CorFuncSAXSNet: deep-learning-driven extraction of nanostructural parameters from small-angle X-ray scattering data of polymeric materials

Abstract

Small-angle X-ray scattering (SAXS) analysis of semi-crystalline polymers remains a labour-intensive process requiring expert interpretation of correlation functions. To address this challenge, we present CorFuncSAXSNet: a deep neural network framework designed to directly predict nanostructural parameters – including lamellar crystalline thickness (dc) and amorphous layer thickness (da) – from 1D raw SAXS curves. Building upon SAXS datasets collected at the Shanghai Synchrotron Radiation Facility's BL19U2 beamline, we developed three neural architectures: a convolutional neural network, a residual network and a q-space attention network. Data augmentation strategies, including Gaussian noise injection and q-shift interpolation, improved model robustness against experimental uncertainties. Cross-validation results demonstrate that all networks achieve mean absolute errors of 0.109–0.112 nm for dc and 0.459–0.499 nm for da. Though amorphous layer predictions at large values exhibit higher errors due to dataset skewness (83.3% of data clustered at 4.5 < dc < 6.5 nm, 5.0 < da < 20.0 nm), our framework enables rapid parameter extraction (<1 s per curve), reducing reliance on manual graphical methods. CorFuncSAXSNet bridges the gap between AI and synchrotron-based structural analysis, establishing a foundation for real-time smart beamline architectures.