Abstract: Identifying architectural elements from well-log curves is an essential task in fine-scale reservoir characterization. In recent years, machine learning methods have been widely applied to architectural element identification; however, the identification accuracy is generally low for lower-order architectural element with low-proportion. In this study, we propose a hierarchical architectural element identification method based on a bidirectional stacked long short-term memory network and random forest (RF), and apply it to the third- and fourth-order architectural element in braided-river reservoirs of the Chengtan 3 block, Ordos basin. First, the temporal network model is employed to extract rhythmic features from well-log curves and identify fourth-order architectural element such as floodplain, channel, and braid bar. On this basis, an RF method with gamma-threshold constraints is used to further identify braid bar intercalated mud layers. In the application of identifying third- and fourth-order architectural elements in the 8th member of the Lower  Shixihezi Formation,  the comparison with existing machine learning approaches shows that the optimized hierarchical identification method achieves an accuracy of at least 94% on the validation set for reservoir architectural element characterization, outperforming direct machine learning approaches and providing higher identification precision.

Key words: well log curves, machine learning, braided river, architectural elements, reservoir, Ordos basin