Geotechnical correlation field-informed and data-driven prediction of spatially varying geotechnical properties

Geotechnical measurements are often limited, leading to the use of interpolation techniques for interpreting spatial variations in geotechnical properties from sparse geo-data. Traditional geostatistical methods suffer from significant computational complexity. On the other hand, data-driven approaches often lack integration with geotechnical domain knowledge, potentially oversimplifying or complicating predictions related to the spatial variability of geotechnical properties. This study introduces a novel framework that combines geotechnical knowledge with data-driven methods to model inherent soil spatial variability incorporating Geotechnical Correlation Field (GCF) that reflects domain knowledge. The GCF, influenced by Autocorrelation Function (ACF) types and Scale of Fluctuation (SOF), provides a flexible basis for accurately representing spatially varying geotechnical properties. Using a large synthetic database comprising known ACF types and SOFs, we constructed a series of specialized neural networks. These networks identify random field parameters at different sites based on sparse data, and the estimated parameters can be directly used to calculate GCFs for a given site. The performance of the proposed method is validated using a set of synthetic data and a real case history in New Zealand. The results demonstrate the model can accurately predict random field parameters for irregularly spaced geo-data, even with limited information. Significantly, the GCFs offer improved physical interpretations and enhance the performance of subsurface modeling. The computational complexity of this method is independent of the number of soil cells, making it highly efficient and scalable.