Resolving Thin Beds Using AI-Based Petrophysical Property Prediction

Thin-bed reservoirs are challenging to characterize due to their sub-wavelength thickness, interbedding, and stratigraphic complexity. These issues are compounded in complex stratigraphy and areas with sparse well control, where conventional seismic inversion techniques often fail to resolve thin pay zones. In this study, we apply a deep learning (DL) method based on a temporal convolutional network to directly predict elastic and petrophysical properties from seismic data.

The model is trained using well-calibrated impedance, Vp/Vs, gamma ray, porosity, and gas indicators, and incorporates stratigraphic priors including interpreted surfaces and relative geologic time volumes to improve vertical resolution and geological consistency. Fluid indicators are estimated following rock physics crossplotting methods.

Applied to the fluvial Mungaroo Formation in the Carnarvon Basin, the model achieves strong blind-well performance and accurately resolves the gas-bearing M16 sand, which is approximately 10 milliseconds thick. Compared to deterministic inversion, the DL approach improves vertical localization, sharpens lithological contrasts, and enables full-volume prediction suitable for interpretation and risk screening.

This work demonstrates how AI-based workflows, when guided by stratigraphy and rock physics, can enhance thin-bed resolution and support more informed subsurface decisions in exploration, development, and CO2 storage contexts.