(535f) Accuracy of PR+Cosmosac Eos in Predicting Vapor Pressure with Machine Learning

Thermodynamic properties are an important piece of information for the design of chemical processes. The Peng-Robinson+COSMO-SAC equation of state [1], denoted as PR+COSMOSAC EOS, has been shown to provide reasonable prediction for thermodynamic properties of pure substances and fluid phase equilibrium of mixtures without the problem of missing parameters. Recently, there are some researches using machine learning to predict thermodynamic properties or combining it and group contribution methods to improve prediction accuracy [2]. In this work, we combine PR+COSMOSAC EOS and machine learning to improve the accuracy of it in predicting vapor pressure. The machine learning model is developed to correlate the prediction result of PR+COSMOSAC EOS. To confirm the reliability of this model, we use 10 different data sets to train the model, and each of the data sets would be trained 10 times. It is found that no matter which set it is, combining PR+COSMOSAC EOS and machine learning is more accurate than PR+COSMOSAC EOS. It indicates this model can learn the diversity of compounds and the relationship of temperature and vapor pressure. Besides, the standard deviation of 100 models is low, so the stability of this model is high. We also change the ratio of training sets to observe whether this model can still learn the diversity of compound. It is found that although the standard deviation would increase, even if the ratio of training set is changes to 0.1, the accuracy of the model is close to that of the ratio of training set is 0.9, and it also learn the diversity of compounds. Our results indicate that machine learning combined with a theoretically based model can be more effective than the machine learning model alone and also improve the accuracy of theoretical model.

References

[1] Wang L.-H., Hsieh C.-M., Lin. S.-T. Improved prediction of vapor pressure for pure liquids and solids from the PR+COSMOSAC equation of state. Ind. Eng. Chem. Res. 2015; 54: 10115−10125

[2] Gharagheizi, F.; Eslamimanesh, A.; Ilani-Kashkouli, P.; Mohammadi, A. H.; Richon, D., “Determination of vapor pressure of chemical compounds: a group contribution model for an extremely large database”, Ind. Eng. Chem. Res., 51 (20), 7119-7125, 2012.