Geological information-driven deep learning for lithology identification from well logs

Lithology identification is crucial for characterizing complex unconventional reservoirs, where thin interlayers significantly influence hydrocarbon accumulation. Although deep learning-based methods utilizing well logs have become prevalent, most approaches treat well logs as generic 1D time series, frequently neglecting the multi-scale geological information inherent in the data. This oversight limits their accuracy and generalizability, especially in geologically complex environments. To overcome this limitation, we propose a novel geology-driven deep learning framework. Our key contribution is the transformation of 1D well logs into 2D multi-scale feature maps through multiresolution wavelet decomposition, a process designed to explicitly represent geological features resembling sedimentary cycles. These feature maps are subsequently processed by a novel Geology-Guided Hybrid Network with channel-spatial attention, which integrates a 2D CNN to capture geological patterns and a Bidirectional long short-term memory to model sequential dependencies. Evaluated on field data from complex reservoirs, our method achieves an outstanding F1-score of up to 0.966, outperforming four established deep learning benchmarks. Importantly, the approach demonstrates improved accuracy in identifying thin layers and enhanced generalization across wells with differing lithological distributions, attaining an F1-score of 0.885 on a challenging test well exhibiting significant data drift. This study validates the robustness of our geology-informed approach and offers an effective framework for high-precision lithology identification.