2020 NVIDIA Global Technology Conference

Online event hears from Halliburton on adding machine learning to DecisionSpace 365 real time well engineering. at the edge. Equinor’s U-Saltnet and feature pyramid attention network for salt body interpretation. Baker Hughes/C3.ai ‘Reliability’, deep learning-based anomaly detection for oil and gas. Total on GPU-based compute acceleration in seismic applications.

Speaking at the online/virtual 2020 Nvidia Global Technology Conference, Vanessa Kemajou and Joseph Winston teamed to present Halliburton’s work on real-time machine learning and quantitative analytics ‘at the edge’. Applying deep learning models in near real time requires the right hardware and software. On the hardware front, the Nvidia Jetson Nano enables cost-effective deployment of field accelerators. For the software, Rapids (for ETL), GPyTorch (machine learning) and TensorRT (for optimization) have added data analytics to the DecisionSpace 365 Real-Time Well Engineering cloud application. The solution provides automatic rig state detection for non-productive time analysis, and friction factors calibration through reverse torque and drag analysis. Halliburton is now planning to add natural language generation to alert drillers to changing conditions. The authors referred to Gardner et al.’s 2018 paper on ‘GPyTorch: Blackbox matrix gaussian process inference with GPU acceleration’.

Equinor researcher Hongbo Zhou traced the history of AI from ‘GOFAI, ‘good old fashioned AI’ involving symbolic reasoning and logic that ‘did not work well’, through machine learning with decision trees, support vector machines and self-organizing maps (OK for some applications) to the ‘current revolution’ of deep learning with deep neural nets. Equinor’s U-Saltnet leverages Unet’s ‘semantic segmentation’, and a feature pyramid attention network to automate salt body interpretation. The Unet algorithm was tuned with Nsight’s profiling tools. Tensor cores allowed for automated mixed precision calculations. Trials on the SEG’s Gulf of Mexico SEAM dataset ‘seem to match people’s perceptions’ (of where the salt boundary is).

Henrik Ohlsson and Nikhil Krishnan, on behalf of the Baker Hughes C3.ai joint venture, presented BHC3 Reliability, a productized, deep learning-based anomaly detection framework for large-scale oil and gas applications. Reliability uses a ‘system-of-systems’ model to predict failures. More from the Reliability JV.

Total’s Lionel Boillot and Long Qu extolled the merits of GPU-based compute acceleration in seismic applications such as reverse time migration and full wave inversion. Pangea3 comprises 1116 IBM Power9 CPUs with 3348 Nvidia GPUs. The machine came in at N° 11 in the Nov 2019 TOP500 list. Software is developed in a collaboration between Total, IBM, Nvidia and Altimesh. Seismic workloads involve moving very large amounts of data between memory and disk. The IBM Power9 on-chip accelerator provides the ‘fastest on-chip gzip/gunzip compression engine in the industry’ and an ‘80 to 125x speedup’ over a CPU single thread. Total contrasts its legacy, CPU-based Pangea II (CPU) with the GPU-based Pangea III. Overall Pangea III provides an 8x speedup over its predecessor.

Read these and other presentations and videos on the GTC2020 website.

Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.