Real-time virtual circuits for plasma shape control via neural network surrogates: dynamic validation in closed-loop simulations

This paper demonstrates that neural network emulators of virtual circuits can robustly and effectively control MAST Upgrade plasma shapes in real-time closed-loop simulations, offering a low-latency alternative to traditional physics-based controllers for future fusion devices.

The Gist

Imagine you are trying to steer a giant, invisible, super-hot balloon made of fire (a plasma) inside a donut-shaped machine called a tokamak. This is the goal of fusion energy research. The problem? This "fire balloon" is incredibly unstable. If you don't steer it perfectly, it wobbles, hits the walls, and the reaction stops.

To keep it stable, scientists use giant electromagnets (like invisible hands) to push and pull the plasma into the exact shape they want. This is called plasma shape control.

The Old Way: The "Pre-Planned Route"

Traditionally, scientists had to calculate exactly how to move the magnets before the experiment started. They would draw a map: "At 1 second, move magnet A. At 2 seconds, move magnet B."

The Problem: This is like driving a car using a map drawn for a specific day with no traffic. If the wind changes, or a road is closed (the plasma behaves slightly differently than expected), that pre-drawn map becomes useless. The car (the plasma) starts to drift off course, and the magnets have to work harder and harder to fix it, sometimes causing stress or failure.

The New Way: The "AI Co-Pilot"

This paper introduces a new, smarter way to steer the plasma using Artificial Intelligence (Neural Networks).

Instead of a static map, the scientists built a super-smart AI "co-pilot." This AI has studied millions of simulated scenarios (a library of "what-if" situations). It knows exactly how the plasma will react to every possible move the magnets can make.

Here is how the new system works, using some everyday analogies:

1. The "Virtual Circuit" (The Magic Remote)

Imagine you have a remote control with many buttons. In the old days, pressing "Up" might also accidentally move "Left" or "Forward." It was messy.
The scientists created a "Virtual Circuit" (VC). Think of this as a magic remote where pressing "Up" only moves the plasma Up, without touching anything else. It perfectly decouples the controls.

• The Innovation: Usually, calculating this "magic remote" takes a long time (like solving a complex math puzzle). But their AI can calculate the perfect "magic remote" settings in a split second, based on what the plasma is doing right now.

2. The "Dynamic Validation" (The Flight Simulator)

You wouldn't fly a real plane for the first time without a simulator, right?
The researchers built a Flight Simulator (called the FreeGSNKE Pulse Design Tool). They put their AI co-pilot inside this simulator and let it steer the plasma through a virtual experiment.

• The Test: They didn't just check if the AI knew the math (Static Validation). They let the AI steer the plasma while the plasma was moving, changing shape, and reacting to the magnets (Dynamic Validation).
• The Result: The AI kept the plasma perfectly stable, even when the "wind" (plasma conditions) changed.

3. Handling the "Real World" Mess

In the real world, sensors aren't perfect. Sometimes they give a slightly wrong reading (like a speedometer that says 60 mph when you're actually going 61).

• The Test: The researchers added "noise" (fake errors) to the data the AI received.
• The Result: The AI didn't panic. It was robust. Even with imperfect information, it kept the plasma on track. It also handled situations where the "fuel" (plasma heating) changed unexpectedly, like if one of the heating lasers broke down.

Why This Matters

Think of the old method as driving with a paper map that you can't update. If you miss a turn, you're lost.
This new method is like driving with Google Maps in real-time. It constantly recalculates the best path based on current traffic (plasma behavior).

The Big Picture:
This research proves that we can use AI to steer fusion plasmas in real-time. This is a huge step toward making fusion power plants that are:

1. Safer: They won't crash as easily.
2. Smarter: They can adapt to changes instantly.
3. Easier to run: We won't need teams of experts to manually calculate every single move before the experiment starts.

In short, the scientists have taught a computer how to be the ultimate pilot for the "star in a jar," bringing us one step closer to clean, limitless energy.

Technical Summary

1. Problem Statement

Accurate real-time magnetic control of tokamak plasma shape and position is critical for maintaining confinement and accessing high-performance operating regimes. Current Plasma Control Systems (PCS) rely on Virtual Circuits (VCs) to decouple the control of multiple shape parameters (e.g., X-point location, strikepoints, radii) by calculating the necessary adjustments to Poloidal Field (PF) coil currents.

However, traditional VC implementation faces significant limitations:

• Offline Calculation: VCs are typically pre-calculated offline for specific reference equilibria using linearization of the Grad-Shafranov (GS) equation.
• Local Validity: These pre-set VCs are only valid near their reference equilibrium. As the plasma evolves during a discharge, the VC matrix degrades, leading to coupled (non-orthogonal) responses where adjusting one parameter inadvertently affects others.
• Complexity: To maintain accuracy, operators must use complex, segmented control strategies with numerous pre-computed VCs, requiring significant expert intervention and limiting flexibility.

