next up previous print clean
Next: Conclusions and future work Up: Extracting Lags Previous: Another approach with similarities

Cooling schedule

As () points out, the selection of the cooling function is very important and can greatly speed convergence. In my method, I treat the sample interval function as a cooling function. To find a good cooling function, I first created a linear cooling function and plotted how the energy decreased as a function of sample interval. This is shown in Figure [*]. In this particular example, there appear to be three sample intervals associated with large drops in energy: 185, 130, and 70. I decided to create a cooling schedule that treats the sized 70 sample interval as the critical temperature. This is shown in Figure [*]. The critical temperature is where convergence is most significant.

 
trialschedule
trialschedule
Figure 12
Trial cooling schdule. Top: input cooling schedule drops linearly with iterations. Middle: energy(617#617) drops in steps with iterations. Bottom: energy(617#617) drops in step with sample interval. This is used to determine the more efficient cooling schedule in Figure [*].
view burn build edit restore

 
coolingschedule
coolingschedule
Figure 13
Cooling schedule. Top: input cooling schedule based on trial cooling schedule designed to spend more iterations near the critical tempurature. Middle: energy(617#617) drops in steps with iterations. Bottom: energy(617#617) drops in step with sample interval.
view burn build edit restore

Notice that in Figure [*], the energy drops off much quicker than in Figure [*] and therefore requires fewer iterations.

Figure [*] shows the result of applying the cooling schedule in Figure [*]. It has converged to the desired event. Figure [*] shows the application of the calculated displacements to the left side of the fault. Its results are about the same quality as Figure [*], which shows the results of applying the actual known displacement to the left side of the fault.

 
anneal_final
anneal_final
Figure 14
Cross-correlagram with overlay of final solution and known displacement.
view burn build edit restore

 
makefinalmod
Figure 15
Applied result, the ``model'' in the center shows the result of applying the calculated displacement to the left side of the fault.
makefinalmod
view burn build edit restore

 
makefinalideal
Figure 16
Ideal applied result, the ``ideal'' in the center shows the result of applying the known displacement to the left side of the fault.
makefinalideal
view burn build edit restore


next up previous print clean
Next: Conclusions and future work Up: Extracting Lags Previous: Another approach with similarities
Stanford Exploration Project
6/7/2002