8MLIGD

Eighth Machine Learning in Geotechnics Dialogue (8MLIGD) on “Data ground truths and 3D ground modelling”

In conjunction with the Fourth Workshop on the Future of Machine Learning in Geotechnics (4FOMLIG), Dongguk University, Seoul, South Korea, 26-28 August 2026

Organizers:
Prof. Kok-Kwang Phoon, Singapore University of Technology and Design, Singapore
Prof. Stephen Wu, The Institute of Statistical Mathematics, Japan

Date: 27 Aug 2026
Time: 2:30 – 5:30 pm
Venue: TBC

INTRODUCTION
The first Editorial for the journal Geodata and AI defines the term “geodata” as: (1) data relevant to geo-disciplines that include soil science, geotechnical engineering, rock engineering, earthquake engineering, geoenvironmental engineering, engineering geology, mining, geo-hazard analysis, geophysics, and remote sensing and (2) data for ML/AI development, training, validation, and inference. The second part of the definition is often overlooked. There is a need to emphasize a distinction between “physical ground truths” and “data ground truths”. The former is to support decision making based on the current physics-centric paradigm. The latter is to support decision making based on the emerging data-driven paradigm. This difference in intent is crucial, because it has far reaching implications on how data is viewed, what constitutes “data”, how it is analysed, and how it is applied. In fact, Geodata and AI was launched in 2024 in part to promote cross-pollination of research ideas between geo-disciplines, because geodata for the data-driven world does not respect the traditional disciplinary boundaries that were demarcated primarily based on physics.

In geotechnical engineering, “ground truths” refer to the physical ground reality that is verified by sampling or direct testing “on location”. One common ground reality is to identify soil types and stratigraphic profiles. “Ground truths” constitute the ultimate benchmark for physical model validation. It is part of deterministic physics-centric thinking that does not address spatial variability explicitly. Hence, soil layers are either horizontal or at most inclined, although there such idealized layers are rare exceptions in practice.

To make explicit that “data ground truths” is a distinct concept, this dialogue termed conventional “ground truths” as “physical ground truths”. “Data ground truths” refer to all measured (and possibly qualitative and experiential) data related to: (1) characterization of an in-situ volume of ground, (2) processes that modify it (natural or man-made), and (3) its responses to these processes. “Data ground truths” constitute training, validation, and inference datasets for ML/AI. In contrast to “physical ground truths” that reveal high quality information about the ground at a few specific points typically through direct sampling, “data ground truths” seek to unveil the entire 3D ground reality (evolving at different time scales) which is not possible through limited sampling. It is important to see “data ground truths” as foundational to a data-driven future in geotechnics. In the context of the journal Geodata and AI, geodata is viewed as “data ground truths” collected by various geo-disciplines for ML/AI. As noted above, “physical ground truths” is a small subset of this much larger dataset and its purpose is much more restricted.

The basic goal of 3D ground modelling is to create an image of the subsurface almost always based on site characterization and numerical data. The key challenge is data sparsity. The key opportunity is spatial data in the context of site characterization has been shown to exhibit certain patterns that can be exploited. It may be argued that 3D ground modelling originates from the empirical work of Danie G. Krige in the 1950s in the South African gold mining industry to solve biased estimations of ore reserves. Krige’s empirical methods were formulated into a rigorous statistical framework by Georges Matheron. He coined the term “geostatistics” and named the interpolation technique “kriging” to honor Krige’s pioneering work. 3D ground modelling is now part of practice and has proven its value in geotechnical engineering and many geo-disciplines.

The role of machine learning and AI in 3D ground modelling and the foundational questions on what types of data ground truths are useful for training/validation, how to train/validate effectively and efficiently with complex attributes and context in mind, and the role/responsibility of an engineer in decision making are unanswered as of today.

The 8MLIGD is a continuation of a pre-workshop entitled “From 304dB to Geoworld-dB – Challenges and Solutions in Geotechnical Database Sharing” that was organized before the First Workshop on the Future of Machine Learning in Geotechnics (FOMLIG), 5–6 Dec 2023, Okayama, Japan and the ISSMGE TC309/TC304/TC222 Third Machine Learning in Geotechnics Dialogue (3MLIGD) (https://www.tandfonline.com/doi/abs/10.1080/17499518.2022.2105366)

1. Availability of big indirect databases – e.g. New Zealand Geotechnical Database (NZGD) (https://www.naturalhazards.govt.nz/resilience-and-research/data-and-modelling/new-zealand-geotechnical-database/)
2. Complex attributes of geodata – e.g. MUSIC-3X-G (M = multivariate, U = unique and uncertain, S= sparse, I = incomplete, C= potentially corrupted, 3X = spatial variability, G = geologic uncertainty)
3. Data fusion across geo-disciplines, between geodata and performance data, etc.
4. Quantification of uncertainties

5. Role of machine learning and AI
6. New sensing methods

7. Value to decision making in practice
8. Regulations such as design codes, building codes.
9. Engagement with engineers according to trustworthy data-centric geotechnics (https://www.sciencedirect.com/science/article/pii/S3050483X25000073)

PROGRAMME
The purpose of a “dialogue” is to talk about issues that are consequential to the full-scale deployment of trustworthy data-centric geotechnics in practice, rather than research issues. Issues are not restricted to models, computations, and other technical advancements.
The format of this dialogue is to invite academics and practitioners to talk about their challenges in implementing data-centric geotechnics in general and 3D ground modelling to a project in particular. This will set the scene for a dialogue with the audience. The focal point of the discussion is to update the data-centric geotechnics agenda and data-driven site characterization agenda so that geotechnical engineering practice can participate more rapidly and more extensively in a data-driven future.
The audience will be invited to submit questions online during the presentations. The organizers will cluster the questions and use these clusters to organize the discussion segment.

