Knowledge-guided deep learning inversion of electromagnetic look-ahead logging-while-drilling data

Inversion of electromagnetic look-ahead logging-while-drilling (LWD) data enables real-time estimation of formation properties ahead of the drill bit, providing critical support for proactive geosteering. However, conventional model-parameterized inversion methods rely on predefined layer structures, which limits their adaptability to complex geological settings and leads to non-uniqueness when constrained solely by electromagnetic responses. To overcome these limitations, this study proposes a Knowledge-Guided Residual Network (KG-ResNet) that integrates prior resistivity information to effectively constrain the inversion process and enable efficient quasi-pixel-based resistivity imaging. Performance evaluations on test datasets demonstrate that KG-ResNet accurately recovers both resistivity distributions and layer interfaces by leveraging prior information. The proposed quasi-pixel representation allows the network to flexibly model formations with arbitrary numbers of layers. Synthetic experiments further confirm the robustness of the method under different well inclinations and its capability to resolve both thin and thick layers. Overall, the proposed framework provides a computationally efficient, stable, and geologically consistent inversion solution for electromagnetic look-ahead LWD, enhancing real-time formation evaluation and decision-making in heterogeneous reservoirs.