Plasma GraphRAG: Physics-Grounded Parameter Selection for Gyrokinetic Simulations
Ruichen Zhang, Feda AlMuhisen, Chenguang Wan, Zhisong Qu, Kunpeng Li, Youngwoo Cho, Kyungtak Lim, Virginie Grandgirard, Xavier Garbet · 2026
Read Plasma GraphRAG: Physics-Grounded Parameter Selection for Gyrokinetic Simulations by Ruichen Zhang, Feda AlMuhisen, Chenguang Wan, Zhisong Qu, Kunpeng Li, Youngwoo Cho, Kyungtak Lim, Virginie Grandgirard, Xavier Garbet (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-06279.txt.
Introduction
Zhang and eight colleagues introduce Plasma GraphRAG, pairing graph retrieval with LLMs for physics-grounded parameter picks in gyrokinetic runs. Manual paper reviews make picks slow and uneven. It builds a knowledge graph from plasma literature and retrieves over graph entities and relations, yielding context-aware ranges. Across comprehensiveness, diversity, grounding, hallucination, and empowerment metrics, it beats vanilla RAG by over 10 percent and cuts hallucination rates by up to 25 percent. The authors offer it as a method for discovery across data-rich domains.
Catalog record
| CREATOR | Ruichen Zhang, Feda AlMuhisen, Chenguang Wan, Zhisong Qu, Kunpeng Li, Youngwoo Cho, Kyungtak Lim, Virginie Grandgirard, Xavier Garbet |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39017 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.06279
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Plasma GraphRAG: Physics-Grounded Parameter Selection for Gyrokinetic Simulations Ruichen Zhang, Feda AlMuhisen, Chenguang Wan, Zhisong Qu, Kunpeng Li, Youngwoo Cho, Kyungtak Lim, Virginie Grandgirard, Xavier Garbet Abstract Accurate parameter selection is fundamental to gyrokinetic plasma simulations, yet current practices rely heavily on manual literature reviews, leading to inefficiencies and inconsistencies. We introduce Plasma GraphRAG, a novel framework that integrates Graph Retrieval-Augmented Generation (GraphRAG) with large language models (LLMs) for automated, physics-grounded parameter range identification. By constructing a domain-specific knowledge graph from curated plasma literature and enabling structured retrieval over graph-anchored entities and relations, Plasma GraphRAG enables LLMs to generate accurate, context-aware recommendations. Extensive evaluations across five metrics, comprehensiveness, diversity, grounding, hallucination, and empowerment, demonstrate that Plasma GraphRAG outperforms vanilla RAG by over 10 % 10\% in overall quality and reduces hallucination rates by up to 25 % 25\% . Beyond enhancing simulation reliability, Plasma GraphRAG offers a methodology for accelerating scientific discovery across complex, data-rich domains. I Introduction Understanding turbulence and transport in magnetically confined plasmas remains one of the central challenges in fusion energy research [ 2 ] . Gyrokinetic (GK) simulations are indispensable for modeling these multiscale phenomena, as they capture the kinetic behavior of charged particles while reducing the dimensionality of the full six-dimensional Vlasov–Maxwell system [ 19 ] . Modern GK codes are widely used to investigate plasma stability, turbulence-driven transport, and performance limits in devices ranging from tokamaks to stellarators [ 11 ] . These tools form the computational backbone of fusion research, providing insight into confinement optimization, transport barrier formation, and scenario development for future reactors. A critical prerequisite for accurate GK simulations is the selection of appropriate input parameter ranges, including normalized temperature and density gradients, safety factors, magnetic shear, and collisionality. These parameters have a decisive impact on turbulence characteristics and transport predictions, and their accurate specification is essential not only for predictive modeling but also for the construction of surrogate models and databases. Traditionally, identifying suitable parameter ranges has relied on expert judgment and manual reviews of experimental and theoretical literature. This manual process is time-consuming, error-prone, and difficult to reproduce, which results in inconsistencies across benchmarking studies and undermine confidence in simulation outcomes. Recent progress in artificial intelligence (AI), particularly large language models (LLMs), has opened new possibilities for automating knowledge extraction from unstructured scientific literature [ 7 ] . An LLM is a type of deep learning model trained on massive corpora to capture statistical patterns of language, enabling it to perform tasks such as summarization, reasoning, and question answering. To enhance domain specificity, LLMs are often combined with Retrieval-Augmented Generation (RAG), a paradigm where relevant documents are first retrieved from a knowledge base and then supplied as context to the LLM during generation [ 10 ] . While RAG has demonstrated effectiveness in many fields, standard implementations treat the literature as a flat corpus, failing to capture the structured interdependencies among physical variables that are central to plasma physics [ 29 ] . This limitation often leads to hallucinations or incomplete recommendations, particularly in highly technical applications such as gyrokinetic modeling. To address these challenges, we introduce Plasma GraphRAG , a novel framework that integrates Graph Retrieval-Augmented Generation (GraphRAG) with LLMs to automate the identification of physics-grounded parameter ranges for gyrokinetic simulations. Plasma GraphRAG builds a domain-specific knowledge graph, where nodes represent plasma parameters, device metadata, and bibliographic sources, while edges encode physical couplings and co-occurrence relations. Structured retrieval over this graph provides the LLM with explicit relational context, enabling accurate, transparent, and reproducible parameter recommendations while significantly mitigating hallucinations 1 1 1 In this study, hallucination refers to cases where the model generates information that is inconsistent with the retrieved evidence or established physical principles. A lower hallucination rate therefore indicates higher factual consistency and physical reliability in the model’s responses. . The key contributions of this study are summarized as follows: • We introduce Plasma GraphRAG, which constructs a domain-specific graph for GK modeling by harmonizing diverse literature sources into a unified, code-facing parameter space. By adopting standardized gyrokinetic notation and explicitly linking parameters with bibliographic evidence, the graph resolves long-standing inconsistencies across simulation codes and provides a reproducible foundation for parameter integration. • Plasma GraphRAG employs a retrieval mechanism that encodes physical couplings and co-occurrence relations among plasma parameters. Compared to standard RAG, this structured retrieval provides the LLM with richer context, enabling more accurate parameter extraction and improved interpretability through transparent evidence paths, while substantially reducing hallucinations. • We evaluate Plasma GraphRAG on controlled GK parameter identification benchmarks, showing improved response quality, diversity, and grounding relative to baseline methods. The framework streamlines the preparation of surrogate model databases, alleviates expert workload, and offers a scalable pathway toward reproducible, data-driven discovery in plasma turbulence and transport studies. However, the current benchmark remains limited in scope, and the evaluation metrics are primarily heuristic, suggesting that larger-scale and quantitatively validated studies will be essential in future work. II Related Work This section reviews prior work in two relevant areas: (i) GK plasma physics and turbulence modeling, and (ii) LLMs with retrieval-based reasoning. The former outlines the evolution of GK simulation tools and surrogate modeling techniques, while the latter focuses on RAG and GraphRAG approaches for scientific question answering. II-A Gyrokinetic Plasma Physics GK theory has been the cornerstone of turbulence and transport modeling in magnetically confined plasmas for several decades. Foundational work has established the nonlinear GK formalism for low-frequency waves, which underpins modern turbulence simulations [ 8 ] . Since then, dedicated local solvers such as GS2 , GENE , CGYRO , and GKW have been developed and widely adopted to model ion- and electron-scale turbulence, trapped electron modes, and zonal flow dynamics [ 14 , 3 ] . In parallel, global gyrokinetic codes such as GYSELA , ORB5 , and GT5D have been introduced to simulate turbulence and transport across the entire tokamak volume, capturing finite-radius effects and global profile variations [ 6 ] . These codes have enabled benchmarked studies of core transport phenomena, including the Dimits shift, and are now integral to scenario development for fusion devices. In parallel, reduced-order models such as Trapped Gyro-Landau Fluid (TGLF) [ 26 ] and quasilinear models like QuaLiKiz [ 11 ] have been introduced to bridge detailed GK physics and integrated transport modeling. These models have achieved substantial speedups at the cost of some fidelity, enabling broader exploration of design spaces. More recently, the field has embraced hybrid and data-driven modeling to address computational bottlenecks. Neural network surrogates trained on large GK datasets [ 3 ] have been developed to provide flux predictions several orders of magnitude faster than first-principles simulations. Generative models [ 5 ] and transformer-based 5D surrogates [ 9 ] have further enabled direct emulation of nonlinear turbulence with preserved spatiotemporal dynamics. Multi-fidelity methods have been proposed to combine reduced models with sparse high-fidelity GK data for improved predictive accuracy [ 21 ] . Ongoing benchmarking efforts have continued to validate solvers such as GENE and CGYRO under emerging physics regimes [ 16 ] . This evolution reflects a growing trend toward surrogate-augmented, data-driven workflows for turbulence-informed fusion scenario design. II-B LLMs with Retrieval LLMs have demonstrated impressive capabilities across a range of