Automatic Differentiation-Based Full Waveform Inversion of Anisotropic Parameters in TTI Media

Conventional full waveform inversion (FWI) requires explicit computation of adjoint source backpropagated wavefields and derivation of adjoint equations to obtain the gradient of the objective function with respect to model parameters. However, derivation of gradient formulas for complex equations involving multiple parameters often presents significant mathematical challenges, hindering the development and application of multiparameter FWI in anisotropic media. This study applies automatic differentiation (AD) to anisotropic parameter FWI in tilted transversely isotropic (TTI) media, reformulating FWI as an optimization problem that minimizes the objective function. By directly computing gradients through chain rule propagation, our method eliminates the need for explicit adjoint source calculation and backward wavefield propagation, thereby significantly simplifying the implementation workflow. The enormous memory consumption is the main factor limiting the application of AD technology in FWI, and the AD FWI of multi-parameters in TTI media further exacerbates this issue. To address the excessive memory consumption during AD-based gradient computation for complex TTI equations, a boundary-saving strategy is integrated into the AD framework. Furthermore, the Adam gradient optimization algorithm from deep learning is introduced to enhance both inversion efficiency and accuracy for multiparameter reconstruction in TTI media. The proposed approach ultimately achieves efficient and high-accuracy multiparameter FWI in TTI media.