(143e) Increasing Resource Efficiency in Process Engineering on Planetary Roller Extruders By Use of Scaling Algorithms, CFD-Simulation and Prediction Models

The reduction of raw material consumption is a contribution to global environmental protection and can be achieved in two ways. On the one hand, by recycling products and recovering the materials used. On the other hand, by optimizing production processes. The potential of the latter is illustrated by the production volume of fossil based virgin polymers from 2021, which was 352 million tons per year [Pla22]. The vast majority of polymers are produced using extrusion processes. For each product manufactured, various preliminary tests and extensive experimental studies were carried out to determine the optimum process conditions to meet ideal product properties and highest quality standards. Thus, reducing the amount of material used in process development represents a mighty tool to reduce the overall consumption of polymers. This work contributes one way to optimize operating conditions of polymer processing by extrusion using planetary roller extruders PREs.

PREs feature a modular set-up. Each module consists of a roller cylinder with one central spindle and multiple planetary spindles (see figure attached). All components feature a 45° helical gearing which provides radial mixing and axial material conveying at low shear rates, due to the rotating spindles. PREs offer a multitude of configuration options. This includes, for example, a variable number of modules, which allows a precise adaption of the plant to the requirements of the production process. In addition, the geometrical dimension of the plant can also be varied. This is relevant for the process design as the plant size influences possible mass flows. PREs are available from laboratory scale (throughput: 0.5 kg/h - 10 kg/h) to large production scale (throughput: more than 8 t/h). Due to its small dimensions and thus less mass that needs to be maintained at a certain temperature, the laboratory scale extruder has a significantly lower energy consumption. Although this is quite advantageous in terms of resource efficiency in process development, operating parameters of PREs on laboratory scale cannot be transferred directly to PREs on pilot scale. Since the plant dimensions are so different, the maximum possible throughput for example differs considerably, too.

To avoid this problem and to determine comparable process settings, scaling algorithms are needed. These algorithms have been newly developed for PREs and will be presented as part of this contribution. Using them, experimental results from laboratory scale can be transferred to production scale. Thus, process development becomes more sustainable, as experiments can be carried out in small scale with low energy demand and less material consumption.

The functionality of these algorithms was verified with a series of physical experiments processing polymer granules, both carried out on laboratory and pilot scale. The identical experimental setup (see attached figure) consists of PREs with two modules [Rad23]. In the first module, the granules are melted because of external heat supply and internal friction between the rotating plant components. The second module contains a variable spindle configuration in order to be able to analyze the influence of the plant setup on the operating parameters. Various sensors for measuring pressure, temperature and residence time allowed a detailed analysis of the extrusion process. In addition to the plant setup, an analysis of the process settings is also carried out. The throughput (3.75 kg/h – 8.75 kg/h in laboratory scale, 30 kg/h – 70 kg/h in pilot scale), the set temperatures of external temperature control units (185 °C – 220 °C) and the conveying speed (60 rpm – 200 rpm) are varied according to a statistical design of experiments. The polymers processed as received are high and low density poly(ethene) (PE-HD 53090, Total Polymers and PE-LD MP20, Versalis) and poly(propene) (PP-H 9081, Total Polymers), showing different viscosity values.

Although the sensors mentioned above determine all measurable process conditions inside the PRE within physical experiments, non-measurable variables such as the shear rate cannot be determined. Nevertheless, an analysis of these properties is relevant as well, as it allows a more detailed investigation of the influence of extrusion on the material quality. Polymer damage for example may occur as a result of extensive shear rates. Due to their high relevance, the non-measurable variables and the flow profiles within the roller cylinder are calculated within 3D-CFD-simulations.

The results from about 2,000 individual physical experiments and digital simulations are collected in a database. This allows further analyses of statistical significance of the influence of input variables (such as throughput or conveying speed) on output variables (such as pressure or temperature).

In addition, artificial intelligence is used to investigate the scalability of PREs. With the large amount of data recorded, process parameter prediction models were trained using MATLAB-based Machine Learning algorithms.

From the selection of available Machine Learning algorithms, Gaussian Processes and Support Vector Machines (including the respective submodels) have proven to be most suitable for describing pressure, temperature and residence time of PREs. This can be described using the example of process pressure and the Machine Learning quality parameters RMSE and R². Support Vector Machines predict the processing pressure over a test dataset with a RMSE_Pressure,SVM of 0.604 ± 0.125 MPa. The pendant of the Gaussian Processes is a RMSE_Pressure,GP of 0.389 ± 0.016 MPa. The coefficient of determination is higher for the Gaussian Processes (R²_Pressure,GP = 0.64 ± 0.03) compared to the Support Vector Machines (R²_Pressure,SVM = 0.10 ± 0.38). To increase the prediction accuracy, especially for Support Vector Machines, a preliminary Machine Learning-based approach was further enhanced by developing a hybrid approach of Machine Learning algorithms including physical models. Thus, the new prediction model architecture consists of parallel structures using both physical and prediction models. If physical models are used, usually they show several simplifications that yield distinct deviations between calculated results and measured values. In the hybrid approach, the prediction model is no longer trained to predict process variables, but the deviation between the physical model and the measured values. The overall result, which is the prediction of the process variables, results from the addition of the physical model and the predicted deviation value. This complex algorithm improves the accuracy of the prediction results in comparison with real measured values. This is particularly evident in the coefficients of determination: the hybrid Gaussian Processes accuracy increases to R²_{Pressure,hybridGP} = 0.86 ± 0.05, and the accuracy of Support Vector Machines increases to R²_{Pressure,hybridSVM} = 0.57 ± 0.18.

The consolidation of the developed hybrid Machine Learning prediction models and the 3D-CFD-simulations have proven to be suitable for reducing resource-intensive physical experiments. Thus, physical experiments can be carried out on a laboratory scale. By using the scaling algorithms that have been newly developed, they can be transferred to pilot scale or industrial scale. Thus, this approach with its innovative tools helps to transform process engineering towards more sustainability by digitalization.

[Pla22] PlasticsEurope AISBL, Plastics – the Facts 2022, Brussels 2022.

[Rad23] Radwan, M.; Frerich, S.; Chem. Eng. Technol. 2023, 46 (00), 1-8. DOI: 10.1002/ceat.202200523.