Learning 4D Material-Interface Dynamics From Few X-RAY Projections

Reconstructing the time-varying 3D material-interface state of a two-material system from only a few X-ray projections remains an unsolved challenge. Conventional tomography breaks down in this regime, and physics-driven trial-and-error is slow and biased. We present a learning-based framework for dynamic sparse-view tomography that addresses both data scarcity in real experiments and the ill-posedness of time-resolved 3D recovery. First, we introduce the Synthetic Dynamic Tomography (SDT) dataset: a scalable pipeline that lifts inexpensive 2D simulations of interfacial instability into 4D ground-truth volumes and physically realistic radiographs while varying noise, view sparsity, and symmetry breaking. Second, we adapt neural radiance fields to represent evolving 3D states as continuous 4D functions (3D space + time), using SDT to learn priors and enabling optimization from sparse projections when volume supervision is unavailable. On Rayleigh-Taylor benchmarks, our model reconstructs coherent spike-bubble dynamics from as few as one view per timestep and remains robust to moderate symmetry breaking. Code and data: https://doi.org/10.5281/zenodo.16996679