2021 Nvidia Global Technology Conference

Chevron: Cloud-native 3D volume visualization with Nvidia Index. Abyss trains flare stack corrosion model on GPUs. C3.ai’s deep learning/graph search parses engineering documents. PCPC Direct serves Petrel from Lenovo RVIZ/Leostream VDI. Halliburton Decision Space 365 on T4 Tensor Core in Azure cloud. Fu2re Smart Solutions ‘Pangea’ DMS for Petrobras. Quantiphi’s TensorFlow estimator for Chevron’s seismic imaging. Beyond Limits’ ‘soft actor critic’ deep learning for field planning. Russia’s Data Analysis Center: ‘Open AI/CUDA beats human interpreters’.

Kadri Umay (Microsoft) and Josephina Schembre-McCabe (Chevron) demonstrated the use of Nvidia’s ‘Index’ 3D volumetric visualization framework on large-scale ‘super-resolution’ digital cores. Today, the bottleneck in oil and gas data visualization is software, as supplied by multiple vendors, in different formats and access paradigms. This makes it hard to develop custom, cross-vendor views of the data*. There is a need for central (cloud) data accessible and viewable by all. Umay described the shift from early steps in the journey to the cloud that involved a ‘lift and shift’ of storage, architecture and apps. The future will bring ‘cloud native’ visualization of data served from object storage in the cloud and in-cloud GPU computing ‘leveraging industry standard data and APIs’. Today, the cloud architecture is somewhat complex involving Nvidia Index, Azure Kubernetes services, blob storage, the Helm package manager, Azure pipelines, Docker, RBAC. An ‘Azure blob viewer’ launches Nvidia Index, allowing visualization of digital rock image in the cloud.

Comment: Whether this is up to tools from Thermo Fisher Scientific or really achieves cross vendor visualization would appear moot. The approach might even sacrifice app functionality and add considerable IT complexity.

* This problem was identified some 20 years ago and resulted in the development of Dynamic Graphics’ CoViz!

Zarif Aziz from Abyss Solutions showed how offshore asset inspections are performed using drones, with data post-processed with convolutional neural networks. Abyss has been using drones to perform photographic studies of offshore flare stacks. Drones are said to offer more complete coverage of complex structures than is possible with human inspection. Image processing (semantic segmentations with CNNs) was trained with manually collected corrosion data and classified as low, medium and high risk. Abyss claims a ‘best in industry’ corrosion model, trained on thousands of high quality labels from different offshore platforms. Machine learning was performed on Nvidia T4 Tensor Core GPUs running in the AWS cloud. Software included CuDNN and TensorFlow. The flare stack corrosion model represented a fine-tuning of Abyss’ topside corrosion model, incorporating some 1,000 high quality flare stack images. Corrosion identified by the model is displayed over the original imagery, resulting in ‘faster and more targeted’ remediation work. Abyss is now working to add drone-based Lidar scanner data into the model and, by repeat surveys, perform change detection of corrosion.

See also Abyss’ work for Anadarko on offshore corrosion identification.

Shouvik Mani and Michael Haddad showed how C3.ai showed how deep learning and graph search are used to parse and digitize engineering diagrams. Piping and Instrumentation (P&ID) diagrams are paper-based representations of plant components and their relationships. These are scanned and parsed with image recognition software to identify components and connections. Symbol detection leverages a CNN/rectified linear unit (ReLU) in a parsing pipeline. An EAST text detection network grabs textual information from the diagram. A LeNet-architected CNN was trained to classify components and tag numbers. The graph search is used to trace connections. The authors concluded that the diagram digitalization pipeline enables applications such as diagram search and equipment-to-sensor mapping, supporting the creation of a facility-wide digital twin.The methodology is a component of the Baker Hughes/C3.ai Reliability application.

Mike Walsh showed how PCPC Direct managed to serve high-end 3D graphics apps to remote workers from a private cloud during the pandemic. Walsh believes that most virtual desktop infrastructure (VDI) offerings fail to give a satisfactory user experience. All oil and gas 3D remote VDI solutions that PCPC has tested ‘do not solve common end user requirements when tested in a true remote manner, on a laptop, over an internet connection’. PCPC has developed a solution certification process to test remoting of Schlumberger’s Petrel, including metrics (FPS, latency, image quality …) and different user profiles. PCPC Direct’s solution is built on a Lenovo RVIZ VDI server. When suitably configured (Petrel requires 8GB GPU memory per user), remote VDI Performance was ‘indistinguishable from a workstation’. Leostream’s OpenStack VDI also ran.

Shashank Panchangam* presented Halliburton’s graphics-intensive geoscience applications, now deployed either on-premises or in the cloud via the Microsoft Azure Stack Hub (ASH). Halliburton/Landmark’s Decision Space 365 geoscience suite has been tested running on an Nvidia T4 Tensor Core on the ASH with applications running remotely on a virtual desktop interface. A speedup in seismic attribute computation (2.5 to 1.5 hours) was observed and a 70% decrease in deployment time on the Azure cloud infrastructure (Ansible, Terraform). Of note was the Azure Arc offering, a ‘single control plane’ spanning Azure, AWS and the Google Cloud, ‘bringing Azure services to any infrastructure’. More from Landmark.

* With Gaurang Chandrakant (Microsoft)

Andre Sih demonstrated’ the AI-powered ‘Pangea’ document management system that his company, Fu2re Smart Solutions, developed for Petrobras. ‘Pangea’ is an ‘intelligent’ document classification system that extracts parameters and images from documents and classifies unstructured data. The company also markets SmartVision.AI, a general purpose GUI and API for ML model development.

Vishal Vaddina* presented Quantiphi’s work for Chevron on ‘generative models for seismic image enhancement’. The work was performed with Quantiphi’s TensorFlow estimator API on the Google AI platform and a ‘super resolution’ general adversarial network (SRGAN). The approach has been published.

* With Aravind Subramanian (Quantiphi) and Chanchal Chatterjee (Google).

Ananthan Vidyasagar (Beyond Limits) presented on the application of a ‘soft actor critic’ method using GPU-accelerated deep reinforcement learning for field planning. The approach is described on the BL website as ‘Cognitive AI for Field Management’.

Anna Dubovik from Russia’s Data Analysis Center gave a wide-ranging presentation on geological interpretation with Open AI and CUDA tools. These are now claimed to ‘outperform human experts’. The DAC methodology is described in a paper by Dubovik’s co-author Roman Khudorozhkov on seismic horizon detection with neural networks.

Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.