Artificial Intelligence in Geosciences最新文献

Logs are valuable information for oil and gas fields as they help to determine the lithology of the formations surrounding the borehole and the location and reserves of subsurface oil and gas reservoirs. However, important logs are often missing in horizontal or old wells, which poses a challenge in field applications. To address this issue, conventional methods involve supplementing the missing logs by either combining geological experience and referring data from nearby boreholes or reconstructing them directly using the remaining logs in the same borehole. Nevertheless, there is currently no quantitative evaluation for the quality and rationality of the constructed log. In this paper, we utilize data from the 2020 machine learning competition of the Society of Petrophysicists and Logging Analysts (SPWLA), which aims to predict the missing compressional wave slowness (DTC) and shear wave slowness (DTS) logs using other logs in the same borehole. We employ the natural gradient boosting (NGBoost) algorithm to construct an Ensemble Learning model that can predicate the results as well as their uncertainty. Furthermore, we combine the SHAP (SHapley Additive exPlanations) method to investigate the interpretability of the machine learning model. We compare the performance of the NGBosst model with four other commonly used Ensemble Learning methods, including Random Forest, GBDT, XGBoost, LightGBM. The results show that the NGBoost model performs well in the testing set and can provide a probability distribution for the prediction results. This distribution allows petrophysicists to quantitatively analyze the confidence interval of the constructed log. In addition, the variance of the probability distribution of the predicted log can be used to justify the quality of the constructed log. Using the SHAP explainable machine learning model, we calculate the importance of each input log to the predicted results as well as the coupling relationship among input logs. Our findings reveal that the NGBoost model tends to provide greater slowness prediction results when the neutron porosity (CNC) and gamma ray (GR) are large, which is consistent with the cognition of petrophysical models. Furthermore, the machine learning model can capture the influence of the changing borehole caliper on slowness, where the influence of borehole caliper on slowness is complex and not easy to establish a direct relationship. These findings are in line with the physical principle of borehole acoustics. Finally, by using the explainable machine learning model, we observe that although we did not correct the effect of borehole caliper on the neutron porosity log through preprocessing, the machine learning model assigned a greater importance to the influence of the caliper, achieving the same effect as caliper correction.

Frosts are one of the atmospheric phenomena with one of the larger negative effects on the agricultural sector in the southern region of Brazil, therefore, an earlier forecast can minimize their impacts. In the present work, artificial neural networks (ANNs) techniques were applied in order to improve the predicting capabilities of frost events in southern Brazil. In the study, two multilayer perceptron (MLP) ANNs were built, one with ADAM optimizer and the other with SGD. The input parameters MLP-ANNs were numerical predictions of the Eta model. The ANNs were trained using four years (2012–2015), while validation and testing were performed using 2016 and 2017, respectively. An episode of frost that occurred on May 21st, 2018, related to an intense cold air mass, was also utilized to evaluate the performance of the ANNs. The best configurations (topologies and hyperparameters) of the ANNs were identified through experiments, using the highest accuracy obtained during the validation period as a metric. The results of the ANNs with ADAM and SGD optimizers were compared with the predictions of the Eta model. For the case study, an additional comparison against the operational frost index (IG) from the National Institute for Space Research (INPE) was also included. The performance of both ANNs (properly configured) with ADAM and SGD optimizers are comparable one to the other. And both are significantly better compared to the Eta model. The ANNs were able to drastically reduce the underestimation trends of frost events caused by the warm bias of the Eta model. The ANNs also indicated more satisfactory performances when compared to the INPE IG. In general, the ANNs were able to identify deficiencies in Eta predictions, and consequently improve their results. In this sense, the use of ANNs to predict frost events can be a very useful tool in an operational environment.

