Joint prediction and inversion of porosity and oil saturation based on MAE-MoE

To improve the efficiency of joint porosity and oil-saturation interpretation, enhance the use of parameter coupling, and improve model generalization under complex reservoir conditions, this study proposes an MAE-MoE joint prediction-inversion model. The method is developed using logging data from the L-well series in Block N of the Daqing Oilfield. Multi-source logging curves, including acoustic transit time, compensated neutron, compensated density, gamma ray, deep and shallow resistivity, and spontaneous potential, are used as inputs. By introducing shale-correction terms and deep-shallow resistivity separation features, a joint inversion framework for continuous depth-series modeling is established. In the proposed framework, a Masked Autoencoder (MAE) is first employed for self-supervised pretraining, where random masking and reconstruction are used to learn intrinsic cross-channel relationships and depth continuity from multi-channel log responses. Then, a Mixture of Experts (MoE) module dynamically routes shared representations to expert-specific sub-networks, enabling collaborative prediction of porosity and oil saturation. The results show that the proposed method outperforms comparative models in prediction accuracy, cross-well adaptability, and robustness to disturbance. In particular, using engineering-tolerance-based accuracies, the proposed model achieves 91.2% accuracy for porosity and 91.0% for oil saturation, where the tolerance thresholds are defined as an absolute porosity error within 2 percentage points and an absolute oil-saturation error within 0.10. The corresponding RMSE values are provided for standard regression-based evaluation. The main contribution of this study does not lie in treating MAE or MoE as new algorithms individually, but in adapting and integrating masked sequence reconstruction, expert-specialized routing, uncertainty-weighted dual-task learning, and petrophysical feature construction into a unified framework for continuous well-log-based joint inversion. This work offers a practical reference for intelligent well-log interpretation and fine reservoir evaluation in sandstone–mudstone reservoirs with similar geological and logging-response conditions.