Machine learning method puts the “super” in super-resolution spectroscopy

Spectroscopy and microscopy methods with high spatial resolution could enable scientists to visualize and chemically identify individual molecules. Spatial resolution, however, is fundamentally limited by the optical diffraction limit.

Overcoming the diffraction limit is especially important for the “fingerprint” region in the infrared zone, where chemicals can be identified by their unique spectral signature. In this region, resolution is restricted to about 5 microns. Zhang et al. have developed a machine learning method to improve the spatial resolution of an already sub-diffraction, mid-infrared microscopy technique called Infrared Photothermal Heterodyne Imaging (IR-PHI). An enhancement in feature resolution, from 300 to 150 nanometers, has been realized.

“Infrared spectroscopic studies of materials within spatially congested environments are difficult,” says author Kirill Kniazev. “IR-PHI therefore represents an advance in high spatial resolution infrared microscopy and spectroscopy. It, however, is still limited by the diffraction limit to a resolution of about 300 nanometers. With the developed network, it is now possible to realize a feature resolution of about 150 nanometers.”

To develop the network, the authors first trained their algorithm with thousands of ground truth and degraded ground truth images so it could learn how features were degraded by diffraction induced blur. In tandem, the network was trained to remove experimental noise artifacts from acquired super resolution infrared images.

The authors plan to extend their research to study how heat diffusion influences the ultimate resolution of images restored by their network.

Source: “Deep image restoration for infrared photothermal heterodyne imaging,” by Shuang Zhang, Kirill Kniazev, Ilia M. Pavlovetc, Shubin Zhang, Robert L. Stevenson, and Masaru Kuno, Journal of Chemical Physics (2021). The article can be accessed at https://doi.org/10.1063/5.0071944 .

This paper is part of the Chemical Imaging Collection, learn more here .