🔬 physics

SuperCond-GNN: Scalable Graph Neural Network Surrogate for Superconducting Circuit Simulations

This paper introduces SuperCond-GNN, a scalable graph neural network surrogate that accurately and rapidly predicts voltage distributions and current redistribution in high-temperature superconducting magnets by modeling them as lumped-element circuits, offering a topology-agnostic alternative to conventional solvers for design exploration and real-time monitoring.

Original authors: Nandana Menon, Giorgio Vallone

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nandana Menon, Giorgio Vallone

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Traffic Jam" in Superconductors

Imagine you are trying to manage a massive highway system made of superconducting wires (wires that carry electricity with zero resistance until they get too hot). These wires are used in giant magnets for things like particle accelerators and fusion reactors.

The problem is that if one part of the wire gets a little hot, the electricity has to "reroute" around it, like cars swerving into other lanes to avoid a stalled truck. This is called current redistribution. If the system can't predict where the electricity will go, the magnet can overheat and get damaged.

Traditionally, engineers use a digital tool called SPICE (like a very precise, but very slow, traffic simulator) to calculate exactly how the electricity will flow. However, as the magnet gets bigger and more complex, this simulator becomes incredibly slow. It's like trying to calculate traffic for a whole country by simulating every single car one by one; it takes too long to be useful for real-time safety checks.

The Solution: A "Smart GPS" (SuperCond-GNN)

The authors created a new tool called SuperCond-GNN. Think of this not as a simulator that calculates every step, but as a highly trained "Smart GPS" that looks at the map and instantly predicts the traffic flow.

Instead of doing heavy math for every single wire, this tool uses a type of Artificial Intelligence called a Graph Neural Network (GNN).

The Analogy: Imagine the magnet's circuit as a social network. Each wire segment is a "person" (a node), and the connections between them are "friendships" (edges).
How it learns: The AI watches thousands of simulations where it sees how electricity flows through different network sizes. It learns the "rules of the road" (the physics) by passing messages between the nodes, just like people sharing news in a network.

Why This is Special: The "Lego" Advantage

Most AI models are like a custom-built house: if you want to add a room, you have to tear the whole house down and rebuild it. They can only handle one specific size of magnet.

SuperCond-GNN is like Lego. Because it is built on a "graph" structure, it doesn't care how many pieces you use.

You can give it a small circuit with 2 wires.
You can give it a massive circuit with 100 wires.
The magic: The AI uses the exact same brain to understand both. It doesn't need to be retrained or rebuilt; it just looks at the new connections and applies the same rules it already learned. This makes it incredibly flexible for designing different types of magnets.

How Accurate is It?

The team tested this "Smart GPS" against the slow, traditional simulator (SPICE).

Speed: On a computer chip designed for AI (a GPU), the new tool is up to 2.5 times faster than the old simulator for large circuits. More importantly, as the circuits get bigger, the old simulator gets slower and slower, while the new tool stays fast.
Accuracy: It predicts the voltage (the electrical "pressure") with an average error of only 4.3%. In the world of superconductors, this is like predicting a storm's path with enough accuracy to save a city. The error is so small it is less than 1% of the safety limit engineers use to detect problems.

The "Physics" Check

To make sure the AI didn't just "guess" but actually understood the laws of electricity, the researchers added a special rule called Kirchhoff's Current Law.

The Analogy: Imagine a teacher checking a student's homework. The student (the AI) solves the problem, and the teacher (the physics rule) checks: "Did the total amount of water flowing into this pipe equal the amount flowing out?" If not, the student gets a penalty.
This helps the AI stay honest, especially when it encounters a magnet design it has never seen before.

Can It Handle New Designs? (The "Zero-Shot" Test)

The researchers asked: "What if we show this AI a magnet design it has never seen in its training?"

The Result: Without any extra training, the AI got the answer roughly 15–30% wrong (which is a big jump from its usual 4% error).
The Fix: However, if they showed the AI just 5 to 10 examples of this new design (a process called "few-shot fine-tuning"), it quickly learned the new rules and became accurate again. This proves it can adapt to new, unseen magnet shapes very quickly.

The Bottom Line

The paper presents a new way to predict how electricity moves in superconducting magnets. Instead of using a slow, rigid calculator, they built a flexible, fast, and adaptable AI "Smart GPS."

While the paper focuses on tape stacks (simple bundles of wires) as a proof of concept, the goal is to use this tool to:

Explore new magnet designs faster.
Analyze how electricity shares itself across complex cables.
Eventually, monitor real magnets in real-time to prevent them from burning out.

The authors emphasize that this is a proof of concept for tape stacks, but the "Lego-like" nature of the system means it could eventually be scaled up to handle the most complex magnets in the world.

Technical Summary: SuperCond-GNN

Problem Statement

High-temperature superconducting (HTS) magnets, particularly those based on REBCO conductors, face significant challenges in protection and diagnostics due to slow normal-zone propagation and complex current-sharing behaviors. Traditional protection strategies often rely on reactive voltage-based detection, which can be too slow to prevent damage from localized heating. To transition toward predictive protection and digital twin frameworks, there is a need for real-time, high-fidelity models capable of capturing electromagnetic behavior at the cable or subscale level.

Current standard tools, such as SPICE-based lumped-element circuit models (e.g., ngspice), are accurate but lack scalability. As circuit complexity increases from individual tapes to full cables and magnets, the number of components grows, making parametric studies and real-time simulations computationally prohibitive. Furthermore, existing machine learning surrogates often rely on fixed-dimensional input spaces (e.g., feedforward neural networks), rendering them unable to generalize across variable circuit topologies or sizes. This limitation prevents the development of a universal surrogate capable of handling the inherently variable dimensionality of superconducting circuit models.

