(520c) Development of Property Models with Uncertainty Estimate for Process Design Under Uncertainty

Physical and
thermodynamic properties of pure compounds and their mixtures play an important
role in design, simulation and optimisation of chemical processes. The accuracy of process design and simulation largely depends on
the accuracy of the underlying physical and thermodynamic data and property
prediction models. While use of
experimentally measured values for the needed properties is desirable in
process design, the experimental data for the compounds of interest may not be
available in many cases. Therefore, development of efficient and reliable
property prediction methods and tools that can also provide estimates of
uncertainties in predictions of properties and their effects on process design
becomes necessary. For instance, the accuracy of design of distillation column
to achieve a given product purity is dependent on  many pure compound
properties such as critical pressure, critical temperature, acentric factor
etc. In such cases, accurate property values along with uncertainty estimates
are needed to perform sensitivity analysis and quantify the effects of these uncertainties
on the process design.      

The objective
of this work is to develop a systematic methodology to provide more reliable
predictions with a new and improved set of model parameters for GC (group
contribution) based and CI (atom connectivity index) based models and to quantify the
uncertainties in the estimated property values from a process design
point-of-view. This includes: (i)
parameter estimation using available GC based and CI based property prediction
models and large training sets to determine improved group and atom
contributions; (ii) uncertainty analysis of property prediction models to
establish statistical information such as covariance, standard error and
confidence intervals; and (iii) use the results of uncertainty analysis to
predict the uncertainties in process design. For parameter estimation, large data-sets of experimentally
measured property values for a wide range of pure compounds are taken from the
CAPEC database. Classical frequentist
approach i.e., least square method is adopted for the estimation of model
prediction uncertainties. The developed
methodology provides property values along with uncertainties for the following 20 properties: normal
boiling point, critical temperature, critical pressure, critical volume, normal
melting point, standard Gibbs energy, standard enthalpy of formation, standard
enthalpy of fusion, standard enthalpy of vaporization at 298 K and at the
normal boiling point, entropy of vaporization at the normal boiling point,
surface tension at 298 K, viscosity at 300 K, flash point, auto ignition
temperature, Hansen solubility parameters, Hildebrand solubility parameter,
octanol/water partition coefficient, aqueous solubility, acentric factor, and
liquid molar volume at 298 K. The
performance of property models for these properties with the revised set of
model parameters is highlighted through a set of compounds not considered in
the regression step. The comparison of model prediction uncertainties with
reported range of measurement uncertainties is presented for the properties with
related available data. The application of the developed methodology to
quantify the effect of these uncertainties on the design of different unit
operations (distillation column, liquid-liquid extraction, heat exchanger,
crystallizer, equilibrium reactor etc.) is presented. The results show that
depending on the chemical systems involved and the operating conditions being
considered, some of the input uncertainties can result in significant
uncertainties in design.  The most sensitive properties for each unit operation
are also identified. This analysis can be used to reduce the uncertainties in
property estimates for the properties of critical importance (by performing additional
experiments to get better experimental data and better model parameter values).
Thus, the developed methodology can be used
to quantify the sensitivity of process design to uncertainties in property estimates; obtain rationally the
risk/safety factors in process design; and identify additional experimentation
needs in order to reduce most critical uncertainties.