Observation Based Modeling of Liquid Slosh in Drop Tests
Abstract: Tetra Pak® is a world leading food packaging company with a large variety of paperboard beverage packages. In the development of paperboard packages, an essential component is utilizing simulation models long before physical prototypes are manufactured. This speeds up the development process and allows quick evaluation of new designs. For this to be an effective approach, confidence in the virtual models is a key aspect. In the context of paperboard beverage packages, liquid product sloshing in drop tests is currently not adequately modeled in existing simulations. In this thesis, a framework for gathering data related to liquid product motion in drop tests, in a controlled and repeatable manner, is built, using image analysis and computer vision algorithms. Both the quantitative and observational results are successfully used to improve and verify new liquid sloshing simulation models at Tetra Pak®. In parallel, the collected data is utilized to create approximate data driven models using neural networks. Three different model architectures are implemented and evaluated, U-Net, convolutional LSTM, and a graph convolutional LSTM. The resulting U-Net data driven model achieved the best performance and is shown to sufficiently approximate liquid behavior in drop tests in the setting of dropped transparent bottles. The U-Net model is found to be an adequate complement to the physics based simulations, offering faster run times but with reduced accuracy.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)