AI-DINO: Artificial Intelligence for the Dynamic Imaging of Nanoscale Objects

Coherent X-ray scattering techniques provide complementary information about nanoscale systems with inherent trade-offs. Bragg Coherent Diffraction Imaging (BCDI) provides high-resolution 3D structural information but requires computationally intensive phase retrieval and cannot capture diverse temporal dynamics. X-ray Photon Correlation Spectroscopy (XPCS) offers time-resolved dynamics but lacks spatial resolution. Sequential measurements are often impractical due to beam damage, sample drift, and experimental time constraints. To address these limitations, we present a machine learning framework that combines neural ordinary differential equations (NODEs) with physics-based forward scattering models to model time-evolving structures directly from speckle patterns without requiring frame-by-frame phase retrieval. The NODE framework learns the underlying differential equation governing structural evolution, enabling extrapolation beyond the temporal range of experimental measurements. We demonstrate recovery of real-space structural dynamics from simulated diffraction patterns under realistic conditions, including Poisson noise characteristic of synchrotron measurements. Validation studies show consistent performance across varying noise levels and temporal sampling rates. This approach offers potential advantages in reducing computational overhead compared to traditional phase retrieval methods while providing access to temporal dynamics that extend beyond measurement windows. The framework could help reduce beamtime requirements and enable studies of samples sensitive to prolonged X-ray exposure. Future work focuses on experimental validation and application to more complex structural evolution scenarios.