Shear Wave Velocity Estimation by Deep Reinforcement Learning: A Case Study

Abstract

Optimization algorithms used in underwater geoacoustic inversion are time-consuming since they need many iterations to approach a global minimum. An efficient geoacoustic inversion approach based on deep reinforcement learning is proposed to estimate seabed shear wave velocity profiles. The model is built upon the deep-Q network with a specially designed environment and an agent. The performance of this approach has been validated by simulation cases in our previous work. In this paper, the model is applied to real-world scenarios to estimate seabed shear wave velocity profiles. The approach is assessed on dispersion curves of the ocean seismic interface waves generated by real data against several optimization algorithms commonly used in underwater acoustics for geoacoustic inversion. Two cases are considered for extracting the interface wave dispersion curves. One set of the dispersion curves extracted from the data generated by an active shear source detects shallow sediment layers, while the second set of the dispersion curves estimated from passive ocean ambient noise can reach about reservoir depth. The assessment results demonstrate that the proposed approach is efficient in terms of accuracy and computational time.