MPEX AI Digital Twins

The MPEX AI Digital Twins project aims to maximize the scientific output of the MPEX device by training AI models on experimental and physics simulation data to create a digital twin for material assessment, operational control, and PMI simulation.

The Gist

Imagine you are trying to build the ultimate, indestructible shield for a star that has been captured inside a machine. This machine is called MPEX (Material Plasma Exposure eXperiment), and its job is to blast special materials with super-hot, super-fast particles to see if they can survive the conditions inside a future fusion power plant.

The problem is that this "star" is unpredictable. It can suddenly shoot a laser-like beam of heat at the wrong spot, cracking the machine's windows or melting the test materials. Testing every possible material by hand would take forever and cost a fortune.

This paper proposes a solution: The MPEX AI Digital Twin. Think of this as a "video game" version of the real MPEX machine, but one that is so smart it can learn from the real thing and predict the future. Here is how the team plans to build it, broken down into simple steps:

1. The "Photo Album" and the "Smart Eye"

First, the team needs to organize the data. MPEX will take thousands of high-speed photos of the materials before and after they get blasted.

• The Analogy: Imagine a detective trying to solve a crime by looking at millions of blurry photos. The team is building a "Smart Eye" (AI) that can instantly zoom in on these photos to find the tiniest cracks, melted spots, or rough patches.
• The Goal: Instead of a human staring at screens for days, the AI automatically measures exactly how much damage occurred, creating a standardized "damage report" for every test.

2. The "Virtual Physics Lab"

Real experiments are expensive and slow. So, the team is building a complex computer simulation called STRIPE.

• The Analogy: If the real MPEX is a wind tunnel where you test a real airplane wing, STRIPE is a super-advanced flight simulator. It doesn't just guess; it uses real physics equations to calculate how the wind (plasma) hits the wing (material), how the metal heats up, and how it might chip away.
• The Upgrade: They are using AI to make this simulator run faster. Normally, simulating every single particle takes forever. The AI acts like a "speed dial," learning the patterns so it can predict the results in seconds instead of weeks.

3. The "Traffic Cop" (The Hot Spot Controller)

One of the biggest dangers is a "hot spot"—a tiny area where the heat gets too intense and could crack the machine's windows.

• The Analogy: Imagine driving a car with a blind spot. You need a co-pilot who can see the danger before you hit it. The AI Hot Spot Controller is that co-pilot. It watches the real-time camera feeds and the machine's settings.
• How it works: If the AI sees a dangerous heat spot forming, it instantly suggests a new setting for the machine's magnets (like turning the steering wheel slightly) to steer the heat away from the windows and back onto the target material. It learns by trial and error, but much faster than a human could.

4. The "Material Oracle" (The Digital Twin)

This is the grand finale. The team wants to combine the real photos (from the "Smart Eye") and the computer simulations (from the "Virtual Lab") to train a master AI model.

• The Analogy: Imagine a master chef who has tasted every dish in the world and also knows the chemistry of every ingredient. If you ask, "What happens if I mix this new spice with that new meat?" the chef doesn't need to cook it to know the answer.
• The Goal: This "Material Oracle" will be able to look at a brand-new, never-before-tested material and predict exactly how it will hold up against the plasma. It can "dream up" thousands of virtual materials, test them in the simulation, and tell the scientists: "Don't waste time testing these 999; go test this one specific alloy instead."

Why Do This?

The paper argues that guessing which materials will work is inefficient. By building this Digital Twin, the scientists can:

1. Save Time: Find the best materials much faster.
2. Save Money: Avoid building and testing materials that will fail.
3. Stay Safe: Prevent the machine from breaking by steering heat away from sensitive parts in real-time.

In short, they are building a super-smart, virtual assistant that helps them design the shields for the world's first fusion power plants, ensuring they can survive the heat of a star.

Technical Summary

Technical Summary: MPEX AI Digital Twins

Problem Statement
The Material Plasma Exposure eXperiment (MPEX) is a high-power, steady-state linear plasma device designed to replicate the extreme divertor conditions of future magnetic confinement fusion power plants, specifically targeting energy fluxes of 20 MW/m², ion fluences of 10³¹/m², and pulse durations up to 10⁶ seconds. A critical challenge in achieving these operational milestones is the control of plasma flux to prevent "hot spots" that can damage antenna windows and vessel walls, while simultaneously qualifying materials for long-term survivability under intense plasma-material interaction (PMI). Traditional empirical methods for material selection and operational control are inefficient given the finite number of samples available for testing. Furthermore, the complex physics spanning atomistic sputtering to device-scale impurity transport requires a unified, predictive framework to guide experimental campaigns and accelerate the discovery of optimal plasma-facing materials.

Methodology
The paper proposes a comprehensive "MPEX AI Digital Twins" framework that integrates experimental data, advanced physics simulations, and artificial intelligence (AI) to create a closed-loop system for data processing, operational control, and materials assessment. The methodology is structured across six primary technical domains:

1. Data Infrastructure and Curation:

   • American Science Cloud Integration: Experimental data (from proto-MPEX and future MPEX runs) and physics simulation data will be organized using standard schemas (similar to IMAS) and transferred to the American Science Cloud.
   • Data Backbone: A system links raw images, metadata, AI outputs, and twin states. Images and manifests are ingested into versioned object storage, while AI inferences (masks, features, uncertainties) are stored in columnar formats (HDF5).
   • Interface: The NOVA-Galaxy framework is customized to provide a web-based interface for analyzing scientific data and running workflows on remote High-Performance Computing (HPC) resources.

