Ученые Кореи создали ИИ для мгновенной оценки сейсмических рисков на атомных станциях

Previous Next Sort Images/Video Newswise — The Korea Research Institute of Standards and Science (KRISS, President Dr. Lee Ho Seong) has developed a deep learning technology that predicts, in real time, the vibrations at multiple locations inside a nuclear power plant using the signal from just a single seismometer, and that quantifies the risk at each location to enable the prioritization of inspections. In particular, because the AI model can be flexibly designed to match the vibration characteristics of a structure, it is expected to be widely applicable to major industrial facilities such as semiconductor plants and data centers in the future. After the magnitude 5.8 Gyeongju earthquake in 2016, Wolsong Nuclear Power Plant Units 1 through 4 underwent 80 days of performance testing and detailed inspections before being restarted. Likewise, during the powerful earthquake that struck Kumamoto, Japan, this past July, TSMC's Kumamoto Fab 1 temporarily halted operations and then returned to normal in stages after equipment inspection and adjustment. As these cases show, even when there is no major damage after an earthquake, the process of verifying safety can prolong shutdowns, leading to considerable downtime and economic losses   The problem is that immediately after an earthquake, there is not enough data to determine where to check first. Installing sensors at every major equipment location would allow the vibrations at each point to be identified quickly, but this is practically impossible in a nuclear power plant due to constraints such as cable routing, licensing, and maintenance within radiation zones. As an alternative, computational analysis can be used to calculate the vibrations at locations without sensors, but this takes anywhere from several hours to several days, limiting the ability to respond quickly. To solve this, the KRISS research team developed a virtual sensing technology in which AI analyzes the seismic wave signal measured by a single seismometer to infer, in real time, the structural response at 139 points inside a nuclear power plant that have no sensors. Without installing sensors at every location, it can quickly narrow down the areas most affected by an earthquake, and it demonstrated excellent predictive performance even on records from real earthquakes that were not used in training. Furthermore, by incorporating uncertainties arising from factors such as structural characteristics, the technology quantifies, on a scale of 0 to 100 percent, the probability that the structural response at each location will exceed a predefined risk threshold. Going beyond a binary classification of simply "safe" or "at risk," it ranks locations by their risk, helping experts quickly determine inspection priorities. The team also greatly improved the efficiency and versatility of the AI design. The researchers developed a design equation that derives the optimal AI architecture based on a structure's natural frequency, reducing the trial and error of repeatedly designing and comparing multiple models. The AI model designed this way achieved accuracy comparable to that of a state-of-the-art deep learning model more than 200 times larger in terms of the number of parameters, while substantially lowering the computational burden and thus improving its usability even in resource-limited environments. This has opened up the possibility of extending the technology to a wide range of industrial facilities, including not only nuclear power plants but also semiconductor fabs, data centers, and industrial plants. Dr. Lee Jaebeom, Senior Research Scientist at KRISS, said, "The key to this research is that it quantifies even the uncertainty in the seismic response prediction results, allowing experts to determine which equipment to inspect first." He added, "In the field of safety, AI must not only make accurate predictions but also recognize the limits of its own judgment. Building on this, we will develop trustworthy AI technology that flags uncertain situations to support additional human judgment." This research was conducted jointly by KRISS Senior Research Scientist Dr. Lee Jaebeom and the research team of Professor Lee Young-Joo of the Ulsan National Institute of Science and Technology (UNIST), with Lee Jingoo, a student researcher, as the first author. The work was supported by an individual basic research program of the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (MSIT) and by the KRISS project "Development of Digital Safety Measurement to Enhance the Availability of Smart Structural Monitoring of Facilities", and other programs. The related results were published in a total of three papers, including one in Reliability Engineering & System Safety (IF 13.7), a leading international journal in the top 1.4 percent of the civil engineering field. Journal Link: Reliability Engineering & System Safety, Mar-2026 MEDIA CONTACT Register for reporter access to contact details Article Multimedia CITATIONS Reliability Engineering & System Safety, Mar-2026 TYPE OF ARTICLE Research Results CHANNELS All Journal News Artificial Intelligence Energy Engineering Nuclear Power Technology