Using Machine Learning to Differentiate Types of Particle Jumps in DNS Data of Sediment Transport

In this study, machine learning is used to identify types of jumps that share motion patterns of saltating particles in sediment transport based on DNS data. For this purpose, a framework is proposed, consisting of a partitioning of jumps and a subsequent calculation of prototypical jumps. The partitioning is achieved through cluster ensembles that aggregate the results from numerous k-means clustering models. The inputs for these models consist of large numbers of features per jump, extracted via time series analysis. As a result, four different types of jumps, associated with different stages of transport, are identified. Corresponding prototypical jumps are computed using a modified variant of dynamic time warping barycenter averaging and serve as models for the types. The framework contributes to improving the understanding of particle saltation.