Controllers

Classical proportional-integral-derivative. Anti-windup (clamping + back-calculation). Gain-scheduled interpolation on (I_p, β_N) operating point. Derivative filter. Rust-accelerated.

Gradient-based model predictive control. 20-step prediction horizon. Quadratic cost with terminal penalty. Trajectory optimisation via iterative QP. Warm-start from previous solution.

Nonlinear MPC via sequential quadratic programming (SQP). Box constraints on actuators. Slew-rate limits. Nonlinear plant model in prediction. RTI (real-time iteration) variant.

Robust H-infinity synthesis via DARE (Discrete Algebraic Riccati Equation). ZOH discretisation. Padé scaling-and-squaring for matrix exponential. Guaranteed worst-case performance.

D-K iteration for structured singular value μ. D-scaling optimisation. Handles parametric uncertainty (density, temperature, geometry). Robust stability + performance.

Continuous control law with dead-band saturation to eliminate chattering. Designed for vertical stability control. Disturbance rejection without plant model accuracy.

PID gains interpolated across operating regime. Scheduling variables: plasma current I_p, normalised beta β_N, elongation κ. Smooth bumpless transfer between regimes.

Innovation monitoring for sensor/actuator fault detection. Reduced-rank reconfiguration on fault. Graceful degradation. Kalman-based residual generation.

Stochastic Petri Net → LIF neuron compilation. Bitstream encoding of control signals. Event-driven computation. Designed for neuromorphic hardware deployment.

Proximal Policy Optimisation. 500K-step cloud-trained on Vertex AI. Gymnasium-compatible tokamak environment. Reward 143.7 vs MPC 58.1 vs PID -912.3. 0% disruption rate.

Full-state tokamak simulator. Couples GS equilibrium, transport, MHD, and controller in a single time-stepping loop. Configurable physics fidelity per tier.

Real-time operator training mode. Scenario injection (VDE, disruption, loss of heating). Dashboard visualisation of plasma state and actuator commands.

Equilibrium reconstruction from synthetic diagnostics. Magnetic probe + flux loop + MSE constraint fitting. State estimator for controller feedback.

Hardware-in-the-loop test harness. Controller runs on target hardware, plant model runs in software. Deterministic timing via ZMQ transport.

Gradient-boosted classifier on 12 plasma features. Trained on synthetic disruption database. Configurable alarm threshold. False positive rate tracking.

Shattered Pellet Injection model. Neon/deuterium mixed pellets. Radiation fraction, current quench time, halo current mitigation assessment.

Halo current model (toroidal peaking factor). Runaway electron avalanche (Rosenbluth-Putvinski). RE beam termination dynamics.

Federated learning framework for multi-device disruption prediction. Privacy-preserving gradient aggregation. Cross-machine transfer learning.