Deep Learning for Electromagnetic Inversion in LWD: A Collaborative Optimization Framework With Enhanced UNet++

This study proposes a collaborative optimization framework for solving directional electromagnetic logging while drilling (LWD) inverse problems that integrates a physics-constrained sample generation strategy, a hybrid feature compression transformation, and an enhanced UNet++ architecture. First, a Markov chain perturbation strategy is employed to generate formation samples with physical continuity, ensuring that the generated samples closely resemble real geological variations. Second, a hybrid compression transformation method is designed to address the heterogeneous distribution of electromagnetic responses, enhancing feature separability. Finally, the proposed network is improved by dynamically fusing multiscale and directional features through multiscale convolutional branches and preserving vertical resolution using an asymmetric downsampling strategy. Experiments demonstrate that resistivity-boundary inversion network (R-BInvNet) reduces the training loss by 57.5% compared to U-Net and further reduces it by 43.1% compared to UNet++. It also exhibits higher inversion accuracy in complex synthetic formation models and MNIST-derived formation models. The framework offers an efficient solution for real-time geosteering and demonstrates potential in high-dimensional electromagnetic inversion.