Innovative applications of visualization technologies for scientists and the public in magnetic confinement fusion

Abstract

Magnetic confinement fusion research heavily relies on massive data generated by large-scale experimental facilities. Researchers need to process experimental data and interpret theoretical results using complex mathematical and physical tools; however, the intuitive presentation of internal physical processes within these devices remains challenging. To address these challenges, this study focuses on innovative visualization technologies for magnetic confinement fusion. Based on high spatiotemporal resolution physical simulation data from the Experimental Advanced Superconducting Tokamak tokamak device, we constructed a dynamic presentation of tokamak edge instability evolution processes with cinematic visual effects. Furthermore, we developed three-dimensional reconstruction algorithms for parallel computing results, effectively extracting structural characteristics of Weakly Coherent Modes (WCM) in I-mode and filamentary structures of Edge Localized Modes (ELM) in H-mode. Through key information-preserving mapping and dimensionality reduction, we achieved model visualization in Extended Reality (XR) head-mounted displays, enabling researchers to intuitively compare the differences between these two instability structures and gain deeper understanding of simulation results. Furthermore, taking the “Keda Torus eXperiment” (KTX) device as the research object, we established an intelligent conversion system from CAD models to lightweight 3D models, employing adaptive mesh simplification algorithms based on geometric feature recognition. Under the premise of ensuring key geometric feature precision, the model face count is compressed to 1% of the original scale; we further developed a cross-platform multimodal interaction system supporting Apple Vision Pro and Meta Quest 3 devices. Meanwhile, through 3D Gaussian Splatting technology, we constructed a laboratory-level digital twin system that high-fidelity reconstructs the “Keda Torus eXperiment” laboratory in virtual space, providing a high-precision virtual platform for remote experimental rehearsal.

1 Introduction

Magnetic confinement fusion, as the primary technical pathway to achieve controlled nuclear fusion energy, is gradually transitioning from fundamental plasma physics research to the critical stage of reactor engineering integration. In 2025, the Experimental Advanced Superconducting Tokamak (EAST) in China successfully achieved 1,066 s of steady-state long-pulse high-confinement-mode plasma operation at 100 million degrees Celsius [1], marking a significant breakthrough in this field. However, modern tokamak devices can generate terabyte-scale multi-physics high spatiotemporal resolution data in a single discharge. The complex magnetohydrodynamic instabilities inside the plasma involve coupled problems spanning more than ten orders of magnitude in spatial and temporal scales. Traditional two-dimensional static visualization methods struggle to intuitively present the three-dimensional dynamic evolution processes within the devices, constraining researchers’ intuitive understanding of key physical phenomena.

Scientific visualization serves as an important bridge connecting numerical simulations, experimental data, and physical understanding. In recent years, Extended Reality (XR) technology has broken through the perspective limitations of traditional desktop visualization through immersive interaction, demonstrating significant potential in scientific exploration. Feibush et al. [2] highlighted the macro opportunities of XR technology in scientific discovery, while Liu et al. (2022) [3] further elaborated on the advantages of interactive XR technology in information visualization. However, existing research has primarily focused on medical imaging, geographic information, and other fields. XR visualization research for magnetic confinement fusion remains relatively scarce, particularly regarding the processing of terabyte-scale large scientific datasets.

To address these challenges, developing efficient large-scale data rendering and lightweight model construction technologies is crucial. In terms of rendering algorithms, 3D Gaussian Splatting (3DGS) technology achieves high-fidelity real-time rendering through explicit Gaussian primitive representation. Han et al. (2024) [4] developed distributed 3DGS workflows for high-performance computing environments, providing new approaches for efficient processing of large-scale scientific datasets. At the engineering application level, digital twin technology is gradually being applied to fusion energy research: Bhatia et al. (2025) [5] constructed a fusion power plant digital twin system based on the NVIDIA Omniverse platform; Yao et al. (2024) [6] proposed an agent-model-based cognitive digital twin framework. However, existing research still faces technical challenges in high-precision model lightweighting for XR devices and cross-platform multimodal interaction: CAD models for fusion device engineering design typically contain tens of millions to hundreds of millions of triangular facets, which present significant performance bottlenecks when directly used for XR rendering; existing mesh simplification algorithms struggle to achieve sufficient compression ratios while preserving key geometric features (such as solenoid winding topology).

