Analyzing 3D Seismic Data using Intel Distribution of OpenVINO Toolkit

[MaryT_Intel]

Use of Deep Learning (DL) in exploration workflows is rapidly attracting the interest of many geophysicists in the oil and gas industry. The fast, automatic interpretation of seismic data can be used to accelerate the process of finding oil and gas. Recently, there have been multiple studies published^1,2,3 that demonstrate how convolutional neural networks (CNNs) can be used to identify sub-surface faults or deep sub-sea salt bodies. One interesting development is that CNN models that have been trained on synthetic seismic datasets are producing acceptable accuracy in identifying faults using real datasets^1,2. This has huge potential to further accelerate oil and gas exploration since scientists do not need to train models on newly acquired seismic datasets in order to automatically interpret them. 2^nd Generation Intel® Xeon® Scalable Processors and the Intel® Distribution of OpenVINO™ Toolkit enable fast inference on 3D seismic datasets from a pretrained model. To show this, we used data from the F3 Dutch block in the North Sea to perform identification of salt bodies — an important step in locating oil and gas deposits.

The seismic deep learning workflow optimized with the Intel Distribution of OpenVINO toolkit is shown below. This workflow utilizes full precision (32 bit) of seismic data and performs accelerated AI inference for seismic datasets on 2^nd gen Intel Xeon Scalable CPUs.

The model used in this exercise was PyTorch-based with 5 layers of 3D convolutions developed by the University of Oslo³. The model training was performed using a single slice of annotated salt body. The Intel Distribution of OpenVINO toolkit was used to increase the inference performance of a pre-trained model. The toolkit enables significant improvements in AI inference performance on Intel CPUs running high-memory capacity DL models without any significant loss in accuracy. The industry-leading memory capacity of 2^nd Gen Intel Xeon Scalable processors make them well-suited to performing deep learning work on voluminous datasets such as 3D seismic. The Intel Distribution of OpenVINO toolkit allows HPC users to preserve their existing HPC data flow while providing a significant improvement in deep learning for seamless integration of HPC and AI in the hybrid and -cloud compute space. Following the research on adverse effect of tiling performed on MRI data and satellite data,⁴ it would be interesting to test INT8 convolutions for seismic data that utilize INT8 data type acceleration powered by Intel® Deep Learning Boost (Intel® DL Boost) while performing inference on larger tile sizes to boost accuracy in automatic seismic interpretation. Available in 2^nd generation Intel® Xeon® Scalable processors and based on Intel® Advanced Vector Extensions 512 (Intel® AVX-512), Intel DL Boost’s Vector Neural Network Instructions (VNNI) deliver a significant performance improvement for inference workloads. In seismic interpretation, geographical context is important and larger tiles will likely provide better definition in interpretation. 

References:

1.       Xinming Wu, Luming Liang, Yunzhi Shi, and Sergey Fomel, (2019), "FaultSeg3D: Using
         synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault
         segmentation," GEOPHYSICS 84: IM35-IM45.

2.       York Zheng, Qie Zhang, Anar Yusifov, and Yunzhi Shi, (2019), "Applications of supervised deep
          learning for seismic interpretation and inversion," The Leading Edge 38: 526–533.

3.       Anders U. Waldeland, Are Charles Jensen, Leiv-J. Gelius, and Anne H. Schistad Solberg, (2018),
          "Convolutional neural networks for automated seismic interpretation," The Leading Edge
          37: 529–53

4.       Reina, G. A., Panchumarthy, R., Thakur, S. P., Bastidas, A., & Bakas, S. (2020). Systematic
          Evaluation of Image Tiling Adverse Effects on Deep Learning Semantic Segmentation. Frontiers
          in neuroscience, 14, 65. https://doi.org/10.3389/fnins.2020.00065 ;

Backup: Configuration Details
Test by	Intel
Test date	08/06/2019
Platform	S2600WFT
# Nodes	1
# Sockets	2
CPU	6252
Cores/socket, Threads/socket	24/48
ucode	0x5000021
HT	On
Turbo	On
BIOS version (including microcode verison: cat /proc/cpuinfo | grep microcode –m1)	SE5C620.86B.02.01.0008.031920191559
System DDR Mem Config: slots / cap / run-speed	12 slots / 32GB / 2666
Total Memory/Node (DDR+DCPMM)	384+0
Storage - boot	2TB NVME SSD
OS	Ubuntu 18.04.2 LTS
Kernel	4.15.0-55-generic
Mitigation variants (1,2,3,3a,4, L1TF) https://github.com/speed47/spectre-meltdown-checker	Mitigated

_{Notices and Disclaimers:}

_{FTC Optimization Notice}

_{Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors.}  

_{Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions.  Any change to any of those factors may cause the results to vary.  You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products.   For more complete information visit www.intel.com/benchmarks.}

_{Performance results are based on testing as of dates shown in configurations and may not reflect all publicly available ​updates.  See backup for configuration details.  No product or component can be absolutely secure.} 

_{Your costs and results may vary.} 

_{Intel technologies may require enabled hardware, software or service activation.}