Dimensionality Reduction for Plasma Data
Compressing many correlated diagnostic channels into a few informative coordinates.
Why reduce dimensions
A plasma is described by many correlated signals, but its state often lives on a much lower-dimensional manifold. Dimensionality reduction finds a compact set of coordinates that retains the important structure, aiding visualization, speeding downstream models, and revealing regime structure.
Linear methods
Principal component analysis (PCA) finds the directions of greatest variance and projects onto them. It is fast, interpretable, and often a strong baseline: a handful of principal components can capture most of the variation in a set of profiles or signals. Its limitation is that it only captures linear structure.
- PCA for variance-preserving linear projection
- Independent component analysis for separating mixed sources
- Nonnegative matrix factorization for additive, interpretable parts
Nonlinear methods
When the data manifold is curved, nonlinear methods help. Autoencoders learn a compact latent code; t-SNE and UMAP produce low-dimensional embeddings good for visualization. The latter two are for exploration, not for producing features to compute on, because their embeddings are not stable or metric-preserving in a usable way.
Uses in fusion
Reduced representations help visualize how shots cluster by regime, compress data for storage or transmission, provide inputs to classifiers and controllers, and reveal when a new shot falls outside familiar territory. They are a step toward, not a substitute for, physical understanding.
Cautions
Reduced coordinates rarely map cleanly to single physical quantities, and discarding low-variance directions can drop rare but important signals. What looks like noise in variance terms may be the precursor that matters. Reduction is chosen with the downstream task in mind, not as a blind preprocessing default.