Recent progress in prediction of vapor pressure: Addressing the trade-off between accuracy and generalizability

Accurate vapor pressure is of fundamental importance for the design of chemical engineering equipment and process optimization. Over the past century, researchers have developed various vapor pressure models based on theory, semi-empirical methods, corresponding states principles, and data-driven approaches. However and unfortunately, they have always faced a trade-off between accuracy and generalizability. In this work, we present a systematic review of the current research in the prediction of vapor pressure. Firstly, we categorize and summarize the four main types of existing vapor pressure models, analyzing their theoretical bases and ranges of applicability. Secondly, by evaluating these models against a comprehensive database of 629 compounds, we expose the shortcomings of each approach. For example, the empirical equations achieve very high accuracy but rely heavily on fitted experimental data. The corresponding state models are relatively general, but offer only limited accuracy. And the emerging graph neural network models remain in their early stages, showing poor adherence to thermodynamic constraints and requiring further improvements in modeling. To surmount the current bottleneck, we believe that the crucial approach lies in organically incorporating physical knowledge into machine learning algorithms, which in turn facilitates the construction of novel models with intrinsic satisfaction of thermodynamic constraints. We envision that future efforts will yield a new generation of vapor pressure models combining high accuracy with generalizability, thereby being applied to the accurate and precise prediction of the vapor pressure of different types of substances.