The Hindu Kush-Pamir region (HKPR) is characterized by complex ongoing deformation, unique slab geometry, and intermediate seismic activity. The availability of extensive seismological data in recent decades has prompted the use of deep learning algorithms to extract valuable insights. In this study, we present a fully automated approach for augmenting earthquake catalogue within the HKPR. Our method leverages an attention mechanism-based deep learning architecture to simultaneously detect events, perform phase picking, and estimate magnitudes. We applied this model to a ten-month dataset (January 2013–October 2013) from 83 stations in the region. Utilizing a robust criterion to evaluate the model's probabilities, we associated phases at different stations and pinpointed earthquake locations in the region. Our results demonstrate a significant enhancement, revealing nearly four and a half times more earthquakes than previously documented in the International Seismological Center (ISC) catalogue. A notable portion of these newly detected events falls within the category of very low-magnitude earthquakes (<3), which were absent in the ISC catalogue. Notably, our spatiotemporal analysis reveals a concentration of crustal seismicity along poorly mapped neotectonic north and northeast-oriented faults in the western Pamir, as well as the Vakhsh Thrust System and the Darvaz Karakul Fault. These findings underscore potential sources of future seismic hazards. Furthermore, our expanded earthquake catalogue facilitates a deeper understanding of the interplay between crustal and intermediate seismic activity in the HKPR, shedding light on the deformation and active faulting resulting from Eurasian-Indian plate interactions.

Evolution in geoscientific data provides the mineral industry with new opportunities. A direction of geochemical data generation evolution is towards big data to meet the demands of data-driven usage scenarios that rely on data velocity. This direction is more significant where traditional geochemical data are not ideal, which is the case for evaluating unconventional resources, such as tailing storage facilities (TSFs), because they are not static due to sedimentation, compaction and changes associated with hydrospheric and lithospheric processes (e.g., erosion, saltation and mobility of chemical constituents). In this paper, we generate big secondary geochemical data derived from Sentinel-2 satellite-remote sensing data to showcase the benefits of big geochemical data using TSFs from the Witwatersrand Basin (South Africa). Using spatially fused remote sensing and legacy geochemical data on the Dump 20 TSF, we trained a machine learning model to predict in-situ gold grades. Subsequently, we deployed the model to the Lindum TSF, which is 3 km away, over a period of a few years (2015-2019). We were able to visualize and analyze the temporal variation in the spatial distributions of the gold grade of the Lindum TSF. Additionally, we were able to infer extraction sequencing (to the resolution of the data), acid mine drainage formation and seasonal migration. These findings suggest that dynamic mineral resource models and live geochemical monitoring (e.g., of elemental mobility and structural changes) are possible without additional physical sampling.

The stratigraphic correlation of well logs plays an essential role in characterizing subsurface reservoirs. However, it suffers from a small amount of training data and expensive computing time. In this work, we propose the Attention Based Dense Network (ASDNet) for the stratigraphic correlation of well logs. To implement the suggested model, we first employ the attention mechanism to the input well logs, which can effectively generate the weighted well logs to serve for further feature extraction. Subsequently, the DenseNet is utilized to achieve good feature reuse and avoid gradient vanishing. After model training, we employ the ASDNet to the testing data set and evaluate its performance based on the well log data set from Northwest China. Finally, the numerical results demonstrate that the suggested ASDNet provides higher prediction accuracy for automated stratigraphic correlation of well logs than state-of-the-art contrastive UNet and SegNet.

In this study, a deep learning algorithm was applied to two-dimensional magnetotelluric (MT) data inversion. Compared with the traditional linear iterative inversion methods, the MT inversion method based on convolutional neural networks (CNN) does not rely on the selection of the initial model parameters and does not fall into the local optima. Although the CNN inversion models can provide a clear electrical interface division, their inversion results may remain prone to abrupt electrical interfaces as opposed to the actual underground electrical situation. To solve this issue, a neural network with a residual network architecture (ResNet-50) was constructed in this study. With the apparent resistivity and phase pseudo-section data as the inputs and with the resistivity parameters of the geoelectric model as the training labels, the modified ResNet-50 model was trained end-to-end for producing samples according to the corresponding production strategy of the study area. Through experiments, the training of the ResNet-50 with the dice loss function effectively solved the issue of over-segmentation of the electrical interface by the cross-entropy function, avoided its abrupt inversion, and overcame the computational inefficiency of the traditional iterative methods. The proposed algorithm was validated against MT data measured from a geothermal field prospect in Huanggang, Hubei Province, which showed that the deep learning method has opened up broad prospects in the field of MT data inversion.

