- Course News & Geophysical Curiosities
To test whether your knowledge of geophysics, geology , statistics and Artificial Intelligence is sufficient got to Quiz 0
Beschikbare cursussen
AVO + Inversion
I will start with “Geophysical Methods” or “Multi-Physics” presentation, in which I show the importance of considering the whole range of geophysical data (seismic: velocity & density, gravity: rock density, magnetics: magnetic susceptibility, electromagnetics: resistivity) that could help in understanding the subsurface. The next step after a structural and stratigraphic interpretation is quantitative interpretation (QI). For QI, special, relative-true-amplitude processing is needed to assure that the reflections can be interpreted quantitatively in terms of lithology and pore fluids. For QI the final resolution after processing is shown by the compactness of the Point-Spread-Functions (PSFs).
The complexity of the subsurface is looked at in terms of inhomogeneity and anisotropy. Inhomogeneity is well known, the earth consists of rock bodies with different lithologies and pore fluids, but in addition some properties can only be properly described by including anisotropy. Clear examples are shales, where the stacking of clay particles cause wave propagation to be dependent on the direction relative to the orientation of the clay platelets. This is also in the case of cracks and fractures, which when “organised/systematic”, say perpendicular to the minimum horizontal stress influences the wave propagation. Thinking the other way around, these rock properties can be derived from seismic AVA observations. The way the reflection amplitude changes with azimuth can be used to determine the presence of vertical fracture systems. In addition, amplitude changes with offset, or angle-of-incidence to be precise, can be used to determine changes in rock and fluids across interfaces. As the seismic waves are band-limited they interrogate only the “wavelength-average” rock properties. Different ways of averaging can be used to calculate so-called effective media. Acoustic Impedance (AI) is the rock property usually considered. AI is related to rock properties seen by normal incidence reflections. More information on the rock properties is captured in AVA related Elastic Impedance or even better Extended Elastic Impedance (EEI). AVA can be calculated using Zoeppritz equations, which are often linearized to provide inside into the influence of rock properties. Alternatively, it can be calculated using the wave equation, which will be demonstrated.
Then I concentrate on Rock Physics. Rock Physics deals with the relationship between elastic wave propagation properties (velocity, density, attenuation) and reservoir properties (porosity, permeability, fracture systems, saturations). Two basic kinds will be discussed. One for clastic rocks, the other for carbonate rocks. These are conventionally given by standard well-based statistical relationships, but recent case studies show that better results can be obtained using Machine Learning. For direct calculation of the seismic response to a reservoir filled with a pore fluid, it is necessary to know the dry rock skeleton properties. This is near to impossible, except based on a core. Therefore, an alternative is applied, namely having log measurement of the rock filled with a determined (known) fluid mixture, Gassmann’s equation allows the calculation of its properties when the rock is filled with another fluid mixture (Gassmann fluid replacement algorithm). But the “averaging” of fluid properties in case of a mixture also depends on the distribution of the saturation (uniform or patchy). Exercises will help to see the significance of the topics discussed.
Finally, I will discuss a topic that is revolutionizing our workflows, namely that part of Artificial Intelligence that is called Machine. An important characteristic of Machine Learning that it allows a much use of the multi-feature space describing the rock properties. This will lead to better rock models.
Many exercises will deal with the use of AVA for PP, SS and PS data.
Inversion
The aim of inversion is to derive the rock properties from seismic data. However, there are several ways of inverting seismic data.
One kind is to invert the seismic reflection amplitudes to elastic parameters such as acoustic impedance. Examples are Bandlimited (Recursive), Sparse Spike, Coloured and AVO inversion. As seismic data is bandlimited, constraints are added to obtain absolute elastic values. A clear example is the use of a low-frequency (low-wavenumber) background velocity model. Apart from AVO, these are, although useful, all “poor man’s solution” as the result is only approximately correct.
