Introducing Thermodynamics-informed Neural Network: A Smart Equation of State for CCUS Application

Abstract

Accurate prediction of thermophysical properties of fluids is crucial for optimising processes involving CO2-rich mixtures, particularly in CCUS. The Soave-Redlich-Kwong equation of state exhibits deviations when modeling non-ideal CO2-rich mixtures in the supercritical phase, necessitating advanced optimisation techniques. In this study, a novel hybrid SRK-neural network model was developed through integrating a vectorised SRK model with deep learning to enhance the accuracy of physical properties predictions, consists of a preprocessing scaling layer, a vectorised SRK EoS layer, and one or four hidden layers, trained using the MAPE loss function. The SRK-NN model was validated against some experimental datasets for CO2-N2 mixtures with CO2 mole fraction ranging from 0.50365 to 0.9585, and pressure and temperature ranges of 8–99.93 MPa and 245–673.15 K, respectively. The results show reduction in the average absolute relative deviation (AARD%), from 1.25 to 0.49% at high CO2 concentrations. The one-hidden-layer SRK-NN variant provides more accurate predictions at various pressures compared to deeper architect. These new findings highlight the innovative application of integrating data-driven optimisation with classical EoS for accurate estimation of CO2-rich streams physical properties. This work offers a promising direction for future thermodynamic modelling of CCUS fluids with introducing a new and reliable technique.