Smallest holographic microscope needs smallest silicon LED ever

This form of microscope does not have conventional optics (see diagrams), but uses maths to reconstruct an image of the object from interference patterns. In this case, SMART created a deep neural network architecture to do the reconstruction – of plant seeds and tissue samples in this case

The room-temperature LED was fabricated on a 300mm wafer using an unmodified commercial 55nm bulk CMOS process along with electronics and other photonic components, and emits 1.1μm infra-red at over >50mW/cm² from an area below 0.14μm² (~400nm in diameter).

In the microscope, it faces a 10 x 12mm 9.5Mpixel CMOS image sensor. 20μm diameter beads could be imaged.

Surface-passivation proved important for the LED, as non-radiative recombination due to surface defects becomes more of a problem as dimensions shrink. Carriers were confined by a gate oxide layer and the electric field of the carrier-injecting top contact, which was made from transparent polysilicon instead of opaque metal to improve emission.

On image reconstruction, SMART said: “Traditional reconstruction methods require detailed knowledge of the experimental setup for accurate reconstruction and are sensitive to difficult-to-control variables such as optical aberrations, the presence of noise, and the ‘twin image’ problem.”

The team’s neural network takes account of system variables, and can be used with no prior knowledge of the spectrum nor beam profile of the light source. It needs no training data, and instead has a physics model embedded within its algorithm.

“In addition to holographic image reconstruction, the neutral network offers blind source spectrum recovery from a single diffracted intensity pattern, which marks a departure from all previous supervised learning techniques,” according to SMART, which sees similar LED-neural network microscopes being used for live-cell tracking or spectroscopic imaging of biological tissues such as living plants.

Of the LEDs it said: “Further applications include arraying these LEDs in CMOS to generate programmable coherent illumination for more complex systems.”

Full details of the LED have been published in the freely-available Nature Communications paper ‘A sub-wavelength Si LED integrated in a CMOS platform‘ and details of the novel untrained neural network can be found in ‘Simultaneous spectral recovery and CMOS micro-LED holography with an untrained deep neural network‘, published in Optica, and also available without payment.

Images: Singapore-MIT Alliance for Research and Technology (SMART)

View more : IGBT modules | LCD displays | electronic components