(431f) When Is It Appropriate to Kiss? a Case Study on the Adverse Consequences of Overfitting GE Models

In his presentation at the 19th International Conference on Chemical Thermodynamics (ICCT-19, Boulder, July/August 2006), Marco Satyro emphasized the importance of developing effective models guided by experience: "In this paper, I describe the different ways in which incomplete data and models are used to do process engineering." He opened his presentation by illustrating the negative consequences of a heat-capacity model where the 24 data points were fit to machine precision using a 24-coefficient model. The "perfect model" was impossible to converge due to absurd temperature derivatives. This presentation shows an analogous case study where the simulation of a simple system - purification of methanol from a mixture containing C2 to C4 alcohols - gave highly inaccurate results compared to established licensor data. It turned out that the two asymmetric NRTL parameters for the various binaries came from a database that fit the near-ideal binary data "exactly," but unintended consequence was that the model predicted strong - and utterly wrong - departures from ideality for the multicomponent system. This is a case where the KISS (Keep it Simple and Straightforward) principle is a superior approach to blind nonlinear fitting of the data. The same principle applies when activity-coefficient parameters are determined from estimation methods such as UNIFAC.