Neural Network-Based Prediction and Sensitivity Analysis of CO₂ Solubility in Ionic Liquids

In this study, the prediction of the dissolution of CO₂ in the ionic liquid BMIM was achieved using Artificial Neural Networks (ANNs) together with a feedforward multilayer perceptron design constructed in MATLAB. The models were trained with two recognized learning algorithms: Levenberg-Marquardt (LM) and Scaled Conjugate Gradient (SCG). A proper data set made up of experimental results and measurements was used for the training and validation aspect. These experimental results cover temperatures ranging from 228 to 322 K and pressures up to 90 bar. The ANN trained with the SCG algorithm was less effective than the ANN trained with the LM algorithm as this algorithm visibly outperformed the SCG model by getting a regression coefficient of R = 0.9999 and low error (MSE ∼10^-1). The results showed that these models did not just memorize the already given data; instead, they were trained smoothly and were able to make accurate predictions on new sets of data given without any signs of overfitting. Thermodynamics principles have established that the solubility of gases depend heavily on temperature and pressure, which in this case, is the solubility of CO₂ in [BMIM]-[BF₄]. Sensitivity analysis also confirms this principle, and well-trained ANNs are in alignment with this theory, therefore they are reliable for the prediction of complex and nonlinear chemical processes.