|
<< Click to Display Table of Contents >> Navigation: Modeling > Full waveform inversion (distributed, 2D) |
Full Waveform Inversion (FWI) is a high-resolution velocity model building technique that updates an initial depth-domain velocity field by iteratively minimizing the difference between observed seismic data and synthetic data generated by acoustic finite-difference modeling. Unlike traveltime-based tomography, FWI uses the full waveform — amplitudes, phases, and traveltimes — to recover velocity structure at a resolution approaching the seismic wavelength.
This module implements the distributed variant of 2D acoustic FWI, designed to offload the computationally expensive gradient and misfit calculations to remote worker nodes. The inversion is performed in the slowness domain and uses a nonlinear conjugate gradient algorithm with an automatic line search to determine the optimal update step at each iteration. After every iteration, the updated velocity model is clipped to the user-specified minimum and maximum velocity bounds, and a misfit chart is generated so that convergence can be monitored visually.
The gradient at each iteration is computed using Reverse Time Migration (RTM): forward-modeled synthetic shot records are compared with preprocessed observed data, and the residuals are back-propagated through the velocity model. Source shots are optionally grouped into multisource super-shots to reduce the number of forward modeling runs per iteration. Each iteration model update is stored and can be inspected in the Intermediate Models viewer alongside the misfit-versus-iteration chart.
This module requires a distributed (multi-node) g-Platform environment. For single-machine FWI, use the standard Full wave inversion (2D) module instead. Before running, connect an Input seismograms preprocessing sub-flow to apply any required data conditioning (muting, filtering, normalization) to the observed shot records, and connect an RTM image postprocessing sub-flow to smooth or condition the RTM-based gradient image before each model update.
The SEG-Y file handle providing access to the observed shot-record seismic data. This must be connected together with the trace headers item. The data is read shot by shot during each FWI iteration. Ensure the sample rate and record length in the SEG-Y file are consistent with the modeling parameters set in the Modeling params group.
The trace geometry headers associated with the input SEG-Y data, providing source and receiver coordinates for each trace. These headers are used to organize traces into common-shot gathers and to position sources and receivers in the finite-difference modeling grid. Every trace must have valid source and receiver picket assignments — traces with missing pickets will cause an error at startup.
The starting velocity model in the depth domain (interval velocity as a function of lateral position and depth), used to initialize the FWI iterations. The quality of this initial model directly affects convergence speed and the risk of cycle skipping. A smooth, kinematically consistent model from tomography or depth migration velocity analysis is recommended. The model is re-interpolated to the grid spacing defined by the Model resolution parameter before processing begins.
The total number of FWI update cycles to perform. Default: 10. Each iteration involves one complete pass over all source groups: forward modeling, residual computation, RTM-based gradient calculation, line search, and model update. Monitor the Misfit chart after a test run to determine whether convergence is reached before the maximum iteration count; if the misfit is still decreasing at the end, increase this value. Typical FWI workflows require tens to hundreds of iterations depending on model complexity and starting model quality.
The lower bound applied to the updated velocity model after each iteration, in m/s. Default: 1500 m/s. Any model cell updated below this value is clipped to this bound. Set to the minimum physically plausible P-wave velocity in your study area (for example, the water velocity for marine data). This prevents the inversion from producing geologically unrealistic low-velocity artefacts.
The upper bound applied to the updated velocity model after each iteration, in m/s. Default: 3550 m/s. Any model cell updated above this value is clipped to this bound. Set to the maximum expected P-wave velocity in the target interval. Together with Min velocity, this constraint stabilizes the inversion and keeps the velocity model within a geologically meaningful range.
This group defines the source wavelet used to generate synthetic seismograms during forward modeling. The wavelet shape, dominant frequency, and length must closely match the effective wavelet of the observed data. A mismatch between the modeled and actual source wavelet will degrade the quality of the residuals and slow convergence.
The dominant (peak) frequency of the source wavelet, in Hz. Choose a value that matches the central frequency of the observed seismic data. For a Ricker wavelet, this is the peak of the amplitude spectrum. Using a frequency higher than what is present in the data may introduce artifacts; using a frequency that is too low will reduce resolution. A multi-scale FWI strategy typically starts with a low frequency and progressively increases this value across successive runs.