Methodology

The paper introduces SuperCond-GNN (SC-GNN), a graph neural network-based surrogate model designed to predict steady-state nodal voltage distributions in HTS tape stack cables.

1. Circuit Representation

The authors model HTS tape stacks as 2D networks of lumped elements, where each tape is divided into segments governed by a power-law current-voltage relationship ( $V \propto I^n$ ). To overcome the fixed-input limitation of standard neural networks, the circuit is mapped to a graph structure $G=(V, E)$ :

Nodes: Represent circuit junctions. Features include boundary conditions (ground/input), normalized degree, and normalized distances from ground and current sources.
Edges: Represent circuit components (resistors, superconducting elements, current sources). Features encode element type (one-hot) and physical parameters (resistance $R$ , critical current $I_c$ , power-law exponent $n$ , and operating ratio $I_{in}/I_c$ ).
Topology Agnosticism: This representation allows the model to process graphs of arbitrary size and topology without architectural changes.

2. Model Architecture

SC-GNN is based on the MeshGraphNet architecture, comprising three stages:

Encoder: Projects raw node and edge features into a shared latent space using MLPs.
Processor: Consists of $L$ sequential message-passing layers. It iteratively updates edge and node embeddings by aggregating information from neighbors, utilizing residual connections to stabilize training.
Decoder: Maps the final latent node representations to scalar nodal voltage predictions. The ground node is masked out as its voltage is a known boundary condition.

3. Training and Physics-Informed Regularization

Data Generation: A dataset of 2,000 circuits (1–10 tapes, 1–6 segments) was generated using ngspice DC simulations across a defined design space of parameters ( $R_t, R_c, I_c, n$ ).
Loss Function: The primary loss is a scale-normalized Mean Squared Error (MSE) to handle sparse voltage values.
Physics-Based Loss: To enforce physical consistency, a Kirchhoff's Current Law (KCL) regularization term is added. A unique challenge arises with superconducting elements, where the voltage-to-current mapping is ill-conditioned in the sub-critical regime. The authors employ a "supernode merging" technique: if the predicted voltage drop across a superconducting edge is below a threshold, the connected nodes are merged into a single supernode, treating the element as an ideal short circuit. This allows for stable KCL residual calculation.

Key Contributions

Topology-Agnostic Framework: A graph-based representation for HTS tape stacks that enables surrogate modeling across variable circuit sizes and topologies, addressing a fundamental limitation of conventional fixed-input neural networks.
Physics-Informed GNN: The integration of KCL regularization via a supernode merging strategy to enforce current conservation, enhancing the physical plausibility of predictions.
Comprehensive Benchmarking: Extensive evaluation against ngspice across the full topology space (1×1 to 10×6), demonstrating high accuracy and scalability.
Generalizability Assessment: Evaluation of zero-shot inference and few-shot fine-tuning capabilities on out-of-distribution (OOD) topologies (e.g., 15×6, 10×10, 15×10), demonstrating the model's adaptability to unseen configurations.

Results

Predictive Accuracy

The baseline SC-GNN (trained without physics loss) achieved a mean RMSE of 0.04 µV and a mean MAPE of 4.30% across the test set.
Performance remained robust across mid-to-large topologies. Minor degradation was observed in very small circuits (1×1), attributed to limited local context for message passing, and in very large circuits, likely due to the fixed number of message-passing layers limiting the receptive field.
The error margins correspond to an electric-field error well below 1% of the conventional 1 µV/cm critical-current criterion.

Computational Performance

Scalability: SC-GNN exhibits near-constant scaling with graph size ( $\propto N^{0.008}$ ) on GPU, compared to the linear scaling of ngspice ( $\propto N^{1.08}$ ).
Speedup: On GPU, SC-GNN achieves a peak speedup of 2.5× over ngspice for larger circuits. While CPU inference did not consistently outperform ngspice, GPU inference remains largely insensitive to circuit size, offering significant advantages for complex magnet configurations.

Generalizability and Fine-Tuning

Zero-Shot: Direct application to OOD topologies resulted in higher errors (e.g., MAPE ~30–44%), but the physics-informed variant consistently outperformed the baseline in zero-shot scenarios.
Few-Shot Fine-Tuning: The model successfully adapted to unseen topologies with as few as 5–10 training samples. Freezing earlier layers generally preserved performance while reducing trainable parameters, though optimal strategies varied by topology.

Physics-Regularization Impact

Incorporating KCL loss provided a modest improvement in validation loss (~4% reduction) but yielded comparable test set performance to the baseline. The authors suggest the graph topology itself provides a strong inductive bias toward electrical physics.
However, the physics-informed model demonstrated superior robustness in zero-shot generalization to OOD topologies, suggesting that physical priors improve out-of-distribution stability even if they do not drastically improve in-distribution accuracy.

Significance and Claims

The paper positions SuperCond-GNN as a scalable, topology-agnostic alternative to conventional circuit solvers for HTS magnet simulations. Its primary significance lies in enabling real-time inference and design space exploration where traditional SPICE solvers are too slow or non-scalable.

The authors claim that the framework is a critical step toward an AI-enabled, physics-informed digital twin for HTS magnet systems. By accurately predicting nodal voltage distributions, the surrogate supports voltage-based quench detection and current-sharing analysis. The work demonstrates that GNNs can effectively learn the underlying physical behavior of variable-size superconducting circuits, offering a pathway to real-time magnet monitoring and protection strategies that are currently unattainable with standard simulation tools. The authors remain modest, noting that while the current proof-of-concept is limited to tape stacks, the framework is naturally extensible to more complex cable and magnet configurations.