Machine-learning (ML) surrogates are increasingly used to accelerate plasma simulations by learning compact, reusable models that capture essential physics at a small fraction of the computational cost. In this lecture segment on Surrogates, Digital Twins, and Emulators, I will introduce core ideas in reduced-order modeling: compressing high-dimensional simulation outputs into a low-dimensional latent space, and learning interpretable latent dynamics (e.g., LaSDI – Latent Space Dynamic Identification) that enable rapid emulation for design, optimization, and control. I will then connect these concepts to my recent work on time-dependent atomic kinetics in laser-driven plasmas, where high-fidelity inline coupling to radiation–hydrodynamic simulations is often computationally intractable. We develop a LaSDI-based physics-informed surrogate that (i) compresses code-generated atomic populations using an autoencoder trained with a mixed loss that preserves both microscopic populations and macroscopic observables, (ii) identifies latent dynamics using symbolic-regression methods driven by time-dependent temperature and density histories, and (iii) enforces convergence and steady-state constraints—together with stability conditions—to ensure physically meaningful long-time behavior. The result is a fast, interpretable surrogate that illustrates how constraint-aware latent dynamics can bring ML surrogates closer to practical workflows for plasma modeling.

Min Sang Cho is a staff scientist at Lawrence Livermore National Laboratory (LLNL), where he has conducted research since 2022 after earning his Ph.D. from the Gwangju Institute of Science and Technology (GIST), South Korea. His research focuses on the non-equilibrium dynamics of laser-driven and high-energy-density plasmas, with particular emphasis on atomic-physics modeling. More recently, he has been developing physics-informed surrogate models for atomic plasma kinetics, leveraging latent-space dynamics identification and reduced-order modeling to enable fast, interpretable emulators of complex plasma systems for simulation, design, and analysis.

Data-driven methods (DDMs), such as deep neural networks, offer a generic approach to integrated data analysis (IDA), integrated diagnostic-to-control (IDC) workflows through data fusion (DF), which includes multi-instrument data fusion (MIDF), multi-experiment data fusion (MXDF), and simulation-experiment data fusion (SXDF). These features make DDMs attractive to nuclear fusion energy and power plant applications, leveraging accelerated workflows through machine learning and artificial intelligence. Here we describe Physics-informed Meta-instrument for eXperiments (PiMiX) that integrates X-ray (including high-energy photons such as -rays from nuclear fusion), neutron and others (such as proton radiography) measurements for nuclear fusion. PiMiX solves multi-domain high-dimensional optimization problems and integrates multi-modal measurements with multiphysics modeling through neural networks. Super-resolution for neutron detection and energy resolved X-ray detection have been demonstrated. Multi-modal measurements through MIDF can extract more information than individual or uni-modal measurements alone. Further optimization schemes through DF are possible towards empirical fusion scaling laws discovery and new fusion reactor designs.

Dr. Zhehui (Jeph) Wang is a scientist and team leader in the Physics Division at Los Alamos National Laboratory (LANL). Since earning his Ph.D. from Princeton University, he has devoted nearly thirty years to experimental physics and scientific instrumentation at one of the world’s premier research institutions. Dr. Wang has made pioneering contributions across a broad range of scientific areas, including plasma dynamos, plasma flow diagnostics, hypervelocity dust technology and applications, novel applications of dust in fusion plasmas, and fundamental neutron physics, with a particular emphasis on precision measurements of the neutron lifetime. Many of these advances were enabled by his innovations and leadership in measurement technologies, lately enhanced by machine learning and AI. He has also played a key role in dynamic materials research through the development and application of ultrafast synchrotron X-ray imaging and tomography techniques, breaking critical technological barriers at state-of-the-art light sources, such as fourth-generation synchrotrons. Dr. Wang has authored and coauthored more than 100 peer-reviewed technical publications and holds a growing portfolio of patents in advanced radiographic imaging and tomography, spanning plasmas, neutrons, and X-rays. In recent years, his scientific and technical leadership has expanded to data-driven experiments and instrumentation, integrating classical and quantum-enabled methods with cutting-edge instruments, and experimental facilities.

