Prediction of self-diffusion coefficients via a hybrid PCP-SAFT+ANN model incorporating COSMO-SAC sigma-profile descriptors

Abstract

prediction remains difficult due to the strong influence of thermodynamic conditions (temperature and pressure) and molecular characteristics such as size, shape, and intermolecular forces. In this study, a hybrid predictive model is introduced, combining the PCP-SAFT equation of state with an artificial neural network (ANN) to estimate self-diffusion coefficients over a broad range of conditions. The model is developed using a dataset, comprising of 2263 experimental measurements for 67 compounds, spanning temperatures between 93.0 and 973.2 K, pressures up to 3036 bar, corresponding to self-diffusion coefficients, spanning nearly five orders of magnitude from 10-12 to 10-7 m2 s-1. For a rigorous assessment of predictive performance, the dataset was partitioned into 30% reserved for independent validation and 70% for training. The proposed model incorporates thermodynamic inputs, namely density and dimensionless form of residual entropy obtained from PCP-SAFT, together with molecular descriptors derived from COSMO-SAC sigma-profiles. The selected ANN architecture, comprising two hidden layers with 14 and 7 neurons, respectively, provides high predictive performance, achieving R2 values of 0.9937 and 0.9763 and AARD values of 8.89% and 15.89% for the training and testing datasets. Overall, the proposed framework offers a unified reliable model for predicting diffusion behavior under diverse thermodynamic conditions.