Matvey Nikitenko-Valiakhmetov: Clustering Methods for Atmospheric Vortices in the Barents and Kara Seas

At the 60th seminar of the MIPT Center for Earth Sciences, held on 13 August 2026 at 16:30 Moscow time (13:30 UTC), Matvey Nikitenko-Valiakhmetov, a student at the Department of Oceanology of Saint Petersburg State University, presented preliminary results from the study “Clustering Methods for Atmospheric Vortices in the Barents and Kara Seas.” The research is supervised by Tatiana V. Belonenko of Saint Petersburg State University (Saint Petersburg), with Alexander V. Gavrikov of the Shirshov Institute of Oceanology, Russian Academy of Sciences (Moscow), serving as scientific consultant.

The source dataset comprised atmospheric vortices tracked manually in output from a regional WRF simulation covering part of 2019. The talk described about 1,700 trajectories over eight and a half months at 6 km horizontal resolution and hourly time steps. Thermodynamic, kinematic, and geometric characteristics were calculated for each trajectory at genesis and at maximum development, defined using the Rortex criterion. About 15 features were retained from the initial set, including measures of wind, pressure, temperature, humidity, precipitation, vertical motion, vorticity, vortex radius, and propagation. Correlated pairs were reduced using a Pearson linear-correlation matrix and a threshold of 0.7, which the speaker explicitly described as a subjective choice.

Matvey reviewed the principles of k-means, Ward hierarchical clustering, Gaussian mixture models, and self-organizing maps. K-means was used for the current preliminary calculation. Cluster counts were assessed using inertia and several quality and reproducibility measures, repeated subsampling, and comparisons with noise datasets. A seven-cluster solution was examined in detail. t-SNE and UMAP projections were used to explore how objects were arranged, but only as diagnostic visualizations rather than independent evidence that the groups were separable.

In the reported calculations, some physical cluster characteristics were similar at vortex genesis and maximum development. Trajectories matched in space and time to two independent polar-low databases were concentrated in more than one cluster. This was discussed as a possible indication that the vortices are heterogeneous and as a working hypothesis for further study, not as an established physical classification. The matching is also limited by differences among the model, satellite products, and independent catalogues, so trajectory proximity alone does not establish that cases are identical.

The main discussion focused on methodology. Participants suggested defining one quality criterion in advance and applying it consistently when comparing algorithms and cluster counts, testing alternative methods and hyperparameters, and evaluating whether groups are statistically distinguishable. They also emphasized that a clustering algorithm optimizes its own objective and does not guarantee a physically objective typology: expert knowledge may guide interpretation but can also introduce bias. The seven clusters, the possible subdivision of polar lows, and the similarity between development stages therefore remain preliminary observations requiring further validation and physical justification.