Navigating Geotechnical Uncertainties: Quantification, Propagation, and Management

This is the written version of the Bright Spark Lecture, titled “Machine Learning in Geotechnical Engineering: Opportunity or Overstated Hype”, presented at the Third Workshop on the Future of Machine Learning in Geotechnics in Florence, Italy, on October 17, 2025. It discusses the development and application of machine learning solutions for the quantification, propagation, and management of the multifaceted uncertainties present in geotechnical engineering systems. Included in the discussion are uncertainties encompassing the inherent and spatially variable physical and mechanical properties of engineering materials, the spatial uncertainties in subsurface stratigraphy, and construction and measurement errors. This lecture demonstrates the interconnectivity of the quantification, propagation, and management of uncertainties in engineering workflows. It starts with a case study on the statistical characterization of the spatial variability of marine clay at a unique land reclamation site in Singapore, highlighting the impact of site-specific conditions on material uncertainty. The finite element simulation of a deep excavation project on this reclaimed land is then presented to demonstrate the impact of uncertainties on the construction and performance of the excavation support system. Managing uncertainties constitutes an important downstream task once there is an understanding of how uncertainties impact system-level behavior, and the use of sensors for continuous monitoring and Bayesian updating offers a viable channel for uncertainty management. These examples not only illustrate the importance of a holistic consideration of uncertainty quantification, propagation, and management but also highlight the need for advanced techniques to improve the workflow of these individual components. Machine learning (ML) offers vast potential in these areas. The development of two ML architectures for subsurface characterization and the concept of surrogate modeling for accelerating simulation-based reliability analysis and Bayesian updating are presented to demonstrate the wide applicability of ML in supporting the quantification, propagation, and management of uncertainties, thereby offering a viable direction for harmonizing and streamlining workflows for addressing uncertainties in engineering systems.