Structure and understand data
The problem - not only in the chemical industry - is the quality of the data. Data is collected during the process with sensors that are installed at various points of the production facilities. It is not entirely clear whether the sensors are always fully functional, whether there is a sufficient number of sensors to map the process and whether they are located in the right places on the plant. In addition, there are external factors that can distort data collection. Frequently, purely data-based algorithms cannot be sufficiently trained because the data from which they are to learn are too inhomogeneous or too poor.
Capturing different operating states
Dr. Michael Bortz, head of the Optimization - Technical Processes department and responsible at the ITWM for the main content of the workshop, makes it clear: »If you want to learn reliably from data, you need as complete information as possible about the variants of the object under consideration. If, for example, an algorithm in image processing is to be trained to recognize a tap, it must be shown many different images of taps.« Applied to process engineering production processes, this means that the variety of possible operating states must be mapped completely in the form of measurement data - which is neither possible nor sensible in industrial practice. This is precisely where Greybox models are used to obtain a reliable overall picture through the combination of physical process knowledge and existing data.
Beyond Number 42: Asking the right questions
Prof. Dr. André Bardow, head of the Chair of Technical Thermodynamics at RWTH Aachen University, demonstrated the great progress that has already been made in digital methods for optimizing energy supply, including in the chemical industry. Algorithms and calculation results alone are not enough. »We know this from The Hitchhiker's Guide to the Galaxy«, says Bardow. »It also depends on asking the right questions. Otherwise we are at a loss for a result 42.«
In his lecture he explained methods which allow to ask questions or to derive systematic insights from the results. He emphasized that it is not enough to know the best possible result as the result of a mathematical optimization; one must rather understand why this result is the best among the assumptions made, and what changes about it if the assumptions change.
Scientific discussions and personal networking form the basis for communicating and shaping the topics of future process engineering. There was plenty of input on the two days, especially for linking different scales in process simulation as well as for new methods for process optimization.