Abstract:

Reliable geological carbon storage requires estimates of subsurface CO₂ saturation and pressure despite incomplete observations and uncertain reservoir properties. This dissertation develops uncertainty-aware Digital Shadows and Digital Twins that combine physics-based simulation with amortized Bayesian inference for monitoring evolving reservoir states.

The proposed framework uses ensembles propagated through multiphase flow, rock physics, seismic modeling, and imaging to train conditional normalizing flows that generate posterior samples from time-lapse seismic and well observations. The dissertation examines three extensions of this framework: multimodal monitoring, robustness to uncertain rock physics, and context-aware inference. Rock-physics augmentation improves reconstruction under tested model mismatch, while a three-phase extension enables joint estimation of oil and CO₂ saturation. A sensitivity-aware Digital Twin treats rock-physics assumptions as explicit context, supporting assumption-specific inference and what-if analysis. Finally, a collaborative case study demonstrates how probabilistic Digital Shadow estimates can be supplied to a separately developed injection-control workflow. Together, these studies provide a simulation-based framework for uncertainty-aware subsurface monitoring and decision support.

Defense date: November 6, 2026

Defense time: 10:30 am - 01:30 pm EST

Defense location: ES&T L1114

Advisor (Chair):

Felix Herrmann, Professor, Georgia Institute of Technology

Committee members:

Andrew Newman, Professor, Georgia Institute of Technology

Zhigang Peng, Professor, Georgia Institute of Technology

Sheng Dai, Professor, Georgia Institute of Technology

Olav Moyner, Senior Research Scientist, SINTEF, Norway