Data-Driven Optimisation of Superconducting Magnets at CEA Paris-Saclay

This paper presents a new AI-based platform developed at CEA Paris-Saclay to optimize superconducting magnet design for particle accelerators by leveraging machine learning and advanced techniques to manage complex datasets and address challenges in multiphysics modeling, topology optimization, and anomaly detection.

The Gist

Imagine you are trying to build the world's most powerful, complex, and delicate machine: a superconducting magnet. These aren't your average fridge magnets; they are the massive engines that steer particles in accelerators, power fusion reactors, and create the super-clear images in MRI machines.

Building one is like trying to bake a cake where you have to simultaneously:

1. Design the recipe (physics).
2. Build the oven (mechanics).
3. Control the temperature so it doesn't burn (thermal).
4. Ensure the ingredients don't explode (safety).
5. Do all of this while the ingredients are changing their properties as you mix them.

For decades, scientists at CEA Paris-Saclay have been doing this by running thousands of slow, expensive computer simulations. It's like trying to find the perfect cake recipe by baking one cake, eating it, writing down the notes, and then baking another one, hoping the next one is slightly better. It takes years.

This paper introduces a new, game-changing tool called ALESIA (think of it as a "Super-Chef's AI Assistant") that uses Artificial Intelligence (AI) to speed up this process, make it smarter, and prevent disasters.

Here is how ALESIA works, broken down into simple concepts:

1. The "Digital Librarian" (Data Management)

The Problem: In the past, every time a scientist ran a simulation, the data was scattered in different folders, different formats, and different computers. Finding a specific piece of information was like looking for a needle in a haystack made of other needles.
The ALESIA Solution: ALESIA is a central "Digital Librarian." It automatically collects every single piece of data from every simulation, organizes it perfectly, and stores it in a way that AI can read instantly. It's like having a library where every book is automatically sorted by author, genre, and page number, and you can ask the librarian, "Show me all the books about chocolate cakes from 2023," and they hand them to you in a second.

2. The "Smart Architect" (Multiphysics Optimization)

The Problem: Designing a magnet involves balancing conflicting needs. You want a strong magnetic field, but if it's too strong, the metal coils might snap under pressure. You want it small, but it needs space to cool down.
The ALESIA Solution: Instead of a human guessing the balance, ALESIA uses Active Learning. Imagine a video game character that learns by playing thousands of levels in seconds. ALESIA runs thousands of virtual designs, learns which ones fail, and quickly figures out the "sweet spot" where the magnet is strong, safe, and efficient.

• Real-world example: They used this to design magnets for the Electron-Ion Collider. The AI figured out the perfect shape and material mix for the coils, saving years of trial and error.

3. The "Shape-Shifter" (Topology Optimization)

The Problem: Sometimes, we don't know what the best shape for a magnet should be. We usually just copy old designs.
The ALESIA Solution: ALESIA uses Topology Optimization, which is like a digital sculptor. You tell the computer, "I need a magnet that fits in this box and holds this much weight," and the AI starts eroding material from a solid block, carving away everything unnecessary until only the most efficient, organic-looking shape remains. It's like a river carving a canyon; the water (the AI) finds the path of least resistance to create the perfect structure.

4. The "Crystal Ball" (Surrogate Models)

The Problem: Running a full physics simulation is like running a marathon; it takes a lot of energy and time.
The ALESIA Solution: ALESIA trains a "Surrogate Model" (a fast, simplified AI version of the simulation). Once the AI has seen enough real simulations, it can predict the outcome of a new design in milliseconds instead of hours. It's like a weather forecaster who has studied decades of data; they can predict tomorrow's rain without needing to wait for the clouds to actually form first.

5. The "Fire Alarm" (Quench Detection)

The Problem: A "quench" is when a superconducting magnet suddenly loses its super-power and becomes a normal, resistive wire. This releases a massive amount of stored energy instantly, which can melt the magnet. It's like a pressure cooker exploding.
The ALESIA Solution: The team is building Digital Twins (virtual copies) of the magnets. These twins constantly listen to the magnet's "heartbeat" (voltage and temperature). If the AI hears a tiny irregularity that might lead to an explosion, it sounds the alarm and shuts the system down before the disaster happens. It's like a smoke detector that can smell a fire before a single spark appears.

