Hierarchical Framework of Runaway Electrons using Deep Learning

This paper presents a novel adjoint deep learning framework combined with physics-informed neural networks to create fast, accurate surrogate models for predicting runaway electron kinetics across diverse plasma scenarios, offering orders-of-magnitude speedups over traditional solvers.

The Gist

Imagine you are trying to predict the behavior of a chaotic crowd of people (electrons) in a giant, invisible stadium (a fusion reactor). Some of these people are running away so fast they become "runaway electrons," which can damage the stadium walls.

Traditionally, to predict how this crowd moves, scientists have to simulate every single person individually. It's like trying to predict traffic by tracking every single car on the highway with a stopwatch. It's incredibly accurate, but it takes so much computing power that it's too slow to use in real-time emergency planning.

This paper introduces a new, much faster way to do this using a "smart shortcut" powered by Artificial Intelligence (Deep Learning). Here is how they did it, explained simply:

1. The "Reverse Movie" Trick (The Adjoint Method)

Usually, to know where a crowd ends up, you have to watch them move forward from the start. The authors used a clever mathematical trick called the Adjoint Method.

Think of it like watching a movie of the crowd in reverse. Instead of asking, "If I start here, where will I end up?", they ask, "If I want to know the total energy of the crowd at the end of the movie, what did the people need to be doing at the start?"

By solving this "reverse movie" problem once, they can instantly calculate the final outcome for any starting situation. It's like having a single map that tells you the total traffic jam at 5:00 PM, no matter where the cars started at 4:00 PM.

2. The "Physics-Brained" AI (PINNs)

They didn't just use a standard AI that learns by memorizing thousands of examples. Instead, they used a Physics-Informed Neural Network (PINN).

Imagine teaching a student to play chess.

• Standard AI: You show the student 10,000 games and say, "Memorize these moves." If they see a new board setup they haven't seen before, they might get confused.
• Physics-Informed AI: You give the student the rules of chess (the laws of physics) and say, "You can't move a knight like a bishop. You must follow these rules."

The AI in this paper was taught the "rules of the universe" for electrons (how they collide, how electric fields push them, how they lose energy to light). Because it knows the rules, it doesn't need to memorize every possible scenario. It can figure out the answer for a situation it has never seen before, instantly.

3. What They Predicted

Using this "Reverse Movie + Physics-Brain" combo, they built three specific tools (neural networks) to predict:

• The Current: How much "electric flow" the runaway electrons are carrying (crucial for keeping the reactor stable).
• The Average Energy: How fast, on average, these electrons are moving (important for knowing how much damage they could do).
• The Energy Distribution: A detailed breakdown of how many electrons are moving at slow speeds, medium speeds, and super-fast speeds.

4. The Results: Speed vs. Accuracy

The authors tested their new AI against the traditional, slow method (which they call a "Monte Carlo solver," essentially a super-accurate simulation of every single particle).

• The Old Way: Takes about 3.5 minutes on a powerful computer to simulate 10 million particles.
• The New Way: Takes milliseconds to give the same answer.

They found that for most situations, the AI's predictions matched the slow, accurate simulation almost perfectly. However, they noted one small catch: if the electrons are moving so fast they "escape" the stadium (the computer's simulation limits), the AI makes a slight assumption that they stop at the wall. In reality, they keep going. But for most practical scenarios, the AI is incredibly accurate and millions of times faster.

The Bottom Line

This paper presents a new "super-fast calculator" for fusion scientists. Instead of waiting hours to simulate how dangerous runaway electrons will behave, they can now get an answer in a blink of an eye. This allows them to quickly test different scenarios and keep fusion reactors safe, without needing to run heavy, slow simulations every time.

Technical Summary

Technical Summary: Hierarchical Framework of Runaway Electrons using Deep Learning

Problem Statement
The description of runaway electron (RE) formation and evolution in magnetized fusion devices presents a significant computational challenge, particularly when coupling RE dynamics with magnetohydrodynamic (MHD) models. While first-principles kinetic models based on the relativistic Fokker-Planck equation offer high fidelity, they are computationally intensive, limiting their utility in multi-physics simulations. Conversely, reduced fluid models are computationally efficient but discard critical information regarding the energy and pitch-angle distributions of REs, thereby reducing the physical fidelity of integrated models. There is a need for a surrogate model that retains kinetic fidelity while enabling rapid inference of RE moments and energy distributions for arbitrary initial conditions and plasma parameters.

