David Duarte
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
David Duarte has a total of 8 co-authors. Here are the top co-authors by number of shared publications:
- Matthew Jude Pranter
University of Oklahoma 1 , School of Geosciences, Norman, Oklahoma, USA
3 shared publications - Rafael Pires de Lima
Geological Survey of Brazil, São Paulo 01304-010, Brazil
2 shared publications Benmadi Milad
1 shared publication- Deepak Devegowda
University of Oklahoma, Department of Petroleum and Geological Engineering 1 , Norman, Oklahoma.;;;;;
1 shared publication Javier Tellez
University of Oklahoma 3 , Department of Geology and Geophysics, Norman, Oklahoma, USA and Colorado Mesa University, Department of Physical and Environmental Sciences, Grand Junction, Colorado, USA
1 shared publicationR. Douglas Elmore
1 shared publication
Latest Publications
- 2 November 2025
- Published by Society of Exploration Geophysicists in Interpretation
Semi-supervised Workflow to Generate Petrofacies Logs from Thin Sections and XRF Data
Core data provide valuable in situ information on the chemical and physical characteristics of subsurface formations. Thin sections, for instance, allow us to define rock types with similar mineral composition, lithologies, and pore types (i.e., petrofacies). Petrofacies logs illustrate the stratigraphic variability within a reservoir and are often used to constrain 3-D facies and petrophysical-property models. However, core data are often scarce, prompting us to explore alternative sources. Well logs, with their different vertical resolution, are commonly used to characterize mineralogy, lithology, and porosity. The challenge lies in reconciling the differences between thin sections and well logs. To address this, we developed novel Machine Learning (ML) methodologies. Our goal was to bridge the resolution gap between these two data types and identify subtle variations in rock properties. Specifically, we focused on collocated cores that provided thin-section-based petrofacies and X-ray fluorescence (XRF) data. Our approach involved two semi-supervised methods: self-training and labeled clustering. By combining XRF data with dimensionality reduction techniques, we achieved a reliable classification of thin-section-based petrofacies. Remarkably, both approaches achieved accuracies exceeding 90% on Sycamore Formation data. Among the dimensionality reduction techniques tested. Uniform Manifold Approximation and Projection (UMAP) yielded the best results. The outcome? Petrofacies logs that bridge the resolution gap between core-based thin sections and well-log data. Furthermore, integrating semi-supervised methods into routine core analysis offers substantial cost and time savings. These methods enhance stratigraphic correlation, aid in identifying target zones, design horizontal wells, and constrain subsurface models.
- 2 November 2025
- Published by Society of Exploration Geophysicists in Interpretation
Semi-supervised Workflow to Generate Petrofacies Logs from Thin Sections and XRF Data
Core data provide valuable in situ information on the chemical and physical characteristics of subsurface formations. Thin sections, for instance, allow us to define rock types with similar mineral composition, lithologies, and pore types (i.e., petrofacies). Petrofacies logs illustrate the stratigraphic variability within a reservoir and are often used to constrain 3-D facies and petrophysical-property models. However, core data are often scarce, prompting us to explore alternative sources. Well logs, with their different vertical resolution, are commonly used to characterize mineralogy, lithology, and porosity. The challenge lies in reconciling the differences between thin sections and well logs. To address this, we developed novel Machine Learning (ML) methodologies. Our goal was to bridge the resolution gap between these two data types and identify subtle variations in rock properties. Specifically, we focused on collocated cores that provided thin-section-based petrofacies and X-ray fluorescence (XRF) data. Our approach involved two semi-supervised methods: self-training and labeled clustering. By combining XRF data with dimensionality reduction techniques, we achieved a reliable classification of thin-section-based petrofacies. Remarkably, both approaches achieved accuracies exceeding 90% on Sycamore Formation data. Among the dimensionality reduction techniques tested. Uniform Manifold Approximation and Projection (UMAP) yielded the best results. The outcome? Petrofacies logs that bridge the resolution gap between core-based thin sections and well-log data. Furthermore, integrating semi-supervised methods into routine core analysis offers substantial cost and time savings. These methods enhance stratigraphic correlation, aid in identifying target zones, design horizontal wells, and constrain subsurface models.
