Generation of Local Geological Structure Ahead of the Bit Using Diffusion-Based AI Models

Recent advances in generative AI have shown strong potential in learning and reproducing complex spatial structures. This study explores the latest diffusion models to generate multiple plausible realizations of subsurface geology ahead of drilling. We aim to preserve complex geological structures while reflecting the uncertainty of the subsurface environment. As new geological information is acquired during drilling, the model updates its predictions and progressively reduces uncertainty. We employ a denoising diffusion probabilistic model (DDPM) trained to reconstruct structural geological realizations forward in space while honoring the data acquired behind the bit during drilling. This process is similar to handwriting, where each stroke builds on the previous one without complete context. Similarly, our model iteratively denoises random fields into coherent subsurface representations, combining both local and global spatial dependencies. These dependencies are learned prior to geosteering using geological training representative of the target structural setting. This framework shows how diffusion models can be used as a probabilistic tool for data-driven structural modeling. We apply the same machine learning framework to two distinct datasets: handwriting and geological models. Trained diffusion models can produce visually consistent structures in two scenarios: (1) generating complete realizations and (2) extending realizations conditioned on available information with natural smooth transitions from the pre-defined region. The ability to generate conditioned geological models in real-time makes this approach suitable for uncertainty-aware geosteering workflows. This work presents early insights into the use of diffusion models for probabilistic subsurface geomodeling. The resulting AI geomodeling system can produce coherent, geologically consistent realizations that honor input constraints and exhibit spatial continuity across generated samples, providing a new, constructive, data-driven approach to real-time subsurface characterization. The analogy to handwriting generation offers a new conceptual benchmark for further development and verification of generative-AI subsurface modeling techniques.