Improved vapor pressure prediction from PR + COSMOSAC EOS using normal boiling temperature

Abstract

We evaluate the predictive power of the PR + COSMOSAC Equation of State (EOS) for vapor pressure using a large dataset of 19,081 compounds. The PR + COSMOSAC EOS uses results of quantum mechanical solvation calculations to determine the energy and molecular volume parameters in the Peng-Robinson EOS and thus does not require experimental critical temperature (T c ), pressure (P c ), and acentric factor (ω) as in the conventional approach. The prediction accuracy (average absolute relative deviation) is 141%, about seven times that of the PR EOS. A new approach is developed to improve the prediction accuracy from PR + COSMOSAC EOS by incorporating experimental normal boiling temperature. The prediction accuracy improves to 58%, about three times that of the PR EOS. The PR + COSMOSAC EOS is an effective method for vapor pressure prediction when no or just some experimental data is available.

Advancing Vapor Pressure Prediction: A Machine Learning Approach with Directed Message Passing Neural Networks

“…In our experiments, the D-MPNN model demonstrated superior performance over the PR + COSMOSAC method without the inclusion of experimental boiling temperatures, achieving an impressive AARD of 0.617. This result is notably competitive with the enhanced PR + COSMOSAC method that includes experimental boiling points [32]. The results show that the D-MPNN architecture may achieve accurate vapor pressure predictions with input of only molecular structures (atom connectivity).…”

“…In this section, we focus on evaluating the performance of our optimal model, the mfp-sum Wagner EE model, against three previously proposed methods: PR-pred, PRpred-expTb, and PR-exp, as delineated by Tsai and Lin [32]. Each of these methods represents a distinct approach to predicting vapor pressure, ranging from reliance on purely computational inputs to the integration of extensive experimental data.…”

“…The PR-pred method stands out for its use of quantum mechanical solvation calculations to estimate energy and molecular volume parameters for the PR EOS, eliminating the need for experimental 𝑇 𝑐 , 𝑃 𝑐 , and 𝜔 [32]. This approach offers a theoretical model that bypasses the requirement for experimental critical properties.…”

“…Following the methodology of Tsai and Lin [32], we gathered vapor pressure data for 19,081 unique organic molecules from the NIST-TRC databank available in Aspen Plus V11 [64]. This extensive dataset, documented using Wagner25 coefficients [63], critical temperature, and critical pressure, serves as the foundation for our machine learning models.…”

“…This led to the successful development of a model adept at learning molecular features essential for accurate vapor pressure predictions, circumventing the need for additional inputs like critical properties, acentric factors, or manually crafted descriptors typically required in traditional models. To evaluate the D-MPNN model's performance, we compared it with the PR + COSMOSAC method [32]. While the PR + COSMOSAC method is independent of experimental critical properties, it necessitates quantum mechanical calculations for each molecule.…”