Deep learning-driven conversion of scanning superlens microscopy to high depth-of-field SEM-like imaging

Scanning electron microscopy (SEM) enables nanoscale imaging but requires vacuum environments and coating samples with conductive films. We present a deep learning approach to transform optical super-resolution (OSR) microscopy images into high-resolution images resembling SEM, specifically optimized for chip samples. Utilizing our custom-designed scanning superlens microscopy (SSUM) system, we acquire OSR images with a resolution down to ~80 nm without the need for coatings or vacuum conditions. Notably, the SSUM system achieves an effective depth-of-field (DoF) of approximately 2 μm through Z-stack scanning, enabling clear visualization of multilayer chip structures across a larger axial range than conventional optical imaging. Our algorithm further enhances the nanoscale microstructures observed with the SSUM platform, significantly improving the visibility of structures that are otherwise less distinct. A cycle-consistent generative adversarial network (CycleGAN) model is trained on paired OSR and SEM images to learn the mapping between these imaging modalities. The model is then applied to unseen OSR test images from silicon wafer samples. Quantitative analysis shows that the reconstructed images achieve a mean peak signal-to-noise ratio (PSNR) 1.64 dB higher than the input OSR images. Qualitative assessment further demonstrates the model's ability to generate results with high structural detail, specifically in chip-level applications. This technique overcomes key SEM constraints while preserving nanoscale resolution, offering the potential for advanced chip manufacturing and inspection tasks where traditional SEM requirements pose challenges.