Abstract
Summary: Depletion delineation is of critical importance to the success of recovery of unconventional reservoirs. Existing analytical and semianalytical approaches provide rapid pressure estimates but are generally limited by simplified physics, whereas full-physics embedded-discrete-fracture-model (EDFM)-based reservoir simulation captures nonlinear multiphase depletion behavior at high computational cost. Recent deep-learning surrogates improve efficiency, but standard encoder-decoder models often under-resolve the sparse and highly localized pressure and saturation changes around hydraulic fractures in low-permeability reservoirs. In this work, we present a convolutional neural network (CNN)–based supervised surrogate for rapid single-well depletion delineation in fractured unconventional reservoirs, targeting the key bottleneck of highly localized pressure and saturation evolution near hydraulic fractures. The method learns a generalized black-oil proxy that maps fracture geometry (image) together with rock/fluid properties and a parameterized production history to transient pressure and saturation fields on a fixed 512×512 grid using U-Net. The novelty of this work lies in two localization-aware improvements: (1) a physics-guided focal regression loss that adaptively shifts between mean-squared-error (MSE)-like behavior and focal emphasis using an image-specific diffusion indicator β, forcing learning to concentrate on the small, hard-to-predict depleted region; and (2) a semianalytical pressure “initial guess” input derived from a fast semianalytical model that replaces thin fracture geometries with a physically meaningful field, increasing input coverage while preserving fracture connectivity information. Trained on large-scale synthetic fracture realizations with an industrial-standard simulator, the surrogate achieves high-fidelity pressure and saturation reconstruction with sound accuracy quantified by R2 score and error maps. The model significantly outperforms traditional U-Net and reduces the time for depletion-mapping by two orders of magnitude. The proposed workflow provides a promising basis for rapid depletion-delineation screening and well-placement analysis, subject to further validation with field-calibrated fracture models and field-inspired blind test cases.

This publication has 58 references indexed in Scilit: