Answers are grounded in this paper and cite their sources.
Machine learning (ML) has achieved great success in production prediction for unconventional shale gas reservoirs. However, these methods mostly rely on the discrete data collected from the wells, such as drilling, completion, and production data. In this study, a multimodal ML approach is proposed to incorporate not only the aforementioned tabular data but also the geological property distribution maps surrounding the production wells. More specifically, a visual parameterization method was applied to preprocess the unstructured data from a 3D geological model to account for the geology properties near the horizontal wells. A comprehensive architecture for a multimodal model was then developed, assimilating a convolutional neural network (CNN) module, an artificial neural network (ANN) module, and a fusion module. The CNN module was established to process and extract high-level information from the visual dataset, while the ANN module was devised to learn from traditional tabular datasets. A fusion module combined and interacted with the data from both modalities. Results have shown that the proposed multimodal model achieved the highest testing R 2 of 0.828 by integrating the formation maps with tabular datasets, compared to 0.736 from ANN. This is owing to the fact that two wells with similar porosity values measured at well sites could penetrate formations with different qualities along their thousand meters of lateral length. Visual feature analysis indicates that while integrating more property distribution maps generally increases model accuracy, considerable improvement (from 0.736 to 0.816) is achieved by solely incorporating porosity maps. • Incorporating geological maps by multimodal architecture raises R 2 from 0.74 to 0.83. • Significant improvement can be achieved by solely incorporating porosity maps. • An optimal width exists for geo-patching to maximize model performance. • Alignment test validates synchronization between tabular and visual datasets.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: