Abstract / Summary
Abstract Fast radiography of dynamic material interfaces is blurred, sparse, and difficult to interpret in three dimensions. Restoration, state inference, and sparse-view reconstruction are often studied separately, which makes their outputs difficult to compare against known simulation labels. We address this problem with a curated Rayleigh–Taylor instability (RTI) corpus containing 101,250 interface images from 1,350 simulations, each labeled by Atwood number, acceleration, initial perturbation amplitude, and time. The simulations are sampled on a common uniform time grid, although acceleration and time are only partly independent under the normalized-time scaling of the underlying RTI problem. Controlled temporal averaging and spatial blur generate radiograph-like surrogates with known target states. On these surrogates, a fine-tuned restoration model improves mean SSIM from 0.905 to 0.962 and mean PSNR from 20.67 to 22.05 dB relative to Restormer. The same corpus supports retrieval, regression of Atwood number and time, local ambiguity diagnostics, short-horizon temporal densification, and a lifted 4D benchmark with 54 dynamic 3D sequences and 72 synthetic projections per volume. Sparse-view neural fields reach about 35 dB on held-out views in that synthetic benchmark. At sufficiently late times, some frames are best interpreted as controlled image-processing cases because finite-domain effects can influence the morphology. Within that scope, the RTI corpus provides a shared benchmark for restoration, label-space inference, ambiguity analysis, and sparse-view reconstruction.