Machine Learning meets Process Simulator: Grey Box and Surrogate Modeling to Support Process Optimization


Besides experimentation process modeling, simulation and optimization is one of the key elements of process development. Especially, when comparing process concepts with regard to partly competing sustainability criteria, multicriteria optimization and decision support (cf. for example Bortz et al., 2014, Burger et al., 2014) are indispensable. The applicability of such methods is sometimes restricted either due to missing models for certain process steps to complete a simulation model for the whole process or by costly and hardly solvable simulation models. In both cases machine learning can provide a feasible solution to tackle these restrictions.

For missing scientific process models available process data can be used to set-up data-driven models. In most cases process data are not or only partly available at the interfaces of the process step to be modeled. Therefore, as a first step usually a simplified model for the process step is included in the simulation model to complete the model for the whole process and used as a soft-sensor for the missing information. After this step the necessary information is available to model the process step for example with machine learning methods separately (so-called incremental model identification, cf. e.g. Bardow and Marquardt, 2004) and include it afterwards in the entire process simulation model.

In cases where a process model is too hard to be applied directly for optimization, the use of surrogate modeling (Bhosekar and Ierapetritou, 2018) can be a resort. Here the original model is sampled, and the collected data of these samples can be taken to train for example artificial neural networks reflecting the optimization variables as inputs and the objectives and constraints as outputs. The obtained optimized variables can be used as inputs for the original process model to verify the results of the surrogate model. If the agreement with the predicted results of the surrogate is not satisfying the verification results can be used to retrain the surrogate, run again the optimization and repeat the verification step. This iterative procedure could be repeated until sufficient agreement is obtained.

Both methods will be demonstrated for a process simulation example of a Cumene process.

A. Bardow, W. Marquardt, 2004, Chem. Eng. Sci., 59, 2673 – 2684.

M. Bortz, J. Burger, N. Asprion, S. Blagov, R. Böttcher, U. Nowak, A. Scheithauer, R. Welke, K.-H. Küfer, and H. Hasse, 2014, Comp. Chem. Eng., 60, 354–363. doi:10.1016/j.compchemeng.2013.09.015

J. Burger, N. Asprion, S. Blagov, R. Böttcher, U. Nowak, M. Bortz, R. Welke, K.-H. Küfer, and H. Hasse, 2014, Chem. Ing. Tech., 86 (7), 1065–1072. DOI: 10.1002/cite.201400008

A. Bhosekar, M. Ierapetritou, 2018,Comp. Chem. Eng. 108, 250 – 267.


This paper has an Extended Abstract file available; you must purchase the conference proceedings to access it.


Do you already own this?



AIChE Members $695.00
AIChE Graduate Student Members $695.00
AIChE Undergraduate Student Members $695.00
Non-Members $895.00