Prediction of Pure-Component Thermodynamic Properties Using the Peng–Robinson Equation of State with Machine Learning-Derived Input Parameters

A predictive framework combining machine learning and a cubic equation of state (EoS) is proposed for estimating thermodynamic properties directly from molecular structure. Ensembles of artificial neural networks trained with Mordred molecular descriptors were used to estimate critical temperature (Tc), critical pressure, acentric factor, and a volume translation parameter. These parameters were subsequently employed in the translated Peng–Robinson EoS to calculate vapor pressure, saturated liquid density, saturated liquid isobaric heat capacity, and enthalpy of vaporization. The results show that machine-learning-derived parameters provide accuracies comparable to those obtained using experimental inputs for density, heat capacity, and enthalpy of vaporization. Vapor pressure predictions were found to be highly sensitive to small deviations in Tc and therefore constituted the most challenging property to predict accurately. Despite this sensitivity, the proposed machine-learning approach demonstrates robust predictive capabilities and significantly outperforms traditional group-contribution methods. The models are accessible through a web application at https://lrgp-thermoppt.streamlit.app/ .