22 April 2022 in PLOS ONE
The Ordinary Kriging method is a common spatial interpolation algorithm in geostatistics. Because the semivariogram required for kriging interpolation greatly influences this process, optimal fitting of the semivariogram is of major significance for improving the theoretical accuracy of spatial interpolation. A deep neural network is a machine learning algorithm that can, in principle, be applied to any function, including a semivariogram. Accordingly, a novel spatial interpolation method based on a deep neural network and Ordinary Kriging was proposed in this research, and elevation data were used as a case study. Compared with the semivariogram fitted by the traditional exponential model, spherical model, and Gaussian model, the kriging variance in the proposed method is smaller, which means that the interpolation results are closer to the theoretical results of Ordinary Kriging interpolation. At the same time, this research can simplify processes for a variety of semivariogram analyses.
21 September 2023 in SPE Journal
Summary: Low-resistance reservoirs have been of great interest as a key topic in the study of low-permeability reservoirs. Accurate identification of fluid properties is a challenging problem in the effective exploration of such reservoirs. Logging data, as a common identification tool, can provide rich and highly accurate geological information. Here, we combine extreme gradient boosting (XGBoost) and MAHAKIL’s oversampling method for fluid property identification using logging data from low-permeability sandstone reservoirs. The MAHAKIL method is used to solve the class imbalance problem due to unbalanced training samples, and the data are fed into XGBoost to build a geological model with complex abstract feature weights related to fluid properties through multiple decision trees. We first demonstrate that MAHAKIL can improve the XGBoost model accuracy using four evaluation criteria, namely, the F1-score, recall, precision, and accuracy, among which the F1-score is most applicable to the classification problem of reservoir fluid prediction. Then, the receiver operating characteristic (ROC) and area under the curve (AUC) values are used to demonstrate that MAHAKIL improves the XGBoost model performance. Finally, the results and performance of this method and the traditional XGBoost model with unbalanced real data are compared. The results show that the proposed method is superior for fluid property identification in low-permeability sandstone reservoirs with imbalanced learning samples.
17 April 2026 in IEEE Access
To address the challenges of low accuracy, poor noise resistance, and inefficiency in seismic first arrival picking, this paper proposes Physical-Constraint SwinUNETR (PCS-UNETR). First, a SoftLabel dataset is constructed by converting seismic traces into single-band 2D images with Gaussian probability density labels to enrich gradient information. Second, a physical prior-constrained SwinUNETR model is designed. By incorporating an amplitude change rate penalty term into the loss function, the model is forced to identify arrivals on the rising edge of the first peak. Finally, morphological operations define a region of interest, where the pixel with the maximum predicted probability is selected as the final pick. Experiments on Sinopec Zhongyuan Oilfield's Puguang seismic data (10.24 million traces) demonstrate that PCS-UNETR effectively reduces outliers, significantly improves accuracy and noise resistance, and achieves full automation without manual intervention. The innovation lies in fusing physical priors with deep learning via a "global probability prediction + local morphological constraint" mechanism, offering a new approach for intelligent seismic processing with high application value.