Performing optical logic operations by a diffractive neural network

Optical computing, using photons instead of electrons, offers advantages such as high speed, low power consumption, and parallel processing capabilities, making it promising for applications like augmented reality and autonomous driving. Optical logic gates are fundamental components of optical computing, and their miniaturization and efficiency are crucial research areas. Existing methods rely heavily on precise control of input light signals (phase, polarization, intensity, beam size), leading to instability and low contrast ratios in the output logic states. The required precise controls also hinder miniaturization. This paper proposes a diffractive neural network implemented by a compound Huygens' metasurface to overcome these limitations. The design uses plane waves as input, eliminating the need for complex control mechanisms and enabling compact optical logic gate implementation. The metasurface is trained to directionally scatter light into designated areas representing '1' or '0', achieving all seven basic logic operations.

Previous research on optical logic gates focused on constructive/destructive interference effects between input light signals. These methods heavily depend on precise control of input light properties, including phase difference, polarization, and intensity, as well as the size of incident beams. This dependence leads to instability and low contrast ratios in practical scenarios, making miniaturization challenging. The authors highlight the need for a solution that eliminates these precise control requirements to facilitate the development of compact and stable optical logic gates.

The proposed method employs a diffractive neural network, analogous to an artificial neural network, to perform optical logic operations. The network consists of an input layer (optical mask), one or more hidden layers (metasurfaces), and an output layer. The input layer spatially encodes the input plane wave according to the desired logic operation. Each region in the mask has two transmittance states (high/low), representing the selected/unselected states for computation. The hidden layers, composed of metasurfaces with subwavelength meta-atoms acting as neurons, decode the encoded light. The Huygens' wavelet from each meta-atom is described by the Rayleigh-Sommerfeld diffraction equation. The transmission coefficients of the meta-atoms are trained using an error back-propagation algorithm, minimizing the mean square error between the output intensity and the target output. A two-layer metasurface was designed and fabricated for experimental validation at 17 GHz (wavelength λo = 17.6 mm). Each metasurface consisted of 30 × 42 meta-atoms, with each meta-atom having a square cross-section and made of F4B dielectric material. The training process involved adjusting the transmission coefficients of the meta-atoms to direct the light to one of two designated areas at the output, representing '1' or '0'.

Numerical simulations demonstrated the successful realization of NOT, OR, and AND gates using a two-layer metasurface. Microwave experiments verified these results, showing that the field intensities are correctly focused into one of two designated regions, with contrast ratios exceeding 9.6 dB. The experimental setup included a horn antenna to excite plane waves and an E-field probe to measure the transmitted field at the output layer. The authors further demonstrate that their approach can, in principle, be extended to implement all seven basic logic gates and cascaded logic gates using a three-layer metasurface. The experimental results show that the intensity peaks are well-confined within the designated regions for all three logic gates, showing good agreement with numerical simulations. The slight deviations are attributed to impedance mismatches at air-dielectric interfaces, which could potentially be reduced using anti-reflection structures.

The proposed diffractive neural network approach offers a significant advancement in optical logic gate design. By eliminating the need for precise control of input light parameters, it facilitates the development of compact, stable, and easily miniaturized devices. The successful experimental demonstration at microwave frequencies validates the design principles. The ability to perform all seven basic logic operations using the same metasurface further highlights the universality and potential of this approach. This work opens pathways for developing sophisticated all-optical devices and systems for various applications.

This paper presents a novel design strategy for optical logic operations using a diffractive neural network, eliminating the need for precise input light control. The successful implementation of NOT, OR, and AND gates at microwave frequencies demonstrates the feasibility and potential of this approach. Future research could focus on extending this method to higher frequencies (terahertz and optical) and exploring more complex logic operations and integrated photonic circuits.

The current experimental demonstration is limited to microwave frequencies. While the theoretical framework suggests applicability to higher frequencies, experimental validation at those frequencies is needed. The impact of fabrication imperfections on the performance of the metasurfaces also needs further investigation.