A convolutional neural network approach for multilayer analysis in infrared nanospectroscopy

\nThe combination of Fourier-transform infrared spectroscopy and scattering-type scanning near-field optical microscopy allows for the spectroscopic investigation of materials and structures at the nanoscale, far below the diffraction limit in the infrared. This resolution is achieved by the use of metallized atomic force microscopy tips which locally illuminate the sample through the creation of near-fields at their apex. The complex interaction between incident light, a metallized tip and a layered sample necessitates the use of sophisticated models. While these models are powerful, using them to fit measured spectra is generally slow and often unstable, thus requiring expert oversight. Neural networks present a fast and often more stable alternative, but their application so far has focused on bulk samples. Here, we present the use of convolutional neural networks for the recovery of the optical and thickness properties from the spectra of samples consisting of one or two layers of polar crystals on silicon.\n