With the launch of ChatGPT by OpenAI in 2022, generative artificial intelligence (AI) technology has attracted significant attention in various disciplines, which involves creation of digital content by mimicking different types of data (e.g., text, image, and video). Recently, generative AI has been applied to development of smart city, such as digitalization of aboveground and underground infrastructures. In geo-engineering practice, two-dimensional (2D) or three-dimensional (3D) subsurface ground models are critical to the design, construction, and management of infrastructures, and they aim to realistically depicts complex site conditions (e.g., spatial stratigraphy and geotechnical properties). Due to the time, budget, technical, or access constraints, measurements from ground investigations of a specific site are commonly sparse and limited, although some prior geological knowledge of the site or similar sites might be available from previous projects or geophysical survey. It is particularly challenging to integrate sparse site-specific measurements with prior geological knowledge to develop subsurface ground models with a high spatial resolution. To tackle this problem and construct realistic subsurface ground models, generative AI methods based on Generative Adversarial Network (GAN) is developed in this study.

The study comprises of geological interpretation, geostatistical prediction, and immersive visualization of subsurface ground models. A generative machine learning methods called multi-scale generative adversarial network (MS-GAN) is firstly proposed for delineating 3D subsurface stratigraphy from limited specific-site borehole data and a 3D training image representing prior geological knowledge. The proposed method is developed to automatically learn multi-scale stratigraphic information extracted from a 3D training image and generate 3D subsurface ground models conditional on limited borehole data in an iterative manner. 3D simulated and real data examples were used to verify that MS-GAN not only effectively generate multiple 3D realizations from sparse boreholes, but also provide best estimate of stratigraphy with quantified uncertainty at unsampled locations. MS-GAN was also applied to integrate limited boreholes with geophysical data from different sites in Hong Kong. The major stratigraphic knowledge interpreted by geophysical data are used to establish 3D training image and subsequently integrated with site-specific boreholes for development of 3D subsurface ground models. MS-GAN is shown to accurately delineate 3D stratigraphic variations with quantified uncertainty.

Random field theory is an effective tool for interpreting spatial variation of geotechnical properties. In this study, a GAN-based random field generator is developed for directly generating random from incomplete measurements. Compared to conventional random field generators, the proposed GAN-based generator is data-driven and bypasses a need of function form selection or parameter estimation. In addition, a user-friendly software, Analytics of Sparse Spatial Data based on Bayesian compressive sampling/sensing (ASSD-BCS), is used for modelling 3D spatial variation of geotechnical properties. A benchmark study is presented to evaluate the performance of ASSD-BCS in four aspects, including accuracy, uncertainty, robustness, and computational efficiency. Moreover, virtual reality (VR) technology is utilized to efficiently visualize spatial variations of geological and geotechnical features within a 3D subsurface model developed. A VR system is developed using related software and hardware devices currently available in the markets for immersive visualization and interaction with subsurface ground models. The results demonstrate that VR visualization of the 3D subsurface models in an immersive environment has great potential to revolutionize the geo-practices from 2D cross-sections to a 3D immersive virtual environment, particularly for development of smart cities.