Fukushima Reactor: 3D Technology to Aid in Decommissioning

Franco-German scientific collaboration has led to the development of an innovative 3D reconstruction method, used recently in the damaged reactor of the Fukushima Daiichi nuclear power plant contaminated by radioactivity. This major scientific and technological advancement is now being utilized to inspect previously inaccessible areas, aiding in the planning of the site's decommissioning, years after the nuclear accident.

Scientists from Inria, the French national institute for research in digital science and technology, and the Max Planck Institute for Informatics (Germany) have jointly developed a reconstruction technology that generates interactive 3D visualizations from video sequences. This method, known as "3D Gaussian Splatting," has enabled the inspection and modeling of the interior of the plant's reactor, a previously inaccessible area, thereby contributing to the targeted planning of the reactor's decommissioning.

On March 14, 2024, the plant operator TEPCO (Tokyo Electric Power Company) sent two micro-drones, each weighing 185 grams, into the "pedestal" area located directly beneath the nuclear reactor vessel, where melted fuel residues are potentially located. The 3D Gaussian Splatting method then transformed these video sequences into an interactive 3D visualization, allowing for a digital exploration of the area, similar to navigating a video game environment.

The method combines the geometric principles of computer graphics with modern optimization techniques derived from artificial intelligence research. Initially, the method uses the camera positions moving through the scene from multiple individual images, such as video images, using an established technique known as “Structure from Motion.” The principles of the method are as follows:

• Generation of a sparse 3D point map with the main features of the scene
• Conversion of points into 3D Gaussian functions, i.e., small three-dimensional "ellipsoid shapes," each with a precise position in space, color, shape, and transparency value
• Iterative optimization process to adjust, complete, or remove these elements until they match the original image data as closely as possible, resulting in a smooth and realistic rendering from any viewing angle

The method was developed as part of the ERC Advanced Grant FUNGRAPH at the Inria Centre at Université Côte d'Azur. "This result is the culmination of 20 years of research in image-based rendering and the main outcome of the ERC FUNGRAPH," explains George Drettakis, head of the Inria GRAPHDECO team. "Our approach allows for much faster, interactive, and realistic rendering of 3D reconstructions, enabling seamless navigation through scenes, even with image quality equal to or better than previous methods," explains Thomas Leimkühler, group leader of the Image Synthesis and Machine Learning research group at MPI for Informatics.

While the Fukushima nuclear power plant decommissioning project is one of the first concrete applications of this technology in a high-risk environment, the potential of this method extends far beyond the nuclear sector and offers many other applications. Several ongoing commercial licenses confirm the interest of sectors such as:

• Cinema and video games for generating ultra-realistic sets
• Virtual reality for culture, tourism, or archaeology, for example, to reconstruct real places to explore
• Online commerce and social networks to create ultra-immersive objects or environments
• Generative AI for creating 3D content, which uses Gaussian Splatting as a 3D representation

The scientists behind the "3D Gaussian Splatting" method—Bernhard Kerbl, Georgios Kopanas, and George Drettakis (Inria, Université Côte d'Azur), along with Thomas Leimkühler from MPI for Informatics—have been awarded the Best Paper Prize at the ACM SIGGRAPH 2023 conference, among other accolades. The project page is available here. First presented at the SIGGRAPH conference in 2023—a global reference in computer graphics—the "3D Gaussian Splatting" method has quickly become a staple in research, with over 8000 scientific publications in two and a half years.

Original publication:
Kerbl, B., Kopanas, G., Leimkühler, T., & Drettakis, G. (2023). 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4). 
https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

Further information:
Reconstructed sequences from the drone material: https://nextcloud.mpi-klsb.mpg.de/index.php/s/NzMm43KRCzDot4H

Press contact MPI-INF:
Philipp Zapf-Schramm
Max Planck Institute for Informatics
Phone: +49 681 9325 4509
Email: pzs@mpi-inf.mpg.de

Press contact Inria:
Magalie Quet, magalie.quet@inria.fr

About the Max Planck Institute for Informatics:
The goal of the Max Planck Institute for Informatics is to advance foundational research and drive innovation in key areas of computer science. The research at the institute covers a broad spectrum, ranging from the study of the fundamental principles of algorithms and logic to the study of systems such as the Internet and multimodal areas such as computer vision, computer graphics, databases and information systems, machine learning and artificial intelligence. By integrating these diverse areas and fostering a culture of collaboration, MPI-INF operates at the highest scientific standards and works to shape the computing landscape of tomorrow.