Founder and CEO Jensen Huang, wearing his signature leather jacket, gave the keynote to the 2022 NVIDIA Global Technology Conference stating that with the advances in AI, ‘data centers are becoming AI factories’. There are now ‘three Nvidias*’, AI, HPC and the ‘Omniverse’ a platform of algorithms, compute power and ‘science’ that is to power the next wave of development. The Omniverse poster child is ‘Earth 2’, ‘the world’s first digital twin’. Earth 2 is to perform ‘physics-informed deep learning’ for weather forecasting building on 40 plus years of EU weather data. Thousands of simulations on the huge training sets ‘obviated the need for supervision’ and the conditions are primed for other scientific breakthroughs. Huang proceeded to describe a bewildering technology offering with some staggering numbers. An new H100 chip sports some 80 billion transistors with a 5 terabyte/s IO bandwidth. ‘20 of these could sustain all the world’s internet traffic’. Chips are combined in DGX Pods to provide exaflops of AI compute bandwidth. Multiple pods connect with InfiniBand switches. EOS is being built with 18 DGX Pods and will provide 18 exaflops, ‘a 4x leap over the current world N°1 supercomputer’.
* Huang omitted to note Nvidia’s fourth revenue stream, crypto mining. The SEC recently fined the company $5.5 million for its failure to disclose that crypto mining ‘was a significant contributor to its 2018 revenue’.
Physics-informed neural networks were evidenced in a presentation, ‘Developing digital twins for energy applications using Modulus’ from Tarak Nandi and Oliver Hennigh (NETL). NNs are trained to ‘minimize the residual form of the same physical equations as in CFD*’ and are claimed to be orders of magnitude faster than CFD for uncertainty quantification/design optimization problems. Nvidia Modulus is a neural network framework that blends physics, in the form of partial differential equations, with data to build parameterized surrogate models. NETL applications address various industrial/chemical process models. NETL twins are being developed for CCUS, direct air CO2 capture and flue gas capture.
* Computational fluid dynamics
In the oil and gas track, Muhong Zhou showed how BP has extended a conditional generative adversarial network (CGAN) from 2D to 3D, targeting execution on a multi-GPU node. The CGAN was developed in-house for seismic attenuation compensation. Data augmentation was developed with the CuPy, open-source array library for GPU-accelerated computing with Python. Tests show the CGAN code can generate images ‘with similar quality as those generated by the existing in-house physics-based attenuation compensation tools’.
Erick Nascimento, Ilan Figueirêdo and Lílian Guarieiro from Brazil’s Senai Cimatec research establishment showed how deep learning can support predictive maintenance activities. Labeled data with annotated failure modes and time-to-failure train deep learning models. But labeled data may be scarce or even nonexistent. The challenge is to combine unsupervised and supervised ML to identify faults in time series data. The team used a public dataset* of an oil and gas offshore platform, including different failure and pre-failure behaviors captured by sensors. An Nvidia GPU Tesla V100-SXM3 was used for the deep learning semi-supervised model. A combination of C-AMDATS** and a deep neural networks was able to detect anomalous patterns The authors concluded that ‘predictive maintenance solutions can be further developed, even in scenarios with few, or even no, labeled data’.
*The 3W Dataset
is said to be a realistic oil well dataset with rare undesirable real
events that can be used as a benchmark dataset for development of
machine learning techniques. The theory behind the approach is given in
the Journal of Petroleum Science and Engineering paper, ‘A realistic and public dataset with rare undesirable real events in oil wells’.
** A Cluster-based Algorithm for Anomaly Detection in Time Series.
BP Launchpad company Fotech reported that the arrival of Nvidia’s Jetson AGX Xavier was a game changer for its proprietary Dataflow processing that allowed for ‘fully embedded’ real time processing. Fotech is a DAS fiber optic specialist whose technology is used to monitor intrusion and damage along pipelines, power lines or fiber cables themselves. DAS generates a ‘fire hose’ of data that needs to be processed in real time. Fotech’s architecture prepares and manages processing and disturbance detection in near real time using Nvidia GPUs. Edge AI can ‘turn a fiber optic cable into 60,000 vibration sensors’.
Tore Kvam and Serhiy Todchuk from Norwegian Visco described their high fidelity renderings as ‘a new digital twin concept’ of oil and gas installations spanning surface, sub-sea and subsurface. Animating these huge (700 million triangles) models requires many software smarts. Sharing them between multiple users is enabled with an Nvidia Tesla-based vGPU server that allocates virtual PCs to any device via RDP. Visco is now looking into the Nvidia DLSS (deep learning super sampling) for efficient resolution scaling, the use of mesh shaders for geometry pipeline and memory efficiency and to ‘move everything to the GPU!’
© Oil IT Journal - all rights reserved.