ML-based visual analysis of droplet behaviour in multiphase flow simulations

Abstract

Modern multiphase flow solvers can simulate flows with increasing domain size and precision. This produces large simulation results which need to be analyzed, and to this end visualized. Because of the amount of data, classical visualization approaches become more and more unfitting. Therefore, it is hard to find interesting regions because of visual clutter which is produced by too much data. One solution could be semi automatic assistance systems to support the observer of the visualization. Over the last years, machine learning has grown to a widely researched area. Development not only brought many use cases in research and industry, but highly sophisticated programming frameworks. This makes it much easier to use machine learning in a wide area of applications, such as visualization. In this work we are interested in analyzing multiphase simulations with thousands of droplets. We use machine learning to train artificial neural networks with the droplet data gained from simulations. These trained models are used for finding interesting droplet behavior in the simulation, which is then visualized. Our trained models can predict the development of physical properties and quantities over time, and therefore errors in prediction can guide us to areas of interest which then can be investigated further. The prediction error is visualized as colored dots directly within the 3D simulation dataset using ParaView. Additionally we can plot the properties and their predictions of single droplets over time and show the prediction error separated by property within a spider chart. Finally we show the results, which cover an evaluation of the learning process and an analysis of the used datasets with our method, as well as give an outlook on possible improvement in future work.