2. AI-Driven Image Analysis:

   • A data-driven pipeline ingests pre- and post-exposure high-resolution images to output quantitative morphologies with calibrated uncertainties.
   • Features: The system focuses on crack detection (length, width, branching, fractal dimension), surface melting (bead geometry, area fraction), and redeposition metrics.
   • Training Strategy: The approach utilizes self-supervised pretraining on a small labeled seed set, followed by fine-tuning with active learning to flag high-uncertainty regions for operator adjudication.

3. Physics Simulation Framework (STRIPE):

   • The Simulated Transport of RF Impurity Production and Emission (STRIPE) framework is upgraded to model MPEX operations. It couples plasma transport, RF heating, sheath modeling, ion-material interactions, and impurity transport.
   • Kinetic Plasma Simulation: PICOS++ (Particle-In-Cell) serves as the kinetic engine, extended to model multi-frequency RF heating (Helicon, ECH, ICH). Acceleration strategies include GPU acceleration, time-splitting, and the use of Pseudo-Reversible Normalizing Flows (PR-NF) to improve velocity-space sampling efficiency for non-Maxwellian distributions.
   • Hybrid Modeling: A kinetic-fluid hybrid model couples PICOS++ with the C1 fluid code to handle neutral particles and plasma moments efficiently.
   • Turbulence Modeling: LLNL contributes high-performance 3D turbulence simulations using the gyrokinetic code COGENT and the fluid code BOUT++/Hermes-3. Machine Learning is integrated to develop Reduced-Order Models (ROMs) via Latent Space Dynamics Identification (LaSDI).
   • PMI Modeling: Advanced sputtering models (BCA, MLFF-MD, DFT) and the WallDYN code are integrated to predict erosion, redeposition, and surface composition evolution for complex ceramics and alloys.

4. AI Hot Spot Controller:

   • Real-time Detection: AI image processing analyzes real-time camera feeds (up to 18k fps) to identify hot spots on antenna windows and walls.
   • Control Logic: An AI model, trained on a database of coil currents and heating powers versus hot spot locations/intensity, predicts optimal coil current adjustments. The goal is to direct plasma to the target while minimizing wall heating, constrained by magnetic resonance points.

5. Target Materials Simulation:

   • Thermomechanics: Multi-scale physics models validate material responses. Continuum simulations (ANSYS/COMSOL) and the meshfree peridynamics code CabanaPD are used to simulate crack initiation, propagation, and fragmentation under thermal shock.
   • Calibration: AI-identified damage features are mapped to state variables (stress, strain, temperature) and used in a Bayesian assimilation loop to calibrate material parameters.

6. MPEX Material Assessment AI Digital Twin:

   • Training Data Generation: The twin is trained on a combination of curated experimental data and synthetic data generated by physics simulations. Synthetic data expands the training domain to include material properties and compositions not yet experimentally tested.
   • Optimization: The trained AI Digital Twin performs multi-dimensional interpolation to identify optimum material candidates with reduced PMI damage, which are then validated via physics codes or experimental testing.

Key Contributions

• Unified Simulation Framework: The integration of the STRIPE framework with AI-accelerated kinetic solvers (PICOS++), advanced sputtering models (MLFF-MD), and turbulence codes (COGENT/BOUT++) provides a predictive, multi-fidelity environment for MPEX.
• Operational Control AI: The proposal of an AI Hot Spot Controller that utilizes real-time camera data and coil current optimization to prevent hardware damage during long-pulse, high-power operations.
• Data-Centric Materials Discovery: A workflow that converts raw experimental images into standardized, quantitative metrics and uses them to train an AI Digital Twin capable of searching high-dimensional material spaces for optimal candidates.
• Infrastructure for AI Readiness: The development of a data backbone and interface (NOVA-Galaxy) specifically designed to link experimental data, simulation outputs, and AI models within the American Science Cloud ecosystem.

Results and Claims
The paper does not present final experimental results from the full MPEX device, as commissioning is scheduled for the end of FY26. Instead, it outlines the proposed technical architecture and preliminary capabilities based on existing tools and proto-MPEX data:

• Proto-MPEX Validation: Surrogate modeling has already been applied to proto-MPEX data (14,666 discharges), serving as the initial cohort for AI development.
• Algorithmic Readiness: The image processing pipeline is based on an existing LDRD effort for electron-beam exposed materials, adapted for MPEX.
• Simulation Capabilities: The STRIPE framework has been successfully applied to other linear devices (PISCES-RF) and tokamaks (WEST, DIII-D). The proposed upgrades (PICOS++ with RF modules, PR-NF acceleration) are presented as necessary evolutions to handle MPEX-specific conditions.
• Pilot Study: The authors propose starting the AI Digital Twin development with a lower-dimensional dataset of tungsten cracking from an electron beam heating experiment (ORNL LDRD) to test model scalability before applying it to full MPEX data.

Significance
The paper positions the MPEX AI Digital Twins project as a critical enabler for the MPEX mission. By integrating AI with physics simulations, the project aims to:

1. Accelerate Material Selection: Move beyond inefficient empirical searches by using AI to interpolate multi-dimensional parameter spaces and identify optimal materials for fusion divertors.
2. Ensure Operational Safety: Enable the MPEX device to achieve its high-power, long-pulse goals by actively controlling plasma flux to prevent catastrophic hot spots on critical components.
3. Maximize Scientific Output: Create a "curated experimental database" and a validated simulation framework that allows for retrospective meta-analysis and the generation of synthetic data for materials not yet physically tested.
4. Collaborative Advancement: Leverage the Transformational AI Models Consortium (TAIMC) and the American Science Cloud to establish unified data standards and reduce the cost of data generation for the broader DOE complex.

The authors emphasize that the ultimate success of the project relies on the iterative feedback loop between experimental validation, physics simulation, and AI model refinement, ultimately providing a predictive capability for next-generation fusion power plant materials.