Experimental, Computational and Machine Learning Methods in Geophysics

A thermal plume in syrup imaged by thermochromic liquid crystal. Credit: Carolina Lithgow-bertelloni

I am also very interested in developing and applying novel methodologies, including both end-to-end and generalizable approaches to address various challenging problems in geophysics.

Specifically, I focus on:

Cartoon of the 3D-scanning Stereoscopic Particle Image Velocimetry (SSPIV) system I built at UCLA

SSPIV measurement that covers 68% volume (left), full velocity reconstructed with hybrid-adjoint (middle), and reconstruction of unmeasured temperature field (right) Ra=1e6~1e8

Architecture of PI-DIFNO, a physics-informed neural operator that works directly on scattered ocean and ice points. The lower block learns the map from one field to another; the upper block passes the prediction through a frozen forward model and checks that it reproduces the input, so that the physics constrains the learning. From Bao et al. (submitted)

Workflow of SLEvolve compared with a conventional solver. Instead of re-reading the whole loading history at every step, SLEvolve carries a compact memory state forward in time; the numbered orange boxes mark the new shortcuts. From Bao et al. (submitted)

3D view of a numerical simulation of Rayleigh Benard convection at Ra=1e5, viscosity contrast up to 10 (hotter less viscous)

2D central-plane view of the simulation above. Left panel is prediction from PINNs, right is the misfit from the reference simulation.