(481f) The Symbiosis of Measuring and Predicting Property Data

Property data can be determined from measurements at every single state point of interest such as temperature and pressure. While measurements are the most precise method to determine property data, the experimental effort can be tremendous. Experimental effort can be reduced by correlating property data with models like equations of state while maintaining very good accuracy. In case of the PCP-SAFT equation of state, at least 3 experiments must be performed per pure component. Predicting property data can eliminate the need for experiments altogether, but the accuracy is not sufficient for many applications. Thus, we can choose between measurement, correlation and prediction trading-off effort against accuracy.

We propose combining information from several approaches to increase accuracy at low effort. The PCP-SAFT equation of state is used as model. As a starting point, PCP-SAFT parameters are predicted using methods developed in our group (J. Phys. Chem. B 112, 5693-5701). For this purpose, molecular descriptors are obtained from quantum mechanics. The predicted PCP-SAFT parameters are then refined using information from 1) the predictive model COSMO-RS or 2) a single measurement point of the vapor pressure. The new information is used to refit the PCP-SAFT parameter set using a log-likelihood function combining all available information.

For a database containing 26 components, the combination of the predictive models for PCP-SAFT and COSMO-RS halves the errors to ~20% error in vapor pressure from about ~40 error for each individual prediction method. Combining predictive PCP-SAFT with a single vapor pressure experiment leads to almost the accuracy of pure correlation (<5% error in vapor pressure). Model-based Optimal Experimental Design can be used to determine the experiment which should be conducted for the largest gain of accuracy in our proposed combination method.

The proposed combination of predictive models with selected experiments thus allows to greatly improve accuracy of property data at low effort.