Deep Learning Architectures for Predicting Geological Parameters of Pinch-Outs Using Deep Sensing Borehole Electromagnetic Instruments

This study explores the capability of Deep Neural Networks (DNNs) for predicting geological parameters in pinch-out reservoirs using Ultradeep Azimuthal Resistivity (UDAR) logs. The research aims to complement real-time geosteering decisions by accurately estimating pinch-out reservoir characteristics - resistivity, dip angles, and distances to bed boundaries - complex geological structures that are common in Norwegian Continental Shelf (NCS). Traditional inversion is computationally expensive and rely on subjective expert interpretation, resulting in inconsistent decision-making. The application of the DNN showed its effectiveness in interpretation of pinch-out reservoir parameters.

The comparative analysis of three DNN architectures was performed: LSTM-FCNN without data sampling, LSTM-Conv1D-FCNN, and Conv1D-FCNN with sampling. Quantile regression was used to address the prediction uncertainty. Results indicate that models using sampled data (LSTM-Conv1D-FCNN and Conv1D-FCNN) provided the best forecast, with normalized pinball loss (NPL) less than 2% for most parameters, except for the dip angle β (∼7%). The unsampled LSTM-FCNN model showed less reliable prediction, with NPL exceeding 5% for most of the parameters.

Despite accurate forecast for resistivity predictions and distances to bed boundaries, dip angle prediction remained uncertain.

Future work will analyse the impact of sampling strategies and model architectures to improve prediction accuracy and reduce uncertainty.