Recurrent Networks for Plasma Time Series
Recurrent neural networks model the temporal dynamics of plasma signals for forecasting, event detection, and control.
Sequential plasma data
Plasma diagnostics produce long multi-channel time series whose meaning depends on temporal context: the same instantaneous value can be benign or alarming depending on its trajectory. Recurrent neural networks (RNNs) maintain an internal state that summarizes history, making them suited to this sequential data.
Gated architectures
Plain RNNs struggle to remember long-range dependencies because gradients vanish over many steps. Gated variants, long short-term memory and gated recurrent units, use learned gates to retain relevant information over longer spans, which matters when a precursor precedes an event by many milliseconds.
- Forecasting: predict signal evolution to anticipate events
- Detection: flag disruptions or mode onsets from evolving patterns
- State estimation: track a latent plasma state through time
- Control: provide temporal features to a policy
Real-time use
RNNs process data one step at a time, updating their state as new samples arrive, which fits streaming operation. Their fixed per-step cost supports deployment in real-time monitoring, provided the model is small enough to meet the latency budget.
Comparisons and limits
Transformers and temporal convolutions now rival RNNs on many sequence tasks, sometimes training faster and capturing longer contexts. RNNs remain attractive for strictly streaming, low-latency inference. Across all sequence models, the challenge is limited labeled events and transfer between devices.
Temporal modeling underlies disruption warning and event detection for any device. For concepts like the Hyperion breeder, RNN pipelines are trained on simulated and archival sequences before hardware exists; they process modeled signals because the machine is not yet built, and their outputs are computational estimates.