Session 9 — The Inversion Problem and the Climate Problem

Session 9 — The Inversion Problem and the Climate Problem#

Format A — Paper Autopsy Week 9 Relevance: CO₂ storage · ML limits · Ill-posedness

4D seismic CO₂ monitoring · What ML can and can’t replace · Two faces of ill-posedness

Pre-read required

Students receive a 2-page excerpt 48 hours before the session: a short excerpt from a 4D seismic CO₂ monitoring paper (SEG 2024 CCUS special section or LLNL geophysical monitoring report). Come prepared to discuss the figures and the inversion approach.

Hook (0 – 7 min)

Show two synthetic seismic images of the same subsurface: one with a clear bright spot (CO₂ present), one ambiguous. Ask:

“How confident are you CO₂ is present in the second image? What would change your confidence?”

Let students argue before any explanation. Note where they use qualitative vs. quantitative reasoning.

Discussion (7 – 42 min)

Three organizing questions:

1. Tikhonov regularization controls model smoothness. In CO₂ monitoring, do you want a smooth model or a rough one — and why? What does the choice reveal about the physics you believe?

2. This paper uses ML to improve the inversion. What does the ML replace, and what physical knowledge does it use or ignore? Is that a problem?

3. If you were advising a CO₂ storage operator on what geophysical monitoring program to design, what would you recommend? What is your single biggest source of uncertainty?

Push hardest on question 2 — the relationship between data-driven and physics-driven approaches is genuinely contested in the literature.

Relevance

Climate/Energy: Carbon capture, utilization, and storage (CCUS) is a major component of net-zero pathways. 4D seismic monitoring is the primary method for verifying that injected CO₂ stays where it was put. Getting this wrong has regulatory and climate consequences.

Basic science: The ill-posedness of inverse problems — the fact that many models fit the same data — is one of the deepest ideas in geophysics. It connects directly to how we think about model uncertainty in climate projections: many climate models also fit historical observations while predicting different futures.

Go Deeper

SEG 2024 CCUS special section · Anjom et al., Geophysics 2024 (ML in seismic exploration review)

One name: Dr. Felix Herrmann, Georgia Tech — seismic inversion with deep learning. His group’s work exemplifies the physics-ML interface this session explores.