Breakaway Learning

Pooling

The aim of pooling is to reduce the input image in order to reduce the compute load. Each neuron in a pooling layer is connected to a limited number of neurons in the preceding layer, located within a small rectangular area (see figure on the left). A pooling neuron does not use connection weights, just combines the values in the area according to specification: average or maximum. In the figure only the maximum in the 2x2 area makes it to the next layer. Also a stride of 2 is used, reducing output image by a factor 4. (No padding at the edges).

Pooling is usually done between the convolutional layers and is applied to all the outputs of the convolutional filters.

Reference: Hands-on Machine Learning with Scikit_Learn, Keras & TensorFlow by Aurélien Géron

In Machine Learning 2 techniques are of great importance: Dropout and Transfer Learning. Dropout was discussed under Geophysical curiosities, here we will dicuss Transfer learning.

In the future software companies will offer for sale trained Deep Learning Networks for specific tasks like seismic interpretation, reservoir characterization, etc. These DNN’s, although trained on specific data sets are a good starting point for building your own models/algorithms for your data, as training from scratch requires a large and expensive compute effort. Those models are based on relevant but still different training instances, thus how do I adapt the model for my own data? The technique is called “Transfer Learning”.In Transfer Learning you take an existing DNN trained for a task A (shown on the left) and freeze the lower layers, as they will handle the most basic characteristics of for example a picture and only replace a few upper hidden layers with new to be trained layers for your task B (shown on the right). The more similar the tasks the fewer layers need to be retrained. For a very similar task only the output layer needs to be replaced.Train your new model on your dataset and evaluate its performance. Then unfreeze the upper frozen layers, so that the training can also tweak the hyper parameters. This will improve the performance without demanding extensive compute time.

Reference: Hands-on Machine learning with Scikit_Learn, Keras & TensorFlow, Aurélien Géron

Course News 20

In my Machine Learning courses I have used so far, the open-source package Weka. The reason is that it is a user-friendly and easy to run package with most relevant Machine learning algorithms, except truly Deep Learning. This suffices for most exploratory applications, where we like to learn the workflows and applications of Machine learning. Weka also has an option to build your own sequence (KnowledgeFlow) in such a way that data is not stored in memory and therefore allows applications to very large data sets. A disadvantage of Weka is that deep learning can be done with only one hidden layer, but in practice it takes too long using a CPU.

Therefore, I have included in my courses an introduction to Google Colab. This runs on the Cloud and allows use of a GPU or a TPU. It is “the way” to learn using a whole range of open-source Machine Learning algorithms. In an exercise you get acquainted with using interactive python notebooks, how to get algorithms using sklearn and if you restrain yourself from using it in earnest on large datasets, it is free.

Course News 19

As part of a more extensive course on Fractured Reservoirs, I have developed the geophysical part called "Seismic Fracture Detection". It discusses the seismic acquisition requirements for multi-azimuth analysis using NMO velocity anisotrop one way to determine the fracture orientation of vertical fracture systems and Amplitude versus Angle of Incidence for more resolution and determination of fracture densisty.. P-wave, PS wave or Converted wave and S-wave data are discussed and used in AVA exercises. The program, which could be fully interactive, is as follows:

Course News 18

Compressive sensing, a method with which significant savings in acquisition costs can be achieved, is getting more and more attention as 3D surveys get bigger and continuous monitoring is used more often, also for CO₂ storage and geothermal.

Compressive sensing circumvents the Shannon-Nyquist criterium of taking 2 samples per wavelength. It utilises sparsity of data in a transform domain, for example Fourier or Radon domain.

The crux of the method is that the data / wavefield will be on purpose (lower cost) under-sampled, but the resulting sampling (aliased) noise can be “filtered out / removed / suppressed” at the benefit of the needed data / wavefield. This allows the reconstruction of the wavefield in the state it would be acquired using full Shannon-Nyquist sampling. There are a few assumptions, however: there must exist a transform domain where the data is sparse, and the necessary random sampling must include the sampling of the desired reconstructed wavefield.

