Addressing phase ambiguity in single-frame coherent X-ray diffraction imaging through ensemble phase retrieval

Single-frame Coherent X-ray Diffraction Imaging (CXDI) enables rapid nanoscale structural characterisation with minimal radiation exposure. However, phase loss in diffraction measurements introduces fundamental ambiguities, causing traditional retrieval algorithms to converge to local optima that appear numerically valid but may not represent true structure. We present an ensemble phase retrieval framework that mitigates overfitting in single-solver methods through spatially aware cross-validation and ensemble aggregation. Unlike arbitrary data partitioning that violates the deterministic Fourier relationship, our approach employs systematic spatial masking via sliding-window skipped regions tailored to CXDI constraints. Multiple reconstructions are then aggregated to generate two outputs: a consensus image through pixel-wise averaging and a spatial uncertainty map quantifying local reconstruction ambiguity. This strategy transforms phase retrieval from deterministic point estimation into probabilistic inference with explicit reliability quantification. Operating algorithm-agnostically, the framework applies across diverse solvers – from iterative algorithms to neural networks. Results demonstrate that ensemble aggregation achieves an optimal balance at the 8×8-pixel skip configuration, maintaining high reconstruction accuracy (SSIM 0.81) while preserving sufficient ensemble diversity (0.18). This ensemble methodology advances uncertainty-aware computational imaging, providing researchers with enhanced reconstruction fidelity and interpretable confidence metrics for single-frame CXDI applications.