IS-RAFT: geometry-informed sub-pixel displacement estimation for X-ray speckle tracking

X-ray speckle-based wavefront sensing and phase-contrast imaging are indispensable tools at synchrotron facilities, particularly for weakly absorbing samples where phase contrast dominates over attenuation contrast. Among the available approaches, single-shot X-ray speckle tracking (XST) is attractive because of its experimental simplicity, but its broader use is hindered by noise sensitivity, limited sub-pixel accuracy, and high computational cost. To address these limitations, we developed IS-RAFT, a geometry-informed speckle-tracking network built on RAFT (recurrent all-pairs field transforms) and designed to combine geometric pre-alignment with learned sub-pixel residual refinement. Trained exclusively on simulated X-ray speckle data and validated experimentally, IS-RAFT achieves a noise floor comparable to that of RAFT and substantially lower than that of XST. It not only resolves fine displacement structures that XST fails to recover but also restores displacement amplitudes to a level broadly consistent with scanning-based X-ray speckle vector tracking (XSVT), thereby achieving sub-pixel displacement fidelity beyond that of single-shot XST and RAFT. It also delivers a speedup of more than two orders of magnitude over XST. These results establish IS-RAFT as a robust, real-time, and low-noise solution for high-precision speckle metrology under weak-attenuation conditions, with applications in dynamic imaging, online beam monitoring, and wavefront sensing at fourth-generation synchrotron sources.