2 Visualization and construction of tokamak instabilities

2.1 Overview of the EAST tokamak device

The Experimental Advanced Superconducting Tokamak (EAST) is the world’s first fully superconducting tokamak device with a non-circular cross-section, independently designed and built by the Institute of Plasma Physics, Chinese Academy of Sciences. It was completed in 2006 and achieved its first successful discharge, with the design goal of providing key experience and technical reserves for the construction and steady-state operation of the International Thermonuclear Experimental Reactor (ITER) and the China Fusion Engineering Test Reactor (CFETR) [7].

In April 2023, EAST successfully achieved 403 s of steady-state high-confinement-mode (H-mode) plasma operation, marking the first time internationally to obtain hundred-second-scale repeatable steady-state H-mode [8]. To date, EAST has set more than 15 world records, systematically advancing high-parameter long-pulse plasma physics research.

2.2 Visualization of fusion processes in the EAST device

To enable EAST device images to be rendered on equipment, this study applied the lightweight conversion system mentioned in Section 3.2 (developed based on the KTX device) to process the EAST device CAD model. The original mesh model obtained from discretizing the EAST device CAD model contained 165,909,086 triangular facets. The lightweight system developed in this study employs an edge collapse algorithm that achieves a processing rate of approximately 2,000,000 faces per hour. Ultimately, while preserving the original model features, the EAST device mesh model was compressed to 287,033 faces to accommodate Vision Pro device requirements, and further compressed to 58,070 faces to meet networked mode bandwidth requirements, as shown in Figure 1. The model is first exported into the universal. fbx format and the Apple-specific. usbz format, and then displayed on the Apple Vision Pro headset through Apple’s official APIs and the Unity engine. After compression processing, the EAST device model can be rendered and displayed in VR headsets, allowing users to interact with the model through gesture interactions, such as grabbing the model using a “pistol-style” gesture, as shown in Figure 2.

FIGURE 1

FIGURE 2

To present the fusion process within the EAST device to users, we developed an interactive control panel in the VR environment, displaying various plasma parameters including plasma current, particle density, ion temperature, and energy confinement time, and providing a series of device operation buttons, such as vacuum pump evacuation control buttons, as shown in Figure 3. Fusion process visualization first requires constructing the magnetic field that confines the plasma. To this end, we constructed animations of the EAST device electromagnetic coils and their generated magnetic fields, as shown in Figure 4. When users activate the cooling system, the coils turn blue, indicating they have been cooled to 4.2 K (−268.95 °C) liquid helium temperature and entered the superconducting state. The magnetic field lines of the earth-like magnetic field generated by the ohmic heating coils change color dynamically, indicating the rapid change of the coil current: exciting toroidal current in the vacuum chamber to ionize the injected neutral deuterium-tritium gas to form plasma; the toroidal current flows through the plasma to achieve ohmic heating and excite the poloidal magnetic field.

FIGURE 3

FIGURE 4

The visualization of the helical magnetic field produced by the superposition of poloidal and toroidal magnetic fields in tokamak devices is a practical challenge. After experimenting with various approaches, this study ultimately calculated the magnetic field equations based on EAST device parameters, generated helical magnetic field line data in MATLAB, and plotted helical magnetic field line animations in Blender software based on the same formulas, as shown in Figure 5. Furthermore, we created particle mode animations, using white curves to represent ions and purple curves to represent electrons, demonstrating the motion of both particle types around helical magnetic field lines. Users can switch to particle mode to intuitively observe this motion.

