Rapid Inverse GISAXS Analysis of Nanoparticle Assemblies with Simulation-Trained Deep Learning

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

High Resolution Image Download MS PowerPoint Slide Nanostructured materials, particularly those formed through nanoparticle deposition or self-assembly on thin film surfaces, are critical to numerous advanced applications due to their exceptional physical and chemical properties. Grazing-incidence small-angle X-ray scattering (GISAXS) has become an indispensable technique for characterizing the morphology of these nanostructures, offering detailed insights into electron density distributions at both the surface and within the film. However, extracting structural information from GISAXS data remains challenging, largely due to the phase problem. Conventional methods typically involve fitting to experimental data using predetermined, simplified models. The process is both time-consuming and constrained by the limited variety of the available models, often resulting in oversimplified descriptions parameterized by a single size and a polydispersity parameter. Moreover, convergence difficulties become more challenging for GISAXS data fits when using traditional regression algorithms compared with transmission SAXS data. To address these limitations within a well-defined model system, we use the distorted wave Born approximation (DWBA) to simulate a diverse library of 2D GISAXS patterns for supported gold nanoparticle assemblies. These simulated datasets are used to train a convolutional neural network (CNN) that predicts the joint nanoparticle height–radius distribution from GISAXS patterns. On simulated validation data, the form-factor branch achieves mean relative errors of 13.8% for particle height and 11.2% for radius, while experimental CNN-predicted mean radii deviate from SEM-derived values by 18.5% and 17.9% for two Au nanoparticle films. After offline training, CPU inference requires approximately 44 ms per 256 × 256 GISAXS pattern, enabling millisecond-scale analysis without iterative fits. These results demonstrate a rapid, model-specific workflow for GISAXS-based nanoparticle size-distribution analysis.

Links and sources

• PDF from original source

Comments