Source-linked AI summary

Design of Task-Specific Optical Systems Using Broadband Diffractive Neural Networks

Yi Luo, Deniz Mengu, Nezih T. Yardimci, Yair Rivenson, Muhammed Veli, Mona Jarrahi, Aydogan Ozcan

arXiv:1909.06553v1cs.NEphysics.comp-phphysics.optics

TL;DR

Broadband optical components must operate across a continuum of wavelengths, yet this remains challenging. The paper uses multilayer diffractive neural networks designed with deep learning and demonstrates tunable filtering and wavelength demultiplexing, with targeted Q-factors maintained across output ranges.

  • Problem

    Operating with a continuum of wavelengths remains challenging, motivating broadband optical-component designs that address this operating regime.

  • Method

    The approach uses multilayer diffractive neural networks whose neuron transmission coefficients are determined by physical thickness and whose output position can be tuned axially.

  • Results

    Multilayer designs provide targeted Q-factors across a large output range, with enhanced and more uniform Q-factors across the targeted axial-displacement range.

  • Takeaways & Limitations

    The results indicate that the D2NN framework can realize task-specific broadband optical components with controlled spectral or spatial outputs.

  • Takeaways & Limitations

    Numerical simulations and experimental results showed relatively minor discrepancies attributed to multiple contributing factors.

Abstract

from arXiv · show

We report a broadband diffractive optical neural network design that simultaneously processes a continuum of wavelengths generated by a temporally-incoherent broadband source to all-optically perform a specific task learned using deep learning. We experimentally validated the success of this broadband diffractive neural network architecture by designing, fabricating and testing seven different multi-layer, diffractive optical systems that transform the optical wavefront generated by a broadband THz pulse to realize (1) a series of tunable, single passband as well as dual passband spectral filters, and (2) spatially-controlled wavelength de-multiplexing. Merging the native or engineered dispersion of various material systems with a deep learning-based design strategy, broadband diffractive neural networks help us engineer light-matter interaction in 3D, diverging from intuitive and analytical design methods to create task-specific optical components that can all-optically perform deterministic tasks or statistical inference for optical machine learning.

1 Electrical and Computer Engineering Department, University of California, Los Angeles, CA, … 4 Department of Surgery, David Geffen School of Medicine, University of California, Los

The work presents broadband diffractive optical neural networks that process temporally incoherent, continuous-wavelength light for deep-learning-designed optical tasks. The architecture engineers three-dimensional light–matter interactions for deterministic optical functions and statistical inference.

  • 4 Department of Surgery, David Geffen School of Medicine, University of California, Los: Previous diffractive approaches employed monochromatic coherent light.The broadband architecture is framed against this earlier illumination regime.
  • 4 Department of Surgery, David Geffen School of Medicine, University of California, Los: Broadband diffractive optical networks simultaneously process a continuum of wavelengths from a temporally incoherent source.This extends diffractive optical processing beyond monochromatic coherent illumination.
  • 4 Department of Surgery, David Geffen School of Medicine, University of California, Los: The networks all-optically perform task-specific functions learned using deep learning.The design strategy links learned computational objectives to optical wavefront transformations.
  • 4 Department of Surgery, David Geffen School of Medicine, University of California, Los: The approach engineers light–matter interaction in three dimensions.Its design strategy departs from intuitive and analytical component-design methods.
  • 4 Department of Surgery, David Geffen School of Medicine, University of California, Los: The resulting optical systems can perform deterministic tasks entirely optically.This capability is identified as an application of broadband diffractive neural-network design.
  • 4 Department of Surgery, David Geffen School of Medicine, University of California, Los: The architecture also supports statistical inference for optical machine learning.Statistical inference is presented alongside deterministic optical task execution.

Introduction

The paper introduces broadband diffractive optical neural networks that process a continuum of wavelengths in parallel for task-specific optical operations. It demonstrates experimentally validated spectral filtering and spatial wavelength de-multiplexing using multilayer diffractive systems designed with deep learning.

