Estimation and Uncertainty Analysis of Flammability Properties of Chemicals using Group-Contribution Property Models

Process safety studies and assessments rely on accurate property data. Flammability data like the lower and upper flammability limit (LFL and UFL) play an important role in quantifying the risk of fire and explosion. If experimental values are not available for the safety analysis due to cost or time constraints, property prediction models like group contribution (GC) models can estimate flammability data. The estimation needs to be accurate, reliable and as less time consuming as possible. However, GC property prediction methods frequently lack rigorous uncertainty analysis. Hence, there is no information about the reliability of the data. Furthermore, the global optimality of the GC parameters estimation is often not ensured.

In this research project flammability-related property data, like LFL and UFL, are estimated using the Marrero and Gani group contribution method (MG method). In addition to the parameter estimation an uncertainty analysis of the estimated data and a comparison to other methods is performed. A thorough uncertainty analysis provides information about the prediction error, which is important for the use of the data in process safety studies and assessments.

The method considers the group contribution in three levels: The contributions from a specific functional group (1st order parameters), from polyfunctional (2nd order parameters) as well as from structural groups (3rd order parameters). The latter two classes of GC factors provide additional structural information beside the functional group. The contributions of all three factors are then summed up

The method is simple and easy to apply. Taking into account higher order groups increases the accuracy. Furthermore, the application range is high due to the high number of considered functional and structural contributions.

In this study, the MG-GC-factors are estimated using a systematic data and model evaluation methodology in the following way:
1) Data. Experimental flammability data is used from AIChE DIPPR 801 Database.
2) Initialization and sequential parameter estimation. An approximation using linear algebra provides the first guess. Then the 1st, 2nd and 3rd order parameter estimations are performed separately.
3) Simultaneous parameter estimation. The result of the sequential estimation serves then as initial guess for the simultaneous parameter estimation algorithm. Different minimization/search algorithms ensure global optimality.
4) Uncertainty. A rigorous uncertainty analysis that includes asymptotic approximation of covariance matrix for parameter estimators is performed in order to provide information of the model prediction quality (95% confidence interval).