Search

Ongoing works on full waveform inversion (FWI) are yielding an increasing number of objective functions as alternative to the traditional L2-waveform. These studies aim at designing more robust functions and inversion strategies to reduce the intrinsic dependence of the FWI results on (1) the initial model and (2) the lowest frequency present in field data. In this work, we perform a comparative study of five objective functions in time domain under a common 2-D-acoustic FWI scheme using the Marmousi model as benchmark. In particular, we compare results obtained with L2-based functions that consider the minimization of different wave attributes; the waveform-based, non-integration-method; instantaneous envelope; a modified version of the wrapped instantaneous phase and an improved version of the crosscorrelation travel time (CCTT) method; and hybrid strategies combining some of them. We evaluate the robustness of these functionals as a function of their performance with and without low frequencies in the data and the presence of random white Gaussian noise. Our results reveal promising strategies to invert noisy data with limited low-frequency content (≥4Hz), which is the single strategy using the instantaneous phase objective function followed by the hybrid strategies using the instantaneous phase or CCTT as initial models, in particular the combinations [I. Phase + Waveform], [CCTT + Waveform] and [CCTT + I. Phase]. [ABSTRACT FROM AUTHOR]

We present a scale- and parameter-adaptive method to pre-condition the gradient of the parameters to be inverted in time-domain 2-D elastic full-waveform inversion (FWI). The proposed technique, which relies on a change of variables of the model parameters, allows to balance the value of the gradient of the Lamé parameters and density throughout the model in each step of the multiscale inversion. The main difference compared to existing gradient pre-conditioners is that the variables are automatically selected based on a least-squares minimization criteria of the gradient weight, which corresponds to the product of the gradient by a power of the parameter to be inverted. Based on numerical tests made with (1) a modified version of the Marmousi-2 model, and (2) a high-velocity and density local anomaly model, we illustrate that the value of the power helps to balance the gradient throughout the model. In addition, we show that a particular value exists for each parameter that optimizes the inversion results in terms of accuracy and efficiency. For the two models, the optimal power is ∼2.0–2.5 and ∼1.5 for the first and second Lamé parameters, respectively; and between 3 and 6, depending on the inverted frequency, for density. These power values provide the fastest and most accurate inversion results for the three parameters in the framework of multiscale and multishooting FWI using three different optimization schemes. [ABSTRACT FROM AUTHOR]