Visual representation of machine learning for low-contrast pay logging identification using 2D formation water resistivity variables

Owing to their similar resistivity to water layers, low-contrast pay (LCP) zones may be misinterpreted. While machine learning (ML) methods offer advantages in identifying LCP zones over conventional approaches, their interpretability remains a challenge. This study employs novel approaches to construct sensitive parameters and develop visual representation models, aiming to achieve accurate identification and enhanced model interpretability of Jurassic LCPs in the southwestern Ordos Basin. The key points can be summarized as follows: 1) True and apparent formation water resistivities (FWRs) were calculated from conventional parameters. Based on FWR spectra, the innovative 2D FWR variables were constructed, thereby enhancing distinctions among production-layer types. 2) Nine models were trained using three base learners—support vector machine (SVM), decision tree (DT), and artificial neural network (ANN)—with three sets of input features: conventional parameters, true and apparent FWRs, and 2D FWR variables. The ANN model performed best with an F₁-score of 95.55 % on the testing set when using 2D FWR variables. 3) Cross-plots of 2D FWR variables visually represent model performance. The ANN model's superior performance is attributed to its hidden layer neurons generating four demarcation lines that finely divide the cross-plot into nine zones for classification. In contrast, the SVM model divides the cross-plot into three zones using two demarcation lines, and the DT model's stepped demarcation lines lead to overfitting. The novelties can be summarized as follows: 1) During domain-specific feature transformation, the introduction of 2D FWR variables integrates both morphological information from logging curves and the differences between true and apparent FWRs. 2) Application of 2D FWR variables provides visual insights into ML model principles and performances, thereby facilitating a clearer understanding.