PaperlayerView exact revision on arXiv

Source-linked AI summary

Broadband Mid-Wave Infrared Imaging Enabled by All-Silicon Hyperboloid Metalens

Rose Johnson, Debabrata Mandal, Zihao Zou, Jessie Wei, Anna Wirth-Singh, Vladimir Yarmolik, Johannes Froech, Arka Majumdar, Praneeth Chakravarthula

arXiv:2610.29841v1physics.opticscs.AI

TL;DR

Strong chromatic aberration limits broadband imaging with meta-optics, despite their potential compactness. This work pairs an all-silicon hyperboloid metalens with a pretrained restoration model adapted from four captured images, demonstrating computational imaging across the full 3–5 μm MWIR band.

  • Problem

    Strong chromatic aberration makes broadband meta-optic imaging challenging because a hyperboloid metalens designed for one wavelength can blur images at other wavelengths.

    Source: Introduction · p. 3

  • Method

    The authors pair a 2 cm diameter, f/1 all-silicon hyperboloid metalens with a pretrained restoration network adapted to the measured MWIR system using four captured images.

    Source: Introduction · p. 6

  • Results

    At fractional bandwidth 0.50, the hyperboloid has 9% higher integrated MTF volume than the inverse-designed counterpart, and four-image adaptation produces reconstructions close to ground-truth targets.

    Source: Discussion · p. 15

  • Takeaways & Limitations

    A single-layer hyperboloid metalens can support full-band MWIR computational imaging without dispersion engineering when paired with a calibration-efficient reconstruction backend.

    Source: Discussion · pp. 15–16

Abstract

from arXiv · show

The mid-wave infrared (MWIR, 3-5 microns) is indispensable for thermal sensing, gas detection, long-range imaging, and free-space optical communications. Meta-optics offer a compact, lightweight, and potentially low-cost alternative to conventional refractive optics, making them attractive for size, weight, and power (SWaP)-constrained systems. However, their strong chromatic dispersion has largely restricted imaging to narrow spectral bands. Here, we demonstrate broadband imaging across the full 3-5 micron MWIR band using a single-layer silicon metasurface coupled with computational reconstruction. Remarkably, the reconstruction model is adapted using only four image captures, overcoming large training data requirements and enabling broadband image recovery despite the strongly wavelength-dependent optical response. In contrast to approaches that engineer complex, dispersion-optimized metalens profiles for broadband operation, our method leverages the wavelength-dependent information encoded by a simple 2cm diameter hyperboloid metalens. We show that this metasurface captures sufficient broadband intensity information for computational reconstruction across the entire MWIR band. These results establish a simple route toward broadband computational meta-optics without requiring broadband achromatic metalens designs, offering a promising platform for compact MWIR imaging systems.

Introduction

Conventional MWIR refractive optics can be bulky and heavy, while meta-optics offer a thinner alternative. Strong chromatic aberration complicates broadband imaging, motivating approaches that combine optical design with computational reconstruction.

Introduction

  • Conventional MWIR systems use infrared-transparent refractive materials such as germanium and chalcogenide glasses, but their optics can be bulky and heavy for SWaP-constrained platforms.Meta-optics instead use a thin layer of subwavelength structures to control the optical wavefront and can reduce system volume and mass.Source: Introduction · p. 3
  • A hyperboloid metalens designed for one wavelength can become severely blurred away from that wavelength, making operation across 3–5 μm challenging.The phase profile's strong wavelength dependence changes its focal response across the band.Source: Introduction · pp. 3–4
  • Dispersion-engineered metalenses can provide broadband focusing, but achievable bandwidth, numerical aperture, and aperture size are coupled, and MWIR demonstrations often cover only part of the band or use smaller apertures and NA.Computational approaches instead pair optics with reconstruction, including inverse-designed meta-optics optimized across a spectral range.Source: Introduction · p. 4
  • Visible-light demonstrations have used a conventional hyperboloid metalens and shared information among RGB channels to correct chromatic blur, raising whether a simple lens can also support MWIR reconstruction without RGB channels.MWIR detectors can collect photons across the full 3–5 μm band in a single channel, so the system must collect useful spatial information across a larger fractional bandwidth.Source: Introduction · p. 5
  • The paper demonstrates full-band MWIR computational imaging with a 2 cm diameter, f/1 all-silicon hyperboloid metalens and a restoration network adapted using four captured images.The metalens is designed for a nominal wavelength of 4.5 μm and requires no dispersion engineering; the authors report that it captures sufficient spatial information for high-fidelity reconstruction across the band.Source: Introduction · p. 6

Results

The results compare integrated MTF across spectral bandwidths and use that analysis to design a 2-cm-aperture hyperboloid metalens. At full MWIR bandwidth, the hyperboloid retains more integrated MTF volume than the inverse-designed optic.

