Full‐Fourier‐Component Tailorable Optical Neural Meta‐Transformer

Abstract Brain‐inspired optical neural computing (ONC) is the state‐of‐the‐art scheme in modern computing, offering robust strategies to execute advanced inference with a high throughput and large‐scale parallelism. However, the hitherto prevalent diffractive ONC networks have watered‐down competence, which is mostly a phase‐only methodology but fails to precisely handle the Fourier transform of complex fields, thus forfeiting the integrity of the architecture and half the volume of available training weights. Here, a novel neural meta‐transformer (ONM) enabled by an optical rotation‐isolator‐assisted paradigm is proposed, whose meta‐neurons utilize structural birefringence and polarization rotation to achieve independently arbitrary tailoring of full Fourier components, that is the complete learnable parameters of diffractive ONC. The full‐Fourier‐component ONM has great merits over phase‐only counterparts in all representative cases: being a classifier, it improves the recognition accuracy, especially for input with more high‐frequency features; acting as an imager, the background noise of output is effectively diminished; and when engineering as an encoder, both near‐field grayscale nanoprinting and neural meta‐holography are yielded. The mechanism is minimalist, compact, and compatible with nonlinear activation, opening the route to fully parametric intelligent meta‐devices, with far‐reaching implications for optical computing, display, encryption, etc.