Interpreting AI for Fusion: an application to Plasma Profile Analysis for Tearing Mode Stability

This paper presents a physics-based interpretability framework using Shapley analysis to decode an AI model's predictions of tearing mode stability in tokamaks, revealing that core electron temperature and rotation peaking are the primary drivers of stability in DIII-D experiments.

The Gist

Imagine you are trying to keep a giant, swirling pot of soup (the plasma) from boiling over and splashing everywhere. In the world of nuclear fusion, this "soup" is superheated gas that we want to use to create clean energy. The problem is that this soup is incredibly unstable. It wants to form ripples and tears (called Tearing Modes) that can ruin the whole experiment, much like a sudden wave capsizing a boat.

For a long time, scientists have used complex math to predict these waves. But recently, they started using Artificial Intelligence (AI) because it's incredibly good at spotting patterns and predicting trouble before it happens.

However, there's a catch: AI is often a "Black Box." It gives you a prediction ("Danger! Wave coming in 5 seconds!"), but it doesn't tell you why. It's like a weather app saying "Bring an umbrella" without explaining if it's raining, snowing, or if a giant bird is about to drop a bucket on you. For a fusion power plant, we can't just trust a black box; we need to know why it's worried so we can fix the actual problem.

The Solution: The "AI Detective"

This paper introduces a new way to use AI. The team built a "Black Box" AI to predict these plasma tears, but then they added a special tool called Shapley Analysis.

Think of Shapley Analysis as a team of detectives or a fair referee. When the AI makes a prediction, this tool breaks down the decision and asks every single piece of data: "How much did you contribute to this prediction?"

It's like a group of friends trying to guess the ending of a movie. The detective asks:

• "Did the temperature of the soup make you think it would boil over?"
• "Did the speed of the spinning pot make you think it would stabilize?"
• "Did the density of the ingredients make it worse?"

By adding up everyone's contribution, the detectives can tell the scientists exactly which factor caused the AI to sound the alarm.

What Did They Find? (The "Recipe" for Stability)

The team tested this on the DIII-D machine, a giant fusion reactor in California. They ran a special experiment where they tried to stop the "tears" before they happened by changing how they heated the plasma.

Here is what their "AI Detective" revealed about the plasma soup:

1. The Core Temperature is the Villain: The AI learned that if the very center of the plasma gets too hot (like the middle of a pot getting scorching), it becomes unstable and wants to tear.
2. Spinning is the Hero: If the plasma spins fast in the center, it acts like a gyroscope, keeping everything stable.
3. The "Edge" is Different: Interestingly, having a hot "crust" or edge (the pedestal) actually helps stabilize the soup, which is the opposite of what happens in the center.
4. Density is a Minor Player: The amount of "stuff" in the soup mattered less than the temperature and spin.

The Real-World Test: Stopping the Wave

In the experiment, the AI predicted a "tear" was coming.

• Without help: The plasma tore, and the experiment failed.
• With help: The scientists used the AI's warning to change the heating beam. Instead of heating the center, they aimed the heat at a specific spot (the $q=2$ surface) to fix the "spin" and "current" right where the tear was about to start.
• The Result: The AI predicted the danger, the scientists fixed the specific ingredient causing the trouble, and the plasma stayed stable!

Why This Matters

This paper is a big deal because it bridges the gap between magic and science.

• Before: We had a magic box that said "Danger!" but we didn't know why.
• Now: We have a magic box that says "Danger! And here is the reason: The center is too hot and spinning too slow. Let's fix that."

This makes AI safe and trustworthy for future fusion power plants. Instead of blindly following a computer, engineers can use these insights to tweak the "recipe" of the plasma, ensuring we can harness the power of the stars without the soup boiling over.

In short: They taught a computer to predict fusion disasters, then taught the computer to explain its reasoning in plain English, allowing humans to fix the problem before it happens.

Technical Summary

Here is a detailed technical summary of the paper "Interpreting AI for Fusion: An Application to Plasma Profile Analysis for Tearing Mode Stability."

1. Problem Statement

While Machine Learning (ML) and Artificial Intelligence (AI) have demonstrated high predictive accuracy for tokamak instabilities (such as Tearing Modes, ELMs, and disruptions), their "black-box" nature poses significant safety and trustworthiness concerns for future fusion power plants. Next-generation control systems require interpretability, where the causes of control actions can be directly linked to physical plasma observations.

• The Conflict: High-accuracy models (like deep neural networks) are often uninterpretable, while interpretable models (like simple random forests) often lack the accuracy or complexity required for real-world fusion scenarios.
• The Gap: Existing Tearing Mode (TM) prediction models often rely on scalar parameters or lack stability interpretation. There is a need to bridge the gap between opaque ML predictions and physical understanding of why a specific plasma profile leads to instability.

