Enhanced Deep Learning Using a Convolutionally Embedded Transformer Architecture for Reservoir Elastic Property Prediction: Comparison With Physics‐Based Seismic Inversion

Authors:
Publisher website
Google Scholar
Abstract
Deep learning (DL) algorithms are increasingly being used to address geophysical inverse problems, particularly when physical models are only partially known. This study introduces a supervised DL encoder–decoder approach based on a fully convolutional transformer architecture, where convolutional attention and wide‐focus modules are used to estimate reservoir elastic properties from seismic data. The DL method is evaluated against three physics‐based (PB) amplitude‐versus‐offset (AVO) inversion techniques: Bayesian linearized inversion, ensemble smoother with multiple data assimilation, and gradient‐based optimization. The DL model is trained on well‐log data to infer elastic properties from partial‐angle seismic stacks and interval velocity volumes, whereas the PB AVO inversion methods integrate geostatistical priors with seismic data through Bayesian inference, ensemble‐based data assimilation and gradient‐based optimization. Both approaches are applied to seismic and well‐log data from the Troll field in the North Sea. The network is trained using multiple well‐log datasets, using nearby seismic traces and interval velocity as input features, and then applied across the full three‐dimensional seismic volume. In contrast, the PB methods are applied to a two‐dimensional (2D) vertical section. The results are compared at training wells, multiple blind wells and along 2D sections of the predicted elastic properties. The DL approach demonstrates superior accuracy at training wells and competitive performance at blind wells, outperforming PB methods in some cases while producing comparable results in others. At blind well locations, the DL model produces under‐dispersed uncertainty estimates, indicating that it does not fully capture the variability represented by the PB methods. Results from real seismic field data highlight the complementary value of combining data‐driven and PB inversion strategies.
Funding Information
  • Norges Forskningsråd (331644)

This publication has 59 references indexed in Scilit: