Synthetic‐Aperture Reflective Holographic Microscopy With Neural‐Fields for Efficient and Robust Surface Topography Measurement

ABSTRACT Reflective synthetic aperture digital holographic microscopy enables high‐precision surface topography measurement of nanostructures. Conventional approaches require storing a high‐resolution sub‐aperture hologram stack, resulting in memory consumption of up to tens of gigabytes and low efficiency. Meanwhile, the parasitic fringes in the reflective configuration pose significant challenges for high‐precision reconstruction. Deep learning offers a promising avenue for these challenges. However, key components in topography measurement, such as speckle denoising, phase unwrapping, and aberration compensation, are intrinsically coupled. Existing approaches that treat them separately or sequentially often fail to achieve globally consistent optimization. To address these challenges, we propose neural‐field‐driven synthetic aperture digital holographic microscopy (NeSADHM), which combines the physical model of reflective holography with coordinate‐based implicit learning to be designed for cost‐effective measurement of microchip structures. Experimental results show that NeSADHM achieves robust topography measurement of microchip structures with different reflectivity. It reduces the memory footprint by up to 15.7‐fold and decreases the required holograms by 10.4‐fold without compromising measurement accuracy. Our framework establishes a topography measurement paradigm, which is suitable for applications such as semiconductor detection.