  • Diffractive neural-network framework: Diffractive neural networks use successive learned layers to modulate optical phase and/or amplitude, enabling all-optical task computation after physical fabrication.The layers connect through spherical waves according to the Huygens-Fresnel principle and can be fabricated using 3D printing or lithography.
  • Prior work and limitation: Previous diffractive optical networks demonstrated blind object-classification inference and improved accuracy, diffraction efficiency, and signal contrast with additional layers, but used monochromatic coherent illumination.Their input light was both spatially and temporally coherent.
  • Broadband architecture: The proposed broadband architecture processes a continuum of input wavelengths in parallel from a temporally incoherent broadband THz pulse.It unifies deep learning methods with engineered material dispersion to control light-matter interaction in three dimensions.
  • Demonstrated optical tasks: Seven fabricated multilayer networks realize tunable single-passband and dual-passband spectral filters together with spatially controlled wavelength de-multiplexing.The systems provided very good fits to their trained diffractive models, and the framework can balance Q-factor and power efficiency through design.
  • Design scope: Additional layers enable sophisticated broadband multiplexing and de-multiplexing tasks that cannot be successfully achieved using only two learnable diffractive layers.Dispersion engineering and polymerization-based 3D printing extend the approach to different parts of the electromagnetic spectrum.

Discussion

The discussion attributes simulation–experiment discrepancies to alignment, beam, detector, and material-characterization limitations. It also highlights tunable spectral filtering, transfer-learning improvements, and the framework’s extension beyond THz wavelengths.

  • Limitations: Simulation–experiment discrepancies may arise from diffractive-layer misalignment, nonideal beam profile and alignment, unmodeled detector properties, and inaccurate material dispersion characterization.The model assumes a spatially uniform THz pulse propagating parallel to the optical axis, while detector acceptance angle and coupling efficiency are not modeled.
  • Spectral-filter design: Multiple-layer diffractive neural networks can provide a targeted Q-factor across a large range of output apertures, unlike a standard diffractive lens.The final layer of some single-passband designs intuitively resembles a diffractive lens, but the complete multi-layer network enables targeted spectral behavior.
  • Spectral-filter tunability: Changing the detector/output-plane distance tunes the passband: moving closer to the final layer causes a red shift, whereas moving farther causes a blue shift and lowers Q-factor.The tested axial distances were not included during training, explaining the reduced Q-factor away from the ideal output-aperture position.
  • Transfer learning: Training an existing diffractive filter model with a band-tunability constraint enhances and equalizes Q-factors across the targeted axial-displacement range.This transfer-learning-like strategy uses multiple loss terms around the corresponding frequency bands at different propagation distances.
  • Broader applicability: The D2NN framework generalizes to broadband sources and continuous wide frequency ranges, extends beyond THz wavelengths including the visible band, and may be strengthened by dispersion-engineering metasurfaces.These extensions broaden the application space for diffractive optical neural networks where broadband operation is attractive or essential.

Methods

The methods combine broadband THz time-domain measurements with a wavelength-dependent diffractive-network model and 3D-printed implementations. Training spans 0.25–1 THz under a flat-spectrum assumption, while measurements are normalized to the incident pulse spectrum.

  • Experimental setup: A mode-locked Ti:Sapphire laser generates 780 nm femtosecond pulses split between a plasmonic THz emitter and an optical delay-line measurement arm.The THz signal is amplified and lock-in detected; each trace averages 10 pulses over 5 seconds.
  • Experimental setup: The measurement system provides signal-to-noise ratio levels over 90 dB and observable bandwidths up to 5 THz.These properties characterize the THz time-domain acquisition system used for experimental validation.
  • Measurement normalization: Measured spectra from the 3D-printed models are normalized using a reference spectrum acquired with only the input aperture in the optical path.Reported power efficiency is evaluated at the peak wavelength of each passband.
  • Diffractive-network modeling: Each network is modeled as wavelength-dependent thin modulation layers linked by free-space Rayleigh-Sommerfeld propagation.Layer transmission depends on amplitude and phase, while the amplitude and phase are functions of neuron thickness and incident wavelength through the complex refractive index.
  • Training and fabrication constraints: Training assumes unit-intensity, uniform-phase plane waves over 0.25–1 THz, uniformly partitioned into M=7500 discrete frequencies.A 1 cm square input aperture matches the incident THz beam width, and the smallest printable feature size is 0.5 mm.
Loading 1909.06553v1…