Towards better earthquake risk assessments using machine learning and geological survey data

“A building is only as strong as its foundation” is a common saying to demonstrate the importance of having a stable, solid foundation to build. The type of foundation and design are important to ensure the structural safety of the building.

Several factors that can affect the design and laying of the foundation, the depth of the layer, that is, the depth where the underlying layer of the soil or rock is strong enough to support the foundation, is one of the most important things. This is because in areas that are prone to earthquakes and landslides, the depth of the bearing layer, also known as the depth of the bearing layer, serves as an indirect indicator of the risk of soil liquefaction, or as a risk of soil collapse and losing its rigidity and behaving like a liquid.

Naturally, accurate estimates of bearing layer depth are key to designing robust foundations, limiting soil liquefaction risks and reducing soil-related disasters.

Traditional methods for evaluating bearing layer depth, particularly the standard permeation test (SPT), are generally reliable, but are expensive, with both time- and labor intensive processes for obtaining underground soil samples. Therefore, cost-effective alternatives are essential.

To address this issue, scientists at Shibaura Institute of Technology (SIT) in Japan turned their attention to machine learning (ML). A team of researchers led by Professor Shinya Inazumi of SIT's Engineering Department utilized 942 geological survey records and SPT data from the Tokyo metropolitan area and adopted three ML algorithms: Random Forest (RF), Artificial Neural Networks (ANN), and Support Vector Machines (SVM) predicted bearing layer depth.

Their findings were published in the journal Machine learning and knowledge extraction.

“The inspiration for this study stems from urgent challenges in geological engineering in earthquake-disturbable urban landscapes like Tokyo. As a region with a history of devastating seismic events, such as the 1923 kanto earthquake, accurate prediction of the depth of the bearing layer is important, and it will improve safety,” Tanaka said, explaining the motivation behind the study.

In their study, researchers first trained and optimized selected ML models using SPT datasets. We then developed two experimental case scenarios, depending on the set of explanatory variables used for evaluation.

The first case scenario (CASE-1) adopted latitude, longitude, and rise as explanatory variables, while the second scenario (Case-2) contained information about the subterranean soil layer, i.e. information about the subterranean soil layer, in addition to the other three geographical parameters.

During the comparative evaluation, the researchers found that the RF model always outperformed ANN and SVM, particularly in terms of accuracy of depth prediction (mean absolute error in case-2 vs. 1.26 m for case-1) and robustness to noisy data. Furthermore, the prediction accuracy of all three models in the Case-2 scenario, including stratigraphic classification data as additional explanatory variables, was significantly improved.

Inspired by the findings, the researchers went a step further to investigate the effects of spatial data density on predictive performance. For this purpose, they generated six different data subsets with different spatial densities: 0.5, 1.0, 1.5, 2.0, 2.5, and 3.0 points/km².

They found that the prediction accuracy of the RF model in Case-2 improves with increasing data density. This indicates that spatially dense data sets are valuable to accurately predict bearing layer depth.

Overall, team research shows that ML, particularly RF, can provide much-needed alternatives to traditional methods for local disaster risk assessment. Furthermore, unlike SPT, the ML model is cost-effective and further improves integration between computing architectures and advanced real-time platforms can revolutionize infrastructure planning in seismically active regions, reducing reliance on expensive, localized testing while improving safety and efficiency.

Inazumi concludes that it highlights potential applications of research. “Our findings highlight the transformative real-world possibilities of ML models in geotechnology and urban planning, particularly in earthquake-prone regions such as Tokyo.