Multilayer Permittivity and Thickness Extraction in Infrared Scanning Nearfield Optical Microscopy Using Deep Learning

Scattering-type scanning near-field optical microscopy combined with Fourier-transform infrared spectroscopy allows for material identification and characterization more than two orders of magnitude below the diffraction limit. This resolution is achieved by positioning an AFM-tip in close proximity to the sample. However, owing to the complex interaction between incident wave, tip and sample, extensive modelling is needed for the quantitative extraction of material parameters defining the optical response, particularly for layered samples. Due to the strong non-linearity between the optical parameters and the scattered light, fitting of the model is generally necessary, but it is often unstable and time-consuming, requiring human expert supervision. Deep learning algorithms offer a fast alternative but have so far only been applied to bulk samples. Here, we extend those approaches to samples consisting of one and two layers of polar crystals, which exhibit a strongly non-linear scattering response, on top of a Silicon substrate, by training convolutional neural networks with model data. We present the network architectures and resulting optical parameter extraction efficiencies and discuss extensions to other sample systems.