2D photoelectric neuron array can achieve broadband and low loss optical nonlinearity accessible to ambient light

Light can calculate functions during propagation and interaction with structured materials, with fast speed and low energy consumption. The use of all optical neural networks for general computing requires an optical activation layer with nonlinear dependence on the input. However, existing optical nonlinear materials either have slow speeds or very weak nonlinearity at the level of natural light intensity captured by cameras. Therefore, designing and developing new optical activation functions is crucial for achieving optical neural networks that utilize ambient light for computation.

In a paper published in Nature Communications, a research team led by Professor Duan Xiangfeng and Professor Aydogan Ozcan from the University of California, Los Angeles reported a new strategy using a photonic neuron array to achieve strong optical nonlinearity for broadband incoherent light at low light intensity.

Their devices are heterogeneous and integrated with two-dimensional transparent phototransistors and liquid crystal modulators. Under weak light irradiation, TPT has high electrical resistance, and most of the voltage drop occurs on TPT. LC is not disturbed and maintains transparency. However, at high input optical power, TPT becomes conductive, causing most of the voltage to drop on the LC layer, thereby cutting off optical transmission.

In their experimental demonstration, the designed photoelectric neuron allowed for nonlinear modulation of its own amplitude by spatially and temporally incoherent light in visible light wavelengths, with photon loss of only~20%. They created a 100x100 array of photoelectric neurons and exhibited strong nonlinear behavior under laser and white light irradiation.

The nonlinear photoelectric array is further integrated as part of the mobile phone based imaging system, used for intelligent reduction of glare, selective blocking of strong glare, and almost no attenuation of objects with weaker intensity in the imaging field of view.

Device modeling shows that the optical intensity threshold is very low, only 56 μ W/cm2 produces significant nonlinear response for optimized devices, and the energy consumption per photon activation is as low as 69 fJ.

This photoelectric neuron array can perform nonlinear self amplitude modulation on spatially incoherent light, with characteristics such as low light intensity threshold, strong nonlinear contrast, wide spectral response, fast speed, and low photon loss. For image processing and visual computing systems that do not rely on strong laser beams, this performance is very ideal.

In addition to intelligent glare reduction, the cascaded integration of photoelectric neuron arrays and linear diffraction optical processors can be used to construct nonlinear optical networks, which may be widely used in computational imaging and sensing, and also open the door to the design of new nonlinear optical processors that use ambient light.

Source: Laser Net