Energy Industry frequently spends billions of dollars in acquiring seismic data for delineation of potential energy resources and that’s not the end of it. Post-survey comes the real challenge of processing the data worth 100s of GBs. This processing is equally expensive to that of acquiring the survey data, if not more. Certain processing steps can take several days to run and still result may vary depending upon the expertise of the person doing the job. With the advancement of Deep Neural Networks (DNNs), several state-of-the-art models have even surpassed human-level-performance. These processing steps can be made computationally as well as economically lot cheaper if they employe DNNs, given the huge amount of data available. We used Roosevelt Hot Springs Geothermal seismic dataset from Utah to identify major seismic horizons which is one of many applications of AI in the Energy Industry.
To approach the study we developed a deep neural network architecture, EarthAdaptNet (EAN), specially designed for semantically segmenting and classifying the seismic images with a minimal amount of training data. Building blocks of EarthAdaptNet can be broadly divided into Residual Blocks (RBs) and Transposed Residual Blocks (TRBs). In the proposed architecture, RB comprises two convolutional layers, each followed by batch normalization and a downsampling residual connection of a 1×1 convolutional layer. In view of U-Net (Ronneberger et al., 2015), this is what is referred to as the building block of the contracting path. TRB is similar in architecture as RB except with the use of a Transposed convolutional layer instead of a convolutional layer. Upsampling transposed residual connection with a 1×1 convolutional layer is used instead of downsampling residual connection. EarthAdaptNet uses an Encoder-Decoder architecture with RBs and TRBs. The encoder starts with a convolutional Layer and is followed by the RB, and the number of RBs used depends on the input size. The decoder starts with a TRB and the number of TRBs used is kept the same as the RB used in the Encoder. The Transposed Residual Layer is followed by a Transposed Convolutional Layer which outputs the segmented seismic image. A 1×1 convolutional layer also exists in the middle which acts as a bridge (Bottleneck) between the Encoder and the Decoder. Skip connection is present between each RBs and TRBs.
Using our novel architecture, EarthAdaptNet, we trained our model as discussed in our previous blog “AI in Geothermal Dataset“. We used a patch-based model and hence generated patches of size 40×40 with a stride of 10 from the seismic section.
We introduced various data augmentations like random noise addition, horizontal flipping, blurring, and random rotation up to 10 degrees. Data augmentation helps the DL model to significantly increase it’s performance.
With the patches generated and data augmentation applied to our original data, we’ve done all the pre-processing steps in building a DL model and have data in a NumPy array which can be used as an input to the model. We used a batch size of 32 for training our model. Cross-Entropy loss function was used as a loss function. With the help of adam optimizer, having a learning rate of 0.001, optimized our proposed network architecture, i.e., EarthAdaptNet.
With our novel architecture, EarthAdaptNet we achieved pixel accuracy of around 80% and Mean Class Accuracy of about 75%. We finally got an accuracy of about 76%, 54%, 79%, 91%, 45%, 70%, 88%, and 97% for class 0, 1, 2, 3, 4, 5, 6, and 7 respectively.
So in conclusion EarthAdaptNet can be applied to Seismic Images for speedy processing and interpretation. The decoder of EarthAdaptNet can be replaced with fully connected layers to build a classification model for several tasks as well (Stay tuned for our paper which will introduce novel Domain Adaptation methodology in seismic images).
We have an upcoming conference (Based on EarthAdaptNet introducing the novel methodology of Domain Adaptation in Energy Industry)(Stay tuned for our upcoming paper, code, and blog) in GeoConvention, 2020: Maiti, T., et al. (2020). In-depth analysis using state-of-the-art deep learning techniques for semantic classification of Seismic Facies In Geo Convention, Abstract.
We also have an invitation as a keynote speaker in Muscat, Oman. The Artificially Intelligent Earth Exploration: Teaching the Machine How to Characterize the Subsurface Workshop, 23–26 November 2020.
About the Author:
Dr. Tannistha Maiti is a Ph.D. in Geophysics and Seismology from the University of Calgary. She has worked on various topics that range from Earth Science to computation and mathematical modeling. Passionate about machine learning applications in the Energy and healthcare sector. She has extensively worked in mathematical modeling and computational Geophysics during her Ph.D. at U Calgary and MS studies at Virginia Tech. She also holds an undergraduate degree from the prestigious university at IIT Kharagpur, India. In her research career spanning for about 10 years, she has been involved in various projects and has more than 20 peer-reviewed publications in various conferences and journals. She also has extensive teaching experience and is passionate to share knowledge with peers and newcomers. During her Ph.D. studies, she received various accolades for her work. She participates in various research projects and supervises several deep learning internships within deepkapha.ai. In her spare time, she enjoys blogging both technical and non-technical.