Skip to the content

XWITM Revolutionary Earth Model Building on the Cloud

Waveform Inversion

Seismic data is used to perturb a section of the earth. Waveform inversion of seismic data is the search application for locating a loss function minimising subsurface rock property velocity model.

The computation iteratively adjusts earth parameters (training weights) to align numerical simulations with field recordings leveraging the full recorded response of the earth in a data fitting feedback loop which:

  • simulates the field experiment with a PDE per shot location propagating through an assumed set of training weights
  • computes the loss function misfit between generated predictions and field recordings
  • solves a constrained optimisation to find the 'gradient descent' direction which adjusts weights to reduce the misfit
  • updates weights towards loss function misfit minimiser

The XWITM optimisation framework

XWITM next generation waveform inversion is regarded as the industry's leading formulation due to:

  • combining adaptive cost functions (AWITM) with generic least-squares cost functions (FWI), for more accurate model updates starting further from the true answer
  • generating a temporary perturbation to match the shortest offsets and using the misfit at offset to update the macro model (RWI), for the deepest model updates with reflection energy 
  • incorporating total-variation based edge-preserving regularisation (CWI), for geological constraints to narrow the search space

It harnesses the power of data-driven parameter-learning which steers away from false local minimum traps as shown below. 

North West Shelf Australia use case showing global convergence maintaining accuracy down to below 3km applying XWI with an adaptive cost function.

North West Shelf Australia use case showing global convergence maintaining accuracy down to below 3km applying XWI with an adaptive cost function.

Same example showing spurious local minimum convergence going wayward at a depth of 2.2km applying waveform inversion without an adaptive cost function.

Same example showing spurious local minimum convergence going wayward at a depth of 2.2km applying waveform inversion without an adaptive cost function.

XWITM on the Cloud

The S-Cube Cloud environment hosted on Amazon Web Services (AWS) enables you to access XWITM on a PAYG basis in a managed VPC environment configured with cloud-native orchestration. It consists of the following components:                

  • parameter learning: underlying optimisation algorithm for iteratively updating training weights
  • parallelisation layer: master/worker distributed architecture with fault-tolerant MPI communications for leveraging burst-capacity across Amazon EC2 cost-effective spot instances
  • visualisation dashboard: web-based application for real-time result monitoring and hyper-parameter tuning for superior quality-scored convergence

It enables you to:

  • upload your raw field shot gathers to a dedicated AWS S3 bucket with a few commands
  • parameterise a new job and calculate the computational workload per shot
  • select the number of shots per iteration and number of iterations to run
  • monitor and QC the live job on the S-Cube Cloud dashboard accessed via a secure URL

Some Concrete Numbers

In this narrow azimuth towed streamer NW Australia deployment, every 5th shot (individual field experiment) is used per iteration and 2 shots are allocated per worker task.

  • 5186 total shots (9 saillines)
  • Grid size = 25m, nx = 113, ny =1041, nz=161  
  • 20% shots per iterations
  • 550 x 16 core EC2 instances
  • Average runtime per shot = 14 mins
  • Runtime per iteration = 28 mins

Cloud Native Architecture

With a cloud-native architecture at its core, XWITM is designed to scale to thousands of square kilometres with hundreds of thousands of shots affecting the earth parameter learning without computational constraint.

At each shot location, simulation and loss function minimisation occur with the computational load for the survey spread across compute instances working in parallel with a sufficient quantity allocated to allow for one for every two shots. As a result, a 100 iteration run which would take several weeks with a standard on-premises system can be completed in days on the cloud.

Furthermore with the virtually limitless compute available, the job can be run with multiple concurrent cases and multiple initialisations at once allowing for traveling through the search space to achieve optimal convergence at the minimiser of the cost function misfit.

S-Cube Cloud

Run XWI on the Cloud

Discover an unprecedented increase in the resolution of your velocity model.