My PhD final presentation "Convergence of a full waveform inversion scheme based on PSPI migration and forward modeling-free approximation: procedure and validation".
HMCS Max Bernays Pre-Deployment Brief (May 2024).pptx
PhD Final Presentation by Marcelo Guarido
1. Supervisors:
Dr. Larry Lines
Dr. Robert Ferguson
Convergence of a full waveform inversion scheme based on PSPI
migration and forward modeling-free approximation: procedure
and validation
Presenter:
Marcelo Guarido de Andrade
5. Reduce the cost of a gradient descent FWI routine preserving its
resolution at most
Replacing the RTM by a PSPI migration
Reducing cost to estimate the step length
Re-interpretation of the gradient
Objective
5
6. Introduction
Steepest-descent (gradient) method
Trial-and-error line search vs. Pica et al. (1990) approximation
Monochromatic averaged gradient
Initial model
Conjugate gradient
Impedance inversion of the gradient
Forward modeling-free gradient
Pre-stack
Post-stack
Well calibration of the gradient
Same initial model
Conclusions
Future work
Outline
6
12. Steepest-descent (gradient) method
96 simulated shots
(“real model”)
Initial model
(smoothed
velocity)
Synthetic Data
Residuals
PSPI Migration
(Deconvolution
Imaging Condition)
Step length
(linear search and
quadratic fit)
Update ModelConvergence?
Final Model
Yes
No
Mute and Stack
12
13. Steepest-descent (gradient) method
Acquisition
Type Marine
Number of shots 96
Shot spacing 100m
Shot depth 30m
# of receivers 401
Receiver spacing 10m
Receiver depth 0m
Frequency 10Hz
Processing
Starting frequency 1-6Hz
Iterations per frequency < 16Hz 10
Iterations per frequency > 15Hz 5
Frequency increment < 25Hz 1Hz
Frequency increment > 24Hz 5Hz
Mute Yes
Water bottom mute Yes
Scale factor tests 21
13
33. Steepest-descent (gradient) method
Acquisition
Model Marmousi
Number of shots 101
Shot spacing 100m
Shot depth 0m
# of receivers 401
Receiver spacing 10m
Receiver depth 0m
Frequency 10Hz
Processing
Starting frequency 1-4Hz
Iterations per frequency < 13Hz 5
Iterations per frequency > 12Hz 2
Frequency increment < 13Hz 1Hz
Frequency increment > 12Hz 5Hz
Mute Yes
Smooth Yes
Scale factor tests 21
33
39. Monochromatic PSPI migration of the residuals
Gradient for each frequency
One step for each gradient
Average scaled monochromatic gradients
Monochromatic averaged gradient
39
40. Starting: low frequency
Increase frequency band
Convergence
By 2Hz
Maximum: 60Hz
More migrations as
frequency band gets larger
Monochromatic averaged gradient
40
59. Gradient -> reflection coefficients or impedance?
Impedance inversion
FWI as seismic processing tools
I – impedance inversion
S – stacking
M – migration
Impedance inversion of the gradient
59
60. Band-Limited Impedance Inversion (BLIMP)
Depth-to-time and time-to-depth conversions
Initial model as pilot impedance
One migration step per iteration
Impedance inversion of the gradient
60
73. Commuting the Migration and Stacking operators
Forward modeling-free gradient
73
74. Invert frequencies > 4Hz
Step length: biased by position
Select random shot
Can’t control objective function
Forward modeling-free gradient
74
85. Sonic log
Storage difference of sonic log and current model (what the
gradient should be)
Compute gradient
Compute matching filter between gradient and stored
difference of sonic log and current model
Convolve matching filter with the whole gradient
Inverting frequencies > 4Hz
Well calibration of the gradient
85
93. Advantages of a forward modeling-free inversion:
No source (wavelet) estimation
No need to for any forward modeling
Cheaper
More robust to work on real data?
Well calibration of the gradient
93
95. Simulation of a velocity analysis
Divide Marmousi in 5 horizontal areas
Pick one random position inside each area
Linear interpolation
Repeat for all columns
Smooth
Same initial model
95
104. Steepest-descent: works well on simple models
Least squares step length: cheaper and good
Monochromatic averaged gradient: higher resolution with high cost
Conjugate gradient: improved resolution
Impedance inversion of the gradient: great resolution reducing costs
Forward modeling-free gradient: pre and post-stack approximations,
cheaper with good resolution
Well calibration of the gradient: 100% forward modelling-free FWI,
lowest cost with improved resolution
Conclusions
104
105. Reverse time migration (RTM)
Multi-parameter (if possible)
Extend to 3D
Real data
Deep learning (Neural Networks?)
Future work
105
106. Dr. Larry Lines
Dr. Rob Ferguson
NSERC
CREWES Sponsors
Students and staff
Dr. Kris Innanen
Dr. Daniel Trad
Dr. Gary Margrave
Dr. Babatunde Arenrin
Dr. Raul Cova
Dr. Wenyong Pan
FWI group
Laura Baird
Soane Mota dos Santos
Acknowledgments
106