Multifunctional volumetric meta-optics for color and

Multifunctional volumetric meta-optics for color and polarization image sensors: supplementary material PHILIP CAMAYD-MUÑOZ, CONNER BALLEW, GREGORY ROBERTS, ANDREI FARAON*

Kavli Nanoscience Institute and Thomas J. Watson Sr. Laboratory of Applied Physics, California Institute of Technology, Pasadena, California 91125 *Corresponding author: faraon@caltech.edu

Published 31 March 2020

I. Inverse design methodInthisworkwedesignthree-dimensionaldielectricstructures,

optimizedtoperformaspecifiedopticalscatteringfunction:inthiscase,focusingincidentplanewavestodifferentpositionsdependingonthe frequencyandpolarization.Thestructure isdefinedbyaspatially-dependent refractive index distribution within a

cubicdesignregion.Thisrepresentsanexpansivedesignspacewiththe capacity to express a broad rangeof complex opticalmulti-functionality.However,identifyingtheoptimalindexdistributionfor a given target function remainsa challenging inversedesignproblem,particularlyforstronglyscatteringdevices[1].Therefore,weadoptaniterativeapproachguidedbygradientdescent(Fig.S1)in order to efficiently generate complex 3D designs[2]: startingfrom an initial index distribution, we use full-wave simulations(FDTD) to calculate the sensitivity of the sorting efficiencywithrespecttoperturbationsoftherefractiveindex.Thesensitivitycanbe calculated from just two simulations, allowing efficientoptimizationofthree-dimensionaldeviceswithmodestresources.Basedonthesensitivity,wemodifytheinitialdesigninordertomaximize the performance while conforming to fabricationconstraints. Thisupdateprocess is repeateduntil theoptimizeddevicecanefficientlyperformthetargetfunction.

Forthedesignspresentedinthismanuscript,westartwithauniformrefractiveindexdistribution, .This

distribution is continually updated to maximize theelectromagneticintensityatthetargetlocationinthefocalplane,

. This figure ofmerit serves as a proxy for

sortingefficiencywhilesimplifyingthesensitivitycalculation.The

sensitivity, ,iscomputedfromtheelectromagneticfieldsin

twoFDTDsimulations:

(S1)

where aretheelectricfieldswithinthecubewhenilluminated

fromabovewith a planewave, and are the electric fieldswithinthecubewhenilluminatedfrombelowwithapointsourceatthetargetlocation.Thephaseandamplitudeofthepointsourcearegivenbytheelectricfieldatthetargetlocationintheforwardsimulation.Wesimultaneouslycalculatethesensitivityformultiple

n !x( )

n0!x( ) = 1

2 nmax + nmin( )

f n !x( )( ) =!E !x0( )

2

dfdn

!x( )

dfdn!x( ) = 2n !x( )Re

!E fwd ⋅

!Eadj{ }

!E fwd

!Eadj

Fig.S1.Scatteringelementsaretailoredtoperformatargetfunctionthroughgradientdescent.Startingfromaninitialdesign,wecomputethesensitivityofthesortingefficiencytochangesintherefractiveindexusing a pair of FDTD simulations. The design is then updated inproportion to the sensitivity. After many iterations, the resultingstructurefocusesincidentlightwithhighefficiency.

This document provides supplementary information to "Multifunctional volumetric meta-optics for color and polarization image sensors," https://doi.org/10.1364/OPTICA.384228.

incidentwavelengthsandpolarizationsacrossthevisiblespectrum,assigningeachspectralbandtoadifferentquadrant:red(600nm–700nm)green(500nm–600nm)andblue(400nm–500nm).Then we use the spectrally-averaged sensitivity to update therefractiveindexofthedevice:

(S2)

Thestepsize fixedatasmallfraction(typically =0.001)toensurethatthechangeinrefractiveindexcanbetreatedasaperturbation in the linear regime.The sensitivity is recalculatedaftereachupdate.

