Robust Control of ECH Deposition Profiles on DIII-D

The paper presents the development and experimental validation of the real-time ECH Optimization (ECHO) algorithm on DIII-D, which utilizes a neural network surrogate of the TORBEAM code and a genetic optimizer to robustly control electron cyclotron heating deposition profiles despite gyrotron failures and plasma parameter variations.

The Gist

The Big Picture: The "Smart Thermostat" for a Star

Imagine a Tokamak (the machine in the DIII-D experiment) as a giant, super-hot oven trying to cook a star. To keep this star stable and hot enough to create energy, scientists need to shoot beams of microwave energy (called Electron Cyclotron Heating, or ECH) into very specific spots inside the oven.

Think of these microwave beams like spotlights shining into a dark room.

• The Problem: The "room" (the plasma) is constantly moving, changing shape, and sometimes the spotlights (gyrotrons) break. If you aim a spotlight at a wall that suddenly moves, the light hits the wrong spot. If a spotlight breaks, you have a dark patch.
• The Old Way: Scientists used to program the spotlights to aim at a specific spot before the experiment started. If the room moved or a light broke, the aim was wrong, and the experiment could fail.
• The New Way (ECHO): The researchers built a "smart brain" called ECHO. It acts like a super-fast, self-correcting thermostat. It constantly checks where the room is, checks which lights are working, and instantly tells every spotlight exactly where to point and how bright to shine to hit the perfect target.

How the "Smart Brain" Works

The paper describes a two-part system that makes this possible:

1. The Crystal Ball (TorbeamNN)
To know where the light will land, you usually need to run a complex physics simulation. But these simulations are slow—like trying to calculate the weather by hand while driving a car.

• The Innovation: The team trained an Artificial Intelligence (AI) model called TorbeamNN. Think of this AI as a "crystal ball" that has memorized millions of physics simulations.
• The Speed: Instead of taking 50 milliseconds to calculate where the light goes, the AI does it in 0.3 milliseconds. It's like swapping a slow calculator for a supercomputer. This allows the system to make decisions faster than the plasma can move.

2. The Chess Master (The Genetic Optimizer)
Once the AI knows where the light can go, the system needs to decide which lights to use and how to aim them to match a specific shape (the "target profile").

• The Process: Imagine you have 10 spotlights and you need to paint a specific shape on the wall. You could try every combination, but that takes forever. Instead, the "Genetic Optimizer" acts like a chess master.
• Evolution: It tries a few random arrangements of the lights. It sees which ones look closest to the target. It keeps the best ones, mixes their settings (like mixing two good recipes), and makes tiny random tweaks. It repeats this process thousands of times in a split second until it finds the perfect arrangement.

What Happened in the Experiments?

The team tested this system on the DIII-D machine and proved it works in three tricky scenarios:

1. The Moving Target (Changing Plasma)

• The Scenario: The plasma inside the machine moved up and down by 10 centimeters (a huge distance for a particle).
• The Result: The ECHO system noticed the movement immediately. It adjusted the angles of the mirrors on the gyrotrons so the beams stayed locked onto the same spot relative to the plasma, even though the plasma itself was dancing around.

2. The Broken Light (Hardware Failure)

• The Scenario: One of the gyrotrons (a spotlight) suddenly died in the middle of the experiment.
• The Result: In the past, this would ruin the experiment. ECHO, however, instantly realized, "Oh, we lost one light." It immediately recalculated the plan, telling the remaining lights to shift their positions and power to fill in the gap. The target shape was maintained almost perfectly despite the broken part.

3. The Changing Rules (Magnetic Field Shifts)

• The Scenario: The magnetic field holding the plasma together was changed drastically.
• The Result: The system adapted the aim of the beams to compensate for the new physics, showing it could handle extreme changes in the environment.

Why This Matters

The paper claims this system is a major step forward because it is robust.

• Old Systems: If you lose a part, the whole plan fails.
• ECHO System: It treats the gyrotrons as a team. If one teammate drops out, the others instantly adjust to finish the job.

The authors conclude that this technology is ready for future fusion power plants (FPPs). In a real power plant, you can't afford for the machine to shut down just because one heater breaks. ECHO provides the "fail-safe" intelligence needed to keep the fusion reaction running smoothly, even when things go wrong.

Summary

The paper presents a new control system (ECHO) that uses a fast AI to predict where microwave beams will land and a smart algorithm to instantly adjust those beams. This allows the system to hit a precise target inside a fusion reactor, even if the reactor moves, changes shape, or loses a piece of equipment. It turns a fragile, pre-programmed process into a flexible, self-correcting one.

Technical Summary

Technical Summary: Robust Control of ECH Deposition Profiles on DIII-D

Problem Statement
Electron Cyclotron Heating (ECH) is a critical actuator for tokamaks, providing auxiliary heating, localized current drive, and MHD stability control. Unlike other RF heating methods that rely on antennae and fixed deposition locations determined by plasma parameters, ECH utilizes steerable mirrors to adjust ray trajectories and deposition locations in real-time. However, achieving precise control over the total ECH deposition profile is challenging due to the complexity of ray tracing, the need to coordinate multiple gyrotrons simultaneously, and the requirement for robustness against hardware failures and dynamic plasma conditions. Existing real-time ray tracing codes (e.g., reduced TORBEAM) are often too slow to explore multiple configurations for optimization or only calculate peak absorption locations rather than full deposition profiles. Furthermore, traditional control strategies often treat gyrotrons as independent units rather than a collective system, limiting the ability to multitask or recover from actuator failures.

