PASSIVE SEISMIC

Ambient Noise Tomography – Southeastern U.S.

Extracted Rayleigh-wave dispersion from continuous noise records using a MUSIC-based workflow on TensorFlow, suppressing non-coherent arrivals and stabilizing phase-velocity measurements.

Produced high-resolution crustal and upper-mantle structure beneath the southeastern United States, with improved sensitivity to sedimentary basins and Moho topography.

INVERSE PROBLEMS

Dictionary Learning for Unsupervised Seismic Inversion

Learned sparse dictionaries directly from shot gathers to represent coherent seismic energy while suppressing noise and acquisition footprints.

Used the learned basis as a prior in an adjoint-free inversion scheme, recovering velocity structure without requiring labeled training models.

DRL-Based Seismic Multiple Suppression

Deep Reinforcement Learning for Multiple Suppression (DRL-AutoMute)

Python/PyTorch environment and DQN agent for adaptive muting of seismic multiples. The environment defines actions as mute-parameter updates (e.g., slope, intercept, taper length), and the reward is computed from residual multiple energy and preservation of primary amplitudes. Developed originally for Shell’s AutoMatch project to demonstrate that DRL can outperform static Bayesian parameter searches for multiple attenuation.