Second EAGE Digitalization Conference and Exhibition

We propose a methodology that allows a fast, accurate, and unbiased interpretation of well data in terms of labelled discrete logs of stratigraphy at multiscale. These discrete logs can be used for: 1. advanced stratigraphy and facies analysis along the wells, 2. advanced subsurface interpretation: correlations between wells, and 3. 3D static model conditioning.

Two separate methods are proposed to provide an automatic stratigraphic zonation: 1. Automatic grain size trend interpretation from gamma ray logs, using UNet architecture, and 2. Automatic sedimentary geometry interpretation, from borehole images, using conventional neural network (CNN) and ResNet architectures. Both approaches are trained using synthetic and real data for optimal interpretation. These techniques have been combined into a workflow and validated on real data from the Wheatstone Field, Australia.

Geological characterization of the subsurface is important for the success of a wide range of applications - From petroleum and renewable energy activities to civil engineering. In this paper, we present a new deep learning assisted framework to interpret geological features in the subsurface from seismic and well data. Specifically, the workflow integrates multiple 3D deep convolutional neural networks (CNNs) trained to target specific geological features such as salt, faults, and stratigraphic boundaries.

This paper recaps various applications of waterflood optimization using the Aqueon, powered by Data Physics, workflow. Data Physics has been described in various papers (see References) and is a blend of reservoir modelling physics, machine learning and advanced optimization techniques. Data Physics combines the predictive capacity and robustness of numerical simulators with the speed of machine learning applications to provide a tool that can create models in weeks, update them automatically, run long term predictions in minutes and provide the ability to evaluate the impact of operational decisions supported by a predictive model. Leveraging the speed at which the models run, Data Physics uses memetic algorithms to generate the pareto front and prescribe optimum operational strategies whether for steam or water injection, unconventional completions, in-fill drilling and any other operational parameter.

The paper presents four practical cases which have been implemented in the field and include:

1. Increasing production redistributing injection
2. Reducing injection while maintaining production
3. Optimizing an injector reactivation program
4. Optimizing a producer to injection conversion program

Hydrocarbons bearing reservoirs, as any other earth natural resource are part of our natural environment. As such, any exploration or production managing decision is facing natural Uncertainty, that results in unavoidable differences between operational results and their expectations. Moreover, these differences always mean a loss in sustainability and profitability of the resource management.

At the digital age, when high computing power is available at low cost, Big Geo Data sets can be acquired and processed using machine learning algorithms, changing the game of Earth resource modelling and decision-making.

We introduce and define the concept of Earth Intelligence (EI) as the use of probabilistic models for replacing knowledge based numerical models for supporting operational exploration production decision making.

Whereas classical Artificial Intelligence (AI) algorithms work using “cost functions” maximizing the match to the input Data available at the time of decision making, Earth Intelligence maximizes the match to the output Resource recovered after the decision has been made.

Artificial Intelligence then best performs for Data interpretation when Earth Intelligence aims at best supporting operational decision making by enabling a rational uncertainty management and resulting in reliable P10 P50 P90 confidence intervals and scenarios.

A case study illustrates the concept of EI.

Despite being powerful levers to democratize digital technologies in many scientific domains, open-source projects still raise limited interest outside computer science circles. For instance, the ELK open-source framework is increasingly employed in various industries to manage and screen large collections of data and unstructured documents, but it remains under-exploited by the geoscience community. Consequently, in this study, we appraise the potential of this framework to browse and manage geoscience knowledge in two practical situations, corresponding to the routine work of exploration teams in an energy resources company. First, the ELK stack is combined with other open-source tools in a dedicated application architecture. Then this architecture is customized to address each use case more specifically, thus assessing its flexibility and versatility. Our conclusions are very positive about the capacity of the framework to considerably facilitate the handling of massive bases of geoscience knowledge, such as the documents capitalized in past exploration studies or the well information accumulated on mature fields. This work also illustrates how open-source projects can contribute to the digital transformation of subsurface-related industries, as geoscience professionals can rapidly develop solutions adapted to their concrete needs and unlock significant efficiency gains in their daily work.

