Transformer-Based Surrogate Model for Prediction of Hydraulic Fracture Geometry

Le résumé fourni par la source

Summary We developed a computationally efficient and accurate surrogate model to assess key descriptors of fracture geometry during reservoir stimulation associated with unconventional subsurface energy scenarios. The work is motivated by the observation that accurate quantification of fracture length, width, and height is key for reservoir stimulation design, performance assessment, and real-time operational decision-making. Fracture propagation is governed by strong reservoir heterogeneity, complex pumping schedules, and multiparameter coupling. Moreover, the scarcity of reliable field-scale data sets poses substantial challenges to purely data-driven approaches. To address these limitations, we establish an integrated conceptual/operational framework that couples large-scale numerical simulation with a multimodal transformer network. An automated workflow resting on the broadly used hydraulic fracturing simulator GOHFER (grid-oriented hydraulic fracture extension replicator) is constructed to generate a physically consistent and diverse data set spanning a broad range of geological attributes and operational conditions. Static reservoir parameters and injection time-series data are jointly embedded within a transformer architecture to directly evaluate fracture geometry metrics. The approach enables effective learning of long-duration injection sequences while providing insights upon quantifying the relative importance of critical injection stages during fracture evolution. The proposed framework demonstrates high accuracy across diverse simulated operating scenarios. Validation using stage-level, field-constrained fracture lengths from five Mahu wells yields an overall mean absolute percentage error (MAPE) of approximately 8.1%. The trained model also substantially reduces the computational cost of repeated fracture-geometry evaluation compared with conventional numerical simulation. The approach provides a scalable surrogate modeling setting to facilitate real-time decision support in unconventional reservoir development.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.