Calibration of Scanning Microwave Microscopy via Substrate-Referenced Machine Learning

Scanning microwave microscopy (SMM) utilizes near-field microwave radiation to offer unparalleled nanoscale material characterization; yet its widespread adoption is hindered by the sophisticated nature of the manual one-port calibration process. Solving this inverse problem is complicated and prohibitive for real-time use. Furthermore, manual calibration often fails to accurately track dynamic changes in probe tip geometry, which is a critical factor in near-field interaction and can result in persistent measurement uncertainty and operator-dependent errors. To overcome these barriers, this work introduces a substrate-referenced deep learning (DL) framework that automates the solution to this inverse problem. By decoupling static substrate parasitics from dynamic tip geometry, a TabPFN-based regressor chain with a root-meansquare error (RMSE) of 0.0086 infers error factors due to the effective probe shapes from a single approach curve, combining with an error correction factor from static substrate parasitics to realize real-time calibration, which greatly reduces calibration times by orders of magnitude and eliminates human variability, enabling real-time, quantitative SMM imaging that was previously unattainable. This method was validated on 2D tin selenide (SnSe) samples with various thicknesses on SiO2/Si++ and Si++ substrates, demonstrating accuracy comparable to expert manual calibration and reducing processing time from hours to seconds, transforming SMM into an accessible tool for the broader research community.