Experimental modeling and optimization of CO2 absorption into MDEA-sulfolane solution using RSM and ANNs

This research studies the modelling and optimization of CO₂ absorption in a mixed MDEA-Sulfolane solvent system using response surface methodology (RSM), multilayer perceptron (MLP), and radial basis function (RBF) approaches. Experimental data from a stirred reactor were collected with five input parameters, including temperature in the range of 20–70 °C, pressure in the range of 2–8 bar, sulfolane concentration in the range of 10–20 mol/L, MDEA concentration in the range of 10–20 mol/L, and time in the range of 60–3600 s. Two main output responses, CO₂ loading in the range of 0.0094–0.3627 and mass transfer flux of 0.00023–0.00782 mol/m².s, were examined. The modeling results showed that the MLP network provided the highest predictive accuracy, with K-fold-validated R² values of 0.9999 for CO₂ loading and 0.9863 for mass-transfer flux, outperforming both the RBF model (R² ≈ 0.9993–0.9885) and RSM (R² ≈ 0.9800–0.9563). The optimal MLP structure used 55 neurons, selected through systematic evaluation to prevent overfitting. Optimization using RSM and ANN–GA produced closely matching results. For the flux-optimum, both methods identified the same operating point (24.77 °C, 7.995 bar, 19.98 mol L⁻¹ sulfolane, 19.998 mol L⁻¹ MDEA, 60.01 s), with predicted fluxes of 0.0090 (RSM) and 0.0089 mol m⁻² s⁻¹ (MLP–GA). For the loading-optimum (at 20 °C, 8 bar, 10 mol L⁻¹ sulfolane, 13.30 mol L⁻¹ MDEA, 3552 s), RSM predicted 0.4510, while MLP–GA predicted 0.4301, corresponding to a small deviation of 4.63 %. Overall, the strong consistency between RSM and ANN–GA confirms the reliability of the ANN framework for predicting and optimizing CO₂ absorption performance in hybrid solvent systems.