Atomic force microscopy (AFM) is used to image nanostructures and characterize material properties, but point-by-point scanning makes it difficult to improve imaging speed and spatial resolution at the same time. A research team at the Shanghai Institute of Ceramics of the Chinese Academy of Sciences has developed StableAFM, a physics-guided generative super-resolution model designed to address this problem.
StableAFM combines AFM sampling physics with a latent diffusion model and uses the original measurement data to constrain image reconstruction. It is designed to recover fine spatial details from fast, sparse scans. The team first uses an AFM-SwinIR network to reconstruct the basic image structure, then applies a diffusion model adapted to AFM images to refine interfaces and local features.
To keep generated details consistent with experimental data, the researchers introduced a latent-space diffusion posterior sampling strategy. It maps each reconstruction back onto the original low-resolution measurement grid and adjusts it according to the difference between the reconstruction and measured data. In tests, the model’s mean squared error fell by about 31.3% when physical measurement constraints were added.
