AI Plasma Control

Neural Networks in Real-Time Control

Using trained networks as fast surrogates, estimators, or controllers inside the real-time loop - with the verification burden that implies.

Networks as fast function approximators

A trained neural network evaluates in fixed, bounded time - a fixed number of matrix multiplications - which makes it attractive inside a real-time loop where deterministic timing matters. Networks appear in control as surrogates for slow physics, as estimators mapping diagnostics to state, and as learned controllers mapping state to commands.

Three roles

Kronos motion — control room

Surrogate: replace an expensive model (transport, equilibrium) with a fast learned approximation
Estimator: infer profiles or stability proximity directly from raw diagnostics
Predictor: forecast disruptions or limit approach for the predictive layer
Controller: output actuator commands directly, usually trained by RL

The determinism advantage

Unlike an iterative solver whose runtime depends on the data, a fixed network architecture has a fixed operation count and thus a predictable worst-case latency. This suits the hard-real-time budget of control - provided the network runs on hardware (FPGA, dedicated accelerator) that itself has deterministic timing, without unpredictable memory or scheduling behavior.

The verification problem

A network's behavior is defined by its weights, not by readable logic, so it is hard to prove what it will do on inputs unlike its training data. This is acceptable for advisory roles and for controllers wrapped in hard safety limits, but it disqualifies networks from the safety-critical reflex layer, which must remain simple and formally verifiable.

Guardrails

Networks in the loop are bounded by conventional limits, monitored for out-of-distribution inputs, and always subordinate to the deterministic safety layers. A network that sees inputs unlike its training set should reduce its influence, not act confidently. Used this way - fast where speed helps, fenced where safety demands - networks extend the stack without weakening its guarantees.