Course News 17

I have built two short courses of importance to the "Energy Transition".  The first one is called "Geophysics for Geothermal Energy" and deals with the two most relevant geophysical technologies: seismic and Electromagnetic (EM) methods. The second one is called "Geophysics for CO2 sequestration and Enhanced Oil Recovery (EOR)". This course includes the use of Gravity and Gravity Gradiometry, in addition to Seismic and ElectroMagnetism, to monitor the CO2. Both courses are 2-day courses, with presentations, exercises and quizes to reenforce the learning. As with all my courses, they can be followed interactively, using Moodle, or Face-to-Face (F2F). They also can be combined or made part of a package of Transition relevant EPTS courses.

Course News 16

In the past we always talked about Seismic and Non-Seismic methods. But I was never quite happy with the indication Non-Seismic. An equivalent issue exists with the name Vertical Transverse Isotropy. It would be better to call it Polar Anisotropy, as that name directly refers to the relevant property, namely anisotropy. Therefore, I am pleased by the increasing use of the word Multi-Physics instead of Non-Seismic. This directly indicates the use of many different (geo)physical measurements used (gravity, magnetics, electromagnetics, self-potential, etc and seismic). In due time I will replace Non-seismic by Multi-Physics in my courses.

Course News 15

From recent experience with Blended Learning courses, I have decided to increase the interactive parts of the course. In a 5-day course, I will have each day a half day interactive session and the other half will be self-paced reading and doing the exercises. Below an example of the new schedule for the Non-Seismic Data Acquisition and Processing course. All other courses will be updated similarly.

For all my Blended Learning courses, consisting of 5-parts (equivalent to a 5-day F2F course) a choice can be made between the following two schedules:

Course News 14

Course News 13

Mar. 16^th 2021

In the E-Learning Course Programs I replaced the Timetable by Part 1, 2, etc.
The main reason is that as it is a E-Learning, self-paced course the client can adapt the time schedule (say, when staff also must work half days).

Two examples of the alternatives:

Machine Learning for Geophysical Applications (MLGA)

Part 1
Day 1:        3 hours interactive presentation & discussion
Day 1&2:   4 hours self-paced exercises+learning-reinforcement quiz 1
Part 2
Day 3:         3 hours interactive presentations & discussion
Day 3&4:    4 hours self-paced exercises+learning-reinforcement quiz 2

Advanced Seismic Data Acquisition and Processing (ASDAP)
Part 1
Day 1:         3 hours interactive presentation & discussion,
Day 1&2:    4 hours self-paced exercises+learning-reinforcement quiz 1
Part 2
Day 3:         3 hours interactive presentations & discussion
Day 3&4:    4 hours self-paced exercises+learning-reinforcement quiz 2
Part 3
Day 4:         3 hours interactive presentation & discussion,
Day 4&5:    4 hours self-paced exercises+learning-reinforcement quiz 3
Part 4
Day 5:         3 hours interactive presentations & discussion
Day 5&6:    4 hours self-paced exercises+learning-reinforcement quiz 4
Part 5
Day 7:        3 hours interactive presentation & discussion,
Day 7&8:   4 hours self-paced exercises+learning-reinforcement quiz 5

Course News 12

Feb. 19^th 2021

Anisotropy is more common than isotropy. Therefore modelling of seismic phenomenon ignoring anisotropy will most often lead to incorrect interpretations. The same is true for processing seismic data assuming isotropy.

The change in reflection strength with increasing angle of incidence (AVA) can provide important information with respect to the properties of reservoirs. Although formulations exist for strong contrasts across an interface, weak contrasts are more common in sedimentary sequences. This leads to formulations, through linearization, which provides insight in the influence of the elastic parameters across the interface on the AVA behaviour. These expressions can also be extended to include anisotropy of the media and in case of weak anisotropy the expressions still allow insight into the influence of the anisotropy contrasts across the interface on the AVA behaviour.

An interesting question is: Can we express the AVA behaviour in terms of the isotropic expressions using generalised elastic properties?

The answer is yes, we can with so-called pseudo-elastic properties. These are function of the true elastic properties modified by expressions involving the anisotropy parameters.