6. The "Future Builder" (ECR Ion Sources)

The paper also shows ALESIA being used to design Ion Sources (machines that create beams of particles). Here, the AI is helping to design the "extraction system"—the nozzle that shoots the particles out. The AI is literally "drawing" the shape of the metal electrodes, testing millions of shapes to find the one that shoots the particles straightest and strongest.

The Big Picture

In short, this paper is about moving from "Guess and Check" to "Learn and Predict."

By using ALESIA, the scientists at CEA are no longer just building magnets; they are teaching computers how to build them better. This means:

• Faster designs: Projects that took years now take months.
• Better magnets: Stronger, safer, and more efficient.
• New frontiers: They can now tackle problems that were previously too complex to solve, like using new high-temperature materials (HTS) that were too difficult to model by hand.

It's the difference between a craftsman chiseling stone by hand and a 3D printer that knows exactly how to print the perfect stone structure in seconds. The future of particle physics and medical imaging just got a whole lot brighter.

Technical Summary

1. Problem Statement

The design and operation of superconducting magnets for particle accelerators, fusion devices, and MRI systems are increasingly complex due to:

• Multiphysics Coupling: Designs require simultaneous optimization of electromagnetism, mechanics, thermal analysis, cryogenics, and materials science.
• Computational Cost: Traditional design relies on computationally expensive Finite Element Method (FEM) simulations (e.g., Opera, ANSYS, CAST3M), which can take days or weeks for a single iteration.
• Data Management: The design process generates massive datasets (terabytes) with diverse parameters, making data storage, retrieval, and reproducibility difficult.
• Emerging Technologies: New High-Temperature Superconductors (HTS) introduce non-linear behaviors and anisotropic properties that require new modeling and optimization tools beyond traditional methods.
• Safety: Detecting "quench" events (rapid transition to normal resistive state) is critical but challenging, especially for HTS magnets where normal-zone propagation is slow.

2. Methodology: The ALESIA Platform

To address these challenges, CEA Paris-Saclay developed ALESIA, a Python/Cython-based AI-driven optimization and data management platform. It is structured into seven core modules:

• Project Manager: Orchestrates simulation scripts, manages parallel execution, and coordinates interactions between modules.
• Databases:
  • Simulation Database: Uses HDF5 format to store relevant outputs and input parameters, adhering to FAIR principles (Findable, Accessible, Interoperable, Reusable).
  • Material Database: A unified SQL database containing properties (magnetization, thermal, mechanical) for various conductors (Nb-Ti, Nb3Sn, HTS) to ensure consistency across codes.
• Physics Software Interfaces: Connects ALESIA with external codes (Opera 2D/3D, RAT, ANSYS, CAST3M, IBSimu). It uses two approaches: direct database interaction (modifying code to read/write HDF5) and keyword-based script modification.
• Multiparametric Optimizers: Implements heuristic algorithms including Genetic Algorithms, Differential Evolution, Particle Swarm Optimization (PSO), CMA-ES, and Bayesian Optimization. It also supports custom strategies like uniform grid sweeps and Monte Carlo error analysis.
• Supervised Learning: Constructs surrogate models (e.g., Deep Neural Networks, Random Forests) from generated datasets to predict magnet behavior rapidly without running full FEM simulations.
• Data Analysis Tools: Utilizes Jupyter notebooks and automated reporting (.docx, .pptx, .tex) with metadata management for physical units and symbols.
• In-House Modelling: Includes analytical models (e.g., Biot-Savart) and a 3D adiabatic quench model (QAESAR) for rapid estimation of hotspot temperatures and current decay.