AVO inversion provides an opportunity to move away from acoustic impedance (AI) towards elastic impedance (EI) or even better extended elastic impedance (EEI) using the angle dependent reflection strengths. The true angle of incidence θ [0-90°) on a reflector is used in EI, but for EEI a parameter χ [-90°-+90°] is used, which has the unit of degrees, but clearly cannot be an angle of incidence. For specific values of χ the χ-trace correlates very well with certain rock parameters. These angles can be found by correlation with well logs, but also a physical poro-elastic interpretation can be given.
Another kind of inversion is the Full-Waveform-Inversion (FWI), which derives detailed elastic parameter models directly from seismic data. This approach compares synthetic seismic with observed seismic and adjusts the starting model iteratively till the synthetic data fits the observed data according to a set criterion. This method knows two critical steps, namely a forward modelling step (from model to synthetic seismic) and an inversion step (from synthetic-observed data difference to model updates). FWI, although compute intensive, is being applied increasingly to derive elastic parameters and high-quality images of the subsurface. But a second step is still needed to go from elastic to rock parameters unless these are explicitly included in the forward modelling. Note that to convert elastic to rock properties a rock-physics model is needed. For a clastic environment, such models are readably available, for carbonates an informed (using well logs) choice must be made which rock-physics model should be used.
In the last 10 years Machine Learning started to make successful inroads into AVO modelling and inversion. Various successful case studies will be discussed.
More and more Deep Learning will play a role not only in society in general but also in the geosciences. Deep Learning resorts under the overall heading of Machine Learning / Artificial Intelligence. In this domain often the word “Algorithms” is used to indicate that computer algorithms are used to obtain results. Also, “Big Data” is mentioned, indicating that these algorithms need a large amount of training data to produce useful results.
Many scientists mention “Let the data speak for itself” when referring to Deep Learning, indicating that hidden or latent relationships between observations and classes or values of (desired) outcomes can be derived using these algorithms. Examples are in the field of seismic processing (first arrival picking), interpretation (facies prediction), etc. Often, we resort to statistical relationships. Then Deep Learning enters the game. From a range of labelled data (called instances) we can derive a linear/nonlinear relationship (model in DL terminology) that predicts the label or value (supervised learning) of new data (instances in DL terminology). But sometimes it is already useful if an algorithm can define separate groupings / clusters, which then still need to be interpreted (unsupervised learning). Even more sophisticated is Semi-supervised learning: labelled and unlabelled data together are clustered whereby the unlabelled data receives the label of the dominant class present in the cluster.
The Course
The aim of the course is to introduce how Deep Learning (DL) can be applied in geophysics. It is a sequel to the course AI: Machine Learning for Geophysical Applications. Hence, it is highly recommended to do that course first as it uses a user-friendly package, called Weka. In that course you will acquaint yourself with the workflows and algorithms used in Artificial Intelligence /Machine Learning.
In the Deep Learning course, we will predict lithology and pore fluids as well as facies to learn the Deep Learning workflows and algorithms. Use will be made of open-source software: TensorFlow and Keras. Power-point presentations and videos will introduce various aspects of DL, but the emphasis is on computer-based exercises. The exercises deal with pre-conditioning the datasets (balancing the input classes, standardization & normalization of data) and applying several methods to classify the data: Multilayer Perceptron, Support Vector, Nearest Neighbour, AdaBoost, Trees. Non-linear Regression is used to predict porosity. Use will be made of Google Colab and Scikit-Learn. It runs on the Cloud and allows use of a GPU. It is “the way” to learn using a whole range of open-source Deep Learning algorithms. In the exercises you will get acquainted with using interactive python notebooks, how to get algorithms using Scikit-Learn and if you restrain yourself from using it on very large datasets, Google Colab is free.
The course consists of many presentations and exercises as I am a strong believer in the paradigm: Tell me and I will forget, show me and I might remember, involve me (through exercises) and I will truly learn.
Learning methods and tools
At the end of the course participants will have a clear idea how Deep Learning, being part of Machine Learning / Artificial Intelligence will impact the future of Geosciences. This will be evident from the examples discussed and applied to the case of predicting lithology, porosity, pore fluids and facies. Quizzes are provided to enhance the learning.