An example for VTI:
            
with
          
The use of pseudo-isotropic parameters is added to the Quantitative Reservoir Characterisation Blended Learning and F2F Course.

Course News 11

Feb. 27^th 2020

Facies prediction along wellbore using Machine learning
A new short course has been developed on the use of Machine Learning for prediction of facies along a borehole. Based on a reasonable number of interpreted wells various algorithms can be compared in classifying wellbore profiles of “new” unlabelled wells. In the course it will be shown that limited training data (1 well) results in poorer predictions than when more labelled data can be used for training. It also is found that using a Deep Neural Net results in the best classification. In addition, the use of semi-supervised learning, often needed in geoscience applications because of limited availability of labelled data, is discussed. The course is available in a Face-to-Face and an E-learning version.

Course News 10

Jan. 2^th 2020

Course instructor

I am happy to announce that I have been officially nominated as short course instructor with the EAGE.

Course News 9

Jan. 2^th 2020

Course quizzes

To each course a multiple-choice quiz has been added. The aim is not only to provide a test of what has been learned but also as a way of enhancing the participants understanding of the subjects.

Course News 8

Aug. 20^th 2019

Machine Learning course using a realistic geophysical dataset
Based on Open Source software, a one-day (8hrs) Face-to-Face and a one-month duration (8hrs equivalent) E-learning course called “Introduction to Machine Learning for Geophysics” have been developed. Both courses contain power point presentations with references to publications together with computer-based exercises and Machine Learning related videos.
The exercises deal with a genuine geophysical issue, namely predicting lithology and pore fluids, including fluid saturations. The input features are Acoustic and Shear Impedances, Vp/Vs ratios and AVA Intercept and Gradients. Several exercises deal with pre-conditioning the datasets (balancing the input categories/instances, standardization & normalization of features, etc), the others apply several methods to classify the “instances” (Machine Learning terminology for cases): Bayes, Logistic, Multilayer Perceptron, Support Vector, Nearest Neighbour, AdaBoost, Trees, etc. This for supervised and unsupervised applications. Also, non-linear Regression is used to predict fluid saturations

Course News 7

July 24^th 2019

Update on Machine Learning courses for Geophysics
Based on Open Source software I have developed a one-day (8hrs) Face-to-Face and a one-month duration (8hrs equivalent) E-learning course called “Introduction to Machine Learning for Geophysics”. These courses are made up of power point presentations with references to publications together with computer-based exercises and Machine Learning related videos.

Course News 6

June 4^th 2019

Machine Learning
Based on Open Source software I have developed an “Introduction to Machine Learning” together with computer-based exercises. The Introduction will be included in all my courses, but the exercises will be topic specific. Apart from the Face-to-Face / Classroom courses I plan to develop very short E-learning courses (the equivalent of a one-day course) on the use of Machine Learning in Basic Geophysical Data Acquisition and Processing, Advanced Seismic Acquisition/Processing/Interpretation, Quantitative Reservoir Characterization and Non-Seismic Data Acquisition and Processing.

Course News 5

Feb. 20^th 2019

Based on a DISC given by Doug Oldenburg & Linsey Heagy in 2017 at the Delft Technical University, I have designed a new course on Electro-Magnetic methods. It consists of a number of short lectures and as usual for my courses a large set of exercises. These exercises deal with Direct Currents and Electro-Magnetic Fields. They are based on Jupyter Notebooks in Anaconda, which uses Apps in which parameters can be changed using sliders. Each exercise has a short introduction explaining the theory behind the application. Although all applications are done on simple models, they give a direct insight in the influence of medium and measurement parameters on the resulting measured EM data.

Course News 4

Dec. 24^th 2018

In the courses the new approach of “Machine Learning” will be included. First in the form of lectures, followed later by exercises. Machine learning, also defined as Artificial Intelligence or Algorithms, relate to the use of statistical methods to derive from large (learning) data sets relationships (“correlations”) between observations and subsurface properties, for example lithological facies or fluid saturations.

This is a clear break with the traditional approach in which we ‘totally’ rely on understanding the physics of the subsurface processes (say visco-acoustic wave propagation) causing the observations or data (amplitude anomaly) at the surface.