FIGURE 5

Users can increase particle density through dense torus injection or gas puffing operations, and increase ion temperature through wave heating and neutral beam heating systems, thereby transitioning the plasma operation mode from L-mode (low confinement mode) to H-mode (high confinement mode). The EAST device size parameters and magnetic field strength are insufficient to satisfy the Lawson criterion for achieving break-even. To demonstrate the triggering and self-sustaining of the fusion process, the design estimates the effect of size and magnetic field strength changes on energy confinement time based on a square relationship, allowing users to switch on the control panel to a state with larger device size and magnetic field parameters, and thereby further increase plasma density and ion temperature through the aforementioned operations until reaching the break-even point. We developed animations to demonstrate the fusion process under these conditions: deuterium and tritium undergo fusion reactions, releasing alpha particles constrained by the magnetic field and further heating the system, as well as neutrons moving in straight lines away from the system.

2.3 Visualization of plasma operation modes

Tokamak device plasma operation modes have been extensively studied, accumulating large amounts of experimental and simulation data. To present these data to users completely in the VR environment, the raw data used for showing the plasma operation inside the device reached the scale of 10 TB when stored in MATLAB. To visualize these data, this study developed three-dimensional reconstruction algorithms to compress these data and generate visualizable images.

First, the raw data stored in MATLAB was screened, retaining only the position and density information of different points at different times, removing all other useless columns, and retaining only data at 5 m intervals for subsequent generation of 200 frames of visualization images, removing data at other times. Afterward, data were exported in blocks to CSV format and compressed in parallel. After this preliminary compression, the original data was compressed to approximately 100 GB.

The CSV format data were then further processed using Houdini to generate visualization images with cinematic visual effects. First, each frame of data file was imported into Houdini to construct point clouds composed of a series of data points. Then the point clouds were rasterized, using density information as the scalar attribute for voxel mapping, converting each point into a gradient color mapping based on its density. After further smoothing/denoising operations on the obtained voxel field, it was output in VDB format with the advantage of sparse storage suitable for processing large-scale data. The voxel field is configured at a resolution of 10,24³ (1,024 × 1,024 × 1,024). As the target application is video rendering, the dimensions were not deliberately reduced for optimization purposes. The voxel field at each timestep is rendered, producing 200 frames per second of raw data, which is then encoded into a video at a playback frame rate of 30 fps. Through the above operations, massive raw data was compressed and reconstructed into small-size visualization images that can be loaded in the VR environment. As examples, this study extracted characteristics of two typical modes: Edge Localized Modes (ELM) in H-mode and Weakly Coherent Modes (WCM) in I-mode, and transformed them into cinematic-quality visualizations for comparative display, as shown in Figure 6. Comparative analysis shows that WCM involves multi-mode coupling, while ELM exhibits significant perturbation characteristics possibly dominated by the n = 4 mode; only when this dominant mode appears does it cause the entire boundary pressure collapse.

FIGURE 6

3 KTX laboratory digital reconstruction

3.1 Overview of the KTX reversed field pinch device

The Keda Torus eXperiment (KTX) is a large-scale Reversed Field Pinch (RFP) magnetic confinement fusion experimental device jointly designed and built by the University of Science and Technology of China (USTC) and the Institute of Plasma Physics, Chinese Academy of Sciences (ASIPP) [9]. It was officially completed and put into operation on 3 November 2015. KTX is the largest RFP device in Asia, together with RFX in Italy, MST in the United States, TPE-RX in Japan, and EXTRAP-T2R in Sweden, forming the global RFP research network.

In terms of physics research, KTX has successfully achieved three different types of discharge modes: standard RFP configuration, ultra-low q (safety factor) discharge, and low-current tokamak discharge. The flexibility of these multiple operation modes enables KTX to conduct comparative studies between RFP and tokamak, exploring the commonalities and differences between the two configurations in three-dimensional physics, electromagnetic turbulence, density limits, and other aspects. KTX device research also involves exploration of advanced operation modes such as the “Single Helical” state [10].

