Three experiment reports on volumetric correlation and salt edges in velocity reconstruction

This record archives three experiment reports by Wenjie Lyu, each with its supplementary material, written between June and October 2026. The reports were not peer reviewed and will not be submitted to a journal. They are kept as a record of the experiments. The central result is carried into a separate short paper on salt edges, which is in preparation.

Report 1, "Volumetric Agreement Can Coexist with Unresolved Salt Flanks under a Gradient-Magnitude Probe". On the GeoFWI 1.0 test split (n = 5000), a U-Net reconstruction from a gradient magnitude map of compressional velocity reaches a whole grid correlation of 0.916 while an interface score stays at 0.580.

Report 2, "Volumetric Correlation and Salt-Flank Geometry Carry Different Geological Information". Two salt models with matched correlation (0.893 and 0.890) carry a lateral shift and an amplitude bias. On a GeoFWI tile with a salt spire the correlation is about 0.993 while the salt overlap is 0.004. Across the 1507 test tiles that contain salt, the salt carries a median 6 percent of the tile variance.

Report 3, "Salt Edges and Sediments under a Band-Limited Stack". The archived stack correlation of about 0.967 came from a forward model that did not reach the salt. The report itself withdraws the network scores computed from those archived stacks (in-band Marmousi2 and clipped BP2004). With a corrected forward of the sharp models, the image correlation is about 0.72 below the salt and about 0.98 in the shallow arcs.

Reports 1 and 2 share the GeoFWI test split numbers and the same archived Marmousi and SEG/EAGE Overthrust inversion products, so they overlap. The short paper uses the correlation and interface score separation of Report 1 and the salt spire and salt variance share of Report 2. It does not use the archived stacks of Report 3.

The PDFs were compiled on 4 November 2026. On that date, statements that a reproduction package accompanies each report were corrected to say that code and score files are not released, and working labels were removed from section titles. No result was changed. Some figures still carry working labels, such as Tier-1 and niche.

Data: GeoFWI 1.0 (Li et al., Zenodo, https://doi.org/10.5281/zenodo.18603709); the public Marmousi2, OpenFWI, SEG/EAGE Overthrust, and BP 2004 models. The archived Marmousi and Overthrust inversion products were provided by Chao Li and are not redistributed here. Code and score files are not released.