In Machine Learning we mimic human “brain processes” to learn from a very large number of labelled cases how to explain a new case. In analogy with when you have seen more cases, had a longer learning trajectory, you are better able to evaluate a new case. You might not fully understand why events happen, but you have learnt how to deal with them, i.e. you can classify them say as beneficial or threatening.

Course News 3

Aug. 17^th 2018

In the “Advanced Seismic Data Acquisition and Processing” Course two new topics have been introduced: 1) Blended acquisition and 2) Penta-source acquisition.

In blended acquisition sources are fired closely in time such that the records do overlap. To separate the responses of the subsurface two methods can be applied.
One method is called “dithering” in which a “random” variation is applied to the auxiliary sources with the result that their responses are non-coherent and can be considered as random noise and be removed/suppressed in processing. As the “dithering” for each auxiliary source is known, it can be removed resulting in coherent records for each, leaving the other source records incoherent.
The other method uses “seismic apparition”. In this method a periodic variation, be it time delay or amplitude scaling, is applied to each source and each source response can be made to “appear” at the expense of the responses of the other sources this is often done in the KF domain.

The Penta-source (Polarcus), consists of 6 arrays (spaced 12.5 m apart), which are used in more than one source configuration resulting in 5 single sources. For operational reasons they are fired in the following sequence: (1,2), (3,4), (5,6), (2,3), (4,5). They can be used with a shot-point interval of 12.5 m and “dithered” over 1 s, or with a shot interval of 62.5 m and “dithered” over 5 ms, allowing deblending of the shot records. Note that the first part (1 or 5 s) of the record is “clean” and do not require deblending.

In both cases also the streamer spacing might vary, where an increasing separation from centre spread outwards is beneficial for a high-resolution shallow image.

Course News 2

In addition to the “Effective Media and Anisotropy” an extensive AVA modelling exercise has been implemented. It allows investigating the various aspects of AVA analysis, namely what rock and fluid properties can be derived from observing the Amplitude changes with Angle of incidence on the interface between two rocks. It not only deals with density, P- and S-wave velocities, but also with anisotropy. From the anisotropy, not only average properties of sub-seismic-scale fine layering, but maybe even more important fracture orientation and density can be derived. These are important in relation to where to drill production wells relative to injection wells (think of fractured carbonates).

In the exercise the AVA response of different rock interfaces (shale, sand, salt, limestone, etc) can be modelled using Vertical Transvers Isotropy (VTI) for fine layering and Horizontal Transvers Isotropy (HTI) related to fractures. Although more than one fracture set do occur in nature, the exercise is limited to one fracture set, based on the idea that most often it is the present stress regime that keeps fractures open in the minimum horizontal stress direction.

Course News 1

For the “Quantitative Reservoir Characterisation” (QRC) course a new module has been designed concerning Effective Media and Anisotropy. The Earth is often too complicated in terms of inhomogeneities to fully honour it in modelling wave propagation and inversion. The solution has been to replace the real earth by a so-called effective medium. This effective medium allows an “appropriate approximation” of the wave propagation and inversion. Although an approximation, it is still appropriate for solving the issue at hand.

One main application is related to the concept of anisotropy. Anisotropy occurs when, for example, we deal with thin sequences of different rocks or in case of fractures. If these “inhomogeneities” are on a sub-seismic wave-length scale the inhomogeneous medium can be replaced by a homogeneous medium with “anisotropic" character. The anisotropy is then an appropriate replacement of the sub-seismic scale inhomogeneities and allows derivation of the rock properties using seismic observations. These anisotropic properties tell us that the wave propagation is not only a function of location (the commonly considered large scale inhomogeneity), but also at each location a function of the direction of propagation.

In the exercise, you will learn how effective media velocities are calculated. There are namely two ways and how in modelling these effective anisotropy parameters can be calculated. You also will derive anisotropy directly from seismic data.

Please feel free to comment or add points of view and indicate whether you would like it to be mentioned on the website or not
(email: j.c.mondt@planet.nl).

Back to main page

© Breakaway