Automatic {AI} {Well} {Log} {Reconstruction} {Leveraging} {Neighboring} {Analog} {Wells} {To} {Reduce} {Covariate} {Shift}

@inproceedings{ouzineb_automatic_RM2024,
 abstract = {Wellbore log data for performing petrophysical interpretation are not always acquired in each drilled well or might be inconsistent due to a challenging measurement environment. When predicting well logs on wells of interest through machine learning regression, Petrophysicists must manually select analog training wells to avoid covariate shift, which is tedious and does not scale with the potential thousands of wells in a wellbore study. To address this issue, this paper presents a novel method leveraging automatically found optimal analog wells to predict and reconstruct well logs at scale using AI models tailored to the geological environment of each well of interest. In this method, we introduce a novel iterative fine-tuning approach for log prediction which automatically finds the optimal analog training wells for each well of interest. These analog wells are used to fine-tune a pre-trained regression model which is then applied on the corresponding well of interest to predict the desired well logs. To perform log reconstruction, we also introduce a novel segmentation-based algorithm that automatically flags the well segments where the log prediction differs significantly from the existing log. In such flagged segments, a log anomaly detection algorithm keeps, between the existing and the predicted log, the one that is the most consistent with its local neighboring well log data intervals. On a test study consisting of public wellbore data from Wyoming, we achieve with our method a reconstruction error on the bulk density log that is 24.53\% smaller than when using a model fine-tuned on all wells present in the study. When comparing our method with the approach of using a frozen pre-trained model, our approach yields a reconstruction error that is 39.4\% smaller. We also experiment different types of regression models, including XGBoost and transformer models. With the best performing model, the self-attention imputation for time-series transformer model (SAITS), we achieve a reconstruction error on the bulk density log that is 20\% smaller compared to XGBoost on a test set of 170 off-shore wells from the Netherlands. These experiments show a significant increase in well log prediction quality when using our method compared with other classical log prediction methods. Model benchmarking allows us to achieve an even greater prediction performance by selecting the optimal baseline model architecture. Our method is also designed and implemented to perform log prediction and reconstruction automatically and at scale; thus, facilitating the petrophysical interpretation task.},
 author = {Ouzineb, Sohaib and Wlodarczyk, Sylvain},
 booktitle = {Proc. 2024 RING Meeting},
 language = {en},
 pages = {25},
 publisher = {ASGA},
 title = {Automatic {AI} {Well} {Log} {Reconstruction} {Leveraging} {Neighboring} {Analog} {Wells} {To} {Reduce} {Covariate} {Shift}},
 year = {2024}
}