Broadband information transfer favors a hyperboloid metalens

  • The comparison uses integrated MTF volume across varying fractional bandwidths as a proxy for spatial-frequency information available to computational restoration.Fig. 2 compares matched hyperboloid and MTF-engineered meta-optics.Source: Broadband information transfer favors… · p. 7, Figure 2 · p. 10
  • At fractional bandwidth ∼0.10–0.20, the inverse-designed meta-optic achieves a larger integrated MTF volume than the hyperboloid.The passage relates this regime to visible RGB Bayer-filter channels with fractional bandwidths of ∼0.18.Source: Broadband information transfer favors… · p. 7
  • At fractional bandwidth 0.50, the hyperboloid has 9% higher integrated MTF volume than the inverse-designed optic, favoring it as the optical front end for full-band MWIR imaging.The authors attribute this comparison to the broader wavelength range over which the inverse-designed optic must balance its response.Source: Broadband information transfer favors… · p. 7

MWIR metalens design

  • The fabricated hyperboloid has a 2-cm aperture, 2-cm nominal focal length, f/1 ratio, and 4.5 μm design wavelength, with lens thickness reduced tenfold versus a commercial Ge lens.It is not designed to eliminate chromatic aberration; its blurred broadband output is intended to preserve enough spatial information for computational restoration.Source: MWIR metalens design · p. 8
  • Fig. 1 depicts the broadband capture and restoration pipeline, the silicon lens's bulk comparison with a commercial Ge lens, and the fabricated meta-atoms.The meta-atom pillar is 4 μm high with 2 μm periodicity.Source: Figure 1 · p. 9

Few-shot computational restoration

The restoration approach adapts a pretrained model to the MWIR system with four captured images, then refines its output. Experiments report recovery of broadband contrast and target structures despite strong chromatic blur in raw captures.

Few-shot computational restoration

  • The restoration pipeline adapts a pretrained all-in-one image-restoration model to the MWIR system using only four captured images.Its two stages collect few-shot data, then fine-tune the model and refine output to better match MWIR targets.Source: Few-shot computational restoration · p. 11
  • The adapted network is evaluated on previously unseen heated stencil targets, and reconstructed broadband images show substantially improved contrast and recovered object structure.The paper also notes supplementary measurements using a 1951 USAF resolution chart and exposure-bracketed captures.Source: Few-shot computational restoration · pp. 11–12
  • Fig. 3 outlines data preparation, parameter-efficient fine-tuning, and subsequent color thresholding to correct generative artifacts and luminance.Fig. 4 presents raw broadband images, restored images, and a comparison with a U-net baseline.Source: Figure 3 · p. 12, Figure 4 · p. 13

Experimental broadband MWIR imaging

  • Raw broadband captures show strong chromatic blur that reduces contrast and suppresses high-spatial-frequency detail, while restoration recovers target structures and edges close to ground truth.The raw captures use a hyperboloid designed to focus at 4.5 μm and a commercial FLIR camera.Source: Experimental broadband MWIR imaging · p. 14

Discussion

Across the full 3–5 μm MWIR band, a forward-designed hyperboloid preserves enough spatial-frequency information for computational restoration, although inverse-designed optics perform better at intermediate fractional bandwidths. The methods pair this metalens with a restoration model adapted from very few samples and a post-processing refinement pipeline.

Discussion

  • Across 3–5 μm, the hyperboloid preserves sufficient spatial-frequency information and contrast for high-fidelity computational restoration despite chromatic aberration.The proposed design objective is useful information transfer rather than an achromatic point-spread function, avoiding the fabrication complexity of multilayer or multilevel metasurfaces.Source: Discussion · pp. 14–16
  • 9% at Δλ = 2.0 μm and 3% at Δλ = 0.3 μm are the hyperboloid’s MTF-volume advantages over inverse design; inverse design leads at fractional bandwidths ∼0.10–0.20.The intermediate regime corresponds to narrower spectral sub-bands, while the full-band and narrowband cases are relevant to unfiltered thermal imaging and narrowband gas sensing, respectively.Source: Discussion · pp. 14–15, MWIR Meta-optic characterization · p. 17
  • Four captured images suffice to adapt the restoration model, and reconstructed images closely match ground-truth targets despite strong chromatic blur in raw measurements.The few-shot calibration avoids the large paired datasets typically required to train restoration networks from scratch.Source: Discussion · p. 15

