Sequence Machine Learning for Geological Forecasting: Insights from the Open Geology Forecast Challenge

Summary Accurate forecasting of geological structures ahead of the drill bit is critical for safe and efficient geosteering. Such forecasting remains a significant challenge due to uncertainty in plausible layer configurations, which, even in simple formations, may include faults, dips, and folding. To address this, the open international Geology Forecast Challenge introduced a data-driven benchmark focused on probabilistic prediction of stratigraphic geometry. Participants were tasked with generating several layer-boundary-depth realizations that capture geological ambiguity while maintaining predictive accuracy. The dataset consisted of thousands of interpreted one-dimensional lateral stratigraphies from worldwide geosteering operations. Submitted models were evaluated on a hidden test dataset using the Negative Log-Likelihood metric, which emphasizes uncertainty calibration. The top-performing approaches employed advanced training and a combination of modern deep-learning architectures. Results show that probabilistic, sequence-aware machine-learning models can effectively learn geological variability. They provide a foundation for improving uncertainty-aware geosteering decision support in the future.