Control & Fuelling Architectures

Overview

Phase-synchronized fuel injection is the core control mechanism of WAFI and forms part of our patent-pending control architecture. Rather than attempting to maintain fusion conditions across an entire plasma volume, we inject fuel into the thermal wake behind a propagating burn crest, where local conditions are expected to favor entrainment and ignition.

This requires real-time detection of crest position, wake structure, and plasma state, coupled with control systems that respond on timescales faster than the physics they're regulating. Control is not a secondary concern, it's the primary engineering challenge.

Wake Detection and Crest Tracking

Before wake-aligned injection is possible, the wake must be detected and tracked in real time.

Early reduced-order studies indicate strong phase sensitivity, with synchronized injection producing materially different burn dynamics compared to random delivery. These results are treated as hypotheses pending full 3D validation and transport-consistent modelling.

Crest position is estimated using phase-weighted spatial moments of the temperature field, producing a continuous position signal that is robust to noise and grid resolution.

This approach:

• Handles slow drift without discretization artifacts.
• Works on periodic domains (no wraparound discontinuities).
• Produces smooth position estimates suitable for derivative-based control.
• Extends naturally to 2D/3D through surface-aligned tracking.

Wake structure is identified by analysing temperature and density gradients in the vicinity of the crest, allowing detection of trailing thermal deficits characteristic of post-burn regions. Injection is triggered when wake conditions meet predefined entrainment criteria.

Phase-Locked Injection Control

Wake-aligned fuelling is fundamentally a phase control problem. Injection must occur within a narrow, moving window behind the burn crest where entrainment and ignition are physically favoured.

Too early and fuel is lost to ionization.
Too late and the wake has passed.

This makes fixed timing strategies ineffective. Injection must be continuously adjusted in response to burn-front motion, plasma state, and transport dynamics.

Our control architecture treats fuelling as a closed-loop alignment problem, not a scheduled event. Injection timing, rate, and duration are dynamically adapted to maintain phase coherence with the wake under drift, perturbation, and nonlinear response.

Early reduced-order studies show strong sensitivity to phase alignment, motivating a control-first design philosophy where feedback stability is a primary engineering constraint, not an optimisation layer.

All control strategies are developed and validated iteratively. No assumptions are treated as settled until they reproduce under higher-fidelity models and independent checks.

Machine Learning for Adaptive Control

Classical control methods perform well when system dynamics are well-characterised. Plasma systems are not.

We are investigating machine learning–assisted control strategies to handle nonlinear coupling between injection, transport, and burn dynamics where analytic models and fixed-gain controllers become brittle.

Rather than replacing physics with a black box, learned models are used to augment conventional control by capturing behaviours that are difficult to parameterise explicitly.

Current work focuses on:

• Learning phase-sensitive injection responses from simulation data.
• Adapting control policies across operating regimes and transport conditions.
• Handling delayed and noisy state information in crest tracking and wake detection.

Training is performed on reduced-order and ensemble simulation data for rapid iteration, with transfer to higher-fidelity 3D models used for validation and stress testing.

All learned controllers are constrained by physical priors derived from governing transport equations, and are integrated with deterministic control logic to ensure stability and interpretability.

Fail-safe fallback to classical control is maintained at all times. No ML system is permitted to operate without bounded behaviour and explicit stability guarantees.

Injector Hardware Architecture

We are exploring injector system architectures capable of delivering phase-synchronized fuel in magnetized plasma environments, with emphasis on responsiveness, robustness, and compatibility with closed-loop control.

Hardware concepts are developed in parallel with simulation and control design, and are not committed until validated through reduced-order and 3D modelling.

Next Milestones

• Integration of closed-loop ML-assisted control into BOUT++ workflows.
• Parameter sweeps: injection offset, timing jitter, control gain tuning.
• Multi-crest scenarios (what happens when two burn waves collide?).
• Preparation for bench-scale validation in non-fusible plasma systems.
• Preparation of technical communications and preprint material on phase-aligned fuelling concepts.