Predicting CO2 Solubility in Ionic Liquids Using Deep Learning

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jul 4 2024

In a paper published in the journal Scientific Reports, researchers investigated deep learning (DL) models for predicting carbon dioxide (CO₂) solubility in ionic liquids (ILs), which are highly effective for capturing CO₂. They used a comprehensive dataset covering various ILs under different temperature and pressure conditions.

The study implemented robust deep neural network models—artificial neural network (ANN) and long short-term memory (LSTM)—showing high accuracy. The ANN proved more computationally efficient than the LSTM, with global sensitivity analysis emphasizing their utility in CO₂ solubility prediction and IL screening for CO₂ capture.

Background

Past work has extensively explored using machine learning (ML) models to predict CO₂ solubility in ILs, leveraging various algorithms and datasets. Traditional CO₂ capture methods like amine scrubbing face limitations due to high energy demands and significant solvent loss.

In recent years, ILs have emerged as a promising alternative for CO₂   capture, offering benefits such as low vapor pressure and customizable molecular structures. However, ILs often have high viscosity, which poses a challenge for practical applications in CO₂ capture.

CO₂ Solubility Prediction

This study utilizes a dataset of CO₂ solubility in ILs originally collected by Venkatraman and Alsberg and meticulously pre-processed by Song et al. for machine learning model training. The dataset includes 10,116 data points with 53 features predicting CO₂ solubility in 124 ILs, covering a temperature range of 243.2 K to 453.15 K and a pressure range of 0.00798 bar to 499 bar.

The ILs consist of various cations, such as imidazolium, pyridinium, ammonium, and anions, including tetrafluoroborate and hexafluorophosphate. The data is divided into training (80%) and testing (20%) sets, with 10% of the training data reserved for validation to monitor validation loss and prevent overfitting.

The study focuses on developing DL models, specifically an ANN and a LSTM network, for predicting CO₂ solubility in ILs. The ANN model, inspired by biological neurons, consists of one input layer, one output layer, and three hidden layers with 64 neurons each.

Due to their computational efficiency, rectified linear unit (ReLU) activation functions are used for hidden and output layers. In contrast, the LSTM model, an extension of the recurrent neural network (RNN), is designed to handle long-term dependencies and consists of an input layer, two hidden layers with 64 neurons, and an output layer. The LSTM's structure includes memory cells and gates to manage long-term information, making it suitable for sequential data analysis in CO₂ solubility prediction.

Sobol sensitivity analysis, a variance-based method, assesses parameter contributions and interactions by calculating first-order and total-order sensitivity indices. It aims to quantify the relative importance of each parameter in influencing model output variance. Morris sensitivity analysis complements this by using an elementary effects approach, perturbing parameters one at a time to gauge their average influence and variability on CO₂ solubility predictions. Model reliability and accuracy are further evaluated through key statistical indexes ensuring robustness for practical applications.

Neural Network Comparison

The performance efficiency in predicting CO₂ solubility varied across different models, necessitating an optimal choice of optimizer to enhance the neural network models' attributes. Based on Ruder's comprehensive review recommending 'Adam' as a superior optimizer, it was employed with ReLU activation functions to achieve optimized efficiency.

The learning rate was set to 0.001 to ensure effective learning without causing under or overestimation. The ideal number of neurons per hidden layer, starting from 8 and increasing to 64 per layer, resulted in an ANN model with three hidden layers, capturing complex patterns and generalizing well to new data, with an R² of 0.986 and mean absolute error (MAE) of 0.0171.

The LSTM model, configured with dual layers of 64 neurons each and employing the "tanh" activation function with Adam optimizer (learning rate 0.001), performed well over 280 epochs with a batch size 16. Both models exhibited strong predictive accuracy for CO₂ solubility, with minimal deviations and consistent error distributions. Global sensitivity analysis using Sobol and Morris methods highlighted pressure as the predominant factor influencing solubility. The ANN model balanced high accuracy, computational efficiency, and interpretability, contrasting with the LSTM model's higher computational costs.

Conclusion

To sum up, this study explored DL models for predicting CO₂ solubility in ionic liquids (ILs), using a dataset of over 10,116 measurements across 164 ILs under varied conditions. Two models, an ANN and LSTM network, underwent thorough hyperparameter tuning and validation.

The ANN achieved an R² of 0.985 in 4 minutes with 535 MiB memory usage, outperforming prior studies with a 13% lower error rate. Sobol and Morris's sensitivity analyses highlighted pressure and temperature as critical factors influencing CO₂ solubility, supporting experimental findings. Overall, the study advances CO₂ capture research by demonstrating the ANN's efficiency and accuracy, setting a precedent for future material science applications.