XWI^{TM} velocity model optimisation acts on the raw seismic field traces exactly as recorded in the survey. Hence the algorithm is equipped to run in real time as the survey proceeds. Only the following inputs are required:

- basic regional velocity trend
- bandpass filtered shot gathers to low (but no need for ultra low) frequencies
- extracted source wavelet

The requirement for an artificially accurate starting model or artificially low frequencies are eliminated by XWI^{TM}.

A water bottom interface and source wavelet are extracted from the data to ensure the model can accurately predict the direct arrival and water bottom events prior to minimsing the cost function with respect to velocity values.

This is demonstrated in the chart below for a deep water field example. Using a filtered spike source propagated through the velocity model these events are modelled and the simulated traces are compared to the equivalently filtered field traces. A matching filter operation between the two sets of traces is performed and the derived filter is convolved with the filtered spike to obtain a reliable estimate of the field low-frequency source.

At this point, the velocity optimisation feedback loop can commence. For the longest length-scale initial updates the lowest usable frequency in the data is sought, which for streamer acquisition is usually no lower than 4Hz.

The steepest-descent "gradient", which is the model update direction for locally minimising the cost function, can be computed for both adaptive and non adaptive cost functions so it can be demonstrated that the AWI gradient gives the correctly directed update for distant initial models needing macromodel updates.

The velocity model computation is run cycling through the shots in blocks of iterations at sequentially increasing frequencies incorporating progressively shorter length-scale resolution into the model. On a propagation mesh of 25 metres, the frequency range can be increased up to a limit 24Hz based on the required minimum number of grid points per wavelength stability condition.

## XWI^{TM} on the Cloud

The S-Cube Cloud environment hosted on Amazon Web Services (AWS) enables you to access XWI^{TM} 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

## Run Full Waveform Inversion on the Cloud

Use XWI^{TM} on AWS to discover a step change in accuracy and resolution of your velocity model.