Modelling of physical properties - databases, uncertainties and predictive power

Physical and thermodynamic property in the form of raw data or estimated values for pure compounds and mixtures are important pre-requisites for performing tasks such as, process design, simulation and optimization; computer aided molecular/mixture (product) design; and, product-process analysis. While use of experimentally measured values of the needed properties is desirable in these tasks, the experimental data of the properties of interest may not be available or may not be measurable in many cases.

Therefore, property models that are reliable, predictive and easy to use are necessary. However, which models should be used to provide the reliable estimates of the required properties? And, how much measured data is necessary to regress the model parameters? How to ensure predictive capabilities in the developed models? Also, as it is necessary to know the associated uncertainties in the estimated/predicted property values, how to assess the quality and reliability of the estimated/predicted property values? The paper will review a class of models for prediction of physical and thermodynamic properties of organic chemicals and their mixtures based on the combined group contribution – atom connectivity approach.

The development of these models requires measured property data and based on them, the regression of model parameters is performed. Although this class of models is empirical by nature, they do allow extrapolation from the regressed model parameters to predict properties of chemicals not included in the measured data-set.

Therefore, they are also considered as predictive models. The paper will highlight different issues/challenges related to the role of the databases and the mathematical and thermodynamic consistency of the measured/estimated data, the predictive nature of the developed models, the uncertainty estimates of the predicted data.

Related to the database, the consistency of the collected data, the uncertainties in the measured values, and the amount of data to be used in the regression step, will be discussed and analyzed. Related to modelling, the predictive nature of the models, the calculation of the uncertainties of the estimated property values, the theoretical trends will be illustrated with examples.