Methodology
The authors developed the ECH Optimization (ECHO) algorithm, a real-time control framework designed to match a target ECH radial deposition profile by optimizing the mirror angles and power (via duty cycle modulation) of all available gyrotrons. The methodology consists of three primary components:

1. TorbeamNN Surrogate Model: To overcome the computational latency of full-fidelity ray tracing (TORBEAM), the authors trained a machine learning (ML) surrogate model based on the TORBEAM code. This neural network takes real-time plasma state parameters (equilibrium shape, electron temperature, and density profiles) and gyrotron mirror angles as inputs. It outputs parameterized Gaussian profiles (center $\mu$, width $\sigma$, and peak amplitude $A$) for the ECH deposition. The model was trained on over 2.2 million ray trace calculations from 4,168 unique DIII-D shots. To ensure real-time performance, the model runs in parallel for each gyrotron and poloidal angle, generating a lookup table in approximately 0.32 ms (a >100x speedup over optimized real-time TORBEAM).
2. Genetic Algorithm Optimizer: A genetic optimizer searches the pre-computed lookup table to find the optimal combination of mirror angles and duty cycles for the available gyrotrons. The algorithm minimizes the mean-square error between the target profile and the predicted total deposition profile. It incorporates "inertia" by preserving a percentage of the best solutions from the previous cycle to prevent unnecessary hardware movement if conditions are stable, while still allowing rapid adaptation to significant changes.
3. PCS Implementation: The system is integrated into the DIII-D Plasma Control System (PCS). It receives real-time plasma state data from RT-EFIT and RTCAKENN, computes the optimal configuration, and sends hardware commands to the gyrotrons. The entire control loop operates on a 50 ms cycle time.

Key Contributions

• Real-Time Full Profile Optimization: Unlike previous ML surrogates that focused only on peak absorption, TorbeamNN predicts full ECH deposition profiles, enabling the optimization of the entire radial heating distribution.
• Coordinated Multi-Gyrotron Control: The framework treats the gyrotron array as a collective unit, intelligently distributing power and angles to match a single target profile rather than assigning independent tasks to individual gyrotrons.
• Robustness to Faults: The algorithm is designed to be actuator-independent. In the event of a mid-shot gyrotron failure, ECHO immediately re-optimizes the remaining functional gyrotrons to best approximate the target profile, rather than relying on pre-programmed scenarios that would fail.
• Adaptability to Dynamic Conditions: The system automatically adjusts mirror angles and powers in response to changing plasma conditions, such as vertical plasma shifts or variations in the toroidal magnetic field ($B_T$), to maintain a constant deposition profile relative to the normalized flux coordinate ($\rho$).

Experimental Results
The ECHO algorithm was deployed and validated in several DIII-D experiments (e.g., shots 205902, 205838, 205907, 205905):

• Validation: ECH deposition profiles predicted by TorbeamNN and the resulting hardware commands were validated against Electron Cyclotron Emission (ECE) measurements using Flux Fit (FF) and other inverse methods. The centers of the measured deposition profiles matched the target and TORBEAM predictions within $\lesssim 0.05$ in $\rho$.
• Dynamic Tracking: The system successfully tracked a constant target profile while the plasma underwent significant vertical shifts (>10 cm) and rapid changes in the toroidal magnetic field, adjusting mirror angles accordingly.
• Fault Handling: In shot 205907, a gyrotron failure (Gyrotron 5) occurred mid-shot. ECHO detected the fault and re-optimized the remaining gyrotrons (1, 2, and 4) to redistribute the power, successfully recovering the deposition profile. A similar fault mitigation was demonstrated in shot 205838 where Gyrotron 4 failed.
• Profile Shaping: The algorithm demonstrated the ability to rapidly switch between different target profiles (e.g., single-peaked to double-peaked) within a single discharge.

Significance and Claims
The paper claims that ECHO represents a significant advancement in real-time plasma control by bridging the gap between offline integrated modeling and real-time hardware actuation. By utilizing a machine learning surrogate for fast ray tracing and a genetic optimizer for collective actuator management, the system achieves:

1. Precision: Accurate matching of complex ECH deposition profiles in real-time.
2. Robustness: The ability to maintain control objectives despite unpredictable gyrotron failures, a critical requirement for future commercial fusion power plants (FPPs) where actuator redundancy and fault tolerance are essential.
3. Flexibility: The capacity to support diverse physics experiments (e.g., tearing mode suppression, impurity pump-out, sawtooth control) without requiring custom controller designs for each scenario.

The authors note that while the current implementation uses a Gaussian parameterization which may have limitations near the plasma core and edge, the achieved accuracy is sufficient for real-time control and qualitatively superior to typical feedforward approaches. The framework is presented as a scalable solution for future devices like ITER and FPPs, where the number of gyrotrons and the need for failure-robust control will be even more critical.