Enhancing operational decision-making in hydrocarbon exploration drilling using machine learning for gas data interpretation

Abstract Artificial intelligence is increasingly used to support decision-making during hydrocarbon exploration drilling, but mud-gas interpretation remains challenging because gas signatures are influenced by mud properties, drilling parameters, degassing efficiency, and Drill Bit Metamorphism (DBM). Here, we present a machine-learning-assisted workflow for assessing how well reservoir-fluid signals are represented in Advanced Gas (AG) measurements acquired while drilling. A multi-domain dataset from 104 Brazilian exploration wells was quality controlled, harmonized, and integrated with laboratory pressure-volume-temperature (PVT) fluid compositions and expert geological interpretation. Two predictive products were developed: Reservoir Affinity Curves, which estimate the similarity between AG signatures and reference PVT fluids using C2- and C2C-based targets, and a DBM Severity Curve, which quantifies drilling-induced thermal alteration using ethylene-related behavior and operational variables. Kernel Ridge Regression was selected as the primary deployment model because it produced stable, smooth, and interpretable depth-dependent predictions, whereas XGBoost and LightGBM achieved the highest numerical accuracy as benchmark models. The workflow distinguished intervals dominated by representative formation-fluid signatures from zones affected by DBM or other operational artifacts. This approach supports earlier fluid characterization, improves fluid-sampling decisions, and reduces interpretation uncertainty before laboratory results become available.