Neural Operators with Accurate Jacobian for High-Fidelity Image-Domain Seismic Inversion
| Title | Neural Operators with Accurate Jacobian for High-Fidelity Image-Domain Seismic Inversion |
| Publication Type | Unpublished |
| Year of Publication | 2026 |
| Authors | Park, J, Huseyin Tuna Erdinc, Grant Bruer, Richard Rex, Chandramoorthy, N, Felix J. Herrmann |
| Series Title | International Meeting for Applied Geoscience and Energy |
| Month | 3 |
| Keywords | deep learning, derivative-informed, Fisher-information, FNO, IMAGE, Imaging, Inverse problems, reduced-order modeling, RTM, SEG, Uncertainty quantification |
| Abstract | Neural operators (NOs) provide efficient surrogates for wave-equation-based imaging operators, but accurate forward prediction alone does not guarantee reliable inversion. As inversion updates depend on the adjoint Jacobian, errors in learned sensitivities lead to unphysical updates and degraded quality in recovered models. To this end, we propose a Jacobian-informed training strategy that enforces accuracy in Jacobian in the directions where gradient descent occurs during Least-Squares. Numerical experiments with Fourier Neural Operators demonstrate that supervising only a small fraction ($0.1 \sim 0.01%$) of Jacobian information significantly improves gradient quality and leads to more accurate subsurface velocity reconstruction. In particular, the proposed approach reduces mid to deep layer artifacts and improves structural continuity compared to models trained only with forward misfit term. These results highlight that learning inversion-relevant sensitivities, rather than forward mappings alone, is critical for reliable surrogate-assisted seismic inversion. |
| URL | https://slim.gatech.edu/Publications/Public/Conferences/SEG/2026/park2026IMAGEnoa/abstract.html |
| URL2 | |
| Citation Key | park2026IMAGEnoa |