MPEX AI Digital Twins Milestone Report

This six-month progress report outlines the on-track development of two Phase I AI milestones for the MPEX project—a Helicon AI Hot-Spot Controller and an E-beam Damage Assessment Digital Twin—alongside the configuration of the Galaxy software interface to integrate physics simulations with DOE HPC resources and the American Science Cloud for automated data analysis and AI-driven operations.

The Gist

Imagine a massive, futuristic kitchen where scientists are trying to cook the perfect meal: nuclear fusion energy. The "oven" is a machine called MPEX, and the "ingredients" are super-hot plasma and special metal walls. The goal is to test if these metal walls can survive the extreme heat without cracking or melting.

However, cooking in this kitchen is tricky. The heat doesn't spread evenly; it creates "hot spots" that can burn holes in the oven door or the cooking pot. If the walls crack, the experiment fails.

This report is a progress update from a team of scientists (from Oak Ridge and Lawrence Livermore National Laboratories) who are building a "Digital Twin" for this kitchen. Think of a Digital Twin as a perfect, virtual video game version of the real machine. They are teaching an Artificial Intelligence (AI) to be the head chef, using this virtual version to predict what will happen before they turn on the real machine.

Here is a breakdown of their two main "recipes" for success:

1. The "Hot Spot" Controller (Keeping the Oven Door Safe)

The Problem:
In the real machine, the heat comes from a specific type of wave (called "helicon"). Sometimes, this heat gets stuck in one spot, like a magnifying glass focusing sunlight on a leaf, creating a dangerous "hot spot" that could crack the glass window of the machine.

The AI Solution:
The scientists built a smart controller that acts like a traffic cop for heat.

• The Old Way: Scientists used to guess how to adjust the magnetic "roads" (coils) to spread the heat out. It was like trying to herd cats by guessing where they would run.
• The New Way: They created a 3D virtual model of the machine. They taught an AI to look at pictures of the heat (taken by special cameras) and figure out exactly how to tweak the magnetic roads to spread the heat evenly.
• The Analogy: Imagine you are pouring water into a maze of pipes. If you pour too fast in one spot, it bursts. The AI is like a smart system that instantly adjusts the valves (magnetic coils) to make sure the water flows smoothly everywhere, preventing any single pipe from bursting.

They are currently training this AI using data from a smaller test kitchen (called "proto-MPEX") so that when the big machine opens, the AI will already know how to keep the temperature perfect.

2. The "Damage Detective" (Predicting Cracks in the Metal)

The Problem:
The metal walls (made of tungsten) are being tested to see if they crack under extreme heat. To test this, they use a powerful electron beam (like a super-fast hair dryer) to heat the metal. Afterward, they take microscopic photos to count the cracks.

• The Challenge: There are too many different types of metal and too many heat settings to test them all physically. It would take forever to test every single combination. Also, the photos are hard to analyze because the cracks look different depending on the metal's grain structure.

The AI Solution:
The team built a super-smart image analyzer and a crystal ball.

• The Image Analyzer: They taught an AI to look at microscopic photos of the metal and automatically find every single crack, no matter how small or weird it looks. It's like giving the AI a pair of glasses that can instantly spot a hairline fracture in a piece of glass.
• The Crystal Ball (Prediction): Since they can't test every metal, they used a physics simulator (a program that calculates how metal breaks) to generate "fake" data. They combined the real photos with the fake physics data to teach the AI a pattern.
• The Analogy: Imagine you have a few samples of clay that cracked when baked. You want to know if a new type of clay will crack. Instead of baking the new clay (which takes time), you use the AI to look at the "shape" of the cracks in the old clay and the physics of how clay breaks. The AI then predicts, "If you bake this new clay at this temperature, it will likely crack here."

3. The "Kitchen Manager" (Galaxy Workflow)

To make all this work, the scientists built a central control panel called Galaxy.

• The Analogy: Think of this as a master recipe book and a kitchen timer combined. It connects the AI, the physics simulators, and the real machine data. It allows scientists (or even the AI itself) to run complex experiments with a few clicks, ensuring that every step is recorded and repeatable. It's the glue that holds the "Digital Twin" together.

What's Next? (The June Goal)

By June 2026, the team plans to show off a working version of this system:

1. For the Hot Spots: The AI will successfully predict the best settings to keep the heat centered and safe, using data from their smaller test machine.
2. For the Damage: The AI will successfully predict where cracks will form in different metals, using a mix of real photos and computer simulations, proving it can "guess" the outcome without needing to test every single piece of metal physically.

In Summary:
The scientists are building a virtual twin of their fusion machine and teaching an AI chef to manage the heat and a virtual detective to predict metal cracks. This allows them to run experiments faster, safer, and smarter, bringing us one step closer to clean, limitless fusion energy.

Technical Summary

Technical Summary: MPEX AI Digital Twins Milestone Report

Problem Statement
The Material Plasma Exposure eXperiment (MPEX) is designed to qualify plasma-facing materials for fusion power plants under extreme heat and particle fluxes. A critical operational challenge is the formation of localized "hot spots" on plasma-facing components, particularly the helicon antenna window, which can lead to sputtering or structural failure. Additionally, assessing material damage (cracking, erosion) requires extensive experimental testing, which is limited by the finite number of samples and the sparsity of available data across different material compositions and exposure conditions. The project addresses the need for an Artificial Intelligence (AI) advantage to optimize MPEX operations in real-time and to accelerate the assessment of material damage through a combination of experimental data, physics simulations, and generative AI.

