Deep-Learning based surrogate models for plasma exhaust simulations -- SOLPS-NN

This paper introduces SOLPS-NN, a deep-learning surrogate model trained on extensive SOLPS-ITER simulations that utilizes simple fully connected neural networks to efficiently and accurately predict plasma exhaust conditions and detachment access, demonstrating that independent models for specific observables yield higher accuracy and that transfer learning offers no significant advantage over training from scratch.

The Gist

The "Crystal Ball" for Fusion Energy: A Simple Explanation of SOLPS-NN

Imagine you are trying to bake the perfect cake, but the recipe is so complicated that every time you try to calculate the ingredients, it takes a supercomputer days to finish the math. If you want to test 1,000 different variations (more sugar, less heat, different flour), you'd be waiting years for results.

This is exactly the problem scientists face with Tokamaks—the giant donut-shaped machines designed to create fusion energy (the same power that fuels the sun). Specifically, they are struggling to predict what happens in the "Scrape-Off Layer" (SOL), the thin, turbulent edge of the plasma where heat and particles escape.

This paper introduces SOLPS-NN, a "Crystal Ball" (or a Surrogate Model) built using Artificial Intelligence (Deep Learning) that can predict these complex outcomes in milliseconds instead of days.

Here is how they did it, broken down into simple concepts:

---

1. The Problem: The "Slow Cooker"

The current method to simulate the plasma edge is called SOLPS-ITER. It's incredibly accurate, like a master chef tasting every drop of sauce. But it's also a "slow cooker."

• The Issue: It takes hours or days to run one simulation.
• The Consequence: Scientists can't easily explore thousands of "what-if" scenarios to find the perfect settings for future reactors like ITER or DEMO. They also struggle with numerical glitches (the math sometimes breaks).

2. The Solution: The "Fast Food" AI (SOLPS-NN)

The researchers trained a Deep Learning model (a type of AI) to act as a surrogate. Think of it like this:

• The Chef (SOLPS-ITER): Makes a perfect, slow-cooked meal.
• The Student (SOLPS-NN): Watches the Chef cook 8,000 times, learns the patterns, and then can guess the taste of a new dish instantly.

They fed the AI a massive dataset of 8,000+ simulations. The AI learned to look at the "ingredients" (input parameters like gas flow and power) and predict the "taste" (temperature and density of the plasma) without actually running the slow physics engine.

3. The Architecture: Painting the Whole Picture

The team tested different ways to build this AI:

• The "Pixel-by-Pixel" Artist: One approach tried to predict the temperature of the entire plasma map at once (like painting a whole landscape in one brushstroke).
• The "Single-Point" Artist: Another approach asked the AI to predict just one specific spot, then asked it again for the next spot, and so on.

The Verdict: The "Whole Picture" approach (a Fully Connected Neural Network) won. It learned to predict the entire 2D map of the plasma edge in one go, just as accurately as the slow method, but much faster. It's like having a weather app that shows the temperature for every city in the world instantly, rather than calling a meteorologist for each city.

4. The "One-Stop-Shop" vs. "Specialists"

The team also asked: Should we have one giant AI that predicts temperature, density, and pressure all at once, or separate AIs for each?

• The Result: It turns out, specialists are better.
• The Analogy: Imagine a hospital. You could have one "Super Doctor" who tries to be an expert in heart surgery, brain surgery, and dentistry all at once. Or, you could have a Heart Specialist, a Brain Specialist, and a Dentist.
  • The paper found that training separate models for each variable (temperature, density, etc.) was slightly more accurate and much easier to update. If you need to add a new variable later, you just train a new specialist without messing up the existing ones.

5. The "Physics Glitch" and the Fix

There was a tricky problem. When the AI predicted the temperature, it was so good at the general shape but slightly "noisy" on the tiny details.

