The ‘birds of a feather session’ was billed as ‘HPC and ML aspects of the energy transition’. It turned out to be much more of a commercial for the participants’ companies. All are going green and it could be that that the CO2 generated by the HPC devoted to modeling carbon capture exceeds that which is actually sequestered.
BP’s head of HPC, Elizabeth L’Heureux, described BP’s net zero ambition and mutation from oil to energy. BP is working to transition its subsurface expertise to new energy and to develop new skills to differentiate its new energy activities. Uses cases are (will be?) simulation for CCS, reservoir modeling, simulating wind farms, solar panels, batteries and AI for distribution. Today HPC in BP is 85% geophysics and this increasing, with an anticipated 35 petaflops in 2024. As new businesses come in new people are looking for help with HPC expertise. BP’s HPC unit has a tentative agreement to grow its on prem systems but is also exploring cloud and third parties to offload excess demand.
Weichang Li, presented machine learning work underway at Aramco’s Research Center in Houston. Here some 70 researchers work primarily in the upstream on flow measurement, the ‘Sensor Ball’, downhole robots and ML with distributed sensing for CCS. Other activity includes deep learning based completion monitoring with DAS/DTS and core image analysis.
Shell’s Detlef Hohl, came clean ‘I am hardly going to talk about HPC at all’, instead, ‘what Shell is doing in the energy transition’. This involves very ambitious climate goals to ‘avoid, reduce, and mitigate’ CO2 inter alia by ‘digital twin asset optimization’. CCS is important to Shell for example Quest and Gorgon*. Already Shell produces more energy in gas than oil. Shell has some 350 math/data scientists and another 800 ‘citizen data scientists’. Already, 10,000 equipment items are ‘monitored by AI’.
* Although the Chevron-operated Gorgon is struggling as reported in UpstreamOnline.
Following a masked Mauricio Araya (TotalEnergies) was tricky but we did glean that Total plans for less oil liquids down (to 30% by 2030) and more gas (up to 50%). A new OneTech organization has been in place since September 2021. Total does less lab work and more machine learning. Going green includes CCS at Northern Lights, Aramis, NEP, and afforestation of the Batéké plateau. Pymgrid, a Python library is used to generate and simulate a large number of electrical microgrids.
In the Q&A the panel was asked how they (as the environmental ‘bad guys’) were doing recruiting all the new talent needed for data science. BP has indeed struggled to find HPC/CFD specialists especially those with domain skills and is trying to develop people internally. Shell observed that AWS (the questioner’s affiliation) was not having an easy time hiring talent either! Shell is teaching data science. It is really hard to attract talent to oils in the EU although it may be easier in the US. ‘We bring people in, show them what we do. We also show them Amazon Science. It’s a good portal, we don’t have it!’
Shell’s Janaki Vamaraju presented FResNet++. Conventional modeling of complex phenomenon for hydrocarbons, carbon capture and other fluid flow applications is computationally expensive and can take days to run. AI is solving the hardest challenges in scientific simulation and speeds up modeling 1000x. Physics-informed neural networks and surrogate models are used to accelerate 50 or 100 years of CCS simulation. Neural nets working in the Fourier space is not new but the need for high resolution (10 million to a billion cell models) will stretch HPC resources. Enter the FResNet++ model*, a ‘deep residual learning framework’ for multiphase, multi component flow simulations in heterogeneous media.
*This reference comes from the ‘papers with code’ website although curiously, for FResNet++, there is no code!
Visit the conference home page and watch the conference videos.
© Oil IT Journal - all rights reserved.