The total time window length of the source wavelet, in seconds. This should be set long enough to fully contain the chosen wavelet shape at the specified dominant frequency. For a Ricker wavelet at 20 Hz, a wavelet length of 0.2–0.3 s is typically sufficient. Setting this value too short will truncate the wavelet and introduce spectral distortion into the synthetic seismograms.
The mathematical shape of the source wavelet used in forward modeling. Available options include: Ricker1, Ricker2, AKB, Berlage, Gaussian, GaussianDeriv, MinPhase, Klauder, Ormsby, Spike, Zero, Unit. The Ricker1 (second derivative of a Gaussian) is the standard choice for acoustic FWI because it is zero-phase and band-limited. Select a different wavelet type if you have specific knowledge of the source signature characteristics, for example a minimum-phase wavelet for dynamite sources.
The time position of the wavelet peak (reference spike) within the output wavelet window, in seconds. Default: 1 s. This parameter controls where in the wavelet time window the central peak of the wavelet is placed. For zero-phase wavelets it should be set to roughly half the wavelet window length so that the wavelet is centered and both its causal and acausal lobes are fully represented.
The peak amplitude of the source wavelet. Default: 1. This scales the overall amplitude of the synthetic source. In most FWI workflows, the absolute amplitude scaling of the wavelet does not need to match the observed data precisely because the step size is determined adaptively; however, a reasonable scale helps numerical stability. Changing this value effectively rescales the gradient magnitude and interacts with the P param step size control.
This group controls the acoustic finite-difference modeling engine used both to generate synthetic data for comparison with observations and to back-propagate residuals for gradient computation. Parameters here determine grid resolution, boundary conditions, record length, and source encoding strategy.
The spatial grid spacing used for the finite-difference modeling, in meters. Default: 5 m. The same spacing is applied in both the horizontal and vertical directions, so the modeling grid is always square. The input velocity model is re-interpolated onto this grid before inversion begins. To avoid numerical dispersion artifacts in the finite-difference solution, this value should satisfy the criterion: grid spacing < minimum velocity / (10 x maximum frequency). Finer grids yield better accuracy but significantly increase computation time and memory requirements.
When enabled, the polarity of each source wavelet within a multisource super-shot group is assigned randomly (positive or negative with equal probability). Default: off. Random phase encoding is a standard technique in simultaneous-source (multisource) FWI to suppress the coherent cross-talk noise that arises when multiple sources are summed. Enable this option when the Multisource group size is greater than 1 to improve the signal-to-noise ratio of the gradient estimate.
The number of individual shot sources combined (encoded) into a single super-shot for each forward modeling run. Default: 1 (no super-shotting; each source is modeled individually). Increasing this value reduces the total number of forward modeling runs per iteration by a factor equal to the group size, which can dramatically reduce compute time. However, larger groups introduce cross-talk artifacts into the gradient unless random phase encoding is also used. A common starting value is 3–5 for a first test. Cannot exceed the total number of sources in the dataset.
The source index increment used when selecting the individual sources that make up a single super-shot group. Default: 1. With a step of 1, consecutive sources in the sorted source list are combined. Larger step values spread the selected sources farther apart spatially, which can help suppress cross-talk. For example, with a group size of 4 and a step of 10, each super-shot draws sources from positions 0, 10, 20, and 30 in the source list.
The source index offset between the first sources of consecutive super-shot groups. Default: 1. This controls how densely the groups tile the full source array. A value of 1 means each successive group starts from the very next source index, giving maximum overlap between groups. Larger values create sparser, non-overlapping groups, which may be appropriate when the Number of multisource groups is set to cover only a representative subset of the data.
The maximum number of super-shot groups processed per FWI iteration. Default: 1. Increase this value to use more of the source array per iteration. To use all available sources in every iteration, set this to the total number of shots divided by the group size (rounded up). Reducing the number of groups per iteration speeds up each cycle at the cost of using only a subset of the data for the gradient, which may introduce noise into the update direction.