• The Problem: Heat flow depends on how quickly temperature changes over a tiny distance. If the AI's prediction is slightly "wobbly," the calculated heat flow becomes huge and wrong (like trying to measure a mountain's slope with a ruler that has a wobbly edge).
• The Fix: They used a clever trick called "SOLPS in the Loop."
  • They let the AI make a quick guess.
  • Then, they fed that guess into the real slow physics engine for just a few seconds (instead of days) to "smooth out" the rough edges.
  • Result: They got the best of both worlds: the speed of the AI with the physical accuracy of the real simulation.

6. Does it Work on Real Machines? (The "Transfer Learning" Test)

The AI was trained on simulations of a machine called JET (a smaller, existing reactor). They wanted to know: Can this AI predict what will happen in ITER (a much bigger, future reactor)?

• The Challenge: It's like training a driver on a small city car and expecting them to drive a massive semi-truck perfectly. The physics are similar, but the scale is different.
• The Test: They tried Transfer Learning. This is like taking the driver who knows the city car and giving them a quick refresher course on the semi-truck using just a few practice runs.
• The Surprise: The "Refresher Course" (Transfer Learning) didn't actually make the driver much better than just hiring a new driver and training them from scratch on the semi-truck.
  • Why? Because the AI learned the fundamental physics so well from the first dataset that it didn't need much help. However, for the most precise predictions, it's still best to train specifically on the new machine's data.

7. The Big Picture: Why This Matters

The paper concludes that this AI model is a game-changer for fusion energy:

1. Speed: It can explore thousands of scenarios in the time it takes to run one real simulation.
2. Reliability: Even though it was trained on "simplified" physics, it correctly predicts the trends needed to keep the reactor safe (specifically, how to cool the plasma edge to prevent melting the walls).
3. Future Use: It allows scientists to design the exhaust systems for ITER and DEMO much faster, bringing us closer to clean, limitless energy.

In a nutshell: The researchers built a "Fast-Forward" button for fusion simulations. Instead of waiting days to see what happens if you tweak the gas flow, the AI tells you instantly, allowing engineers to design safer, more efficient fusion reactors.

Technical Summary

1. Problem Statement

The Scrape-Off Layer (SOL) of tokamak fusion reactors is critical for heat and particle exhaust management. Accurate modeling of the SOL is essential for designing divertors and ensuring safe reactor operation. However, the standard simulation code, SOLPS-ITER, is computationally expensive, suffers from long convergence times, and is prone to numerical instabilities. These limitations hinder rapid design studies, automatic optimization, and coupling with other plasma models (e.g., pedestal or plasma-wall interaction).

The paper addresses the need for surrogate models (machine learning approximations) that can predict SOL physics with high speed and accuracy, enabling fast exploration of parameter spaces for current devices (JET, ASDEX-Upgrade) and future reactors (ITER, DEMO).

2. Methodology

Dataset Generation

• Source: A large dataset of 8,192 training and 2,048 testing SOLPS-ITER simulations.
• Physics Fidelity: The primary dataset uses a reduced-fidelity fluid neutral model to accelerate simulation generation.
• Parameters: Eight input parameters were varied (deuterium/nitrogen puff rates, cross-field transport coefficients, core density, input power, etc.) covering machine sizes from ASDEX-Upgrade to DEMO.
• Regimes: The dataset covers various physical regimes: sheath-limited, attached, detached, and cold core, categorized by electron temperatures at the outer midplane and divertor target.

Model Architectures Tested

The authors compared several Deep Learning architectures to predict 2D plasma profiles (specifically electron temperature $T_e$, density $n_e$, and neutral densities):

1. NN2D (Fully Connected): Inputs are the 8 simulation parameters; outputs are the $T_e$ values for the entire $104 \times 50$ grid (5,200 neurons in the output layer).
2. NNpos2D (Position-Dependent): Inputs include simulation parameters + $(R, Z)$ coordinates of a specific grid cell; output is a single scalar $T_e$. This requires 5,200 independent forward passes to reconstruct the domain.
3. XGBoost2D: Gradient boosted trees using a similar position-dependent strategy as NNpos2D.
4. Multi-Output Strategies: Comparison between training independent networks for each observable (e.g., one for $T_e$, one for $n_e$) vs. a single "all-in-one" network predicting all quantities simultaneously.