3.2 Intelligent conversion system from CAD models to lightweight mesh models based on the KTX device

CAD modeling uses precise spline interpolation and other methods to describe geometric bodies with smooth surfaces, while many downstream applications cannot directly process such representations; triangular primitives are natively supported by almost all graphics libraries and hardware systems, making triangular meshes the mainstream representation method in 3D modeling. Therefore, the primary step in building an intelligent conversion system is to discretize CAD models to obtain initial triangular mesh models. Mesh models obtained directly from CAD models contain massive triangular facets, requiring large storage space and consuming substantial computational resources when rendered in XR devices. Therefore, mesh models need further compression. This study employs the edge collapse algorithm for mesh simplification. The edge collapse operation refers to merging the two endpoints of an edge in a triangular mesh into a new vertex, thereby removing the edge and its two adjacent triangles. Iteratively performing this operation yields a simplified mesh that maintains the geometric features of the original mesh in overall structure [11].

Because the initial KTX and aforementioned EAST models contain an extremely high number of polygons, we first use the integrated feature-based B-rep geometric simplification algorithm in SolidWorks to reduce them to a model with moderate polygon numbers. We then choose the classic QEM edge collapse algorithm, via self-developed code, to simplify the model mesh [11]. To ensure minimal error between the simplified mesh and the original mesh, the algorithm determines the position of new vertices generated by collapse by summarizing various constraints into a system of linear equations, and uses a cost function to evaluate the degree of mesh change after each edge collapse operation as the cost of the collapse operation. All edges are placed in a priority queue in ascending order of collapse cost. Each time, the edge with the smallest cost is extracted from the queue for collapse. After completing the collapse, the vertex error information of the affected region is updated, and the collapse costs of related edges are recalculated to update the queue accordingly, proceeding to the next collapse until the model face count reaches a preset target or all remaining edge collapse costs exceed a set threshold. Referring to the results reported by Melax et al. [12], we select the cost function as follows:where denotes the Euclidean length of edge and , denote the sets of triangles incident to vertex and the side faces containing or containing both and, respectively, and is the dot product of the unit normal vectors of triangles and . This method can preferentially collapse flatter regions in the model where folding has less impact on the overall shape, while preserving contours, creases, and other regions with higher surface curvature that define the overall shape of the model.

Based on the above algorithm developed in this study, while ensuring solenoid winding topology precision (error <0.05%), the KTX device model face count can be compressed to 1% of the original scale. The compressed model can be rendered in real-time on different devices, so this study further developed a multi-platform multimodal interaction system that supports cross-platform collaboration between Apple Vision Pro and Meta Quest 3 devices.

3.3 KTX laboratory digital reconstruction based on 3D Gaussian splatting technology

To achieve high-fidelity digital reconstruction and networked sharing of the KTX device, this study adopted the 3D Gaussian Splatting (3DGS) technology [13] route to construct a high-performance virtual experimental platform for web deployment. Compared with traditional mesh-based reconstruction methods or Neural Radiance Field (NeRF) implicit representation schemes, 3DGS technology achieves scene representation through explicit anisotropic 3D Gaussian primitives, ensuring rendering quality while possessing real-time rendering capabilities and good web deployment characteristics. The core of 3D Gaussian Splatting technology lies in projecting 3D Gaussian spheres onto the 2D image plane through a differentiable rasterization process, and minimizing the photometric error between rendered images and real captured images during training [13].

The open-source 3D Gaussian Splatting framework adopted in this study is implemented based on the PyTorch deep learning platform. The core algorithm flow contains three main stages: sparse point cloud initialization, Gaussian parameter optimization, and adaptive density control. Specifically, the reconstruction process first uses the Structure from Motion (SfM) algorithm to recover camera poses from input images and generate a sparse point cloud. This point cloud serves as the initial seed points for 3D Gaussians for subsequent Gaussian sphere distribution and optimization. Given that the KTX device has numerous weak texture regions (such as smooth metal surfaces and repetitive structures), this study adjusted the density of the initial point cloud distribution. After preprocessing, accurate camera intrinsic and extrinsic parameters are obtained, along with a sparse point cloud containing a series of 3D points, serving as initialization input for subsequent Gaussian training.

