Meta-optical processors for broadband complex-field image operations

Abstract

All-optical image processing provides a fast and energy-efficient alternative to conventional electronic systems by directly manipulating optical wavefronts. However, metasurface-based optical processors reported to date are often limited in functionality, operating bandwidth, or input modality, which restricts their adaptability across different image processing tasks. Here, we demonstrate a broadband metasurface platform capable of performing diverse analog image processing operations on both amplitude- and phase-encoded inputs. This platform is realized using a single-layer dielectric metasurface designed through an end-to-end, task-driven inverse design framework. By tailoring the spatial-frequency components of incident image wavefronts, the metasurface implements analog operations such as edge detection and pattern recognition across a 200nm wavelength bandwidth in the visible spectrum. Furthermore, we develop a compact processor architecture that integrates imaging and computation within a reduced optical footprint. These results establish a flexible and compact metasurface-based optical processor with strong potential for integration into practical imaging and optical computing systems.

Figures (5)

• Figure 1: Concept and design of meta-optical processors for complex-field image processing.a Schematic of a meta-optical processor performing all-optical complex-field image edge detection and letter pattern recognition. b Unit-cell design of birefringent TiO$_2$ nanopillar meta-atoms on a SiO$_2$ substrate, defined by a height $H = 600\,\mathrm{nm}$, length $L = 350\,\mathrm{nm}$, width $W = 115\,\mathrm{nm}$, period $P = 450\,\mathrm{nm}$, and an in-plane orientation angle $\theta$ that encodes the geometric phase. c Simulated spin-conversion efficiency and geometric phase shift as functions of the nanopillar orientation angle at 530 nm (top), and broadband spin-conversion efficiency of the optimized meta-atoms across the 400--700 nm spectral range (bottom). The shaded region (430--630 nm) indicates the operational wavelength range used in this work. d Differentiable meta-optical processor model for metasurface phase optimization. A forward model simulates the modulation of incident image wavefronts by the meta-optical processor, and gradients of an output-defined loss are evaluated via automatic differentiation in backward propagation to update the phase profile. e Schematics of the experimental setup for characterizing the meta-optical processors using the $4f$ imaging system. f Schematics of the experimental setup for characterizing the compact meta-optical processor.
• Figure 2: Experimental demonstration of meta-optical complex-field edge detection.a Amplitude- and phase-encoded input images based on the USAF-1951 resolution target, along with the corresponding target edge-intensity distribution used for inverse design. b Optimization trajectory showing the normalized loss during 5000 iterations. Insets show the evolution of the metasurface phase profile from an initially flat distribution at iterations 100, 500, and 5000. c Optical microscope image of the fabricated metasurface and corresponding scanning electron microscopy (SEM) images of the TiO$_2$ nanopillar array, shown in top and tilted views. d Experimental edge detection results for an amplitude sample consisting of a transmission mask patterned with the Tampere city logo. e Experimental edge detection results for a phase sample formed by a transparent surface-relief structure. For both samples, the unprocessed outputs and the corresponding edge-detected results at wavelengths of 430, 480, 530, 580, and 630 nm are shown. Scale bars: 200 µ m (optical microscopy in c), 2 µ m (SEM in c), and 500 µ m (d, e).
• Figure 3: Experimental demonstration of meta-optical complex-field pattern recognition.a Amplitude- and phase-encoded input images corresponding to the letters T, A, and U, along with the target output intensity distributions used for inverse design. The target pattern T is mapped to a centered Gaussian output intensity, while the distractor patterns (A and U) are assigned a uniform dark target output to suppress unwanted responses. b Optimization trajectory showing the normalized loss during 5000 iterations. Insets show the evolution of the metasurface phase profile from an initially flat distribution at iterations 100, 500, and 5000. c Optical microscope image of the fabricated metasurface and corresponding SEM images of the TiO$_2$ nanopillar array. d, e Amplitude and phase samples comprising the letters T, A, and U, used for pattern recognition testing (acquired in reflection mode). f--k Experimental recognition results for the amplitude sample, showing the unprocessed output (f) and the processed outputs at wavelengths of 430, 480, 530, 580, and 630 nm. l--q Corresponding recognition results for the phase sample, including the unprocessed output (l) and the processed outputs at the same wavelengths. Intensity distributions are projected onto the coordinate planes to compare output peak heights. Scale bars: 200 µ m (optical microscopy in c), 2 µ m (SEM in c), and 500 µ m (d, e).
• Figure 4: Comparison between analytical cross-correlation filtering and the inverse-designed meta-optical processor.a Amplitude of the complex transmission function of an analytical cross-correlation filter for an amplitude-encoded target letter T. b Phase profile of the corresponding analytical filter. c Phase difference between the inverse-designed metasurface and the analytical filter in b. A global phase offset is removed (RMS: 1.69 rad). d Amplitude of the complex transmission function of an analytical cross-correlation filter for a phase-encoded target letter T. e Phase profile of the corresponding analytical filter. f Phase difference between the inverse-designed metasurface and the analytical filter in e (RMS: 1.85 rad). g, h Simulated output intensities for amplitude- and phase-encoded TAU image inputs (Figure \ref{['fig 3']}d,e), processed using the analytical cross-correlation filter designed for amplitude-encoded inputs. The inset in h shows a top-view projection of the intensity distribution. i, j Corresponding simulated outputs obtained using the analytical filter designed for phase-encoded inputs. The inset in j shows a top-view projection of the intensity distribution. k, l Simulated outputs obtained using the inverse-designed meta-optical processor in Figure \ref{['fig 3']}b.
• Figure 5: Compact meta-optical processor integrating imaging and pattern recognition.a Imaging phase $\phi_I$ designed using ray tracing for the compact meta-optical processor configuration. b Amplitude- and phase-encoded input images corresponding to the target letter Y and the distractor letter U, together with the associated target output intensity distributions and the optimized recognition phase profile $\phi_R$. c Composite phase profile $\phi_C$, obtained by combining the imaging phase $\phi_I$ with the pattern recognition phase $\phi_R$. d Optical microscope image of the fabricated metasurface and corresponding SEM images of the TiO$_2$ nanopillar array. e, k Amplitude and phase samples comprising the letters Y and U. f--j, l--p Corresponding processed outputs at wavelengths of 430, 480, 530, 580, and 630 nm for the amplitude- and phase-encoded inputs, respectively. Scale bars: 200 µ m (optical microscopy in d), 2 µ m (SEM in d), 500 µ m (e, k).