Sweet spot identification of coalbed methane reservoirs by integrating autoencoders and clustering

Deep coalbed methane resources hold significant potential and represent a critical area for achieving reserve growth and production increases of unconventional natural gas in China. To address the challenges in traditional sweet spot evaluation, such as difficulties in multi-parameter comprehensive assessment and heavy reliance on manual expertise, this paper establishes an intelligent sweet spot identification model that integrates a deep autoencoder with unsupervised clustering. Incorporating the coal structure index and ash content—critical factors directly influencing reservoir permeability and gas-bearing potential—along with key well-log parameters such as interval transit time and compensated neutron logs, nonlinear feature extraction and dimensionality reduction are performed using a deep autoencoder to capture complex interrelationships among logging curves. Furthermore, the K-Means clustering algorithm is integrated to automatically classify features within the encoded latent space, enabling intelligent delineation of sweet spot grades. Practical application demonstrates that the study area was automatically classified into Class I, Class II and Class III sweet spot zones. The optimal sweet spots exhibit reservoir characteristics of high interval transit time, high neutron porosity, high structural index, and low ash content. The prediction results from the intelligent evaluation method show an 84.35% consistency rate with field log interpretation. In the actual application of adjacent wells, the model selection method based on the model outperformed the traditional logging interpretation method in indicating the gas production by 29.2%. This research provides a data-driven intelligent solution for sweet spot optimization in coalbed methane fracturing and holds significant guiding value and practical importance for enhancing reserves and production from deep coalbed methane resources in the Ordos Basin.