In the Gaussian parameter optimization stage, this study conducted iterative optimization of Gaussian parameters on a workstation equipped with an NVIDIA RTX 4090 GPU (24 GB VRAM). Each input Gaussian sphere is completely defined by the following parameters: 3D position coordinates, opacity, and spherical harmonic function coefficients used to represent view-dependent color information. The loss function is defined as a weighted combination of L₁ loss and SSIM (Structural Similarity Index) loss L_o-SSIM between rendered images and real observed images:where based on results given in the literature [13], λ = 0.2. Based on continuous iterative comparison between rendered result images and training views in the captured dataset, the optimization process is achieved by gradually reducing the loss function value. To balance the rendering quality and operational efficiency of the virtual platform, an adaptive density control strategy is adopted during training: for Gaussian spheres with large gradients, cloning and splitting are performed to increase detail in corresponding regions; simultaneously, Gaussian spheres with excessively low opacity and minimal contribution to rendering are periodically pruned to effectively control the total number of Gaussian spheres, thereby controlling model complexity.

After iterative optimization through the above process, a 3D Gaussian field containing complete geometric and appearance information of the KTX device is ultimately generated. This field file is encapsulated as the core asset of the lightweight virtual experimental platform. Users can load this asset through web or desktop clients to achieve 360° free-view observation of the KTX device, component-level interaction, and remote experimental simulation rehearsal, thereby reconstructing the KTX laboratory in virtual space. On an RTX 4080 Laptop GPU with 16 GB of RAM, loading a 30–40 GB Gaussian splatting scene takes 30–40 s, with the real-time frame rate maintained at 30–45 fps during runtime. In contrast, on an RTX 4090 with 64 GB of RAM, the loading time is reduced to approximately 5 s, and the frame rate can be sustained above 60 fps. The software is able to run on both Apple and Android systems, accommodate diverse headsets including the Apple Vision Pro, Meta Quest, and Pico, alongside other environments such as smartphones and smart glasses. Its support for multi-device networked interaction leads us to conclude that it possesses cross-platform multi-modal interaction capabilities.

4 Conclusion

This study focuses on innovative visualization technologies for magnetic confinement fusion, systematically describing the innovative applications of cinematic dynamic simulation, Extended Reality (XR) interaction, and intelligent 3D reconstruction technologies in device simulation, theoretical presentation, and science communication. Based on high spatiotemporal resolution physical simulation data from the EAST tokamak device, this study constructed a visualization system for tokamak edge instability evolution processes with cinematic visual effects, introducing multimodal interaction mechanisms that support users in observing the real world and virtual device models simultaneously from mixed reality perspectives. Through gesture operations (such as “pistol-style” grabbing), users can control the model, select particle mode or fluid mode to observe plasma motion during tokamak ignition, and monitor key parameters such as current, plasma temperature, density, and energy confinement time in real-time.

Through three-dimensional reconstruction algorithms for parallel computing results, this study extracted structural characteristics of Weakly Coherent Modes (WCM) in I-mode and filamentary structures of Edge Localized Modes (ELM) in H-mode, achieving model visualization in XR headsets through key information-preserving mapping and dimensionality reduction. This enables users to intuitively perceive plasma motion characteristics under different operation modes and compare the differences between the two instabilities.

