Recently, a series of small earthquakes has been occurring near the Tokara Islands in Kagoshima Prefecture, Japan. Concerns about a major earthquake have spread online, compounded by the so-called 'Nankai Earthquake Theory' depicted in the comic 'The Future I Saw.'
Amid the repeatedly rising concerns about a major earthquake, the scientific community is ramping up its efforts to predict earthquakes. Knowing the time and intensity of an earthquake in advance is a key challenge that could save countless lives and reduce significant economic damage. With the full-scale introduction of explainable artificial intelligence (XAI), deep learning, and machine learning-based analysis systems, predictions about the timing and magnitude of earthquakes, which were previously deemed impossible, are gradually becoming a reality.
Scientists are detecting in real-time how, when, and where tectonic plates begin to move, going beyond simply observing whether the plates are in motion. It is known that gigantic earthquakes primarily occur in areas where one plate is being pushed beneath another due to the interaction of the Earth's tectonic plates.
Researchers from Monash University in Australia combined experimental models simulating tectonic movements with XAI. While general AI only presents predictive results, XAI explains the data on which the results were based. As a result, when AI predicts an earthquake, researchers can closely examine the basis of that judgment.
The model combined with XAI captured subtle deformations and distortions occurring in the lower part of specific tectonic plates or inland areas from a few hours to several months before a major earthquake. The AI detected the small signals sent from the earth's crust before an earthquake occurs.
The research team noted, "If finer observation networks and AI analysis are combined in the future, the feasibility of earthquake prediction will increase." They added, "Although this AI cannot independently predict earthquakes yet, it can provide patterns that help determine if an earthquake is imminent, significantly enhancing forecasting accuracy."
Researchers at the Maulana Azad National Institute of Technology in India introduced an AI-based earthquake prediction model that utilizes both meteorological and seismic data. The model focuses on predicting the magnitude of earthquakes by analyzing environmental factors such as temperature, humidity, and air pressure alongside earthquake-related data.
The key technology utilized in this process is an automated analysis pipeline based on a machine learning operating system. This structured approach manages the entire process from data collection to preprocessing, model training, and deployment using machine learning models, allowing the system to autonomously update and optimize the model as new data comes in. Depending on the situation, it can automatically select and use the most appropriate model.
When testing various machine learning algorithms in the analysis pipeline, 'Gradiant Corporation' produced more accurate results with a smaller data set, while 'LightGBM' yielded better results as the amount of data increased. This demonstrates the capability to flexibly and automatically apply the optimal algorithm based on the quantity and quality of data.
Researchers from India indicated that they built a powerful and scalable system that is flexible enough to be applied in real-world environments, stating, "In the future, it can be directly utilized in real-world earthquake response systems through real-time data integration and the connection of various external variables."
◇From earthquakes to landslides… AI's disaster mapping
When an earthquake occurs, it causes significant damage, but secondary disasters such as landslides that follow can further increase human casualties and economic losses. The challenge is that it is difficult to predict when and where these landslides will occur.
To address this, researchers at Chengdu University of Technology in China compiled data from 398,698 landslides that occurred following 38 major earthquakes since the 1970s to build the world's largest database, using it to develop a deep learning-based landslide prediction model.
First, the research team analyzed 17 landslide-triggering factors based on the data, including terrain, ecology, hydrology, and earthquake-related elements. They found that ground acceleration (intensity of vibrations), slope, and rock type were the most significant factors, while the ruggedness and irregularity of the terrain also had secondary effects. Furthermore, they categorized earthquake-prone zones around the Pacific Rim and the Alps-Himalaya region by climate, determining which factors are more crucial for earthquake prediction. Applying these results, models can be designed to predict landslides caused by earthquakes according to the region.
The deep learning prediction model trained on all data accurately predicted areas where landslides could occur with an 82% accuracy within less than a minute. Professor Pan Xuanmei of Chengdu University of Technology emphasized, "Even without high-resolution satellite images, we can generate real-time landslide risk maps immediately after an earthquake," adding that it is a system that can help emergency responders determine where to go within minutes.
The research team plans to integrate rainfall forecasts and aftershock data into this model to develop a compound disaster prediction system. They are working to connect it to drones or ground sensors and apply it to cloud-based platforms to link with actual field response systems.
References
Geophysical Research Letters(2025), DOI: https://doi.org/10.1029/2024GL114428
Scientific Reports(2025), DOI: https://doi.org/10.1038/s41598-025-00804-x
National Science Review(2025), DOI: https://doi.org/10.1093/nsr/nwaf179