AFM-net: machine learning acceleration of atomic force microscopy nanometrology from scarce data and fast scans

Atomic Force Microscopy (AFM) is a cornerstone of nanometrology but is constrained by slow acquisition and processing speeds. To address this bottleneck, we introduce AFM-net, a machine learning framework that simultaneously replaces conventional post-processing pipelines and enables high-speed imaging by reconstructing high-fidelity data from fast, low-quality scans. AFM-net leverages a unique training strategy: rather than relying on scarce experimental data, we trained the model on massive natural image datasets overlaid with real, extracted instrumental corruptions. This workflow accelerates data processing by up to 500× and demonstrates zero-shot generalization to other modalities like Scanning Tunneling Microscopy. Ultimately, AFM-net illustrates that overcoming bottlenecks in high-throughput nanometrology depends not only on faster hardware but also on smarter, data-centric computational pipelines.