Abstract

In engineering geology and geotechnical engineering, subsurface soils and rocks are natural geomaterials and exhibit inherent variability in stratigraphy due to geological deposition process. Explicit knowledge of subsurface stratigraphy is a critical input for the analysis, design, and construction of geotechnical engineering systems. However, the accurate and reliable modelling of subsurface geological stratigraphy is challenging due to the limited number of available boreholes in practice and the complex nature of soil stratigraphy. This paper presents an innovative machine learning framework built upon the neighborhood aggregation technique for the prediction of digitized subsurface geological stratigraphy. To predict the stratigraphy at a given point of interest, neighborhood aggregation is first performed to intelligently consolidate the stratigraphy information from its neighboring boreholes, resulting in additional features associated with the target location. By combining the extra stratigraphy information with conventional location-specific features, the framework enhances the predictive capabilities of classical machine learning models at a finer scale. The proposed framework is implemented using common machine learning models and is validated using a simulated benchmark 3D example. The results of leave-one-out cross-validation demonstrate that the proposed framework can improve the performance of classical machine learning models, leading to more reasonable stratigraphy transition and associated uncertainty quantification.

Abstract

In engineering geology and geotechnical engineering, subsurface soils and rocks are natural geomaterials and exhibit inherent variability in stratigraphy due to geological deposition process. Explicit knowledge of subsurface stratigraphy is a critical input for the analysis, design, and construction of geotechnical engineering systems. However, the accurate and reliable modelling of subsurface geological stratigraphy is challenging due to the limited number of available boreholes in practice and the complex nature of soil stratigraphy. This paper presents an innovative machine learning framework built upon the neighborhood aggregation technique for the prediction of digitized subsurface geological stratigraphy. To predict the stratigraphy at a given point of interest, neighborhood aggregation is first performed to intelligently consolidate the stratigraphy information from its neighboring boreholes, resulting in additional features associated with the target location. By combining the extra stratigraphy information with conventional location-specific features, the framework enhances the predictive capabilities of classical machine learning models at a finer scale. The proposed framework is implemented using common machine learning models and is validated using a simulated benchmark 3D example. The results of leave-one-out cross-validation demonstrate that the proposed framework can improve the performance of classical machine learning models, leading to more reasonable stratigraphy transition and associated uncertainty quantification.

Abstract

Quantifying uncertainties in subsurface properties and stratigraphy can lead to better understanding of the ground conditions and enhance the design and assessment of geotechnical structures. Several studies have utilized Cone Penetration Test (CPT) data and employed Bayesian and Machine Learning methods to quantify the geological uncertainty, based on the Robertson’s soil classification charts and the Soil Behaviour Type Index (Ic). The incorporation of borehole data can reduce the stratigraphic uncertainty. Significant challenges can arise, however, mainly due to the intrinsic differences between field and laboratory-based soil classification systems, which can potentially lead to inconsistent soil classification. To this end, this study proposes a multivariate Gaussian Process model that utilizes site-specific data and: i) jointly models multiple categorical (USCS labels) and continuous (Ic) variables, ii) learns a (shared) spatial correlation structure and the betweenoutputs covariance, and iii) produces two types of dependent classification outputs. The results indicate that the integration of geotechnical and geological information into a unified model can provide more reliable predictions of the subsurface stratification, by allowing simultaneous interpretation of USCS and Ic profiles. Importantly, the model demonstrates the potential to integrate multiple variables of different types, aiming to contribute to the development of a methodology for joint modeling of geotechnical, geological and geophysical data.