Agentic Large Language Model Interpretation versus Deep Segmentation Networks for Seismic Fault Delineation: A Blind-Test Comparison of Claude Opus 5.5, U-Net, and Swin UNETR
Automated fault interpretation in seismic data is dominated by self-supervised segmentation networks, whose outputs remain fragmented along continuous faults and contaminated by isolated false-positive artifacts where noise mimics reflection discontinuities. General-purpose agentic large language models (LLMs) can now operate desktop software through screenshots and mouse and keyboard actions, but whether such an agent can delineate faults at a quality comparable to trained networks has not been tested. This study compares Claude Opus 5.5, operating the manual-interpretation module of the SeismicFlow platform without task-specific training, with two networks of matched capacity (about 25 million parameters) pretrained with masked reconstruction and trained on the Thebe dataset: a U-Net and a Swin UNETR. Seven 2D sections that were unseen during training formed the blind test: three crossline sections of the Thebe test set with reference labels (fault density 4.07% to 7.83%) and four sections from other surveys (an offshore Middle East volume, Opunake and Parihaka) without labels. The networks were thresholded at 0.85, and the agent was instructed to mark only faults it judged with at least 85% confidence. Each system was evaluated five times, independently, on each labeled section. Averaged over the five evaluations and the three sections, F1 was 0.853 for Claude, 0.858 for the Swin UNETR and 0.831 for the U-Net, and a two-way analysis of variance showed a significant effect of the system (F(2, 36) = 17.64, p < 0.0001) and no significant interaction between system and section (F(4, 36) = 1.38, p = 0.261). Tukey’s post hoc test showed that Claude and the Swin UNETR both exceeded the U-Net in mean F1 (by 0.022 and 0.028, respectively; adjusted p < 0.001 in each case) but did not differ from each other (mean difference − 0.006, 95% confidence interval − 0.018 to 0.006, adjusted p = 0.474). Qualitatively, both networks produced fragmented faults, truncated fault tips and isolated artifacts on all seven sections (for the U-Net also in the water column above the seabed), whereas Claude produced continuous traces without artifacts, an outcome that follows in part from its polyline output. The contrast was greatest on the four sections from surveys that the networks had not seen: there the networks broke the faults into disconnected segments, recovered only short pieces of them on the Opunake sections and added artifacts on every section, whereas Claude traced the visible faults as continuous lines. Claude also stopped traces short in low signal-to-noise zones. An agentic LLM is therefore competitive with a Swin UNETR on the labeled sections and visibly more robust to a change of survey, and the two approaches fail in different ways.
Authors
- Matin Mahzad (ORCID: https://orcid.org/0009-0000-9346-8451)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11498043/v1
- Primary Topic
- Seismic Imaging and Inversion Techniques
- Type
- preprint