A Surrogate model for High Temperature Superconducting Magnets to Predict Current Distribution with Neural Network

Imagine you are trying to design a giant, super-powerful magnet for a fusion reactor or a futuristic motor. This magnet is made of a special material called REBCO (a type of high-temperature superconductor).

The problem is that these magnets are incredibly complex. Inside them, electricity doesn't flow evenly like water in a smooth pipe; it gets "jammed" and uneven due to magnetic forces, a phenomenon called the screening current effect. To design a safe and efficient magnet, engineers need to know exactly how this electricity is distributed inside every tiny layer of the material.

The Old Way: The Slow, Exhausting Calculator

Traditionally, engineers use a method called Finite Element Method (FEM) to figure this out. Think of FEM as a super-precise, but incredibly slow, calculator.

If you want to design a small magnet, it might take a few hours.
If you want to design a meter-scale magnet (huge!), that same calculation can take days or even weeks on a powerful computer.
If you want to test 100 different designs to find the best one, you'd have to wait months. It's like trying to find the perfect recipe by baking a cake, waiting a week for it to cool, tasting it, and then starting over with a slightly different recipe.

The New Way: The "Surrogate" Oracle

This paper introduces a clever shortcut: a Surrogate Model powered by Artificial Intelligence (Neural Networks).

Think of this AI model as a highly trained apprentice or a crystal ball.

The Training: First, the engineers let the slow, old calculator (FEM) do its work on a bunch of different magnet designs. They feed these results into the AI.
The Learning: The AI studies these results and learns the "rules of the game." It learns how changing the size of the magnet, the number of wire turns, or the current affects the electricity flow.
The Prediction: Once trained, the AI doesn't need to do the heavy math anymore. When you ask it, "What happens if I make the magnet slightly bigger?" it answers in milliseconds with incredible accuracy.

The Secret Sauce: The "Residual" Network

The researchers didn't just use a standard AI; they used a specific type called a Fully Connected Residual Network (FCRN).

The Analogy: Imagine a standard AI as a student trying to solve a math problem by writing every single step on a long piece of paper. If the problem is too long, the student gets tired, loses focus, and makes mistakes (this is called the "vanishing gradient" problem).
The Residual Fix: The Residual Network is like giving that student a shortcut. It allows the student to peek at the answer from the previous step and add it directly to the current step. This keeps the "signal" strong and clear, even for very deep, complex problems. This allowed the AI to learn much better than older models.

Testing the Crystal Ball

The team tested their AI in two scenarios:

The Fast Ramp (Case 1): Imagine turning the magnet on very quickly. The AI had to predict how the electricity behaves while things are changing fast.
- Result: The AI was amazing. Even when they asked it to predict magnets 50% larger than anything it had ever seen before, it was still accurate (less than 10% error). It was like the apprentice being able to guess the size of a giant's house just by seeing a dollhouse.
Steady State (Case 2): Imagine the magnet is running at a constant, high power.
- Result: The AI was great at predicting the shape of the magnet (geometry). However, if they asked it to predict what happens with a much higher electrical current than it had seen in training, it got a bit confused.
- Why? Because at very high currents, the material behaves in a totally new, non-linear way (like a sponge that suddenly stops absorbing water). The AI hadn't seen enough examples of this "full saturation" to learn the rule.

The Grand Finale: Designing a Magnet in Minutes

The real magic happened when they used the AI to design a magnet.

The Goal: Find the smallest, most efficient magnet that creates a magnetic field of 16 Tesla (super strong!) without wasting too much material.
The Process: Instead of waiting weeks for the slow calculator to test thousands of designs, the AI tested them all in 3 minutes.
The Outcome: It found the perfect design. When the engineers double-checked this "AI-designed" magnet with the slow, old calculator, the results matched almost perfectly.

The Takeaway

This paper shows that we can stop waiting weeks for magnet designs. By training a smart AI on a smaller set of data, we can create a "surrogate" that acts like a super-fast, highly accurate oracle. It allows engineers to explore "what-if" scenarios and design massive, powerful magnets for fusion energy and advanced motors in the blink of an eye, rather than over a cup of coffee that has gone cold.

Here is a detailed technical summary of the paper "A Surrogate Model for High Temperature Superconducting Magnets to Predict Current Distribution with Neural Networks."

1. Problem Statement

High-temperature superconducting (HTS) magnets, specifically those using Rare-earth barium copper oxide (REBCO) coated conductors, are critical for applications like fusion Tokamaks and ultra-high-field solenoids. However, their design is hindered by screening current effects, which cause non-uniform current distributions within the tapes. This non-uniformity reduces the central magnetic field and induces stress concentrations, necessitating accurate current density calculations for reliable design.

The primary challenge is computational cost:

Finite Element Methods (FEM), using formulations like $T-A$ , are the standard for calculating these distributions but are extremely time-consuming.
For meter-scale HTS magnets, a single operating condition can take tens to hundreds of hours to simulate, especially when considering electromagnetic-thermal or electromagnetic-mechanical couplings.
This computational bottleneck prevents rapid optimization and iterative design of large-scale HTS magnets.

