Monte Carlo full waveform inversion of tomographic crosshole

Monte Carlo full waveform inversion3uncertainty with a standard deviation of 10 − . Accordingly,the standard deviation of the data uncertainties, σ , is set to3. The standard deviation of the amplitude noise isindicated by the red error bar in figure 3 and compared withtwo waveforms recorded at 0 degrees (short offset / highamplitude) and 45 degrees transmitter-receiver angles (longoffset / low amplitude), respectively.1Recorded at short offset0.5Recorded at long offset0-0.5-110 − 100 150 200 250 300 350 400Normalized amplitudeSample no.Figure 3. Dashed line is a waveform recorded at shorttransmitter-receiver offset (0 degrees). Solid line is awaveform recorded at long offset (45 degrees). The heightof the red errorbar indicates 2 times the standard deviationof the noise added to the data.Depth [m]05100 2 405100 2 405100 2 4Distance [m]05100 2 405100 2 4Figure 4. Five statistically independent samples from the aposteriori probability density of the full waveform inverseproblem.The initial model used for the Metropolis algorithm ischosen as an unconditional sample of the training imageusing a different random seed than for the reference model(see figure 2 right). Burn-in was reached after 2000accepted models. Hereafter samples accepted by theMetropolis rule are representative samples of the aposteriori probability density. Figure 4 shows the 2000 th ,6000 th , 1000 th , 14000 th , and 18000 th accepted sample usinga priori information defined by the training image (figure 1)and Gaussian data uncertainty (equation 4). Only a slightdeviation between the individual samples is seen, whichindicates little a posteriori model uncertainty. Figure 5shows the a posteriori mean and variance based on 18000samples from the a posteriori probability density after theburn-in period. From the a posteriori variance it is seen thatthe overall structures of the model is recovered (varianceclose to or equal to zero) whereas the higher variancesalong the edges between the clay and sand deposits aresubjected to uncertainty. A comparison between thereference model (figure 2 left) and the mean of the samples(figure 4 right) confirms that the waveform inversion is ableto recover the sand structures very well. Moreover, itshould be noted that the algorithm is able to reach theseresults even though it is initiated in a model uncorrelatedwith the reference model (compare figure 2 right and left).Depth [m]0246810Mean120 2 4Distance [m]4.64.44.243.83.63.43.232.82.6( ε/ε 0)0246810Variance120 2 4Distance [m]Figure 5. Mean (left) and variance (right) of 18000 samplesdrawn from the a posteriori probability density of thesolution to the full waveform inverse problem.The suggested Monte Carlo inversion strategy is preferablein that it allows for arbitrary antennae geometry.Furthermore, complex a priori inversion can be includedusing the perturbed simulation algorithm. Finally, a fulldata covariance matrix can be included in order to accountfor correlated data errors, which are often present intomographic inverse problems (e.g. Maurer and Musil,2004; Cordua et al., 2008, 2009). However, note that thisexhaustive sampling strategy needs substantially morecomputationally expensive forward calculations comparedto the traditional migration based approach.ConclusionsWe have demonstrated the potential of producing samplesof the solution to a tomographic full waveform inverseproblem using the extended Metropolis algorithm withcomplex a priori information. The methodology provides ameans of evaluating the a posteriori uncertainty, which isnot provided using optimisation based strategies for fullwaveform inversion. Moreover, the present approach isrobust with regard to the initial guess of the solution and thetransmitter-receiver density. Finally, the extendedMetropolis Algorithm is flexible regarding the choice of apriori information and specification of data uncertainties.SEG Expanded abstracts0.80.70.60.50.40.30.20.10( ε 2 /ε 02 )