Spray drying is a fast method for drying liquid products. By spraying the finely atomized product into a hot air stream, the solvent can be quickly evaporated and the solid components remain as a powder. Spray drying is used in the chemical, pharmaceutical and food industries as it is a particularly gentle drying method and the powders have advantageous properties.
The process control of a spray dryer is generally not critical, as there are only a few, but a sufficient number of adjustable process parameters. However, practical process control is difficult due to the non-ideal drying behavior of the products and the dynamic behavior of the spray dryer. If the operating parameters change, e.g. product properties or humidity, the system must be readjusted by trained personnel. Automated control with standard control solutions or data-driven machine learning (ML) approaches or digital twins is not practicable due to the non-linear control behavior and inadequate on-line sensor technology.
The research objective focuses on the start-up of new spray drying systems, process optimization for new products and scale-up or scale-through between systems of different sizes or designs. These are currently very cost-intensive and associated with correspondingly long production downtimes. The scientific problem is to identify the ideal process parameters for the respective product with a given spray dryer and given environmental conditions. With the help of ML approaches, a system is to be developed that enables a “first time right” for the tasks mentioned, i.e. a specifically optimized configuration is already suggested initially on the basis of the digital twin.
The aim is to create a hybrid digital twin as a combination of physical modelling and parameterization with ML methods that describes the system behaviour of a spray tower. By defining the basic functional sequence and the system description in the physical modeling, the ML models should be able to extrapolate to new operating states and the problem complexity for the ML models should be kept low. By integrating transfer learning methods and XAI components, the models are to be transferable between plants and predictions are to be made comprehensible.