Methods

  • A fabricated hyperboloid metalens (20 mm aperture, 20 mm nominal focal length, 4.5 μm central wavelength) retains a forward-design advantage over matched-dimension inverse-designed optics: 9% at Δλ=2.0 μm and 3% at Δλ=0.3 μm.Source: MWIR Meta-optic characterization · p. 17
  • A frozen latent-diffusion backbone and control pathway are specialized with a lightweight MWIR branch trained on augmented data, leaving pretrained restoration capabilities intact.The base model extracts low-level cues, and fewer than five samples can suffice for a usable inverse of the metalens degradation.Source: MWIR Meta-optic characterization · pp. 18–19
  • Four labelled samples adapt the diffusion restoration model, but its outputs show spurious texture, boundary halos, and thermal-scale intensity mismatch, prompting post-processing refinement.The refinement pipeline is intended to suppress these artifacts while preserving fine structural detail.Source: MWIR Meta-optic characterization · p. 19

Supplementary Information for

The supplementary-information front matter identifies the paper and its authors. It also lists their affiliations and equal-contribution note.

Supplementary Information for

  • The supplementary information is for “Broadband Mid-Wave Infrared Imaging Enabled by All-Silicon Hyperboloid Metalens.”Source: Supplementary Information for · p. 24
  • Rose Johnson, Debrabrata Mandal, Zihao Zou, Jessie Wei, Anna Wirth-Singh, Vladimir Yarmolik, Johannes Froech, Arka Majumdar, and Praneeth Chakravarthula are listed as authors.Source: Supplementary Information for · p. 24
  • The listed affiliations are the University of Washington and the University of North Carolina at Chapel Hill; Johnson and Mandal are marked as equal contributors.Source: Supplementary Information for · p. 24

Note 1. Details on the Hyperboloid Metalens Design

The hyperboloid metalens phase profile is implemented with silicon pillars selected from simulated phase and transmission responses. The fabricated lens has a 2 cm diameter and focal length.

Note 1. Details on the Hyperboloid Metalens Design

Note 2. Integrated MTF Area Calculation

The integrated MTF-area analysis compares broadband hyperboloid and engineered phase profiles over spectral bandwidths centered on 4.5 μm. It normalizes each area to the hyperboloid’s monochromatic reference.

Note 2. Integrated MTF Area Calculation

Note 3. Inverse Design Approach

The inverse-design comparison uses differentiable proxies to optimize metasurface phase profiles over spectral bands. The design process maps between phase and structure, then evaluates propagated point-spread functions through normalized MTFs.

Note 3. Inverse Design Approach

  • Four inverse-design profiles are optimized independently for four spectral bands using an Adam-updated phase profile and gradients through proxy and propagation models.The phase loss samples wavelengths and spatial frequencies below an optimization cutoff.Source: Note 3. Inverse Design Approach · pp. 27–29
  • Differentiable polynomial proxies map duty cycle to phase and fit precomputed transmission coefficients as functions of wavelength.The phase-to-structure mapping uses harmonic orders, while the transmission fit uses linear least squares.Source: Note 3. Inverse Design Approach · pp. 27–28
  • The all-silicon inverse-design metalens uses a 2 μm periodicity, 3 μm height, 1.2 cm diameter, and 2.5 cm focal length across 3–5 μm.Its radial phase profile is initialized with a quadratic focusing phase and optimized at discrete radial positions.Source: Note 3. Inverse Design Approach · p. 28

Note 4. Computational Restoration Algorithm

The restoration framework adapts a pretrained all-in-one model to MWIR images using low-level visual cues and diffusion-based restoration. It combines degradation-specific experts and refines outputs to address artifacts from the pretraining-to-MWIR domain gap, requiring limited system-specific adaptation.

Note 4. Computational Restoration Algorithm

  • The framework extracts low-level cues from degraded MWIR images and uses them to condition a denoising diffusion restoration model.The pretrained model draws on features learned across multiple degradation and restoration tasks, including when MWIR images are out of distribution.Source: Note 4. Computational Restoration… · p. 29
  • The diffusion model identifies input degradations in latent space and combines degradation-specific expert adaptors to guide restoration.A final refinement stage reduces artifacts associated with the domain gap between pretraining and MWIR datasets.Source: Note 4. Computational Restoration… · p. 29
  • The staged framework generalizes a pretrained restoration model to the metalens's chromatic degradation while requiring limited system-specific adaptation.Source: Note 4. Computational Restoration… · p. 30

Supplementary Figures

Supplementary experiments show broadband restoration across varying target stencils, exposure times, and temperatures. The restored images show increased intensity contrast and finer stencil-boundary detail, though artifacts remain.

Supplementary Figures

Loading 2610.29841v1…