Good modelling practice for process engineering: pitfalls and requirements to develop fit for purpose models

Abstract

Abstract Many process industries are yearning for more detailed process knowledge mainly envisioning process optimization, both in terms of operation and design. Most processes are complex in nature (physical-chemical and kinetic processes involving multiple phases and in a heterogeneous environment due to process scale) and cannot be optimized using experimentation only. The last decades have proven that mathematical models are powerful tools to achieve this goal. However, many physical-based modelling frameworks of different levels of complexity are at hand and many pitfalls exist in the model building process. It is therefore crucial that Good Modelling Practice (GMP) is applied in any modelling study in order to get the most out of models, i.e. ensuring they are reliable and have a sufficient predictive power. This is crucial to convince people that models are a necessity for any process optimization and worth investing in. Consecutive steps in a model building exercise according to GMP are: objective definition, state of the art, experimental data collection (including optimal experimental design-OED), framework definition, parameter estimation and model selection and model validation. Within GMP, the goal or objective is probably the most important decision as it significantly impacts the subsequent steps such as the framework choice and data collection. In many cases the objective is process optimization, either at unit process level (operational and design) or at process train level (mostly operational). In any case, a model for optimization should have a sufficient predictive power if one wants to use it in decision making, which means it should be validated up to a certain accuracy level (defined in the objective). The latter is not yet common practice for a variety of reasons: insufficient data quality and/or quantity, lack of availability of tools to verify calibration quality, model overparameterisation, insufficient model complexity. In most cases, this leads to uncertainties in inputs and parameters that result in model output uncertainty. One should therefore strive for this as part of the modelling exercise and be self-critical about the real usefulness of a model rather than being pleased by calling a model prediction “quite good”. The uncertainties mentioned are mostly related to insufficient process knowledge. In order to increase process knowledge, one needs to move to more complex models that aim to describe processes in more detail (e.g. CFD, PBM, DEM). This comes at a significant cost as it requires significantly higher computational effort as well as more tedious data collection. Probably the latter is even most often the problem and is a clear point of attention. Another issue is the fact that many model analysis tools such as sensitivity analysis, global optimization, uncertainty analysis and optimal experimental design require a vast number of model simulations which become hard or even impossible when model structures are too complex. The avenue to go here is to translate the obtained knowledge at the complex model scale to models at intermediate complexity scale such as compartmental models (CM). Those models should be less uncertain given the extra knowledge they contain which will reflect in less need for calibration and, hence, a higher predictive power yielding better decisions for operational optimization. When it comes to design optimization, the developed complex models are now a good reference to evaluate where improvement is possible. However, using these models is also not straightforward as they are computationally intensive and there is a myriad of options to alter a design. There is clearly a need for a sound methodology that could also be based on CMs. Alternatively, we need to investigate scale-up vs numbering up and think out of the box. In this contribution I would like to touch on some of these aspects by means of examples.