Elastic full wave-form inversion with recurrent neural networks

We implement elastic full wave-form inversion in the framework of recurrent neural networks. We use a staggered-grid stress-particle-velocity scheme to solve the elastodynamic equations for forward modeling. Automatic differentiation is obtained, with batch gradient descent optimization, for inversions of P- and S-wave velocities simultaneously. We analyze the in?uence of different batch sizes on the inversion, and ?nd that setting a minimum batch size (i.e., 1) has the best convergence rate for both clean and noisy data. The algorithm is tested with two synthetic models and show acceptable inversion results.