- 6 January 2023
- Published by Society of Exploration Geophysicists in Interpretation
- Vol. 11 (2) , T289-T302
Spatial variability of petrofacies using supervised machine learning and geostatistical modeling: Sycamore Formation, Sho-Vel-Tum Field, Oklahoma, USA
The Sycamore Formation at the Sho-Vel-Tum Field primarily consists of clay-rich mudstones (Mdst) and quartz-rich siltstones. The clay-rich Mdst are mainly composed of clays, quartz grains, some allochems, and detrital organic matter. The siltstones are structureless and are divided into two petrofacies: high porosity and permeability massive calcareous siltstones and low porosity and permeability massive calcite-cemented siltstones. Core and well-log data provide mineralogical, lithologic, and porosity information that is useful to define petrophysical facies (petrofacies) and to create facies logs within the Sycamore Formation. We used the data to establish the Sycamore Formation stratigraphic architecture and to map its spatial variability and reservoir properties. To classify the Sycamore Formation petrofacies in noncored wells, we developed a machine learning-based workflow that compares more than 1800 classification models and selects the best combination of well logs, algorithms, and hyperparameters to predict defined petrofacies. The process includes combinations of well logs that were optimized in four classification algorithms: artificial neural network, K-nearest neighbor, support vector machine, and random forest. To adjust each classifier, we used a grid search and a fivefold cross validation to find the best combination of three hyperparameters to improve results of each algorithm. This workflow allows for the efficient extraction of information from cores at a low cost. After we generated the petrofacies logs in noncored wells, we combined them with multiple constraints to create a 3D petrofacies model for the Sycamore Formation at the Sho-Vel-Tum Field and analyze the stratigraphic and diagenetic controls on petrofacies and its impact in reservoir quality.
- 6 January 2023
- Published by Society of Exploration Geophysicists in Interpretation
- Vol. 11 (2) , T289-T302
Spatial variability of petrofacies using supervised machine learning and geostatistical modeling: Sycamore Formation, Sho-Vel-Tum Field, Oklahoma, USA
The Sycamore Formation at the Sho-Vel-Tum Field primarily consists of clay-rich mudstones (Mdst) and quartz-rich siltstones. The clay-rich Mdst are mainly composed of clays, quartz grains, some allochems, and detrital organic matter. The siltstones are structureless and are divided into two petrofacies: high porosity and permeability massive calcareous siltstones and low porosity and permeability massive calcite-cemented siltstones. Core and well-log data provide mineralogical, lithologic, and porosity information that is useful to define petrophysical facies (petrofacies) and to create facies logs within the Sycamore Formation. We used the data to establish the Sycamore Formation stratigraphic architecture and to map its spatial variability and reservoir properties. To classify the Sycamore Formation petrofacies in noncored wells, we developed a machine learning-based workflow that compares more than 1800 classification models and selects the best combination of well logs, algorithms, and hyperparameters to predict defined petrofacies. The process includes combinations of well logs that were optimized in four classification algorithms: artificial neural network, K-nearest neighbor, support vector machine, and random forest. To adjust each classifier, we used a grid search and a fivefold cross validation to find the best combination of three hyperparameters to improve results of each algorithm. This workflow allows for the efficient extraction of information from cores at a low cost. After we generated the petrofacies logs in noncored wells, we combined them with multiple constraints to create a 3D petrofacies model for the Sycamore Formation at the Sho-Vel-Tum Field and analyze the stratigraphic and diagenetic controls on petrofacies and its impact in reservoir quality.
- Open Access
- 11 August 2021
- Published by MDPI AG in Geosciences
- Vol. 11 (8) , 336
Pretraining Convolutional Neural Networks for Mudstone Petrographic Thin-Section Image Classification
Convolutional neural networks (CNN) are currently the most widely used tool for the classification of images, especially if such images have large within- and small between- group variance. Thus, one of the main factors driving the development of CNN models is the creation of large, labelled computer vision datasets, some containing millions of images. Thanks to transfer learning, a technique that modifies a model trained on a primary task to execute a secondary task, the adaptation of CNN models trained on such large datasets has rapidly gained popularity in many fields of science, geosciences included. However, the trade-off between two main components of the transfer learning methodology for geoscience images is still unclear: the difference between the datasets used in the primary and secondary tasks; and the amount of available data for the primary task itself. We evaluate the performance of CNN models pretrained with different types of image datasets—specifically, dermatology, histology, and raw food—that are fine-tuned to the task of petrographic thin-section image classification. Results show that CNN models pretrained on ImageNet achieve higher accuracy due to the larger number of samples, as well as a larger variability in the samples in ImageNet compared to the other datasets evaluated.
- 11 August 2021
- Published by MDPI AG in Geosciences
- Vol. 11 (8) , 336
Open AccessPretraining Convolutional Neural Networks for Mudstone Petrographic Thin-Section Image Classification
Convolutional neural networks (CNN) are currently the most widely used tool for the classification of images, especially if such images have large within- and small between- group variance. Thus, one of the main factors driving the development of CNN models is the creation of large, labelled computer vision datasets, some containing millions of images. Thanks to transfer learning, a technique that modifies a model trained on a primary task to execute a secondary task, the adaptation of CNN models trained on such large datasets has rapidly gained popularity in many fields of science, geosciences included. However, the trade-off between two main components of the transfer learning methodology for geoscience images is still unclear: the difference between the datasets used in the primary and secondary tasks; and the amount of available data for the primary task itself. We evaluate the performance of CNN models pretrained with different types of image datasets—specifically, dermatology, histology, and raw food—that are fine-tuned to the task of petrographic thin-section image classification. Results show that CNN models pretrained on ImageNet achieve higher accuracy due to the larger number of samples, as well as a larger variability in the samples in ImageNet compared to the other datasets evaluated.