Researchers at Japan’s Shibaura Institute of Technology have combined artificial intelligence with the Finite Element Method (FEM) to deliver more reliable predictions of grout behaviour in complex soils. This approach is especially relevant for tackling soil liquefaction—the process where saturated soil loses its structure, as seen during the 2011 Great East Japan Earthquake.
Chemical grouting has long been used to stabilise soil by injecting a hardening chemical, but consistent permeation in heterogeneous, low-permeability soils has remained a challenge. By merging AI-driven predictive modelling with traditional FEM analysis, the team has made significant strides in understanding and forecasting grout dynamics.
Professor Shinya Inazumi, who led the research, notes that previous work depended solely on conventional permeation analysis. In contrast, the team built a soil model featuring low-permeability zones and performed a two-dimensional FEM analysis to measure permeation velocity. They then trained AI models—including a neural network and a gradient boosting decision tree—achieving an impressive R² value of 0.849.
One of the most striking benefits of this new method is speed: while FEM simulations can take up to 40 minutes per prediction, the AI models complete the task in under two seconds. Additionally, the AI approach improved the average permeation rate from 94.5% to 96% and modestly reduced the worst-case performance from 81% to 83%.
This study underlines how AI can enhance numerical modelling when paired with robust, well-structured data. However, the researchers advise that further validation with real-world field data is essential. Expanding the dataset could help fine-tune the models even further, ensuring both accuracy and reliability in practical applications.
Professor Inazumi believes these findings offer engineers a practical framework for predicting grout behaviour even under challenging soil conditions. Looking ahead, integrating additional physical properties—such as grout pressure and grain size distributions—into FEM simulations could further enhance the framework’s applicability.