2. Methodology

The authors propose a framework that retains the high accuracy of complex "black-box" models while using Shapley analysis (a game-theoretic approach) to interpret their outputs.

• Data Source: The study utilizes the DIII-D tokamak database (shots 140,000–195,000), comprising 6,050 shots with 677,494 timesteps. The dataset includes kinetic profiles (electron/ion temperature, density, rotation), magnetic equilibrium reconstructions (EFIT), and actuation data (NBI, ECH).
• Model Architecture:
  • Algorithm: Deep Survival Machines (DSM) from the Auton-Survival package.
  • Task: Survival regression to predict the probability of an $n=1$ Tearing Mode occurring within a specific time horizon.
  • Inputs: Real-time kinetic profiles (via RTCAKENN), actuation values, and EFIT scalars.
  • Training Strategy: The model is trained to predict the time-to-event (TM onset). A "profiles-only" model was also trained to isolate the impact of plasma profiles from scalar inputs (like $\beta_N$) during interpretation.
• Interpretation Technique:
  • Shapley Analysis: Used to fairly distribute the prediction result across model inputs. It calculates the contribution of each feature (e.g., core electron temperature at a specific radius) to the final TM probability relative to a background distribution.
  • Background Distribution: The analysis uses 11 specific shots from a dedicated DIII-D TM avoidance experiment as the reference distribution to ensure context-relevant interpretations.

3. Key Contributions

1. Physics-Driven Interpretability Framework: Demonstrates a generalizable methodology to extract physical insights from complex ML models without sacrificing predictive accuracy by using Shapley values.
2. Validation via Dedicated Experiment: The framework is validated against a real-time, active control experiment on DIII-D where the ML model successfully predicted TMs, and control actions (ECH deposition changes) were taken to suppress them.
3. Novel Physics Insights: The study moves beyond scalar parameters to analyze the spatial structure of plasma profiles, revealing specific mechanisms of TM stabilization and destabilization that were previously obscured.

4. Key Results

A. Model Performance

• The Deep Survival Machine achieved an AUROC score of 0.85 on the full dataset and 0.82 on the profiles-only model.
• Warning Time: The model provides warnings approximately 1000ms in advance, with the majority of true positives predicted >500ms ahead. This allows sufficient time for actuation (e.g., changing Electron Cyclotron Heating deposition).
• Experimental Success: In the dedicated experiment (shots 199597–199607), the model achieved an 82% success rate in predicting TMs and enabling their suppression. Failures occurred only when higher-$n$ modes triggered $n=1$ TMs, a scenario the current model does not account for.

B. Shapley Analysis Findings (Physics Interpretation)

The analysis revealed how specific profile features influence TM stability:

• Core Electron Temperature ($T_e$): Identified as the primary destabilizing factor. Higher core $T_e$ leads to a linear increase in TM risk.
• Core Ion Temperature ($T_i$): Shows a non-linear relationship. TM risk remains stable between 2–3.5 keV but rises exponentially at higher values. This suggests that plasmas with a high $T_e/T_i$ ratio may be more stable than those with high $T_i$.
• Rotation:
  • Core Rotation: Generally stabilizing (suppresses MHD instabilities).
  • Edge Rotation: Slightly destabilizing.
  • Conclusion: The rotation gradient (peaking) is a critical feature for stability.
• Density: A large, peaked density profile is lightly destabilizing, though less significant than temperature effects.
• Control Mechanism (ECH Deposition): The study analyzed a shot where ECH was moved to the $q=2$ surface to suppress a predicted TM. Shapley analysis confirmed that this action:
  • Increased current density ($j$) at $q=2$ (stabilizing).
  • Caused density pumpout (stabilizing).
  • Slightly increased $T_e$ (destabilizing), but the net effect was stabilization.

5. Significance and Future Outlook

• Bridging the Gap: This work proves that "black-box" ML models can be made transparent and trustworthy for fusion control by using post-hoc interpretability tools like Shapley analysis.
• Operational Guidance: The findings offer actionable strategies for future experiments:
  • Tuning the ratio of Electron Cyclotron Heating (ECH) to Neutral Beam Injection (NBI) to manage the $T_e/T_i$ ratio and avoid exponential $T_i$-driven instability.
  • Utilizing off-axis neutral beams to create flatter $T_i$ profiles, which are more stabilizing.
• Scalability: The framework is adaptable to other fusion phenomena (ELMs, disruptions) and different tokamak scenarios (e.g., ITER baseline), provided the background distribution is appropriately selected.
• Safety: By identifying the specific physical drivers of instability, this approach enables the development of safer, physics-informed AI controllers for future fusion power plants.