Improving the Speed and Representation of Physical Properties Regression

Abstract: Improving the Speed and Representation of Physical Properties Regression

One of the more difficult challenges facing organizations that provide separations solutions is the need to improve our productivity in the area of regressing and continuously improving our physical property models.  We are often faced with the broad range of tasks ranging from developing a thermodynamic model from inception to the prospect of including and incorporating new data into existing models.  Rarely, if ever, is the task so simple as to require the regression of one or a few sets of data into an acceptable package.  More often, the regression task requires the regression of many sets of data, with an ill defined picture of which interaction parameters should be used, and which set of parameters gives the optimum fit of the collection of data.  In point of fact, the statement about “returning the optimum set of values” is not at all clear.  Using many parameters may minimize the sum of the squares error, but may return a set of parameters that is so unstable as to be useless in practice.

There are a number of approaches to potentially solving this problem.  Modern computing power and optimization techniques have made assessing all or a very large subset of all of the possible permutations and combinations for a particular property representation possible.    We will present some of our efforts to improve the speed, accuracy, and physical representation of some of our properties work in this session.

John Pendergast ⁽¹⁾           John Dowdle⁽¹⁾   Pradeep Suresh-Babu⁽¹⁾   Abu Hassan⁽¹⁾      Diego Cristancho⁽¹⁾

⁽¹⁾ The Dow Chemical Company