Full waveform inversion of infrasound: Sensitivity kernels from the adjoint method along with non-linear inversion tests

Summary Atmospheric models are based on various types of geophysical data, including lidar and radar. Infrasounds, acoustic waves that can propagate over large distances, have not yet been used in atmospheric models, although they provide valuable information. Besides their sensitivity to atmospheric phenomena such as gravity waves, infrasound also presents the advantage of being omnipresent. Previous studies explored the use of infrasound packet arrival properties for model estimation. However, properties such as arrival times present less information than full waveforms. We aim here to investigate, for the first time, the sensitivity of a full infrasound waveform to model parameters and to use these sensitivities in an inverse problem to recover atmospheric structure. For this purpose, infrasound propagation is modeled by Euler equations (i.e. Navier-Stokes equations in the absence of attenuation effects), and discretization is carried out here using the finite-differences method. Waveform sensitivity to atmospheric parameters is computed through the adjoint method via a novel and optimized double checkpointing-based procedure and validated by comparison with a small perturbation method. As an illustration, these sensitivity kernels are computed for the idealized case of an explosion in Finland, recorded by a CTBT station. These first results demonstrate the high sensitivity of infrasound waveforms to the atmospheric perturbations generated by gravity waves. Moreover, the sensitivity kernels of infrasound waveforms allow us to recover the variations of model parameters by solving an inverse problem. To demonstrate this capability, full waveform non-linear inversions are performed using the Limited Broyden-Fletcher-Goldfarb-Shanno method (L-BFGS): wind and sound speed profiles are inverted for a test case with idealized conditions and a synthetic dataset. These estimates of infrasound sensitivity kernels are closing a knowledge gap that allows the use of infrasound full waveforms to constrain atmospheric models.