Acid fracturing is essential in enhancing recovery efficiency, especially within carbonate reservoirs. Although extensive studies have been conducted on hydraulic fracturing, understanding the intricate dynamics between acid–rock reactions and fracture propagation in heterogeneous layered reservoirs remains limited. This study employs a comprehensive coupled hydro-mechanical-chemical flow framework to investigate acid fracturing processes in layered geological formations. The model incorporates a two-stage homogenization approach to account for rock heterogeneity, a dual-scale continuum framework for fluid flow and acid transport, and a phase field method for examining fracture propagation. We thoroughly examine how treatment parameters, particularly acid concentration and injection rate, affect fracture propagation modes. The analysis identifies three distinct propagation patterns: crossing, diversion, and arresting. These are influenced by the interplay between pressure buildup and wormhole formation. Initially, higher acid concentration aids in fracture crossing by lowering the peak pressure required for initiation, but excessive concentration results in arresting because it causes extensive wormhole development, which reduces fluid pressure. Similarly, the injection rate plays a crucial role in fracture movement across layer interfaces, with moderate rates optimizing propagation by balancing pressure and wormhole growth. This comprehensive modeling framework serves as a valuable prediction and control tool for acid fracture behavior in complex layered formations.

A productivity simulation for hydraulically fractured wells with complex fracture geometry involves a heavy computational burden and is therefore not suitable for engineering-scale fracture-optimization designs and production-analysis applications. This paper develops a productivity-prediction surrogate model based on a deep convolution–bidirectional gated recurrent unit temporal network (DC-BiGRU) framework where a deep convolutional neural network is used to extract features from fracture images, while a BiGRU model was designed to fully capture valuable information from the production sequence. Some additional inputs, e.g., cluster spacing and stage spacing, that account for different fracture-placement designs in horizontal wells were also considered. A large number of shale-gas production data samples at different times were generated using a fractured-horizontal-well productivity simulator under diverse hydraulic-fracture geometries and bottom-hole flowing pressures. The surrogate model had relative errors below 10% with an average error of about 6%. Compared to high-fidelity capacity prediction simulators, the computational efficiency of the deep learning surrogate models was improved by two to three orders of magnitude. The runtime of the high-fidelity numerical simulator was about 20 min, while the surrogate model, which was run on an NVIDIA Tesla P100 GPU (NVIDIA, Santa Clara, CA, USA), took less than 1 s, which is almost negligible. The proposed surrogate model resolved the low efficiency of the productivity simulation for complex-fracture hydraulic fracturing wells in unconventional reservoirs, enabling rapid dynamic forecasting of fractured-well productivity.