← Back to Sessions

From Physics to Scale: Building Foundation Models for Seismic Full-waveform Inversion

Description: 

Seismic full-waveform inversion (FWI) is a central tool for high-resolution subsurface imaging in energy, resource, and Earth science applications. By exploiting the full information content of seismic wavefields, FWI enables detailed recovery of subsurface properties, but it remains computationally expensive, highly nonlinear, and severely ill-posed. Recent advances in artificial intelligence offer new opportunities to address these challenges, while raising a fundamental question: how can learning-based approaches respect wave physics while scaling to realistic seismic problems?

In this talk, I will present our recent progress in applying AI to seismic FWI through two complementary directions that together motivate the development of seismic foundation models. The first direction focuses on physics-informed learning, where wave equations and physical constraints are embedded into model architectures or training objectives to improve stability, interpretability, and data efficiency. The second direction explores large-scale, data-driven learning, including auto-regressive and foundation-style models trained on extensive seismic simulations, which learn transferable representations of wave propagation across acquisition settings and geological regimes. By contrasting physics-guided and data-scaled approaches, I will discuss their strengths, limitations, and emerging hybrid strategies. Through examples in seismic velocity model building and waveform inversion, I will illustrate how these ideas improve robustness, scalability, and imaging fidelity. I will conclude with a perspective on building physics-aware foundation models for seismic inversion as a pathway toward reliable and trustworthy subsurface characterization.

Session: New Frontiers in Seismic Observations and Modeling with Innovative Methods and Emerging Data on Earth and Other Planets - III

Type: Oral

Date: 4/17/2026

Presentation Time: 05:15 PM (local time)

Presenting Author: Youzuo Lin

Student Presenter: No

Invited Presentation: 

Poster Number:

Authors