Tokyo, one of the world's most densely populated megacities, sits on a highly active seismic zone where the threat of major earthquakes is ever-present. One of the most destructive aspects of seismic events in Tokyo is a geological phenomenon known as soil liquefaction. This occurs when the intense shaking from an earthquake causes saturated, loosely packed soil to temporarily lose its strength and stiffness, essentially causing the ground to behave like a liquid. The devastating effects of soil liquefaction have been documented many times, such as in the 1995 Great Hanshin-Awaji Earthquake, the 2011 Great East Japan Earthquake, and the recent 2024 Noto Peninsula Earthquake.
Despite this well-known threat, existing tools for assessing soil liquefaction risks have significant of room for improvement. Traditional hazard maps are typically created using simple geostatistical methods and limited borehole data, producing low-resolution results at grid scales of 500-meter or more. This lack of detail is particularly problematic in a city like Tokyo, with its complex subsurface conditions, extensive reclaimed land, and soft soil deposits. More sophisticated techniques are therefore needed to capture fine-scale changes in soil layers and provide accurate risk assessment.
In response to this challenge, a research team led by Professor Shinya Inazumi from Shibaura Institute of Technology, Japan, is pioneering a new method for creating high-resolution 3D liquefaction hazard maps. Their study, published online on September 12, 2025, in Volume 28 of the journal Results in Engineering , detailed a framework that combines geotechnical data with artificial neural networks (ANNs)—a type of machine learning algorithm—to create a unified model for both geological modeling and liquefaction hazard assessment.
The researchers used an extensive dataset of 13,926 borehole records, making it one of the largest applications of artificial intelligence (AI) for geotechnical hazard assessment in Japan. By training their ANN model on these data, the team was able to accurately predict key subsurface properties at unsampled locations across a 200-meter grid. More specifically, they sought to predict both soil type and N-values, which are important measures of soil density and strength.
The proposed model achieved a high regression accuracy for N-values and soil classification, outperforming traditional methods and other comparable machine learning frameworks. This was in part due to the superior ability of the ANN to capture complex, nonlinear relationships within the data.
Using these predicted values, the team then calculated liquefaction potential indices to quantify risk across the city. The resulting hazard map showcased a level of detail significantly higher than existing official maps, delineating high-risk areas in reclaimed coastal zones and river floodplains with unprecedented clarity and precision. Notably, the model could successfully identify localized high-risk zones in locations like Koto Ward, which are particularly susceptible to liquefaction but may not be outlined as such on conventional maps. "Our study offers a robust, scalable model that not only enhances earthquake risk management in Tokyo but also serves as a transferable methodology for other megacities facing similar geohazard challenges worldwide," says Prof. Inazumi.
The results of this work demonstrate how machine learning can be effectively used in urban planning and civil engineering, enabling more informed decisions about where to build, what kind of foundations to use, and where to prioritize soil improvement measures. Furthermore, the framework's scalability allows it to be integrated into geographic information systems for dynamic and interactive visualizations that can aid in public awareness campaigns and disaster prevention. "By integrating advanced AI techniques with geotechnical data, this research sets a new standard for proactive risk management, supporting safer, more sustainable urban development and helping to protect millions of people in vulnerable regions," concludes Prof. Inazumi.
