Sweet spots discrimination in carbonate reservoirs based on weakly supervised learning

Abstract

The spatial distribution of karst caves in carbonate reservoirs plays a key role in guiding well placement. However, field production data reveal a high risk of drilling low-production wells even in cave-dense regions, resulting in low hydrocarbon recovery rates. Consequently, the identification of high-production “sweet spot” reservoirs has become a priority in optimizing well placement and enhancing recovery. This study proposes a two-step method for sweet spot identification using weakly supervised learning. Firstly, a multi-input convolutional neural network (CNN) is employed to detect caves from seismic data, including depth migration data, ant tracking, impedance, and structure tensor seismic attributes. The detection results, along with the seismic data, are then input into a second CNN to predict reservoir effectiveness. Since the effective reservoir classification can only be validated through production data from drilled wells, the available training samples are limited. To address this limitation, we define a random path crossing multiple wells and extract corresponding 2D seismic profiles, cave detection labels, and well-controlled classification labels. Notably, classification labels are only available at well locations, with no labels between wells. In the reservoir classification phase, a weakly supervised 2D CNN is trained using an adaptive loss, which evaluates the output cave classification profiles at partially labeled targets. The CNN can generate consistent 3D sweet spot predictions along both inline and crossline sections. Field tests and case studies demonstrate the prediction accuracy of proposed workflow can reach approximately 80 %, providing a practical solution for drilling risks and optimizing hydrocarbon recovery in carbonate reservoirs.