Learning the Earth’s Hidden Structure with Sparse Observations and Subsurface World Models

Speaker

Dr Gerrit Olivier

Abstract

LLMs, reasoning models, and agents are transforming how we work with information, but they still have a limited understanding of the physical world. In mineral exploration, this limitation matters. Putting a wrapper around an LLM does not create a world-class geologist, and AI has not yet delivered a step change in mineral discovery. A key reason is that exploration is not simply a pattern-recognition problem. The targets are hidden, the observations are sparse, and the most important geological structures are only indirectly constrained by drilling, geochemistry, and geophysics. Understanding the structure of the Earth’s subsurface is fundamentally an inverse problem with sparse direct observations and abundant indirect measurements. Drillholes provide precise but extremely limited information, while geophysical signals such as gravity, magnetics, and seismic data provide broader but indirect constraints on the underlying geology. This seminar will introduce the concept of subsurface world models: generative AI systems trained to learn the statistical and physical structure of the Earth from large ensembles of synthetic geological models and geophysical simulations. These models aim to learn the three-dimensional structure, geometry, and composition of the subsurface itself, combining sparse hard constraints such as drilling and geochemistry with dense indirect observations from geophysics to infer plausible three-dimensional geological structures while quantifying uncertainty. While much of the current discussion around world models focuses on robotics and autonomous driving, one of the largest opportunities may lie beneath our feet. Subsurface world models offer a new framework for AI-assisted mineral exploration and resource discovery.

Bio

Gerrit Olivier is Chief Scientific Officer at Fleet Space Technologies, where he leads scientific strategy across geophysics, geology and machine learning. His work focuses on translating multiphysics data, AI, and compact sensor technologies into scalable systems for 3D subsurface imaging. Gerrit completed his MSc in Theoretical Physics at Stellenbosch University and his PhD in Geophysics at Université Grenoble Alpes. His early research established passive seismic methods as practical tools for mineral exploration, and he has received numerous awards for his contributions to applied geophysics.