Methodology
The project develops two primary AI Digital Twins: one for controlling helicon heating hot spots and another for assessing electron beam (E-beam) induced material damage.

1. Helicon AI Hot-Spot Controller:

   • Physics Modeling: The team transitioned from 2D azimuthally symmetric models to high-fidelity 3D RF simulations using COMSOL to capture the intrinsic azimuthal asymmetry of helicon heating. These results were coupled with the HERMES-3 3D fluid turbulence code to simulate plasma generation and transport.
   • Flux-Tube Mapping: A reduced-order framework was developed to map heat flux from the helicon window to downstream targets by tracing flux tubes along magnetic field lines, preserving the identity of specific flux bundles rather than just surface coordinates.
   • AI Architecture: Variational Autoencoders (VAEs) were trained on Proto-MPEX infrared (IR) measurements and simulation data. These generative models learn a shared latent space to fuse noisy experimental data with multi-fidelity simulations. The AI system uses a control-oriented loop to detect hot spots, predict heat-flux redistribution under varying coil currents, and recommend operating parameters to minimize window loading while maintaining target flux.

2. E-beam Damage Assessment Digital Twin:

   • Automated Characterization: A fully automated image-processing pipeline was developed for Scanning Electron Microscopy (SEM) images. It utilizes Sato ridge enhancement, Gaussian smoothing, unsharp masking, and triangle thresholding to robustly identify and segment crack networks across heterogeneous materials and resolutions.
   • Feature Encoding & Metric Learning: Image patches are encoded using a pre-trained DINOv2 vision transformer to create high-dimensional feature vectors. Dimensionality reduction via PCA and diffusion maps creates a low-dimensional manifold where morphological similarity is preserved. A metric learning approach aligns the distance in this feature space with the distance in the experimental parameter space (temperature, heat flux, alloy).
   • Prediction & Simulation: Kernel Ridge Regression (KRR) is used to predict crack density based on the learned metric. To address data sparsity, the CabanaPD peridynamics code was extended to model anisotropic continuum mechanics and long-time thermal shock simulations. These physics-based simulations generate synthetic data to fill gaps in the experimental parameter space.
   • Workflow Automation: All tools (physics codes, AI models, data analysis) are integrated into the Galaxy web-based platform, enabling reproducible workflows and automated data ingestion from DOE HPC resources.

Key Contributions

• 3D RF Modeling: First-time 3D COMSOL simulations of the MPEX helicon system revealed that heating is localized in the 2D plane perpendicular to magnetic field lines due to edge collisionality and multi-mode excitation, contradicting previous 2D symmetric assumptions.
• Generative Control Framework: Development of a VAE-based controller that successfully interpolates between experimental conditions to predict IR heat flux maps and optimize coil currents for central plasma heating, thereby reducing azimuthal hot spots.
• Robust Crack Identification: Creation of a fully automated, parameter-free image processing pipeline capable of extracting crack density from diverse SEM images, deployed via Galaxy.
• Physics-Informed AI Surrogate: Integration of CabanaPD peridynamics simulations with AI models to generate training data for regimes where experimental data is sparse, specifically addressing anisotropic mechanical behavior in tungsten alloys.
• Metric Learning for Damage: Implementation of a learned metric in parameter space that aligns with morphological similarity in diffusion space, enabling more physically meaningful predictions of crack density than raw parameter regression.

Results

• Hot Spot Control: The AI controller demonstrated the ability to minimize axial heat asymmetry by varying coil currents, though azimuthal asymmetry remains a challenge due to plasma parallel heat flux variations. The system successfully identified that central plasma heating is the critical factor for reducing azimuthal hot spots.
• Damage Prediction: The automated crack identification tool achieved reliable segmentation across both highly cracked and undamaged surfaces. The KRR model using the learned metric showed good prediction accuracy where data existed, though it exhibited systematic underestimation in low-damage regimes due to data sparsity.
• Simulation Validation: The extended CabanaPD code was validated against experimental stress-strain data for various tungsten alloys, successfully capturing anisotropic mechanical responses and predicting crack initiation under E-beam thermal shock conditions.
• Workflow Integration: A subset of physics simulation codes (DREAM, GITR, CabanaPD) and the AI crack identification tool were successfully deployed on the ORNL Galaxy instance, connecting to HPC resources.

Significance
The paper claims that this work establishes the foundation for a fully functioning science platform where AI agents can autonomously control MPEX operations and assess material damage. By bridging high-fidelity physics simulations with experimental data through generative AI and metric learning, the project enables a rapid, generative search for optimal materials with reduced plasma-material interaction (PMI) damage. The integration of these capabilities into the Galaxy framework ensures that the MPEX facility can operate with AI-agentic decisions to avoid shutdowns during long pulses and efficiently qualify new candidate materials, moving beyond inefficient empirical searches. The report positions these Phase I milestones as a precursor to Phase II, which will expand control to the entire facility heating and cooling systems and incorporate additional plasma damage modalities.