Intended Audience
All those interested in understanding the impact Artificial Intelligence will have on the Geosciences. Hence, Geologists, Geophysicists, Petrophysicists and Reservoir engineers, involved in exploration and development of hydrocarbons or mineral resources , but also those involved in geothermal and CO2 storage, where the relationship between data and targets are often difficult to establish.
Pre-requisites
A basic understanding of geophysics and statistics. A pre-requirement quiz (Quiz 0) can be taken to check whether your knowledge of geophysics and statistics is sufficient to follow the course. But note again, it is highly recommended to follow the AI: Machine learning for Geophysical Applications first.
More and more Deep Learning will play a role not only in society in general but also in the geosciences. Deep Learning resorts under the overall heading of Machine Learning / Artificial Intelligence. In this domain often the word “Algorithms” is used to indicate that computer algorithms are used to obtain results. Also, “Big Data” is mentioned, indicating that these algorithms need a large amount of training data to produce useful results.
Many scientists mention “Let the data speak for itself” when referring to Deep Learning, indicating that hidden or latent relationships between observations and classes or values of (desired) outcomes can be derived using these algorithms. Examples are in the field of seismic processing (first arrival picking), interpretation (facies prediction), etc. Often, we resort to statistical relationships. Then Deep Learning enters the game. From a range of labelled data (called instances) we can derive a linear/nonlinear relationship (model in DL terminology) that predicts the label or value (supervised learning) of new data (instances in DL terminology). But sometimes it is already useful if an algorithm can define separate groupings / clusters, which then still need to be interpreted (unsupervised learning). Even more sophisticated is Semi-supervised learning: labelled and unlabelled data together are clustered whereby the unlabelled data receives the label of the dominant class present in the cluster.
Domain Experts and Data Scientists
In discussions at the EAGE Digital conference 2024, it was emphasized that not only the Subject Matter Experts (SME’s) had to become familiar with the terminology and methods used by the Data Scientists, but also the Data Scientists must understand what geology and geophysics is about. That doesn’t mean they need to know the ins-and-outs of these subjects, but at least know the terminology and the overall context for which they need to provide the Machine / Deep learning tools. Therefore, this course will be a first step in providing the necessary geophysical background.
The Course
As it is assumed that the Data Scientists are familiar with mathematics and statistics, the course will include advanced geophysical subjects. A general overview of seismic and non-seismic acquisition, processing and interpretation will be followed by various uses of Machine / Deep learning for Geophysical Applications. We will predict lithology and pore fluids as well as Facies to learn
the Deep Learning workflows and algorithms needed in geophysics. Use will be made of open-source software: TensorFlow and Keras. Power-point presentations and videos will introduce various aspects, but the emphasis is on computer-based exercises. The exercises deal with pre-conditioning the datasets and applying several methods to classify / cluster the data: Multilayer Perceptron, Support Vector, Nearest Neighbour, AdaBoost, Trees. Non-linear Regression is used to predict porosity. Use will be made of Google Colab and Scikit-Learn. It runs on the Cloud and allows use of a GPU. It is “the way” to learn using a whole range of open-source Deep Learning algorithms for geophysical applications. The course consists of many exercises as I am a strong believer in the paradigm: Tell me and I will forget, show me and I might remember, involve me (through exercises) and I will truly learn.
Learning
At the end of the course participants will have a clear idea of what goes on in Geophysics and how Artificial Intelligence will impact the future of Geosciences. Interactive quizzes using “Mentimeter” are used to enhance the learning.
Intended Audience
Data Scientists who will be cooperating with geoscientists to develop AI methods for exploration and development of hydrocarbons or mineral resources. Also, application for geothermal and CO2 storage are discussed.
Pre-requisites
A good understanding of mathematics, statistics and to some degree of physics. Go to Quiz 0 to test yourself
Note: The course can be adapted to comply with the needs of participants.
