Prediction of pure-component parameters of perturbed chain-statistical associating fluid theory equations of state using graph convolutional neural network-based modeling

• Developed a GCNN to predict PC-SAFT pure-component parameters • GCNN predicts parameters directly from molecular structure information • Systematic comparison clarifies strengths different molecular representation models • Graph-based representations effectively capture molecular size and shape effects • Developed GCNN architecture guides machine-learning-based parameter prediction Substance-specific parameters are required to perform calculations using equations of state (EoS), which are powerful tools for predicting physical properties. These parameters are typically obtained from experimentally measured physical properties; however, predicting physical properties requires prior knowledge of the same properties, leading to a circular dependency problem. To address this issue, machine learning-based models that predict substance-specific parameters directly from molecular information without relying on measured physical properties were previously proposed. Among the existing approaches, we found that the pure-component parameters of the perturbed chain-statistical associating fluid theory (PC-SAFT) EoS strongly depends on molecular descriptors related to size and shape. Several molecular structure representation-based models have since been developed for machine learning applications. In this study, we develop a graph convolutional neural network (GCNN)-based molecular structure representation method that treats molecules as graph structures to predict the pure-component parameters of the PC-SAFT EoS. The developed GCNN model was compared with other existing molecular structure representation-based models, including previously reported group contribution- and molecular fingerprinting-based models. The GCNN predicted pure-component parameters with estimation accuracy comparable to those of other existing models while accommodating the broadest range of molecular structures. This study thereby expands the applicability of EoS-based property estimation.