The width of the absorbing boundary layer added around the edges of the finite-difference modeling grid, in meters. Default: 400 m. This boundary absorbs outgoing waves to prevent artificial reflections from the grid edges contaminating the synthetic seismograms. The thickness should be at least one dominant wavelength. For a 20 Hz wavelet in 2000 m/s material, one wavelength is 100 m, so the default of 400 m (four wavelengths) is generally adequate. Thicker boundaries improve absorption but increase the total grid size and computation time.
Controls whether the top edge of the modeling grid is treated as a free surface (zero-pressure boundary), allowing surface-related multiples to be modeled. Default: off (the top boundary is absorbing). Enable this option when the observed data has not been demultipled and surface-related multiples are present, so that the synthetic data also generates multiples and the residual is minimized correctly. For data that has been preprocessed with surface multiple attenuation, leave this option disabled.
The number of time samples in the synthetic seismograms generated during forward modeling. Default: 4001. The total modeled record length equals this value multiplied by the Output sample rate. Set this to match or slightly exceed the length of the observed data records to ensure that all arrivals of interest are included in the residual computation. Excessively long records increase computation time proportionally.
The time sample interval of the synthetic seismograms, in seconds. Default: 0.001 s (1 ms). This is also the time step used for the finite-difference propagation. The sample rate must satisfy the Courant stability condition for the chosen grid spacing and maximum velocity. As a rule of thumb: sample rate < Model resolution / (maximum velocity x sqrt(2)). If this condition is violated the finite-difference simulation will become unstable. Valid range: 0.0001 to 0.1 s.
This group controls the Reverse Time Migration imaging condition used to compute the velocity gradient from the forward-propagated source wavefield and the back-propagated data residual wavefield. These settings affect the quality and stability of the gradient image, which in turn influences the direction and reliability of each velocity model update.
The power of a time-based amplitude gain applied to the wavefields before forming the RTM imaging condition. Default: 0 (no time weighting). Setting this to a positive value applies a t^n gain that boosts later-arriving energy, compensating for geometric spreading and helping to balance the gradient amplitudes at different depths. Values between 1 and 2 are typical when time weighting is desired. Valid range: 0 to 10.
The power of a depth-based amplitude gain applied to the RTM gradient image. Default: 2. A depth weighting of z^n compensates for the illumination falloff with depth that is inherent in RTM, ensuring that deep model updates are not systematically suppressed relative to shallow ones. The default value of 2 provides moderate depth compensation. Increase for datasets with particularly poor deep illumination, or decrease to reduce over-boosting of deep noise. Valid range: 0 to 10.
Selects between two RTM imaging conditions for constructing the FWI gradient. Default: on. When enabled, the standard cross-correlation image condition is used: the gradient is formed by cross-correlating the forward source wavefield with the back-propagated residual wavefield. When disabled, a Laplacian-based image condition is applied to the gradient, which suppresses low-frequency artifacts (backscattering noise) commonly seen in standard cross-correlation images. If the gradient image shows strong low-frequency noise or bulls-eye patterns, try disabling this option.
Controls how the illumination normalization is applied when combining gradient contributions from multiple source groups. Default: on. When enabled, the gradient image and the illumination map are accumulated separately across all source groups before the final normalization (image / illumination) is applied once at the end. When disabled, each source group contribution is individually normalized by its own illumination before being added to the total gradient, which can prevent any single bright source from dominating the result. The cumulative mode is generally more stable and is recommended as the default.
A small stabilization constant added to the illumination denominator during RTM image condition normalization to prevent division by zero in poorly illuminated regions. Default: 1e-9. This parameter is only active when Make cumulative image is disabled (per-group normalization mode). In most cases the default value is appropriate. Increasing epsilon suppresses noise in low-illumination zones at the cost of reducing gradient amplitude there; decreasing it sharpens the normalization but risks amplifying noise where illumination is very weak. Valid range: 1e-20 to 0.01.
A dimensionless scaling factor that controls the initial trial step size in the automatic line search performed at each FWI iteration. Default: 0.05. The initial step size is computed as P x (RMS slowness / RMS gradient), which normalizes the update to be a fixed fraction of the current model. The line search then refines this step using parabolic interpolation. Increasing this value allows larger model updates per iteration, which can speed convergence but risks overshooting and instability. Decreasing it makes updates more conservative and stable. Valid range: 0 to 1. If the misfit does not decrease between iterations, try reducing this value.