Evaluation of Equations of State and Machine Learning for the Prediction of Thermodynamic Properties for Carbon Dioxide Rich Mixtures

Abstract

Accurate property predictions is an essential factor to efficient and safe process design, and there has yet to be an improvement upon the standard fluid package selection process where current works refer to Eric Carlson (1996) process of selection . Thus, the need for a guideline for accurate prediction of specific mixtures is vital.

With the renewed interest in CCS and to comply with climate change legislation such as Carbon Budget in the UK for Carbon Neutrality by 2050; this has called for the re-evaluation of established EoS for the improvement to property predictions for computational analysis in simulation software such as AspenONE suite With few procedures to accurately select property package selection and the inability to use tailored models, one of the focuses of this work evaluates two suggested EoS, Peng Robinson (PR) and Soave Redlich Kwong (SRK) for their predictive accuracy in describing the phase behaviour and compressibility for CO2 rich mixtures.

The feature article of this work evaluates the use of Machine Learning (such as Artificial Neural Network and Classified Learner) to be applied to CO¬2 mixture thermodynamic property prediction and application to simulation software such as AspenONE and its approachability for to solve other design problems within Chemical Engineering.

The results of this work show that PR has a lower Average Absolute Relative Deviation (AARD) in comparison to SRK with 3 components (CO2 with N2 and CH4 in a) binary mixture achieving 7.46% in contrast to 14.99% for SRK, however with the 5 (CO2 with N2, CH4, Ar and O2) component case 13.5% for PR and 19.50% for SRK. Hence, Peng Robinson is a more suitable package to select during fluid property simulations.

The results of applying Machine Learning to property prediction show promise, where implementing corresponding states principle in conjunction with EoS allows for the lower predictive AARD in contrast to EoS, achieving 2.81% AARD for the abovementioned three components and 12.24% for the respective 5 component mixtures. Hence, in future work for applying this technique to existing software for the predictive simulation may be the impending replacement for these long-standing EoS for prediction of thermodynamic properties.