The challenge is to develop a method that provides state-aware, continuously updated VCs in real-time to handle evolving plasma equilibria without the latency of solving the full GS equilibrium equations online.

2. Methodology

The authors propose using Neural Network (NN) surrogates to emulate the VC matrix in real-time. The methodology involves three main stages:

A. Neural Network Emulation

• Architecture: Feedforward neural networks are trained to map input plasma states to geometric shape parameters ($P$).
• Inputs: Active PF coil currents ($I_{act}$), total plasma current ($I_p$), and current density profile parameters ($\theta$).
• Outputs: Geometric shape parameters ($P$), including midplane radii, X-point positions, and divertor gaps.
• VC Derivation: The VC matrix ($V$) is derived by differentiating the NN output with respect to the inputs to obtain the sensitivity matrix ($S$), then computing the pseudoinverse: $V = (S^T S)^{-1}S^T$.
• Training Data: A library of ~900,000 synthetic plasma equilibria generated using the FreeGSNKE solver, covering a wide operational space of the MAST Upgrade (MAST-U) tokamak.

B. Dynamic Validation Framework (FPDT)

To move beyond "static validation" (checking accuracy on a single snapshot), the authors deployed the NN models within the FreeGSNKE Pulse Design Tool (FPDT).

• Closed-Loop Simulation: The FPDT couples the FreeGSNKE evolutive equilibrium solver with a virtual PCS.
• Interaction: The virtual PCS regulates plasma current and shape using a PID controller. It queries the NN emulator at each time step to generate the current VC matrix based on the latest simulated "measurements."
• Realism: The simulation includes passive conducting structures, actuator constraints, and time-dependent plasma profiles, mimicking a real discharge environment.

C. Robustness Testing

The study tested the emulators under three specific stress conditions:

1. Update Frequency: Testing VC update intervals ($\gamma$) of 2 ms (baseline) and 20 ms (conservative) to simulate hardware latency constraints.
2. Measurement Uncertainty: Adding Gaussian noise to simulated sensor data ($I_{act}, I_p, P$) to mimic real-world diagnostic noise.
3. Profile Mismatch: Using plasma current density profiles ($\theta$) from different reference shots (including scenarios with varying Neutral Beam Injection power) in the simulator while the emulator was trained on a different reference profile.

3. Key Contributions

• Dynamic Validation: This is the first study to validate NN-emulated VCs in a closed-loop, time-evolving simulation rather than just static snapshots. It accounts for the interaction between controllers, passive structures, and evolving plasma currents.
• Seamless Integration: Demonstrated that NN emulators can be integrated into a standard PCS architecture (PID + Virtual Circuits) without modifying the underlying control algorithms.
• Robustness Proof: Provided evidence that the system remains stable and accurate despite measurement noise, latency in VC updates, and discrepancies in plasma profile assumptions.

4. Results

The simulations were performed on MAST-U scenarios (Shot 53152 and Shot 53278) with the following outcomes:

• Update Frequency Robustness: The emulated VCs performed effectively with update intervals of both 2 ms and 20 ms. Even with the conservative 20 ms delay, controlled shape parameters remained within a few centimeters of the reference waveforms.
• Measurement Noise: When Gaussian noise was introduced to sensor inputs, the system successfully tracked the reference waveforms. The NN emulators and the virtual PID controller compensated for the noise, showing no significant degradation in performance compared to noise-free simulations.
• Profile Uncertainty: The system maintained robust shape control even when the simulator used plasma profiles from shots with significantly different heating powers (e.g., 3.4 MW vs. 1.8 MW NBI) compared to the emulator's training reference. Core shape parameters showed expected deviations, but divertor geometry remained stable.
• Passive Structure Handling: The emulators maintained performance despite evolving currents in passive conducting structures, which were not explicitly included in the NN training data.

5. Significance and Impact

• Real-Time Viability: The results confirm that NN-emulated VCs are fast and robust enough for real-time deployment in operational tokamaks like MAST-U.
• Operational Flexibility: This approach eliminates the need for pre-calculated, segmented VC schedules. Operators can change target plasma shapes dynamically during a discharge without offline re-computation.
• Automation: By reducing reliance on expert offline calculations and complex feedforward strategies, this technology paves the way for fully automated, adaptive plasma control systems.
• Next Steps: This work represents a critical "dynamic validation" step, bridging the gap between theoretical NN models and the eventual experimental verification and deployment in a live Plasma Control System.

In conclusion, the paper establishes that neural network surrogates can effectively replace traditional, static VC calculations, offering a flexible, robust, and low-latency solution for next-generation tokamak plasma shape control.