Weakly supervised two-step CNN for karst cave detection and high-production sweet spot prediction in carbonate reservoirs

Accurate karst cave identification is critical for hydrocarbon exploration in carbonate reservoirs, where caves serve as primary storage spaces. Despite dense cave distributions, drilling risks persist as wells often encounter water layers, reducing productivity. This study introduces a two-step weakly supervised learning method for identifying highproduction “sweet spots”. First, a multi-input CNN analyses seismic attributes (depth migration, ant tracking, impedance, structure tensor) for cave detection. These results combined with seismic data then feed a second CNN for reservoir effectiveness prediction. To address limited well-based training samples, we design random 3D survey paths intersecting multiple wells, extracting 2D seismic profiles with partial well-location labels. A weakly supervised 2D CNN employs adaptive loss functions, evaluating classifications only at labeled well positions while maintaining consistency across unlabelled areas. Multi-directional 2D training enables robust 3D sweet spot predictions. Field tests demonstrate 80% accuracy in identifying productive reservoirs. The workflow reduces drilling risks by optimizing well placement in carbonate reservoirs, offering practical solutions for enhanced hydrocarbon recovery.