Superresolution diffractive imaging enabled by artificial intelligence

Abstract

For centuries, resolution of optical imaging systems has been constrained to half of the illumination wavelength (λ/2) owing to the Abbe-Rayleigh limit. For optical microscopy, diffraction limits resolution to ~200 nm, which makes the study of viruses, proteins, molecules, and atoms impossible by conventional optical imaging. While fluorescent labelling allows microscopy to overcome the diffraction limit, such techniques are labour intensive, and their applicability is restricted by phototoxicity.
Recently, artificial intelligence (AI) processing of scattering patterns from deeply subwavelength objects had led to the demonstration of imaging with resolution orders of magnitude beyond the diffraction limit. So far, the method has been demonstrated for simple objects of a priori known shape.
In this thesis, I report for the first time AI-enabled label-free imaging of 2D arbitrary objects with resolution beyond the diffraction limit. In particular:
•I have developed a new method for the shape classification of objects smaller than half of the operational wavelength with accuracy of 90%. Numerically simulated far-field scattering patterns from opaque nanoparticles from 0.1λ to 0.5λ in size are classified into eight shapes using an AI model. Each shape class had 700 realisations, resulting in the dataset size of 5600 samples. High-accuracy classification is possible even in the case of arbitrary orientation (67.5%) or position (72%) of the nanoparticles in the field of view. Dataset of only ~1,500 samples allows classification with accuracy over 80%.
•I have experimentally demonstrated artificial intelligence enabled shape classification of nanoscale objects. Diffraction patterns from apertures of five geometrical shapes and sizes ranging from 100 to 320 nm (λ/6.4 to λ/2, with λ=638 nm) were classified with accuracy over 80%.
•I have developed a new scanning microscopy for imaging of complex 2D objects of arbitrary shape based on artificial intelligence processing of its far-field scattering pattern. The scanned objects contained fine details of various sizes, including those smaller than half of the operational wavelength (λ). The method achieves resolution ~λ/4, surpassing the diffraction limit by the factor of two, and outperforming confocal microscopy under the same conditions by 1.5 times. Resolution improvement is observed for the training dataset as small as ~1000 samples. However, doubling the dataset logarithmically increases accuracy of predictions which linearly increases the resolution by ~30 nm. The resolution stops improving with increase of the dataset size over ~15000.
In summary, the research findings reported here allow non-invasive AI-enabled label-free optical imaging of 2D objects at subwavelength scales previously not achievable with conventional microscopy. Subwavelength resolution is achieved using only ~1000 samples for training, which allows easy experimental implementation. Doubling the dataset linearly improves the resolution until its limit. Further improvement of the resolution can be achieved by decrease of the illumination wavelength (scaling) and improvement of the artificial intelligence processing. The method can be extended to imaging of 3D objects as well as recording moving objects. The results of the Thesis will find applications in life sciences, nanotechnology, and materials science.

More information

Identifiers

ORCID for Sergei Kurdiumov: orcid.org/0000-0002-8262-9915

ORCID for Nikolay Zheludev: orcid.org/0000-0002-1013-6636

ORCID for Nikitas Papasimakis: orcid.org/0000-0002-6347-6466

ORCID for Bruce (Jun-Yu) Ou: orcid.org/0000-0001-8028-6130

Catalogue record

Export record