This study employed adaptive mesh simplification algorithms based on geometric feature recognition, establishing an intelligent conversion system from CAD models to lightweight 3D models based on the “Keda Torus eXperiment” laboratory. The system compresses model face counts to 1% of the original scale while ensuring solenoid winding topology precision (error <0.05%). A cross-platform multimodal interaction system was developed, supporting cross-platform collaboration between Apple Vision Pro and Meta Quest 3 devices, with multi-user collaborative latency controlled within 50 m. Furthermore, a laboratory-level digital twin system was constructed based on 3D Gaussian Splatting (3DGS) technology, high-fidelity reproducing the “Keda Torus eXperiment” laboratory in virtual space and providing a high-credibility virtual platform for remote experimental rehearsal.

The visualization technologies involved in this study are not only applicable to science communication but also assist researchers in understanding theoretical models through high-precision 3D simulations, providing insights for model optimization and research directions. Future research will further incorporate ITER-scale simulation results, committed to building a fusion metaverse platform supporting multi-device data linkage, providing a new cross-scale, cross-dimensional research paradigm for future fusion reactors.

References

• 1.

  YanJKongD. EAST team. China's “artificial sun”: experimental advanced superconducting Tokamak (EAST). Front Eng Manag (2025) 12:442–5. 10.1007/s42524-025-5502-3

  • CrossRef
  • Google Scholar

• 2.

  HanMSewellAInsleyJKnowlesJMateevitsiVAPapkaMEet alToward distributed 3D Gaussian splatting for high-resolution isosurface visualization. arXiv preprint arXiv:2509.05216 (2025). 10.48550/arXiv.2509.05216

  • CrossRef
  • Google Scholar

• 3.

  FeibushE. Opportunities for Extended Reality in Scientific Visualization. ASCR Workshop on Visualization for Scientific Discovery (2022).

  • Google Scholar

• 4.

  LiuRGaoMWangLWangXXiangYZhangAet alInteractive extended reality techniques in information visualization. IEEE Trans Human-Machine Syst (2022) 52(6):1–11. 10.1109/thms.2022.3211317

  • CrossRef
  • Google Scholar

• 5.

  BhatiaNCostaRPamelaSJPDavisAGonzalez BeltranANGopakumarVet alVisualizing digital twins of fusion power plants using NVIDIA omniverse. AIP Adv (2025) 15:045018. 10.1063/5.0261883

  • CrossRef
  • Google Scholar

• 6.

  YaoZWuHSongYChengYPanHWuMet alSurrogate model-based cognitive digital twin for smart remote maintenance of fusion reactor: modeling and implementation. Nucl Fusion (2024) 64(12):126007. 10.1088/1741-4326/ad7b56

  • CrossRef
  • Google Scholar

• 7.

  李建刚,万元熙et alEAST全超导托卡马克核聚变实验装置. 物理 (2006) 35(7):560–4.

  • Google Scholar

• 8.

  SongYTZouXLGongXBecouletAButteryRBonoliPet alRealization of thousand-second improved confinement plasma with super I-mode in Tokamak EAST. Sci Adv (2023) 9(1):eabq5273. 10.1126/sciadv.abq5273

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9.

  HaoWYuntaoS. Nuclear Fusion and Plasma Physics (2015) 35(4):241–8.反场竦缩调研报告

  • Google Scholar

• 10.

  ZhengJXuMYangQMaoWLiH. Design of magnet system for Keda Torus for Experiment (KTX) device. Fusion Eng Des (2017) 125:152–8. 10.1016/j.fusengdes.2017.05.089

  • CrossRef
  • Google Scholar

• 11.

  GarlandMHeckbertPS. Surface simplification using quadric error metrics. In: Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH ’97. USA: ACM Press/Addison-Wesley Publishing Co. 4, 9, 41 (1997). p. 209–16.

  • Google Scholar

• 12.

  MelaxS. A simple, fast, and effective polygon reduction algorithm. Game Developer (1998) 5(11):44–9.

  • Google Scholar

• 13.

  KerblBKopanasGLeimkühlerTDrettakisG3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Trans Graphics (2023) 42(4):139. 10.1145/3582430

  • CrossRef
  • Google Scholar