Current students

Exploration Geophysics aims at estimating accurate physical properties of  the Earth subsurface from seismic data acquired close to the surface.
For physical reasons, data are band-limited and corrupted by a great variety of noises, disturbances, and other phenomena.
Therefore, the fundamental tasks of Geophysics are challenging inverse problems.
Moreover, the acquisition campaigns result in massive datasets, limiting the algorithms to be computationally feasible.

To tackle this challenge, I leverage recent Machine Learning techniques.
Seismic data show a great variety of statistically relevant and independent patterns.
I devise Deep Learning methods to solve several geophysical tasks by learning such patterns.
I first interpolate seismic data using Deep Priors, which are Convolutional Neural Networks (CNNs) that precondition the interpolation problem.
Then, I study how to use deep learning to solve imaging problems, i.e., computing images of the subsurface characteristics out of the data.
Specifically, I devise generative networks as a post-processing operator for refining images obtained with Reverse Time Migration (RTM) techniques.
Moreover, I recast the post-stack seismic inversion in a Bayesian framework that uses Deep Priors to quantify the uncertainty.
Finally, I demonstrate the features extraction ability of CNNs for two interpretation tasks: landmine detection and velocity model building.
The former aims at spotting buried threats (e.g., landmines) signatures in Ground Penetrating Radar (GPR) acquisitions.
The latter is an iterative process that estimates the subsurface P-wave velocity. When in presence of salt bodies, imaging becomes particularly challenging because such bodies are characterized by higher P-wave velocities.
By processing images migrated at different angles, CNNs can leverage this physical information to produce accurate segmentations of the bottom of salt.

Through numerical experiments on both synthetic and field data, I demonstrate the devised machine learning methods to be effective compared to the state of the art. The results suggest that improvements can be achieved by integrating pure data-driven algorithms within general inverse problems theory through a-priori information derived from domain knowledge.