Physics-guided attribute selection for improved direct hydrocarbon indication

The non-uniqueness of seismic responses to rock and fluid properties remains a major challenge in hydrocarbon exploration because similar seismic anomalies may arise from different geological factors. Consequently, identifying seismic attributes that are specifically sensitive to pore-fluid effects is critical for reducing the exploration risk associated with direct hydrocarbon indicators. In this study, we introduce a Physics-Guided Local Variability attribute-selection methodology that integrates synthetic sensitivity analysis with local seismic variability measurements to identify physically meaningful seismic attributes for hydrocarbon detection. The methodology is demonstrated in the Hutton Sandstone reservoir of the Eromanga Basin, Australia which is often targeted on structural highs where dry holes may have seismic responses much like nearby commercial reservoirs. Due to strong lateral stratigraphic variability, conventional direct hydrocarbon indicators sometimes fail. In this case study, statistical attribute-selection approaches, including principal component analysis followed by self-organizing map classification, also failed to distinguish commercial reservoirs from nearby dry holes. Synthetic perturbation analysis at a single calibration well was used to quantify the relative influence of fluid, porosity, and thickness variations on seismic attributes, whereas local variability analysis applied to real seismic data was used to evaluate attribute behavior around wells with known fluid saturations. Spectral components in the 30–35 Hz range were identified as the most diagnostic attributes for pore-fluid detection, with the strongest responses observed in the far-angle data. These attributes improved self-organizing map classification, seismic inversion, and amplitude-versus-offset interpretation, resulting in improved discrimination between hydrocarbon-bearing and water-saturated intervals. Blind validation at an independent well further confirmed the robustness of the selected attributes.