First EAGE Conference on Machine Learning in Americas

The aim of this paper is to present a simple but effective workflow to classify depositional facies using conventional well logging data and supervised learning algorithms. Facies recognition is a time-consuming task and economically expensive. Therefore, depositional facies provide essential information regarding the distribution of rock properties like porosity and permeability. Using seismic attributes is possible to define depositional facies using different attributes extracted from 3D seismic and well logs to construct a detailed distribution map ( Bagheri & Riahi, 2015 ). However, this method is limited by seismic resolution and the available 3D seismic data. Core analysis greatly supports facies recognition within depositional environments, yet it is time-consuming and expensive making it hard to apply for every well. Finally, there are some other works related to facies classification using conventional well logging data and Machine learning algorithms, ( Puskarczyk, 2019 ) uses Artificial neural network (ANN) for pattern recognition and identification of electrofacies. ( Al-Mudhafa & (Basrah Oil Company - Iraq), 2020 ) predict facies classification using two non-parametric supervised machine learning approaches: K-Nearest Neighbours (KNN) and Random Forest (RF) in a carbonate reservoir.

Despite all the studies already explored using machine learning algorithms for several authors, there are still issues that most of the works have not been covered totally concerning data structure, algorithms limitations, and hyperparameters management, making most of those in other conditions not replicable. This paper may drive all these points, given and important emphasis on the different issues that imply the use of machine learning algorithms as a predictor of depositional facies in a transitional environment. Results show a much better performance of SVM over KNN to predict depositional facies.

Summary not available

Summary not available

Summary not available

Summary not available

Summary not available

Summary not available

The development of a Mature Oil field as well as others in the world is a challenging task, and Talara Basin is one of them because of its growing depletion and its structural complexity, the latter made the seismic reflection image a little bit difficult to interpret. However, there are some areas where the seismic data plus well information could be used to get more reliable reservoir characterization using machine learning tools. The relation between different seismic attributes and acoustic impedance derived from well-logs using XGBoost (Extreme Gradient Boosting) regression algorithm is a compelling example of how machine learning could add extra information to predict reservoir sandstones. The aim of this study is to perform a machine learning model throughout the interaction of the extracted amplitude from the nine seismic attribute volumes as a log curves inside the reservoir with acoustic impedance log curve (Ip) in seven wells in the structural block (that contains 10 wells with Ip and 3 out of them are blind wells in the model). As a result, the model gives us the opportunity to predict Ip curves from seismic attributes with higher seismic resolution at each trace of the 3D seismic inside the block. Hence, the acoustic impedance volume from the XGBoost model due to its high resolution could be used to point out isolated sand bodies that will be difficult to predict with stochastic model that only use spaced wells. To be honest, this study does not attempt to replace other workflows of seismic inversion or more robust geostatistical model. On the contrary, the possibility to obtain nonlinear operators from the machine learning algorithm that could learn the anisotropic behavior of the wavefield propagation is the most prominent goal of this study. Furthermore, machine learning models could friendly and swiftly be used as a propagation guide of stochastic seismic inversion. For instance, Zapotal Field, in Talara Basin has a long history with several years of oil production, during its first year of exploitation the wells give enormous volume of hydrocarbon as the development of the fields have been growing, the difficulty to maintain a balance between costs of production by barrels turn out to be difficult which is the main reason why many expenses had to be cut off. Nevertheless, it was the opportunity to use alternative techniques such a machine learning which improve our reservoir models without acquiring commercial software tools that raise the cost, thus there is still remain room for machine learning applications in many areas.

The application of commercial-scale deep learning to mineral exploration is in its infancy, with the intersection of affordable computation, large data volumes, novel applications of 3D deep learning and commercial buy-in all contributing to recent developments. The DeepMine team at IBM have spent the last three years developing and applying 3D Convolutional Neural Networks (CNNs) to the prediction of economic-grade resource in a hard-rock mining context. The iterative process has resulted in successes and shortcomings, yielding valuable insights for the resource exploration community. The approach involves representing subsurface data as point cloud information, which is then input as voxelated training examples into the model. 3D CNN models were constructed on top of PyTorch’s Deep Learning framework, developed by Facebook’s AI research group. The success of the project has been judged primarily by the ability of the model to reliably predict economic mineralization at greater distances from existing drillhole data upon each enhancement. Most recent applications have begun to incorporate non-drillhole data sources, such as Vertical Time-Domain Electro-Magnetic (VTEM) and Airborne Gravity and Magnetic data. It was found that key challenges and opportunities existed in the areas of efficient and accurate data representation, bespoke data augmentation techniques and the 3D CNN formulation itself. New improvements continue to be made to the IBMs body of work on the subject, DeepMine.