Artificial intelligence (AI) and machine learning (ML) are rapidly becoming essential tools in experimental plasma physics, enabling new capabilities in diagnostics, control, and experimental optimization. This mini-course provides an accessible introduction to how ML models are designed, validated, and deployed in real experimental environments, with practical examples from DIII-D and KSTAR tokamaks.

The course begins with the fundamentals: how to frame plasma physics problems for ML, prepare experimental data, select appropriate model architectures, and evaluate model reliability. Key concepts such as supervised learning, neural networks, surrogate modeling, uncertainty estimation, and physics-informed learning will be introduced with minimal mathematical prerequisites. Emphasis will be placed on connecting ML tools to physical intuition rather than treating them as black boxes.

We then walk through the full pipeline from offline model development to real-time deployment in plasma control systems. Practical use cases include real-time anomaly detection for identifying disruptions and diagnostic faults, multimodal data fusion for reconstructing missing or low-resolution diagnostics. Examples will demonstrate how ML enables super-resolution diagnostics, failure-tolerant monitoring, detachment control, and adaptive actuator coordination.

Special attention will be given to real-time constraints: latency, interpretability, validation under distribution shifts, and integration with existing plasma control architectures. Lessons learned from running AI-driven control experiments, where models directly influence actuator commands, will be discussed, highlighting both successes and challenges.

By the end of this mini-course, participants will understand how AI tools are developed and safely implemented in experimental tokamaks, and how these methods can enhance experimental efficiency, resilience, and physics discovery. The goal is to equip students and researchers with a clear conceptual roadmap for applying ML/AI to plasma science problems while maintaining scientific rigor and physical insight.

Azarakhsh (Aza) Jalalvand is a researcher at Princeton University’s Plasma Control Group, where his work focuses on applying artificial intelligence to fusion energy. His research spans advanced plasma diagnostics, real-time control, and predictive models to support reliable operation of next-generation fusion reactors. He is also engaged internationally, collaborating with Ghent University as a voluntary researcher and facilities such as KSTAR in Korea, and contributing to the ITER Control Scientist Fellow Network. His recent work has demonstrated AI-driven super-resolution diagnostics, diagnostic failure mitigation, and scenario optimization, positioning AI as a key enabler for ITER and future Fusion Pilot Plants.

High-fidelity plasma simulations are essential for understanding kinetic phenomena, but they can be too computationally expensive for rapid parameter studies, design exploration, or repeated analysis. Reduced-order modeling (ROM) offers a practical way to accelerate simulation by constructing compact surrogate models that preserve the dominant behavior of the full system.

This lecture introduces the basic ideas of projection-based reduced-order modeling in a way that is accessible to participants without prior background in machine learning or AI. We begin with simple examples (such as diffusion-type problems) to build intuition for key concepts, including snapshot data, low-dimensional basis construction (e.g., proper orthogonal decomposition), and projection (e.g., Galerkin projection). Throughout the lecture, these ideas will be demonstrated using the open-source libROM toolbox, providing a practical, hands-on view of how projection-based ROM workflows are implemented.

We then move to more challenging applications in electrostatic plasma physics, focusing on kinetic models such as the Vlasov–Poisson system. In this setting, a single global reduced basis may struggle to represent multiscale, strongly evolving, or instability-driven dynamics over long time intervals. To address this, the lecture presents local reduced-order modeling, in which the solution manifold is decomposed into multiple regions (or time windows), and a tailored reduced model is built for each region.

A central topic is physics-informed solution manifold decomposition, where physically meaningful indicators (such as physical time and electric-field energy) guide the construction of local ROMs. This approach improves robustness and accuracy while retaining interpretability, especially for challenging electrostatic plasma problems.

Youngsoo is a staff scientist at LLNL’s CASC group, where he develops efficient foundation models for computational science. His research focuses on creating surrogates and reduced-order models to accelerate time-critical simulations in areas such as inverse problems, design optimization, and uncertainty quantification. He has pioneered advanced ROM techniques—including machine learning-based nonlinear manifolds, space-time ROMs, component-wise ROM optimization, and latent space dynamics identification—and currently leads the libROM team in data-driven surrogate modeling. His contributions extend to open source projects such as libROM, pylibROM, LaghosROM, ScaleupROM, LaSDI, WLaSDI, tLaSDI, gLaSDI, NM-ROM, DD-NM-ROM, and GappyAE. Youngsoo earned his BS from Cornell and his PhD from Stanford, and he was a postdoc at Sandia National Laboratories and Stanford University before joining LLNL in 2017.

