Role of site characterization information in data-centric geotechnics
Part Of
Databases for Data-Centric Geotechnics: Site Characterization
Start Page
1
End Page
49
ISBN (of the container)
9781040166406
9781032578958
ISBN
9781040166406
9781032578958
Date Issued
2024-01-01
Author(s)
Abstract
This chapter presents an overview of site characterization information with a focus on its availability, coverage, data attributes, value to decision making, and challenges. In terms of availability, it is accurate to say that site characterization databases mainly reside in industry. However, the majority is not shared for confidentiality reasons (“dark data”). The data-centric geotechnics agenda cannot focus purely on maximizing the value of practice but also on minimizing the leakage of sensitive data. It is suggested that the “data first practice central” agenda in data-centric geotechnics can be expanded to “data first, practice central, and privacy enhanced.” Data-centric geotechnics now encompasses four elements: (1) data centricity, (2) fit for (and transform) practice, (3) geotechnical context, and (4) privacy enhancement. In terms of coverage, publicly available site characterization databases are mainly geotechnical in nature containing borehole, laboratory test, and field test data. Detailed geological data such as soil stratification, rock mass features (weathering, discontinuities, etc.), and real-time data such as geoenvironmental processes (subsurface flow, contaminant transport, etc.) are lacking. This “paucity of ground truth” is a rate-limiting step in the development of machine learning algorithms for geo-related disciplines, such as geotechnical engineering, rock engineering, mining engineering, engineering geology, geophysics, and others in geosciences. In terms of data attributes, MUSIC-X (Multivariate, Uncertain and Unique, Sparse, Incomplete, and potentially Corrupted with “X” denoting the spatial/temporal dimension) is a basic challenge to data-driven site characterization (DDSC) research. This chapter proposes a new taxonomy of site data under “4S” to expand the agenda for future machine learning research. The “4S” are site generalizations, spatial features, sampling characteristics, and smart data. In terms of value to design, the majority of the research is focused on the prediction of soil properties (spatial variability) and soil types (stratification) at a target site. One particularly interesting and significant challenge called “site recognition” has attracted attention in recent years. The site recognition challenge is fundamental to geotechnical engineering because it attempts to quantify the “uniqueness” (or site specificity) attribute of a target site. Recent research has demonstrated that quasi-site-specific transformation models can be developed notwithstanding the well-known limitations - site-specific data is sparse, and a generic database containing data from other sites is not directly applicable to a target site. There are three notable achievements: (1) the site recognition problem is tractable even under MUSIC, (2) inference uncertainty produced by a quasi-site-specific transformation model is smaller than that produced by a conventional generic or site-specific model, and (3) increasingly effective quasi-site-specific transformation models have been constructed, particularly those that exploit clustering. The construction of a quasi-site-specific ground model that includes learning cross-correlations and spatial auto-correlations from generic databases under MUSIC-3X is an ongoing research problem. Research on other site characterization databases such as geophysical and remote sensing data and data fusion are limited. Interest in combining site characterization information with monitoring information collected during construction to guide decision making in real time is emerging rapidly. This new agenda is called the machine learning-guided observational method (MLOM). Its value is arguably higher than DDSC, possibly leading to a Type 3 (disruptive) outcome that can transform geotechnical practice. Other challenges such as explainability and interpretability, transferability, geo-compatibility, frugal HBM, thick data, and data protection are briefly discussed. It is uncertain whether a new discipline called geotechnical data science is needed to create novel theoretical concepts and methods for data-centric geotechnics. But trustworthiness needs to be addressed to comply with the EU AI Act and other regulatory frameworks of AI systems.
Publisher
CRC Press
Type
book part