FULL-WAVEFORM INVERSION Overview & application to field data Mike Warner Imperial College London
Topics Overview of full-waveform inversion Application to an OBC dataset Validation 2
Overview of FWI 3
Full-Waveform Inversion Method for generating high-resolution high-fidelity models of physical properties in the subsurface Seeks a model which can predict the entire recorded wavefield, wiggle-for-wiggle Has become practical for 3D field datasets within the last few years 4
RTM with PSDM model 3600 m depth 5
RTM with FWI model 3600 m depth 6
Full Waveform Inversion Most of what you know about conventional imaging will not apply to FWI workflows uses low frequencies uses transmitted arrivals iterative inversion from starting model details can be critical does not fail elegantly 7
Generic workflow 1. Conventional acquisition, processing, model building & depth imaging 2. Use FWI to improve velocity model in top ~ 2 km of heterogeneous overburden 3. Re-migrate using RTM with the shallow FWI velocity model 8
Heterogeneous overburden FWI recovers shallow heterogeneity for deeper depth migration 9
Conventional tomography uses travel times uses simplified physics fast, cheap well-established & robust low spatial resolution ~ Fresnel zone ~ λd 10
Full-waveform tomography uses the raw wavefield uses the complete physics computationally intensive evolving & not yet robust high spatial resolution ~ wavelength ~ λ /2 11
Acquisition Long offsets 3 to 6 times target depth necessary Low frequencies 2 to 3 Hz desirable Many azimuths desirable narrow azimuth possible 12
Method 1. Field data, starting model & source 2. Forward model predicted wavefield 3. Form residual wavefield at receivers 4. Back propagate residuals residual wavefield 5. Cross-correlate unscaled model update 6. Step length calculation scaled model update 7. Update model and iterate 13
Field example Tommeliten Warner et al (2013) Anisotropic 3D full-waveform inversion. Geophysics, 78, No 2, R59-R80. 14
Tommeliten N 15
Tommeliten 4-component ocean-bottom cable invert pressure data only Vp model above reservoir shallow gas low velocities high attenuation significant anisotropy 16
3D OBC field data acquisition geometry 4C OBC 3 swaths of 8 cables 75 m water depth 6 km cables 25 m receiver spacing 300 m cable spacing 6000 receivers 25 m shot interval 75 m shot-line spacing 100,000 shots full azimuth to 7000 m max offset 11,000 m 180 sq km 17
PP PSDM PZ-summed deghosted and demultipled 18
PP PSDM PZ-summed deghosted and demultipled 19
PP PSDM PZ-summed deghosted and demultipled 20
Raw shot record hydrophone only include all ghosts and multiples 21
Picking the starting frequency 0 5 10 Frequency (Hz) 15 20 10 Power (db) 30 50 70 raw data amplitude spectrum 22
Picking the starting frequency single-frequency phase 2.4Hz 3.0Hz 3.6Hz common receiver gather 23
Picking the starting frequency 0 5 10 Frequency (Hz) 15 20 10 Power (db) 30 50 start at 3 Hz 45 db 70 raw data amplitude spectrum 24
Raw shot record hydrophone only include all ghosts and multiples 25
Pre-processing Start from raw field data Do not mute early arrivals Do not remove direct arrival No deghosting No demultiple No debubble No low-cut filter No deconvolution No PZ sum No AGC No divergence correction This raw data can be difficult to obtain 26
Pre-processing Mute ahead of first breaks Mute Scholte waves Truncate to 5000 ms Cut frequencies above 8 Hz Delete three quarters of receivers Delete two thirds of sources Delete offsets < 100 m Delete geophones hydrophone only Apply source-receiver reciprocity Most of this is to reduce compute time, and to avoid adding noise into the inversion 27
Scholte waves at lowest frequencies hydrophone 28
Pre-processing Mute ahead of first breaks Mute Scholte waves Truncate to 5000 ms Cut frequencies above 8 Hz Delete three quarters of receivers Delete two thirds of sources Delete offsets < 100 m Delete geophones hydrophone only Apply source-receiver reciprocity Most of this is to reduce compute time, and to avoid adding noise into the inversion 29
Raw shot record hydrophone 30
Pre-processed for acoustic FWI hydrophone 31
Starting model reflection tomography 32
Starting model reflection tomography 33
Anisotropy VTI, maximum Epsilon = 20%, maximum Delta = 8% 34
Source wavelet Full bandwidth Contractor s wavelet vs Near-source OBH 35
Source wavelet Full bandwidth Low-pass filtered Contractor s wavelet vs Near-source OBH 36
Inversion parameters Time domain, acoustic 3D, VTI anisotropy Hydrophones only include ghosts and multiples Apply reciprocity 6000 1440 sources 80 sources per iteration Six frequency bands from 3.0 6.5 Hz 18 iterations per frequency Each source used once per frequency Amplitude equalisation Conjugate gradients Approximate diagonal Hessian 37
Starting model reflection tomography 38
FWI model reflection tomography 39
Starting model reflection tomography 40
FWI model reflection tomography 41
FWI results from homogeneous start model 250 m depth horizontal depth slice 42
Starting model 1200 m depth 43
FWI model 1200 m depth 44
Validation 45
FWI results from homogeneous start model 250 m depth horizontal depth slice 46
FWI results + original PSDM from homogeneous start model 250 m depth horizontal depth slice 47
Starting model 1200 m depth well log 48
FWI model 1200 m depth well log 49
PSDM 1200 m depth well log 50
Field data 51
Start model Field data Start model data 52
FWI model Field data FWI model data 53
Match to Synthetics 54
Match to reflection geometry can mislead cycle-skipped start model 55
56
57
RTM with PSDM model 3600 m depth 58
RTM with FWI model 3600 m depth 59
Summary Anisotropic 3D FWI works on field data Needs low frequencies Needs refractions Needs careful QC & validation Can we do better? 60
Pitfalls & Practicalities 61
Pitfalls local minima cycle skipping inadequate low frequencies inadequate starting model Essential that starting model is not cycle skipped low frequencies high-quality start model rigorous QC 62
Local inversion start model misfit best estimate of new model model 63
Local inversion misfit local minimum global minimum model the local minimum may be a worse model, but it provides a better match to the data 64
Low frequency misfit start model global minimum model observed predicted 65
High frequency start model cycle skipped misfit local minimum model observed predicted 66
Workflow choose the right problem & acquire the right data determine start frequency build start model + anisotropy check adequacy of model, wavelet & field data pre-process & reduce data volume modelling & inversion strategy run FWI with QA check synthetic against field data check geometry, wells, image gathers, run RTM on broadband reflection data 67
FWI strategy low high frequencies smooth rough shallow deep refractions reflections phase amplitude early late arrivals primaries multiples acoustic elastic QC often Test with synthetics use variable sub-set of sources each iteration 68