deep_learning_SNOM

This repository gives the core codes for the paper: https://doi.org/10.1063/5.0139517.

Data generation

• We follow the codes provided in https://doi.org/10.1063/1.4941343 to generate all training data. We recommend the reader check the corresponding Supplementary Materials for all of the information.
• Meanwhile, the full signal simulation for any tip aspect ratio is also implemented in iHNN_prediction_spheroid.py, FindBetaR.m, and PolesResidues.m.
• iHNN_prediction_spheroid.py calls FindBetaR.m and FindBetaR.mcalls PolesResidues.m.
• For a special tip aspect ratio $L=25a$, the signal can be rapidly generated using rational approximation [1]. This is implemented in iHNN_prediction_rational.py

Network construction and training

The following file implement the network depicted in the Figure 3 of the main text.

• train_image_predictor.py constructs and trains the deep neural network that extracts the optical constants from the corresponding s-SNOM signal.

Network prediction

• iHNN_prediction_spheroid.py allows one to change the tip aspect ratio during optimization.
• iHNN_prediction_rational.py predicts faster for the fixed tip geometry $L=25a$.

References

[1] B.-Y. Jiang, L. M. Zhang, A. H. Castro Neto, D. N. Basov, and M. M. Fogler, “Generalized spectral method for near-field optical microscopy,” Journal of Applied Physics 119, 054305 (2016).