Land use and land cover (LULC) changes refer to alterations in land use or physical characteristics. These changes can be caused by human activities, such as urbanization, agriculture, and resource extraction, as well as natural phenomena, for example, erosion and climate change. LULC changes significantly impact ecosystem services, biodiversity, and human welfare. In this study, LULC changes in Davao City, Philippines, were simulated, predicted, and projected using a multilayer perception artificial neural network (MLP-ANN) model. The MLP-ANN model was employed to analyze the impact of elevation and proximity to road networks (i.e., exploratory maps) on changes in LULC from 2017 to 2021. The predicted 2021 LULC map shows a high correlation to the actual LULC map of 2021, with a kappa index of 0.91 and a 96.68% accuracy. The MLP-ANN model was applied to project LULC changes in the future (i.e., 2030 and 2050). The results suggest that in 2030, the built-up area and trees are increasing by 4.50% and 2.31%, respectively. Unfortunately, water will decrease by up to 0.34%, and crops is about to decrease by approximately 3.25%. In the year 2050, the built-up area will continue to increase to 6.89%, while water and crops will decrease by 0.53% and 3.32%, respectively. Overall, the results show that anthropogenic activities influence the land's alterations. Moreover, the study illustrates how machine learning models can generate a reliable future scenario of land usage changes.

Resource-rich shale gas plays a pivotal role in new energy types. The key to scientifically and efficiently developing shale gas fields is to clarify the main factors that affect the production of shale gas wells. In this paper, according to the shale gas reservoir characteristic of the Fuling marine Longmaxi Formation, a single-well geological model was established using the reservoir numerical simulation software CMG. Then, 10,000 different reservoir models were randomly generated for different formation physical parameters, completion parameters, and fracturing parameters using the Monte Carlo method, and these 10,000 models were simulated numerically. The machine learning model uses a dataset of 10,000 different geological, completion, and fracturing parameters as input and 10,000 production curves as output. Multiple machine learning regression methods were used to train and test the dataset, and the optimal method (GBDT algorithm) was selected, and the accuracy R² of the test set of the GBDT prediction model is 0.96. A fracturing parameter optimization workflow was constructed by combining a production prediction model with a particle swarm optimizer (PSO). The process can quickly optimize the fracturing parameters and predict the production for each time by targeting the cumulative gas production under different geological conditions. The optimized parameters are Fracture Spacing, Fracture Width, Intrinsic Permeability, Fracture Half-length, Langmuir Pressure, and Langmuir Volume. The initial predicted cumulative gas production was 4.59 × 10⁸ m³, which was optimized to 4.90 × 10⁸ m³. The proposed PSO-GBDT proxy model can instantly predict the production of shale gas wells with considerable accuracy, reliability, and efficiency, which is a vital tool for optimizing fracture design. This investigation provides a solid foundation for predicting the production of unconventional gas reservoirs and for parameter optimization.

Multiple-Point Simulation (MPS) is a geostatistical simulation technique commonly used to model complex geological patterns and subsurface heterogeneity. There have been a great variety of implementation methods developed within MPS, of which Patch-Based Simulation is a more recently developed class. While we have witnessed great progress in Patch-Based algorithms lately, they are still faced with two challenges: conditioning to point data and the occurrence of verbatim copy. Both of them are partly due to finite size of Training image, from which a limited size of pattern database is constructed. To address these questions, we propose a novel approach that we call Generative-Patched-Simulation (GPSim), which is based on Generative Adversarial Networks (GAN). With this method, we are able to generate sufficient (in theory an infinite) number of new patches based on the current pattern database. As demonstrated by the results on a simple 2D binary image, this approach shows its potential to address the two issues and thus improve Patch-based Simulation methods.

Deep-learning (DL) algorithms are increasingly used for routine seismic data processing tasks, including seismic event detection and phase arrival picking. Despite many examples of the remarkable performance of existing (i.e., pre-trained) deep-learning detector/picker models, there are still some cases where the direct applications of such models do not generalize well. In such cases, substantial effort is required to improve the performance by either developing a new model or fine-tuning an existing one. To address this challenge, we present Blockly Earthquake Transformer(BET), a deep-learning platform for efficient customization of deep-learning phase pickers. BET implements Earthquake Transformer as its baseline model, and offers transfer learning and fine-tuning extensions. BET provides an interactive dashboard to customize a model based on a particular dataset. Once the parameters are specified, BET executes the corresponding phase-picking task without direct user interaction with the base code. Within the transfer-learning module, BET extends the application of a deep-learning P and S phase picker to more specific phases (e.g., Pn, Pg, Sn and Sg phases). In the fine-tuning module, the model performance is enhanced by customizing the model architecture. This no-code platform is designed to quickly deploy reusable workflows, build customized models, visualize training processes, and produce publishable figures in a lightweight, interactive, and open-source Python toolbox.