Improved Property Predictions by Combination of Predictive Models

Property predictions are essential
when dealing with molecules
that have not been investigated experimentally yet. The accuracy of
current predictive models like predictive perturbed-chain polar statistical
associating fluid theory (PCP-SAFT) and conductor-like screening model
for real solvents (COSMO-RS) is limited. We propose a combination
of predictive models in order to yield a higher accuracy. Information
from both predictive models are combined in PCP-SAFT parameter space
using a log-likelihood function. Experimental vapor pressures, enthalpies
of vaporization, and liquid densities over a wide temperature range
are used to evaluate the predictions. The average error in the combined
property prediction is lower than the error of the individual models.
In addition, the maximum error is considerably lowered.