Skip to the content

Quantitative understanding of the subsurface - critical to minimising drilling risk and maximising subsurface recovery

S-Cube was spun out from the Earth Science & Engineering Department of Imperial College London in 2015 to advance data-driven automated earth parameter learning from raw exploration seismic data. It was formed on the back of breakthroughs by Imperial College researchers made in the algorithm of Full Waveform Inversion (FWI). This is a form of extreme machine learning for seeing into the earth using seismic field recordings and involving compute intensive wavefield simulations. S-Cube through XWITM has taken the powerful technique to the next level through algorithmic developments and a cloud-native architecture which can consume data directly from the seismic survey and turn them into actionable subsurface volumes.

Background

The first major application of time-domain waveform inversion incorporating anisotropic wave propagation was conducted by founding members of S-Cube in this award-winning study. This was at a site in the North Sea with a known shallow gas accumulation in the clastic section situated above a chalk reservoir. The waveform inversion successfully defines the structure using a data feed from a ocean-bottom seismic survey as shown below. The publication won best Geophysics Paper for 2013 from SEG.

Ocean-bottom survey receiver cables are positioned on the seafloor with 1440 individual hydrophone locations marked as dots shown on the horizontal slices below. The left column is the input model grid with a manually inserted low blob to approximate the gas accumulation whilst the right column shows the iterated model grid converting the low velocity feature into a defined structure.

Also displayed are the simulated and recorded traces from a shot firing into a cable of receivers over the anomaly. The details of the gas charged interbedded layers and faults get captured as low velocity features in the model grid as the data fitting feedback loop allow the simulations to increase their fit with the recordings.

Aerial view through input model at depth 1.2km

Aerial view through input model at depth 1.2km

Aerial view through iterated model at depth 1.2km

Aerial view through iterated model at depth 1.2km

Vertical slice through input model

Vertical slice through input model

Vertical slice through iterated model

Vertical slice through iterated model

What happens if we eliminate the low velocity zone from the starting model completely? The gas cloud is detected but the model iterates to a local minimum trap from which it cannot recover. This is due to a phenomenon known as cycle-skipping where the predicted data shifts in a direction which causes it to misalign with the observed data. 

Vertical view through model without any evidence of low-velocity zone

Vertical view through model without any evidence of low-velocity zone

Vertical slice through intermediate model without AWI

Vertical slice through intermediate model without AWI

Vertical slice through final model without AWI

Vertical slice through final model without AWI

Aerial view through model without any evidence of low-velocity zone

Aerial view through model without any evidence of low-velocity zone

Aerial view through intermediate model without AWI

Aerial view through intermediate model without AWI

Aerial view through final model without AWI

Aerial view through final model without AWI

As a result of cycle skipped mismatch, the wider applicability of the generic technique is limited given reliance on a-priori knowledge of the target, i.e. an accurate starting model. S-Cube was able to specifically tackle this through restructuring of the data-fitting feedback loop optimisation cost function, making it sufficiently convex to be able to start further from the true answer without misconvergence.

Adaptive Waveform Inversion cost function - the breakthrough methodology invented and patented by S-Cube

S-Cube formulates Adaptive Waveform Inversion (AWIusing convolutional filters between the predicted and recorded data and minimising the misfit to send the non zero lag filter coefficients to zero. This solves a long-standing problem in waveform inversion theory - succumbing to false local minima due to oscillatory seismic data.

Repeating the previous example with the adaptive cost function and changing nothing else gives the result below. It enables successful recovery of the gas cloud from a cycle-skipped initial position at which point conventional waveform inversion with a point-to-point difference norm cost function can be applied to return to the previously obtained best fit result.

Vertical view through model without any evidence of low-velocity zone

Vertical view through model without any evidence of low-velocity zone

Vertical view through AWI intermediate model (1)

Vertical view through AWI intermediate model (1)

Vertical view through AWI intermediate model (2)

Vertical view through AWI intermediate model (2)

Vertical view through AWI intermediate model (3)

Vertical view through AWI intermediate model (3)

Vertical view through AWI intermediate model (4)

Vertical view through AWI intermediate model (4)

Aerial view through model without any evidence of low-velocity zone

Aerial view through model without any evidence of low-velocity zone

Aerial view through AWI intermediate model (1)

Aerial view through AWI intermediate model (1)

Aerial view through AWI intermediate model (2)

Aerial view through AWI intermediate model (2)

Aerial view through AWI intermediate model (3)

Aerial view through AWI intermediate model (3)

Aerial view through AWI intermediate model (4)

Aerial view through AWI intermediate model (4)

Vertical view through final model with AWI

Vertical view through final model with AWI

Vertical view through tomography model

Vertical view through tomography model

Aerial view through final model with AWI

Aerial view through final model with AWI

Aerial view through tomography model

Aerial view through tomography model

Enables the global minimum solution to be reached using iterative local optimisation 

AWI was patented and became the recipient of the Bonarelli Award for 2015 by the EAGE, the leading European professional organisation in petroleum geoscience for the long-standing problem in localised optimisation it was able to provide a solution for.

Following intensive algorithmic development, AWI and its family of implementations improved the steepest-descent search directions and ultimately the inversion convergence in numerous examples where conventional FWI was an inadequate technique. The technology is now licensed and in mainstream use within EquinorWoodside Energy and most recently in 2019 within Chevron.

For example, the application of AWI to a narrow-azimuth towed-steamer dataset offshore Myanmar led to a pre-drill prediction of a gas bearing formation. The low velocity interval prediction is confirmed by a subsequent well penetration of the identified zone. 

Vertical slice through featureless starting model

Vertical slice through featureless starting model

Vertical slice through AWI model - pre-drill predictive power demonstrated

Vertical slice through AWI model - pre-drill predictive power demonstrated

Cloud-native S-Cube XWITM launches on Amazon Web Services 

We ran our first major cloud project here at the end of 2018. Following that new users can now access the industry’s leading FWI algorithm on demand in a managed VPC environment hosted on the cloud with the benefit of multiple cases running concurrently for tuning purposes to enhance convergence to the basin of the misfit global minimum. S-Cube Cloud consists of the following components:

  • underlying XWITM optimisation toolbox and FWI engine
  • cloud-native MPI-free parallelisation layer
  • real-time monitoring web-based dashboard 

The technology featured at numerous events in 2019 on S-Cube and AWS platforms.

EAGE 2019 - London

S-Cube exhibition booth

ISC 2019 - Frankfurt

AWS exhibition booth

S-Cube Cloud

Run Full Waveform Inversion on the Cloud

Use XWITM on AWS to discover an unprecedented increase in the resolution of your velocity model.