2. Methodology

The authors propose a data-driven surrogate model based on a Fully Connected Residual Neural Network (FCRN) to predict the spatial current density distribution ( $J_\phi$ ) in REBCO solenoids, bypassing the need for repeated FEM simulations.

A. Data Generation

Source: Training datasets were generated using FEM simulations based on the $T-A$ formulation with a thin-sheet approximation.
Scenarios: Two distinct operating conditions were modeled:
1. Case 1 (Fast Ramping): Dynamic evolution of current density during a 1-second ramp. Parameters varied: Number of turns ( $N$ ) and Pancakes ( $N_p$ ). Magnetic field dependence of critical current ( $J_c$ ) was neglected.
2. Case 2 (Steady-State): Stable operation at target current. Parameters varied: $N$ , $N_p$ , Inner Diameter ( $RID$ ), and Operating Current ( $I_{op}$ ). The magnetic field dependence of $J_c$ (via the Kim model) was included.
Preprocessing: Inputs (spatial coordinates, time, geometric parameters) and outputs (current density) were normalized. An explicit index ( $p$ ) for the specific pancake was added to help the network learn systematic differences in current penetration between central and end pancakes.

B. Neural Network Architecture

Model: A Fully Connected Residual Network (FCRN) was chosen over a standard Fully Connected Network (FCN).
Rationale: Residual connections (skip connections) were implemented to mitigate the vanishing gradient problem, allowing for deeper networks (up to 24 layers) without convergence failure.
Hyperparameters:
- Activation: SiLU (Sigmoid Linear Unit) for smooth derivatives.
- Optimizer: Adam with learning rate decay.
- Loss Function: Mean Squared Error (MSE).
- Training Strategy: Models were saved only when both training and interpolation validation losses decreased, ensuring generalization.

3. Key Contributions

Development of an FCRN Surrogate Model: Successfully adapted deep learning to predict complex, non-linear current density distributions in HTS solenoids, addressing the limitations of standard FCNs in deep architectures.
Systematic Evaluation of Extrapolation: The study rigorously tested the model's ability to generalize beyond the training data range (extrapolation), a critical requirement for designing magnets larger than those in the training set.
Rapid Magnet Design Framework: Demonstrated a practical application where the surrogate model was used to perform a heuristic search for an optimal magnet design, reducing a process that would take days with FEM to a few minutes.
Identification of Limitations: Clearly delineated the boundaries of the model's extrapolation capabilities, specifically regarding operating current vs. geometric parameters.

4. Key Results

Performance Comparison (FCRN vs. FCN)

Convergence: Deep FCNs ( $L=24$ ) failed to converge due to vanishing gradients. FCRNs converged stably across all depths.
Accuracy: FCRN consistently achieved lower validation losses than FCN in both Case 1 and Case 2.
Optimal Architecture: A 24-layer network with 256 neurons per layer (24-256) provided the best balance of accuracy and generalization.

Extrapolation Capabilities

Case 1 (Fast Ramping):
- Geometric Extrapolation: The model achieved high accuracy (relative error < 10%) when extrapolating geometric parameters ( $N, N_p$ ) up to 50% beyond the training set.
- Limit: At 150% extrapolation, errors exceeded 35%, indicating a hard limit on geometric scaling without retraining.
- Efficiency: Inference time was 0.1–0.3 seconds, compared to 1–11 hours for FEM (a speedup of several orders of magnitude).
Case 2 (Steady-State):
- Geometric Extrapolation: The model maintained robust performance for geometric variations, with an average relative error of 1.2% for the central magnetic field.
- Current Extrapolation: Performance degraded significantly when extrapolating operating current ( $I_{op}$ ) beyond the training range (error rose to 9.0% for combined extrapolations). This is attributed to the model failing to capture the physics of full tape penetration under high currents, which was not present in the training data.

Application: Rapid Magnet Design

Task: Design an HTS solenoid minimizing tape length while satisfying $B_0 > 16$ T and field homogeneity < 1%.
Process: The surrogate model scanned the parameter space (including extrapolated regions) in ~3 minutes.
Outcome: Identified an optimal design ( $N=360, N_p=9, R=10$ mm, $I_{op}=222$ A).
Verification: FEM validation of the optimal design showed a central field error of only 0.2% (16.02 T vs. 16.04 T), confirming the surrogate model's reliability for design optimization.

5. Significance

This work provides a computationally efficient pathway for the intelligent design of large-scale HTS magnets. By replacing hours-long FEM simulations with millisecond-level neural network predictions, the surrogate model enables:

Rapid Iteration: Designers can explore vast parameter spaces and optimize for multiple constraints quickly.
Scalability: The model's ability to extrapolate geometric parameters allows for the design of magnets significantly larger than those used for training, reducing the need for massive, expensive FEM datasets.
Future Direction: The paper highlights that while geometric extrapolation is robust, future work must incorporate physical constraints or dynamic current evolution data to improve extrapolation for operating currents and full penetration regimes.

In summary, the proposed FCRN surrogate model is a validated, high-speed tool that bridges the gap between high-fidelity physics simulations and the practical needs of rapid engineering optimization for next-generation superconducting magnets.