Methodology
The authors propose a hierarchical framework combining an adjoint formulation of the relativistic Fokker-Planck equation with Physics-Informed Neural Networks (PINNs).

1. Adjoint Formulation:

   • Instead of evolving the full distribution function forward in time, the framework solves an adjoint problem backward in time.
   • By defining specific terminal and boundary conditions, the adjoint solution allows for the direct inference of arbitrary fluid moments (e.g., density, current, average energy) and the energy distribution for any initial electron momentum space distribution.
   • The adjoint operators ($L^*$) are derived via integration by parts from the forward operators (electric field acceleration, collisional drag, and synchrotron radiation losses).
   • Specific adjoint problems are formulated for:
     • Runaway Probability Function (RPF): To determine RE density.
     • Current Probability Function (CPF): To determine the parallel RE current.
     • Energy Probability Function (EPF): To determine the average kinetic energy.
     • Energy Distribution: By treating the runaway threshold energy ($p_{RE}$) as a parametric input, the framework infers the number of electrons in narrow energy windows, effectively reconstructing the full energy distribution.

2. Physics-Informed Neural Networks (PINNs):

   • Three distinct PINNs are trained to solve the adjoint equations for the RPF, CPF, and EPF (including the energy distribution).
   • Physics Constraints: The neural networks are trained to minimize the residual of the adjoint partial differential equations (PDEs) without relying on pre-generated simulation data for training. The loss function includes terms for the PDE residual, terminal conditions, and boundary conditions.
   • Architecture and Training: The models utilize a 6-layer network with 64 neurons per layer. Training employs the SOAP optimizer and adaptive resampling (inspired by [35]) to focus on regions with high residuals.
   • Hard Constraints: "Physics layers" are implemented to enforce boundary conditions (e.g., ensuring zero solution at the low-momentum boundary) and to normalize outputs within physically meaningful ranges (e.g., using $\tanh$ transformations).
   • Parameters: The models are trained across a broad range of plasma parameters: parallel electric field ($E_\parallel$), effective charge ($Z_{eff}$), and synchrotron strength ($\alpha$).

Key Contributions

• Adjoint-PINN Framework: The paper details a novel method to use a single deep learning model to predict the temporal evolution of RE moments and energy distributions for arbitrary initial conditions, bypassing the need for multiple forward solves of the Fokker-Planck equation.
• Parametric Surrogates: The framework provides parametric solutions that generalize across varying plasma conditions, enabling rapid online inference after a one-time offline training cost.
• Energy Distribution Inference: A unique contribution is the ability to infer the full RE energy distribution by parameterizing the runaway threshold as an input to the PINN, allowing for the calculation of distribution shapes without solving for the full phase-space density explicitly.
• Validation: The models are validated against a high-fidelity Monte Carlo solver (JONTA) implemented in JAX, demonstrating agreement across various scenarios.

Results

• Current and Density: The adjoint-PINN predictions for RE current and density show excellent agreement with Monte Carlo simulations across different initial pitch distributions and plasma parameters. The framework accurately captures the threshold electric field for runaway generation and the saturation of current.
• Average Energy: The model accurately predicts the average energy of REs when the population remains within the computational domain. However, discrepancies arise when REs escape the high-energy boundary; in such cases, the adjoint method assumes particles are "frozen" at the boundary energy, leading to an underestimation of energy compared to a Monte Carlo solver that tracks particles beyond the boundary.
• Energy Distribution: The parametric PINN successfully captures the evolution of the energy distribution, including sharp features at low energies (dominated by collisional drag) and smoother features at high energies. It accurately models scenarios ranging from sub-threshold decay to super-threshold acceleration and partial domain escape.
• Computational Efficiency: While training a parametric PINN takes approximately 14 hours on a B200 GPU, the inference time is on the order of milliseconds. This represents an orders-of-magnitude speedup compared to traditional solvers, which take minutes for a single forward run with ten million particles.

Significance and Claims
The paper claims that this adjoint deep learning framework offers a practical pathway to accelerate the evaluation of RE quantities of interest. By embedding physics directly into the neural network training, the method achieves kinetic fidelity without the computational burden of continuum or particle-based solvers. The authors note that while the current study focuses on specific collision operators and does not yet encompass the full complexity of tokamak startup and disruption scenarios, the framework is designed to be generalizable. The results suggest that such surrogates can be straightforwardly extended to more comprehensive collision operators and broader physics parameters, potentially enabling real-time coupling of RE kinetics with MHD evolution in future fusion simulations. The authors explicitly state that the extension to specific generation scenarios in tokamaks is the focus of future work.