2. 2/10
Motivation/Objective
Ï To improve the GEP-RANS simulations
Ï To improve the agreement with experimental data
Research Questions
Ï What features are useful for clustering turbulence data?
Ï Can continuous statistical field data be clustered?
Ï Can dataset be decomposed into partitions, each encoding a
particular type of turbulence physics
Ï Are machine-learned clusters consistent with our human
understanding of turbulent flows?
Ï Can we automatically identify a relatively small number of
"exemplars" or "prototypes", from each cluster?
Hypothesis: Unsupervised learning algorithms can be applied to
turbulence data to produce an automated partitioning of data that
reconciles with our understanding of turbulence.
4. 4/10
Technical challenges
Ï Find a physical-aware feature
space (clustering)
Ï Any turbulent state is
approximately a combination of
3 limiting states
Ï Clustering with Gaussian
Mixture Modeling (GMM)
Ï A greedy feature search is
wrapped around GMM
Ï Cluster extent set as a
threshold on the
Mahalanobis distance of
each data point
Turbulence-derived feature space
+ greedy clustering
+ dM thresholds
= interpretable clusters
5. 5/10
Ï Qualitative understanding of turbulence’s type revealed by clustering
Ï Quantitative understanding requires prototypes and analysis
Ï Incremental classification of complex dataset using learned clusters1
1Feature Selection, Clustering, and Prototype Placement for Turbulence
Datasets, https://doi.org/10.2514/6.2021-1750