en.Wedoany.com Reported - A white paper on multi-fidelity methods for nuclear fusion has been released, focusing on AI and digital twins to accelerate commercialization. This white paper systematically elaborates on how the integration of physics-driven and data-driven models can address core challenges in fusion energy development. It explores cutting-edge technologies including machine learning, multi-fidelity reduced-order models, and AI agents, aiming to accelerate the commercialization of fusion energy through the construction of predictive digital twins.
Regarding the integration of physics-driven and data-driven models, the white paper points out that fusion plasma physics models form a hierarchy ranging from high-fidelity kinetic and gyrokinetic models to magnetohydrodynamic (MHD) models. However, traditional physics-based simplified models are limited by prior approximations. Data-driven reduced-order models (such as Proper Orthogonal Decomposition (POD), Dynamic Mode Decomposition (DMD), tensor networks, and Gaussian processes), as well as convolutional neural networks (CNNs) and generative diffusion models, can directly learn underlying structures from high-fidelity simulations or experimental data, breaking traditional physical constraints. Data acquisition and standardization are key bottlenecks; the fusion field currently needs to align with the FAIR principles (Findable, Accessible, Interoperable, Reusable), utilizing open databases such as IMAS (Integrated Modelling & Analysis Suite) and FAIR-MAST. Recent advances include the development of multi-mode quasi-linear models for stellarators using DMD, and the construction of nonlinear gyrokinetic reduced-order models via POD-Galerkin projection.
In terms of reducing and characterizing uncertainty, the white paper emphasizes the importance of Verification, Validation, and Uncertainty Quantification (VVUQ) for predicting the performance of future devices. Using control variate techniques such as Multilevel Monte Carlo (MLMC) and Multi-Fidelity Monte Carlo (MFMC), low-fidelity models can significantly reduce the variance of high-fidelity evaluations, which is particularly effective for simulating rare events like neutronics and high-energy particle transport. For the challenges of Markov Chain Monte Carlo (MCMC) inference, the Delayed Acceptance scheme and multi-fidelity data assimilation techniques show promise. Virginia Tech, in collaboration with the University of Colorado Boulder, has successfully applied sparse grid UQ to construct surrogate models for predicting tokamak divertor heat loads.
Advanced numerical methods provide underlying support for the multi-fidelity framework. Using tensor networks to develop Boltzmann collision operators can effectively overcome the curse of dimensionality in high-fidelity kinetics (e.g., Vlasov-Maxwell systems); drawing inspiration from Large Eddy Simulation (LES) combined with machine learning closure terms offers new solutions for magnetized plasma turbulence.
In differentiable programming, porting scientific computing codes to frameworks supporting automatic differentiation, such as JAX or PyTorch, allows solvers to embed neural networks and run on GPUs/TPUs, directly obtaining gradient information. Currently, the differentiable gyrokinetic solver iGENE has been ported to JAX.
Regarding formal correctness proofs, using Domain-Specific Languages (DSL) and the Lean proof language, researchers are constructing numerical methods with formal certificates of correctness in mathematical structure and physical conservation laws for compressible Euler equations and MHD equations.
In fusion device design and optimization, multi-fidelity methods enable deep integrated optimization. There is an urgent need for fast, predictive, and self-consistent core-edge coupling schemes, such as combining 4D gyrokinetic edge/divertor models with neutral particle models. In Inertial Confinement Fusion (ICF), multi-fidelity Bayesian methods and deep learning surrogate models have been used to optimize target designs for the National Ignition Facility (NIF). In magnetic confinement fusion, Monte Carlo methods need to be embedded in optimization loops to assess the impact of engineering uncertainties, such as coil manufacturing and installation errors, on performance.
For multi-fidelity methods oriented towards real-time control, due to the difficulty of frequent physical access inside reactors and sensor degradation in nuclear environments, future plasma control will heavily rely on model-based state estimation. Model Predictive Control (MPC) is becoming a mainstream trend, requiring the development of efficient, control-oriented nonlinear reduced-order models. Furthermore, control co-design needs to be introduced in the early design phase, comprehensively considering transient dynamics and extreme events.
In enabling autonomous scientific discovery, AI agents based on the combination of large language models and tool-calling systems are reshaping High-Performance Computing (HPC) workflows. Agents have been able to fully automate the deployment and environment configuration of the gyrokinetic code GENE on multiple HPC systems, assisted in the code refactoring of GENE-3D, ported the Gkeyll code to AMD GPUs, and fully automatically run OSIRIS on the Perlmutter supercomputer for laser-plasma interaction studies. The application of agents in the fusion field is evolving from assistive to fully autonomous.
Building a predictive-centric fusion digital twin is the long-term goal of this field. This requires not only the strategic fusion of high-fidelity and low-fidelity models but also a robust VVUQ framework to establish the credibility of simulation-based decisions. The directions explored in this white paper are highly consistent with the U.S. Department of Energy (DOE)'s recently released "Genesis Mission National Scientific and Technological Challenge Project Application Guide" (Genesis Mission RFA) and the "Fusion Science & Technology Roadmap" (FS&T Roadmap), which explicitly prioritize AI, multi-fidelity computing, digital twins, and facility design optimization to accelerate the commercialization of fusion energy.