Advanced Strategies

• SOLPS-in-the-Loop: To calculate derived quantities (like heat flux $q_{peak}$) which depend on spatial gradients, the authors proposed a hybrid approach: use the neural network to predict state variables, feed them into SOLPS-ITER as initial conditions, and run the simulation for a few timesteps to compute the dependent quantities.
• Transfer Learning: Adapting a model trained on the large, low-fidelity fluid neutral dataset to a smaller, high-fidelity dataset (ITER kinetic neutral simulations) by freezing early layers and retraining only the final layer.

3. Key Contributions & Results

Architecture Selection

• Best Performer: The fully connected NN2D architecture (predicting the whole domain at once) outperformed position-dependent models (NNpos2D) and XGBoost.
• Efficiency: NN2D achieved median absolute errors of 0.86 eV (attached regime) and 0.18 eV (detached regime) at the divertor target.
• Multi-Output Strategy: Training independent networks for each observable was found to be slightly more accurate and significantly more flexible for adding new observables compared to a single "all-in-one" network.

Handling Derived Quantities (Heat Flux)

• Challenge: Directly predicting heat flux ($q \propto T^{5/2} \nabla T$) from neural network outputs led to large errors in sheath-limited regimes because the networks could not perfectly resolve microscopic temperature gradients in nearly isothermal plasmas.
• Solution: The "SOLPS-in-the-Loop" approach (iterating the simulation for ~1000 steps using NN predictions as initial states) successfully smoothed out gradient errors. This reduced the median error for peak heat flux from >100% to <15%, outperforming direct neural network predictions in detached regimes.

Validation on Real Machines

• Scaling Laws: The surrogate model successfully reproduced experimental scaling laws for the onset of detachment in ASDEX-Upgrade and JET. It correctly predicted the inverse relationship between plasma density and impurity concentration required for detachment.
• Limitations: While trends were accurate, the model exhibited unphysical spatial behaviors (e.g., detachment initiating at the outer edge rather than the strike point) due to the simplified fluid neutral physics and gas puff location in the training data.

Transfer Learning to ITER

• High-Fidelity Adaptation: The model was tested against a small dataset of high-fidelity ITER simulations (kinetic neutrals, different geometry).
• Findings:
  • Simple scaling of input gas puff rates improved agreement but could not fix spatial discrepancies.
  • Transfer Learning (fine-tuning the pre-trained model on ITER data) achieved high accuracy but suffered from catastrophic forgetting (loss of the ability to predict sheath-limited regimes).
  • Crucial Finding: Training a new model from scratch on the small ITER dataset yielded similar or slightly better accuracy than transfer learning, suggesting that for small, high-fidelity datasets, transfer learning offers no significant advantage over training from scratch with simple architectures.

4. Significance and Future Outlook

• Operational Speed: The surrogate model provides predictions in near real-time, offering a speedup of 100–1,000x compared to running full SOLPS-ITER simulations.
• Design Utility: The models are sufficiently accurate for scoping studies, allowing engineers to rapidly explore operational windows, assess exhaust scenarios, and perform uncertainty quantification for next-generation devices (ITER, DEMO, SPARC).
• Methodological Benchmark: The paper establishes that simple fully connected networks are sufficient for 2D plasma surrogate modeling, provided the training data is extensive. It highlights the trade-off between model complexity and the need for physics-informed post-processing (like the SOLPS-in-the-Loop method) to handle derived quantities.
• Future Work: The authors suggest focusing on mixed-fidelity training (combining low and high-fidelity data), active learning to optimize dataset generation, and integrating experimental data directly into the training loop to bridge the gap between simulation and reality.

Conclusion

The SOLPS-NN framework demonstrates that deep learning can effectively surrogate complex plasma exhaust simulations. While the reduced-fidelity training data limits absolute physical accuracy in specific spatial details, the models capture essential physical trends and scaling laws. The study concludes that for specific high-fidelity applications, training from scratch on limited high-quality data is often superior to transfer learning, but the speed and flexibility of these surrogates make them invaluable tools for the design and operation of future fusion reactors.