Data-centric geotechnics (30 Mins)
This session presents a summary of a series of interviews with stakeholders in the geotechnical engineering industry called the “Geotech Digital Future Pulse” and invites two practitioners to respond from their own deployment experience. The aim is not to ask whether AI “works,” but to surface why value and feasibility diverge in geotechnics, and what would have to change—technically, organizationally, contractually, and in regulation—to close that gap. These questions feed directly into the subsequent dialogue on updating the data-driven site characterization (DDSC) agenda.

“Geotech Digital Future Pulse”, Stephen Wu (10 mins)

Two invited discussants respond from practice, 10 mins each:
William Cheang (Seequent) and Rie Wada (Kisojiban Consultants)

Each invited speaker is given 10 mins to cover how current state-of-the-art (SOA) and state-of-the-practice (SOP) is responding to the challenges (examples given above) or to propose new challenges

1. Steve Chai, Rocscience
   AI-agent Approach for Automating Geotechnical Engineering Simulations
2. Kok Hun Goh, Land Transport Authority, Singapore
   Data-Centric Geotechnics and 3D Ground Modelling for Implementing Underground Transport Infrastructure
3. Han-Saem Kim, Dongguk University, South Korea
   Toward a National Digital Ground: Trustworthy 3D Ground Modelling in Korea
4. Takayuki Shuku, Tokyo City University
   3D Subsurface Modeling: State of the Practice in Japan
5. Yu Wang, Hong Kong University of Science and Technology
   Development and Value of 3D High-fidelity Ground Models: Examples from Hong Kong
6. Keiran Wright, WSP Australia Pty Ltd
   The Challenges of Implementing Big Data in Commercial Practice

Discussion on how to update the data-centric geotechnics agenda and data-driven site characterization (DDSC) agenda for faster and more extensive adoption in practice (60 mins)

There are 2 inter-related discussion items:

Data-centric geotechnics – Issues concerning deployment of digital technologies in geotechnical engineering in general.
Background information includes the “Geotech Digital Future Pulse” survey and “The role of geotechnics researchers in the era of rapid AI advancement” (https://www.sciencedirect.com/science/article/pii/S3050483X26000365)
The Geotech Digital Future Pulse is a FOMLIG Council initiative that has, to date, gathered insights from more than twenty structured interviews with geotechnical engineers, technical leads, software developers, and decision-makers across consulting, contracting, software, and asset-owner organizations. The interviews asked where geotechnical decisions become hardest, what evidence is needed to sign off with confidence, and what data-driven and AI-enabled tools would have to demonstrate to earn trust in practice. A single signal runs through these conversations, which we frame as the Value Paradox: the places in geotechnical practice where data-driven and AI approaches could add the most value are systematically the places where they are hardest to deploy, trust, and be rewarded for, while the tasks where such tools deploy most easily tend to return the least engineering value. Put differently, value and feasibility are too often inversely correlated in geotechnics.

• Data is sparsest where insight matters most. The hardest, highest-value problem—understanding the ground and its “unknown unknowns”—is precisely where data is scarcest, so big-data methods are least applicable exactly where they are most needed. Geotechnics remains fundamentally a “little data” discipline.
• Near-zero error tolerance pushes AI away from engineering. The high-consequence, safety-critical decisions where expert judgment matters most are exactly where non-deterministic, non-reproducible outputs are least acceptable, so AI is often confined to administrative efficiency rather than core analysis and sign-off.
• Value is real but uncaptured. Site investigation—the foundational activity whose underinvestment causes most failures—is frequently treated as a commoditized “tick-box” exercise. Those who would pay for better data and characterization are rarely the ones who capture its value, and data sharing is driven by commercial incentive rather than goodwill.
• A regulatory vicious cycle. Reliability-based design remains unapproved in many jurisdictions, removing the incentive to invest in the very methods that could improve reliability and build the evidence base needed for approval. Senior engineers’ caution, far from mere conservatism, is a rational response to genuine liability and consequence.

Data-driven site characterization – issues related to deployment of 3D ground modelling in practice as a specific example of data-centric geotechnics
As background information, the original DDSC agenda was proposed in https://www.tandfonline.com/doi/full/10.1080/17499518.2021.1896005

• Ugly data
• Site recognition
• Stratification

This agenda has been expanded to 4”S” to keep pace with fast evolving technologies (hard & soft): site generalizations, spatial features, sampling characteristics, and smart data in https://www.taylorfrancis.com/chapters/edit/10.1201/9781003441946-1/role-site-characterization-information-data-centric-geotechnics-kok-kwang-phoon-jianye-ching-chong-tang

The goal of 8MLIGD is to update the DDSC agenda as a future research and practice roadmap to accelerate its adoption in practice.