Application of Neural Networks to a Predictive Extended Corresponding States Model for Pure Halocarbon Thermodynamics

The extended corresponding states (ECS) model has been extensively studied for representing the thermodynamic surface of pure fluids and mixtures in the aR(r, T) form, and the most advanced version is currently the one for hydrofluorocarbons, but the shape factors J(r, T) and j(r, T) have yet to be determined as analytical functions for the whole PrT surface of a pure fluid. For a sample of pure halocarbons, this work aims to solve the fundamental problem of determining the individual shape functions over the entire PrT domain through an innovative predictive procedure using a density model requiring only a single saturated liquid density input. An original algorithm using artificial neural networks enables the determination of the J(r, T) and j(r, T) functions from a priori knowledge of their functional forms. The proposed algorithm focuses on the determination of the residual Helmholtz energy aR(r, T) for each fluid, subsequently allowing any other thermodynamic residual function to be calculated through the first and second derivatives of temperature and density. For each fluid studied, the model has been validated for residual functions against the same functions coming from highly accurate dedicated equations of state. The prediction accuracies reach average absolute deviation values ranging from 0.3 to 7.8%, spanning from vapor and liquid regions to supercritical conditions, while the corresponding results of the conventional ECS method range from 0.54 to 20%.