Clustering Plasma Regimes
Grouping discharges or time slices into regimes without labels, to organize data and reveal structure.
The task
Clustering partitions data into groups of similar points without using labels. In fusion it can organize a large shot database into regimes, separate confinement modes, or reveal that a campaign contains distinct operating states, all before any expert labeling.
Methods
- k-means: fast, assumes roughly spherical clusters
- Gaussian mixtures: soft assignments with covariance structure
- Hierarchical clustering: a tree of nested groupings
- Density-based clustering: finds arbitrary shapes, flags outliers
Feature choice drives results
Clusters reflect the features and distance used. Clustering on raw signals may group by machine settings; clustering on physics-normalized features may group by regime. There is no single correct clustering; the useful one depends on the question, so feature choice is a modeling decision, not a preprocessing detail.
How many clusters
Most methods need the number of groups, or a resolution, chosen. Metrics like the silhouette score, gap statistic, or held-out likelihood guide this, but the choice is partly interpretive. A clustering is validated by whether its groups correspond to something meaningful, checked against known regimes where possible.
Uses and limits
Clustering supports data exploration, sampling for balanced training sets, and detecting novel operating states that form their own group. It does not explain why groups differ; that requires follow-up analysis. Treated as a lens for organizing and questioning data, rather than as a source of truth, it is a valuable early step in any data-driven study.