Various kinds of geophysical data are available, sometimes summarised as Multi-Physics data. Usually, they are separated into Seismic and Non-Seismic. Seismic is, without any doubt, the dominant method used in the energy industry. But Non-Seismic data (gravity, magnetics, electrical, electromagnetics, etc) is the main source of information in shallow subsurface applications (engineering, mapping pollution, archaeology). It is also used in the early reconnaissance of new basins and plays and in mapping prospects below salt/basalt (Magneto-Telluric). However, seismic has its limitations and therefore also non-seismic methods are used successfully as complementary tools in subsurface evaluation. In combination with seismic data, they can significantly reduce the uncertainty of subsurface models as they measure different physical properties of the subsurface.
In the first part, various aspects of gravity will be discussed, such as the Earth gravity field, determining anomalies in the global field, establishing the depth of density anomalies, be it spherical or anticlinal and the resolution, which is limited because gravity is a “potential field”. Most promising is the development of gravity gradiometer, whereby gradients in the gravity field can be directly measured with great accuracy. These measurements are less sensitive to airplane and ship movement.
In the second part, the Earth’s magnetic field, also a “potential field” will be studied. Being mainly due to the internal dipole source, the inherently more difficult interpretation is simplified by applying a transformation to a monopole field (Reduction to the Pole or Equator). As different causative sources can produce the same surface measurements, non-uniqueness is allways present. However, promising developments to mitigate these issues will be shown.
Finally, we will discuss and apply the “latest and greatest” that is Machine Learning, a part of Artificial Intelligence, which is step-changing the geoscience world. We will use a standard package called Weka and familiarise ourselves with the “ultimate” open-source software Keras and TensorFlow. Exercises will be done using Weka and Google Colab. After the course, you should be able to try out Machine Learning on your own data. In addition we will experiment with ChatGPT.
Geophysics provides technology with which we can "look" into the subsurface. It is a key enabler of many activities in the search for hydrocarbons, minerals, fresh water, and geothermal energy. Of the many existing geophysical methods, three are important for monitoring CO2. These are Seismic and Electromagnetic and Gravity methods. CO2 injection can be done for sequestration as well as enhanced oil recovery
In the first method, high resolution seismic up to 250 Hz can be acquired with short offsets. For deeper layers, long offset seismic is collected. Long offsets are needed for Refraction Static corrections and in case Full Waveform Inversion is applied for obtaining the diving waves. For CO2 sequestration, the presence of fractures and their orientation, being natural or induced is a significant hazard that can be determined from seismic.
The second important geophysical method is electrical and electro-magnetic methods. Electrical or Direct Current surveys use grounded electrodes for source and receivers. They measure the potential difference using increasing receiver electrode spacing. Changes in measured potentials contain information on the resistivities of the subsurface, which can be related to changing pore fluids, like CO2 replacing brine. Electro-Magnetic can use either grounded or inductive sources (aerial surveys), but also natural sources as used in Magneto-Telluric surveys. These Electro-Magnetic sources can be a harmonic source (using a single frequency) and the measurement of the magnitude and phase delay, or real and imaginary responses are used. The other Electro-Magnetic source option is a step-off function. The subsurface information is then contained in the amplitude decay after shut-off.
The third approach is using Gravity measurements. These could be changes in the vertical gravity component Gz or the Full-Tensor-Gravity (FTG), which measures gradients in the gravity field. With the increasing accuracy of modern instruments, changes in the CO2-brine interface have been measured in oil (Prudhoe Bay, Schrader-Bluff and Troll field) and gas-storage (Izaute) fields
We all have seen displays of seismic data in the form of sections or cubes of data. But what do they show and how are they acquired? In this course you will learn to understand that seismic data represents the movement of the surface, resulting from waves generated by a source, dynamite or vibrator, which are reflected by changes in the subsurface rocks. Hence, what we record is related to the properties of the rocks, not only rocks, but also its pore fluids. All information on the subsurface is contained in these records, but almost impossible to extract and understand. Therefore, the records need to be processed to make it possible to interpret structure and content of the pore space. In this course, the basic principles of acquisition and processing will be discussed. But also, insights in advanced methods will be provided. These methods allow a much more accurate interpretation of seismic data. The aim is not to fully understand these methods, but to understand its importance in certain cases, to enable interpreters, reservoir engineers to formulate requests for these methods.
