Creating The Premier Energy Data Company

PGS and TGS have merged to form a pioneering data and services company in the energy sector, serving customers across the energy value chain and positioning itself as a strategic partner for energy companies. Explore our vision. 

Day Two Update | Some Technology Curiosities at EAGE 2020

As I watched several (pre-recorded) seismic imaging and FWI talks yesterday and for the remaining technical program, I touch on some FWI activity below.

wave imagewave image

The less said, the better, but yesterday was hard work trying to get the functionality working in the virtual booth platform. I truly hope the world can return to traditional face-to-face conferences soon because I really don’t like navigating virtual conference platforms. Technical sessions can only be viewed one at a time, you’re at the mercy of chairpersons struggling with home internet access, and there’s negligible interaction with colleagues. But the good news, like watching sport on TV, is that you can access replays if you weren’t paying attention. So all the conference content is there for you to watch at your leisure. 

Big Ambitions, Interesting Ideas, but Debate Needed

Yesterday in my Day One summary. I possibly gave an unreasonably conservative perspective of where the general industry implementation of FWI is. Whilst it is true that the industry is still on the uphill learning curve, there is a lot of exciting work going on that, in some test examples at least, shows FWI R&D is progressing well beyond acoustic marine applications. I should also issue the caveat that FWI is a highly divisive topic that seems to have polarized and opposing camps of opinion in every company I know.

There is correspondingly a couple of interesting FWI presentations in the EAGE 2020 program that I will mention. Both talks will receive mixed opinions, some violent, but here goes.

The first presentation by Mike Warner in the ‘Full Waveform Inversion 2’ session on Wednesday runs a short 2D towed streamer line of data through acoustic FWI (or a version thereof) to 100 Hz, and applies a vertical derivative to the velocity model to show how increasingly higher frequency FWI models can be enhanced in appearance by the application of the derivative filter. I confess I wasn’t impressed by Mike’s presentation at SEG in September, but this recycled version is delivered better at EAGE.

From a traditional perspective, the low wavenumber background model should be built free from artifacts, and then as FWI is run to increasingly high frequencies the earth reflectivity dominates the inversion.

Anyone familiar with the image processing of potential fields data knows that the application of spatial gradients is an excellent tool to enhance higher resolution components of map-based images, and the same principle applies to the enhancement of velocity models. From one perspective to reduce cost, you can run FWI to typical frequencies, and then apply a derivative to the model that is normal to local dips. This adds no additional information to the velocity model, but it may be an effective interpretation product. Alternatively, by running FWI to ‘high’ frequencies (> 60 Hz), for which the equivalent imaging flow would be to run RTM with a low-frequency background model, the FWI gradient does add extra detail that can be visually enhanced by the derivative filtering. There will undoubtedly be future debate regarding the legitimacy of such details.

Of course, ignoring shortcuts and optimization tricks, the computational cost of FWI increases as the fourth power of the highest frequency. Which is where a lot of claims about the theoretically ‘unlimited’ power of cloud computing comes in. It is true that the containerized orchestration of platforms for deploying, scaling, and management of massively parallel applications in the cloud has proven to be less challenging than initially feared (i.e. you can throw a lot of cloud compute at problems quite easily), but there is no escape from the fact that running FWI in a stable manner to ‘high’ frequencies, and free of spurious artifacts that are enhanced during higher frequency iterations, remains far from trivial. But then you get into the ‘trade secrets’ aspect of making FWI work (and the first of my three bullet points yesterday).

Anyway, if I return to the original FWI publications by the likes of Tarantola almost 40 years ago, the ultimate ambition of FWI was always to recover high-resolution reservoir properties (i.e. an accurate model of reservoir properties), for example, the bulk modulus and density, or the elastic Lamé parameters. So there’s a lot more to the FWI story that is yet to be told.

I also point to the presentation by Tariq Alkhalifah that kicks off the ‘Wave Field Modelling 1’ session on Friday. Unfortunately, or fortunately, Tariq and I have been friends for a long time now, but spend most of our communications disrespectfully baiting each other like an old married couple. But credit where credit is due. Tariq very nicely shows (in his words) how a ‘physics-informed neural network can operate like a function to predict the solution of the wave equation’ (in this case, the Helmholtz equation). He uses a simple toy example, but it’s an enticing indication that machine learning maybe can indeed completely replace large parts of FWI. His consortium at KAUST also has several other machine learning-related presentations at EAGE 2020. Cost and stability are clearly the challenges to any FWI implementation, and there is motivation to explore all options.

If you’re interested in this kind of development, Tariq and I are also teaming up with Sheng Wu from Equinor in a ‘Hot Topics’ public discussion session titled ‘FWI: Future Perspectives Without the Hype’ on Friday at 3-4 PM. Let’s make that lively and consider issues such as those above…

In the Meantime, Solutions That Work Well Today

From the PGS perspective, I hope you find time to view Elena Klochikhina’s presentation on Pre-Stack Data Domain Least-Squares Migration (LSM) in the ‘Imaging Theory 1’ session today (Wednesday). She presents an effective solution for producing high-resolution angle gathers in the sub-surface offset domain. Applications show improvements in resolution and corrections for angle-dependent illumination effects, and therefore provide more reliable amplitudes for AVA analysis. In other words, until FWI cheaply delivers elastic reservoir properties, this is the best-practice platform to pursue that journey today.

There are also two opportunities today to hear about the PGS ‘GeoStreamer X’ platform:

  • At 14:40 PM in the EAGE ‘BizTalk’ session you can watch ‘GeoStreamer X – What, Where, Why’, and
  • At 17:15 PM on the PGS virtual booth we have an ‘Ask an Expert’ live discussion titled ‘GeoStreamer X and Beyond: Flexible Towing Solutions’. 


Talk to you tomorrow…