Artificial neural network prediction of transport properties of novel MPDL-based solvents for post combustion carbon capture

Abstract

Novel N-methyl-4-piperidinol (MPDL)-based solvents have been considered as high potential solvents for post combustion carbon capture, especially for power generation industry. To comprehensively investigate the CO2 absorption-regeneration performance of MPDL-based solvents, transport properties (i.e., density, viscosity, and physical CO2 diffusivity) are required. These data are reported in the literature and can be estimated by conventional predictive correlations. However, the conventional correlation is applicable for an individual solvent at various blended ratios and temperatures. Thus, artificial neural network (ANN) was then applied for prediction of the transport properties of MPDL-based solvents, including aqueous solutions of MPDL, MPDL-monoethanolamine (MEA), MPDL-2-amino-2-methyl-1-propanol (AMP), and MPDL-piperazine (PZ). Three learning algorithms of (i) Levenberg–Marquardt (LM), (ii) Bayesian Regularization (BR), and (iii) Scaled Conjugate Gradient (SCG) were applied to develop the predictive ANN models with various hidden neurons. As a result, 6 hidden neurons BR-ANN model was the most convincible single prediction platform for the three transport properties. The develop model can be very beneficial for further applications associated with the novel MPDL-based solvents.

Keywords

• Neural network
• Prediction
• CO2
• Carbon capture