Three-dimensional(3D) reservoir models offer a detailed portrayal of the heterogeneous structures and properties of subsurface reservoirs, forming a backbone of hydrocarbon reservoir exploration and development. Deep learning has become a prominent topic in the field of complex reservoir characterization due to its exceptional capability to learn implicit features. This paper presents a large-scale three-dimensional reservoir reconstruction approach based on a Transformer-CNN autoencoder network. To tackle the complexities and high computational cost when handling large-scale reservoirs, we replace Transformer decoder with CNN decoder. This reduction in computational cost introduces inductive biases, such as locality and translation equivariance, enhancing modeling precision while optimizing the generation scale of single MAE. To address the limitation of the incremental generation scheme based on pattern stitching, we propose a zonal modeling strategy by training multiple MAE networks to capture multiple-scale features. Experimental results demonstrate that our method maintains accurate characterization of heterogeneous reservoirs even under sparse conditional data. The proposed method mitigates the issue of error propagation inherent in the traditional incremental generation techniques, thereby restoring the comprehensive spatial structure of large-scale reservoirs. • A Transformer-CNN Autoencoder Network was proposed for multi-scale reservoir characterization. • CNN decoder replaced Transformer decoder to reduce computational cost and improve accuracy. • A zonal modeling strategy optimized progressive generation of multi-scale reservoir models. • The proposed method effectively mitigates error propagation and reproduces spatial structure of reservoir.
Chen et al. (Wed,) studied this question.