Results of solving fundamental difficulties of data management at NIS a.d. are discussed. The business’ need for solving mentioned problems became a catalyst for the optimal technical solution search, which led to the Exploration Platform project initiation. During the execution phase, authors identified several challenges, which lie far beyond the simple implementation or creation of a single software. This requires systemic changes in the number of business processes and corporate culture. In response to the challenges, authors proposed a novel approach to managing data across domains, which, according to the stakeholders’ opinion, is a desirable change. The Exploration Platform project has already influenced on exploration business processes and it affects related areas of the company’s activities.

There is a substantial need for exploration and field development workflows that generate 3D stratigraphic models, which can be used to explain well log data for making drilling and measurement acquisition decisions. This abstract presents a workflow designed for this purpose; i.e., locate the ensemble of 3D stratigraphic models that can explain observations. 3D stratigraphic models spanning a spectrum of depositional environments are stored in a datastore. This datastore models have encoded indexes of well log signatures. This encoding is obtained using the Vector Quantization Variational Autoencoder (VQ-VAE) approach. User-supplied well logs are also encoded with VQ-VAE to allow for a quick search in this datastore by looking for the nearest clustering centroids.

Deep Learning methodologies are increasingly used to facilitate image analysis in various domains and raise growing interests in geology, as this discipline heavily relies on visual interpretation. However only few practical use cases on operational datasets are yet documented. Thus, in this work we appraise reference Deep Learning workflows to automate the interpretation of thin sections and core photographs from real-life geological studies. In a first appraisal, we use a dataset of carbonate thin sections from the Graus-Tremp basin (Spain) to assess object detection models. We train four Deep Learning models to automatically spot, delineate and characterize 9 different families of microfossils on these sections. The results are qualitatively assessed by human geologists, and precisions and inference times quantitatively measured. In a second appraisal, we use core samples from the Gulf of Corinth to evaluate the potential of supervised classification models in extrapolating human interpretation from a few segments to the entire wells. We carry out a Transfer Learning methodology to generate and compare multiple models with different neural architectures and training strategies. From this experience, we highlight good practices and recommendations for further use of Deep Learning technologies in similar contexts.

We present a stratigraphic forward modeling platform (SFM) to generate synthetic data at the reservoir scale for subsurface interpretation workflows based on artificial intelligence and machine learning (AIML).

Integrating stratigraphic processes in 3D modeling is the best way to avoid generating unrealistic representations of the subsurface that lead to inaccurate forecasting of reservoir production, CO2 sequestration capacity, or fluid migration for geothermal energy production. Conventional workflows for the interpretation of stratigraphy from well log, seismic, and 3D stratigraphic modeling tools comprise steps that are often tedious, manual, isolated, and time consuming. Using AIML algorithms to automate these different steps is ideal, but to be successful it will require a significant amount of multiscale training datasets. The possibility to use only real interpreted well logs, 3D seismic data, or 3D facies models to train the different AIML engines is not an option because we would need a huge, very diverse, and properly labeled dataset free from sampling and user bias.

The SFM is introduced and all its derived rock properties and synthetic measurements. We present some examples of applying these synthetic data to AIML workflows. The SFM functionality and use cases presented in this paper focus on fluvial systems.

Channelized systems characterization is key for geological context understanding. Moving the point of view from seismic vertical profiles to attribute in surface layers facilitates their detection and becomes more suitable for deep learning techniques. However, the channelized systems depict so complex delimitations than an exact labeling is not realistic in reasonable time. We assumed then a coarse labeling that may be done fast. In that case, some pixels are labelled as channel while they are not. We proposed to automatically modify the label values from deterministic 0 or 1 to probabilistic 0 to 1 with a fixed 1 values area around the channels’ skeletons and a smooth decreasing to 0 values. Then, classical accuracy metrics like Intersection-Over-Union are not relevant in coarse labeling context. Indeed, they act at pixel level without distinction between pixel level of importance. We derived a formula proposed for 1D structures to 2D channelized systems detection. The idea is to compute the distance map where pixel values are their distance to the areas of interest in both label and prediction images. Then, the accuracy metric uses a Gaussian attenuation with a tolerance on the pixel distance to allow geophysical uncertainties.