constraints to construct PEDNN model, and its predicted performance were compared with those predicted by traditional model, including Support Vector Regression (SVR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) models. The results indicate that the PEDNN model exhibits superior performance on both the validation and test datasets, with R 2 values of 0.951 and 0.947, respectively. During the final process of experimental validation, the experimental data served as the test set to assess the generalization ability and predictive accuracy of the four trained models. The PEDNN model exhibited the best performance, with an R 2 value of 0.893. This work provides a novel approach for predicting VOCs adsorption performance with small sample size.