3. Key Contributions and Applications

A. Multiphysics Optimization (SPIN ROTATORS)

• Context: Designing 8 solenoid magnets for the Electron-Ion Collider (EIC) using Nb-Ti and Nb3Sn.
• Approach: An automated workflow links Opera (magnetism), CAST3M (mechanics/forces), and in-house conductor design codes.
• Result: The platform automates a 70-parameter loop. A Deep Residual Neural Network was trained on 8,524 simulations to create a surrogate model.
• Performance: The surrogate model achieved high accuracy ($R^2 > 0.98$) in predicting peak fields, integrated fields, von Mises stress, and quench temperatures, reducing evaluation time from hours to milliseconds.

B. Shim Coil Optimization

• Context: Improving magnetic field homogeneity and reducing fringe fields for detectors and MRI.
• Approach: Bayesian optimization explores a large parametric space (multipole orders, loop counts, geometry) to optimize shim coil configurations. Reinforcement learning is also being investigated for innovative topologies.

C. Topology Optimization (TO)

• Context: Designing ultra-high field (11.7 T) MRI magnets where mechanical stress is a limiting factor.
• Approach: Uses the Solid Isotropic Microstructure with Penalisation (SIMP) method to optimize the density distribution of windings.
• Result: Successfully generated a cross-section design for a 11.7 T magnet with a 1m bore, keeping von Mises stress below 160 MPa. The method is fast (minutes on a PC) and requires no initial topology assumptions, serving as a rapid "first draft" generator.

D. HTS Optimization (ECRIS & FASUM)

• Context: Designing magnets for Electron Cyclotron Resonance Ion Sources (ECRIS) and a 40 T user magnet (FASUM).
• Approach: Optimizing REBCO tape windings for a 45 GHz ECRIS (TOUHTATIS project).
• Result: Used evolutionary computation and surrogate models (Random Forest, Gaussian Process, Neural Networks) to optimize contact pressure distribution in mechanical assemblies, balancing thermal efficiency (20 K operation) with mechanical integrity.

E. Quench Protection and Anomaly Detection

• Context: Preventing damage from quench events, particularly in HTS magnets with slow propagation velocities.
• Approach: Developing "Digital Twins" using Partial Element Equivalent Circuit (PEEC) codes.
  • Classifiers: Decision trees/Random Forests to detect anomalies and predict quench types (rapid, slow, recovery).
  • Regressors: CNNs and LSTMs to forecast voltage and magnetic field evolution.
• Result: Fast models (milliseconds) and accurate models (tens of milliseconds) are being integrated into the QAESAR solver to improve protection strategies.

4. Results

• Efficiency: The ALESIA platform significantly reduces the time required for design iterations by automating data flow and utilizing surrogate models.
• Accuracy: Surrogate models trained on extensive datasets demonstrated $R^2$ values between 0.95 and 0.99 for critical physical quantities (stress, field, temperature).
• Feasibility: The platform successfully identified viable Nb3Sn conductor solutions and optimized complex mechanical assemblies for HTS magnets, which were previously too difficult to model manually.
• Scalability: The architecture supports N interfaces efficiently, avoiding the factorial complexity ($N-1)!$ of direct pairwise software integration.

5. Significance

This paper demonstrates a paradigm shift in superconducting magnet design from manual, sequential workflows to data-driven, automated, and AI-integrated ecosystems.

• Accelerated R&D: By replacing expensive FEM simulations with fast surrogate models for initial exploration, the design cycle is drastically shortened.
• Handling Complexity: The platform effectively manages the high-dimensional, multi-objective nature of modern magnet design (e.g., balancing field quality, mechanical stress, and cost).
• Future Readiness: The integration of AI prepares the field for the next generation of magnets (HTS, 40 T+ systems) and extends applicability beyond accelerators to fusion and medical imaging (MRI).
• Open Science: The development of a unified, modular platform (ALESIA) promotes reproducibility and collaboration across different physics domains.

In conclusion, the work at CEA Paris-Saclay establishes a robust framework where Artificial Intelligence acts not just as a tool for analysis, but as the central engine for the entire design, optimization, and safety monitoring lifecycle of superconducting magnets.