Recent innovations in machine learning have enabled a new era of computational science where models are no longer static simulations but tools that are scalable, differentiable, hardware-agnostic, and faster to develop. This talk explores the shift toward differentiable programming in physics, demonstrating how JAX allows researchers to fuse traditional PDE solvers with black-box ML representations to create systems that are directly optimizable via gradient-based methods. We provide a technical deep dive into the JAX ecosystem, illustrating how to encode numerical algorithms as tensor expressions while mastering functional transformations like vectorization, automatic differentiation, and seamless scaling across CPU, GPU, and TPU clusters. By implementing a structured grid PDE solver from scratch, including stencil evaluation and boundary conditions, we demonstrate how to optimize complex system parameters against target design objectives. We also examine several relevant open source tools, including both TORAX and DESC.

Elliot English is a software engineer at Google, where he leads efforts accelerating ML and HPC applications using JAX and TPUs. His core areas of expertise include frontier AI model development, the design of numerical methods for multiphysics, and high-performance software/hardware codesign. He is particularly interested in applying these to computational engineering problems in fusion energy and sustainable technologies. Prior to joining the ML frameworks and compilers team at Google, he was stellarator optimization lead at Renaissance Fusion, a fellow at Syntiant, CTO at Pilot AI, and software engineer at MetaMind. Elliot was also a postdoctoral scholar at Lawrence Berkeley National Laboratory and holds a PhD in computational mathematics from Stanford University.

Artificial Intelligence is rapidly transforming the fusion research and technology. This presentation will be on one critical application: the machine learning surrogate model for the turbulent transport simulations, which is central for Whole Device Simulation in Tokamaks. We will discuss our recent work TGLF-SiNN that improves the robustness and resolves the data scarcity challenge for training this surrogate by careful feature tuning, physics-guided learning, and active learning. TGLF-SiNN is deployed into GA’s simulation pipeline, it reduces training data requirement by 75% and accelerates whole device simulation by 45x.

Yadi Cao is a postdoctoral researcher at UCSD with Professor Rose Yu. He develops novel AI methods to scale up engineering research by reducing experimental costs and expert labor. Leveraging surrogate modeling and agentic workflows, his work has driven major advances in fusion energy and fluid dynamics, including TGLF-SiNN—which delivers a 45x speedup and is integrated into General Atomics’ tokamak pipeline—and BSMS-GNN for realistic turbomachinery simulations on million-node meshes. Yadi is currently seeking a tenure-track faculty position and collaboration opportunities.

Inertial Confinement Fusion (ICF) plays a critical role in national security, the study of exotic states of matter such as those found in planetary interiors, and in the pursuit of fusion energy. However, predictive capabilities in ICF remain limited because it is a highly coupled, multiphysics, and energy-constrained system. Progress in the field has largely relied on simulations to guide experimental design, combined with systematic semi-empirical tuning of experiments, including emerging machine learning approaches. This talk introduces Prospero; a large language model (LLM) developed specifically for ICF applications. Prospero is designed to contextually retrieve and synthesize information from scientific literature. When integrated with experimental data and used interactively with subject matter experts through natural language interfaces, the system has the potential to generate and refine hypotheses that explain observations. The development process behind Prospero is described, with emphasis on building a trusted collaborative aide capable of supporting ICF researchers. Key technical and scientific challenges associated with this effort are discussed. Finally, several use cases are described to illustrate how combining literature knowledge, experimental data, and expert interaction could potentially accelerate progress in the field.

Radha Bahukutumbi leads the Plasma Theory and Applications Group at Los Alamos National Laboratory. Her research focuses on high-power laser applications, including inertial confinement fusion (ICF), nuclear astrophysics, hydrodynamics, and laser–plasma interactions. Prior to joining LANL, she was a Distinguished Scientist at the Laboratory for Laser Energetics at the University of Rochester, where she wrote the radiation hydrodynamic code DRACO and led ICF simulation efforts and experiments on the OMEGA and National Ignition Facility lasers for over two decades. Radha is a Fellow of the American Physical Society and a recipient of the Fusion Power Associates Leadership Award. She earned her PhD from the California Institute of Technology.