Prediction of CO₂ solubility in ionic liquids via convolutional autoencoder based on molecular structure encoding

Abstract

In this study, novel molecular structure encoding descriptors composed of feature encoding and one‐hot encoding was developed and then convolutional autoencoder was used to denoise based on the structure of ionic liquids (ILs). It could be used to predict the CO2 solubility in ILs at different temperatures and pressures, when combined with three different machine learning algorithms (multilayer perceptron [MLP], random forest [RF], and support vector machine [SVM]). Statistics of the prediction results show that the newly proposed molecular structure‐based coding has better regression prediction performance than the conventional molecular cheminformatics descriptors. SE‐MLP model with R2 of 0.9873 and mean square error of 0.0007 has the best performance in predicting the CO2 solubility in ILs. In addition, the relationship between features and dissolved CO2 capacity was analyzed through model interpretation to retrieve physical insights for the underlying system. This work provided a new predictive tool for enriching and refining data on CO2 solubility in ILs and for solving phase equilibrium problems.

Interpretable Feedforward Neural Network and XGBoost-Based Algorithms to Predict CO2 Solubility in Ionic Liquids

“…The XGBoost−GC−MSD model specifically displays a remarkable performance for the testing set, exhibiting an R 2 value of 0.98963, MAE of 0.01480, and RMSE of 0.02369. Comparing our best model, XGBoost− GC−MSD, with established models, 23,24 it surpasses both the ANN−GC and SE-MLP models, which recorded R 2 values of 0.9836 and 0.9873 for the testing set, respectively. Even our least-performing XGBoost model (i.e., XGBoost−GC) outperforms the ANN−GC and SE-MLP models from previous work with an R 2 value of 0.98891.…”

“…Table showcases the effectiveness of the six models studied for predicting CO 2 solubility in ILs. Additionally, we sought to compare the performance of these six models against established ones used for similar purposes (i.e., predicting the solubility of CO 2 ) in previous research. , Among our six models, the XGBoost models consistently outperform the FNN models regardless of the descriptor types utilized. The XGBoost–GC–MSD model specifically displays a remarkable performance for the testing set, exhibiting an R 2 value of 0.98963, MAE of 0.01480, and RMSE of 0.02369.…”

“…Among the three ML techniques investigated, the SE-MLP model exhibited the best performance, achieving an R 2 of 0.9873 and a mean square error of 0.0007 in predicting CO 2 solubility in ILs. Until here, it seems that Song et al 23 have illustrated that ANN performs better than SVM, whereas Liu et al 24 indicated that MLP outperforms RF and SVM. Therefore, among the many common ML techniques that we have introduced earlier (e.g., SVM, DT, RF, GB, XGBoost, and NN), it appears that there are still several other ML techniques yet to be explored.…”

“…In this work, our aim is to investigate the performance of using XGBoost and feedforward NN (FNN) (i.e., as another type of NN) models for screening ILs for CO 2 capture. We aim to compare their performance against ANN–GC and SE-MLP, which were identified as the best ML techniques in previous studies. , Our focus excludes DT and GB methods because, as per our understanding, XGBoost already integrates both DT and GB functionalities into a single powerful algorithm . Its capacity to enhance tree boosting methods enables the swift and accurate processing of diverse data types.…”

An Overview of Flue Gas SO2 Capture Technology Based on Absorbent Evaluation and Process Intensification

“…This will enable the rapid identification and selection of ILs with the highest potential for effective SO 2 capture. Wang research group used a machine learning algorithm to predict the solubility of acidic gases CO 2 and H 2 S in ILs, and obtained the structure–activity relationship between the molecular structure and solubility of ILs. − Wang et al had developed a screening method for ILs based on the UNIFAC-IL model extension to remove SO 2 from flue gas (see Figure ). The screening method consists of four steps.…”