Overview Of Mechanistic Modelling Techniques

OVERVIEW OF MECHANISTIC MODELLING TECHNIQUES

APPROACHES TO MODELLING
Process models are very useful. They can be used for operator training; safety analysis and
design of safety systems; process design and process control systems designs. There are 2 main
approaches to developing process models,
a) empirical or data based modelling
b) building models based on the underlying physics andchemistry governing the
behaviour of the process
The following will discuss differences between the two approaches.
Developing Empirical Models
In empirical modelling, the following procedure is usual
(1) collect data from the process
(2) specify the correlation structure between variables, e.g. polynomials; time-series;
artificial neural networks
(3) use a numerical technique to findparameters for the structure such that the correlation
between the data is maximised
(4) validate the model against an ‘unseen’ data set
(5) if model is not satisfactory, go to step (2)
Thus, we can see that in empirical modelling,
• much depends on the availability of representative data for model building and
validation
• apart from cause-and-effect between variables, not much else is requiredin terms of
process knowledge
• a trial and error approach is adopted
Developing Mechanistic Models
The development of mechanistic models follow a different procedure:

© 1998-2000 Ming Tham
Department of Chemical and Process Engineering
University of Newcastle upon Tyne

Page 1 of 6

(1) use fundamental knowledge of the interactions between process variables to define the
modelstructure
(2) perform experiments to determine the parameters of the model
(3) collect data from the process to validate the model
(4) if the model is not satisfactory, go to step (1) and re-examine process knowledge
Mechanistic modelling therefore
• does not require much data for model development, and hence is not subject to the
idiosyncrasies in data
• requires a fundamental understanding ofthe physics and chemistry governing the
process
• can be very time consuming
As an alternative to step (2), once the structure of the model is defined, numerical techniques can
be applied to parameterise the model. In this case, although the structure has been determined
from process knowledge, the modelling procedure becomes an empirical one. The numerical
techniques that are used are alsovery different from those usually encountered in purely
empirical modelling. They tend to be iterative, and are more complex.
Empirical versus Mechanistic Models
When available, mechanistic models can provide more realistic predictions, and more can be
done with it in terms of analyses. For example, the details contained within a mechanistic model
offer the opportunity to test thesensitivities of the process to meaningful entities such as heat
transfer coefficients; activation energies; catalyst poisoning, etc. With very few exceptions, the
parameters of data based models are just numbers encapsulating combined effects. This it is very
difficult to attach physical meaning to them, and hence such sensitivity studies cannot be
performed.
Thus, while mechanistic models are used todesign processes, empirical models are frequently
used as the bases for process controller designs. The argument here is that model based
controllers only require the models to represent with some accuracy, the trends in process
behaviour. Conservative tuning, together with the feedback mechanism are usually sufficient to
overcome any inaccuracies. Nevertheless, if a mechanistic model isavailable, it will be foolish
to discard it in favour of a data based one.
Another comparison that is always made between the two modelling approaches is that of cost.
Due to the complexity of many processes, mechanistic modelling is indeed very expensive in
terms of human effort and expertise. As the mechanistic modelling approach forces a detailed
examination of fundamental process behaviour,...