Autoencoders for Plasma Data
Compressing high-dimensional diagnostic data into a low-dimensional latent space for detection, denoising, and analysis.
What an autoencoder does
An autoencoder is a neural network trained to reproduce its input after squeezing it through a narrow bottleneck. The bottleneck forces a compact representation, the latent code, that captures the dominant structure of the data. It learns without labels, using only the reconstruction objective.
Uses in fusion
- Dimensionality reduction of profiles or image diagnostics
- Anomaly detection via reconstruction error
- Denoising noisy signals by reconstructing the clean structure
- Feature extraction feeding downstream classifiers
Latent structure
The latent space often organizes plasma states meaningfully: nearby codes correspond to similar plasmas, and trajectories through latent space track a shot's evolution. Variational autoencoders add a probabilistic structure to the latent space, supporting sampling and smoother interpolation.
Anomaly detection link
An autoencoder trained on normal operation reconstructs normal data well and anomalies poorly, because it never learned them. The reconstruction error becomes an anomaly score. This is a common backbone for unsupervised diagnostic monitoring.
Cautions
A bottleneck that is too wide can copy the input, including anomalies, defeating detection. Reconstruction error can be high for rare-but-normal states, causing false alarms. And the latent space, while useful, is not automatically interpretable; its axes rarely map to single physical quantities without extra effort. Autoencoders compress and flag; they do not by themselves explain.