Conclusion
This thesis contributes machine learning applications in geoscience with a focus on field data applications in 4D seismic, bsem, and asi. Additionally, the introduction contains a published review of the history of machine learning in geoscience with insights into the recent interest around the topic.
The book chapter in 11.2 discusses the historic development of machine learning in geoscience. It highlights key papers and developments through the decades, relating the developments to larger developments in the field of ai and machine learning. In the book key algorithms are detailed including svm, rf, gp and the development from kriging, as well as, key neural network developments and dl architectures that enable modern applications throughout many scientific disciplines including geoscience as a whole.
The exploration of bsem data in 12 introduced a novel unsupervised method to extract chalk grain boundaries from image data and shows the improvement of subsequent morphological filtering (Jesper Sören Dramsch, Amour, and Lüthje 2018). These methods reduce labour-intensive manual tasks, introducing varying degrees of automation in geoscience workflows. Following the extraction of the boundaries in the bsem images, computational granulometry can be performed. This includes statistics about grain size and circularity of the grains and the orientation of grains. Commonly this data had to be obtained by manual measurement of every grain. The unsupervised nature of this application means that no training data is necessary; in turn, it can be used to obtain high-quality training data for subsequent supervised machine learning tasks.
The research in 13 showed that transfer learning could alleviate the necessity for large amounts of labelled data, by re-using a neural network trained on natural images. This study showed that neural networks can be transferred to seismic data and outperform smaller networks trained from scratch. The smaller network size was necessary to avoid overfitting. The source code for this research was made available and has been of use to multiple researchers (Jesper Sören Dramsch 2018h). This has broad applications in industry and research settings performing asi. The limited availability of labelled data and wide availability of pre-trained network architectures makes this a viable option to obtain improved results and more robust models. Moreover, this insight is applicable to pre-training geoscientific neural networks.
Jesper Sören Dramsch, Lüthje, and Christensen (2019) shows that
explicitly using phase information as input in a complex-valued neural
network can stabilize the reconstruction of compressed seismic data. The
smaller complex-valued network in 14 outperforms
larger real-valued networks; however, a very large real-valued network
that does not compress the seismic data can implicitly learn partial
phase information. The paper touches on deficits of current metrics
applied to geoscience and exposes a periodic dimming effect of
frequencies from neural networks that should be further investigated,
particularly in the context of aliasing. This paper led to the creation
of the open-source software package keras complex
to enable
complex-valued deep learning in tf
(Manual in
22.5). Considering the modularity of neural
networks, this insight can be transferred to other deep learning tasks
on physical data like seismic data. Additionally, this research could
lead to further investigation of including known physical information in
neural networks not limited to explicitly using the phase information as
input.
15 introduces a novel method to perform pressure-saturation inversion on amplitude difference maps (Jesper Sören Dramsch, Corte, et al. 2019a). This work incorporates basic physical relationships directly as features into the neural network architecture, which was shown to stabilize the training result. Moreover, this work shows the possibility of training dnns on simulation data and subsequently transferring the network to field data. This particularly was enabled by applying Gaussian noise within the network. The dnn results were compared to results from the Bayesian inversion showing a promising application of dnns in 4D qi (Jesper Sören Dramsch, Corte, et al. 2019a). While this work has attracted interest in a sponsors meeting and the workshop presentations (Jesper Sören Dramsch, Corte, et al. 2019a, 2019d), further investigation into model explainability and lower complexity baseline models is necessary (Côrte et al. 2020; G. Corte et al. 2020).
In 16 a novel method for time-shift extraction is presented. This method combines recent advancements in diffeomorphic mapping, dl and unsupervised learning to introduce a 3D time shift extraction method including uncertainty values, where 1D extraction is the standard (Jesper Sören Dramsch, Christensen, et al. 2019). The method is shown to work on 3D seismic post-stack data with strongly differing acquisition parameters, without supplying any time shift information. After applying the method, the 3D seismic volumes are well aligned, with the diffeomorphic constraint performing well on seismic data. This work tests the trained network on two other 3D seismic volume pairs to test the generalization of the convolutional neural network after training. The two test sets show that the trained model on a single 3D seismic volume pair transfers well to the same field with different acquisition parameters and even a different field with a vastly different geological setting.
Overall, this thesis shows dl applications in seismic geophysics and resulted in multiple workshop, conference, journal papers, and a book chapter, including reproducible Python code for all publications. The publications, developed through interdepartmental and international collaboration, have been disseminated at international workshops and conferences. Two novel methods for 4D seismic analysis were introduced and compared to conventional methods. Moreover, transfer learning as a viable application in asi was shown and has found wide application. The Python code in this thesis has been open-sourced for all published papers for reproducibility including the open-source package "keras complex".