II. Fabrication constraints

A. Binary index Duringtheoptimizationprocess,wedirectlyenforceasetof

constraintsontheindexdistributionasrequiredbythefabricationprocess. In particular,we are designing high-contrast scatteringelementsconstructed fromtwomaterials.Althoughthegradientdescentalgorithmdetailedaboveproducesoptimizeddeviceswithgradientindex,wecanenforcethebinaryconditionbyintroducinganauxiliarydensity rangingfrom[0,1].Thedensityisrelated

to the refractive index distribution via a sigmoidal projectionfilter[3]:

(S3)

wheretheparameter controlsthefilterstrength.Forsmall ,theindexdistributionisequaltothedensityscaledtotherangeofavailable refractive index. For large , the sigmoid filterapproximatesaHeaviside function,and the indexdistribution ispushedtowardeitherextreme. Importantly, the filter function iscontinuouslydifferentiable,suchthatthesensitivitycanbewrittenin termsof thedensity: .Duringoptimizationwe can

parameterizethedesignusingthedensity and ,gradually

increasingthestrengthofthefilter.Atearlystagesandsmall ,thisis equivalent to the unfiltered case. Over time as the strengthincreases, the optimized index distribution is gradually pushedtowardabinarydesign,evenasthedensityremainscontinuous.

B. Minimum feature size In addition to material constraints, device designs must

conformtotheresolutionlimitsimposedbythefabricationprocess.Forexample,diffractionandproximitydosingeffectslimitelectronbeamlithographytoapproximately10nmfeatures.Weenforcethisminimumfeaturesizefordevicedesignsbyintroducinga“dilated”density , which represents the maximum density

withinaneighborhood ofeachpoint [4].

(S4)

For a sufficiently large exponent 𝑀, this operation

approximates morphological grayscale dilation. However it iscontinuously differentiable with respect to the arguments.Therefore,thesensitivitycanbewrittenintermsoftheun-dilated

density: !"!#(𝒙&&⃗ )

= . The neighborhood is

taken to be a circle,where the radius represents theminimumfeaturesize.

When combined with the projection filter, the dilationoperation eliminates small features in the final binary indexdistribution.Duringoptimization,thedeviceisparameterizedbythedensity ,whiletheindexisdefinedbythedilateddensity

:

(S5)

The neighborhood is taken to be a circle, where the radiusrepresentstheminimumfeaturesize.

ni+1!x( ) = ni

!x( )+α dfλ

dn!x( )λ

∑

α α

ρ!x( )

n !x( ) = P ρ!x( )( )

=12+tanh 2βρ

!x( )−β( )

2tanh β( )

⎛

⎝

⎜⎜

⎞

⎠

⎟⎟nmax − nmin( )+ nmin

β β

β

dfdρ =

dfdn

dndρ

ρ!x( ) β

β

!ρ!x( ) ρ

!ʹx( )

Ω!x

!ρ!x( ) = D ρ

!x( )( ) = 1

Mρ!x '( )( )

M

Ω∑M

dfdρ !x( ) =

dfd !ρ !ʹx( )

d !ρ !ʹx( )dρ !x( )Ω

∑ Ω

ρ!x( )

!ρ!x( )

n !x( ) = P !ρ !x( )( ) = P D ρ!x( )( )( )

Ω

Fig.S2.Thedesignofoptimizedscatteringelementsconsistsofiterativeupdatestotherefractiveindexdistribution,guidedbythesensitivitytoperturbationsintheindex,andconstrainedbymaterialsandfabricationprocessesasoutlinedabove.

2

C. Connected layered designs Someof the device designs are intended for fabrication by

multi-layer2D lithography, consistingof severalpatterned slabsthatareinvariantintheverticaldirection(Fig.2,3).Inthiscase,werestricttheoptimizationbyaveragingthecalculatedsensitivityintheverticaldirectionwithineachlayer.Ineffect,voxelswithineachlayeraregovernedbyashared2Dprofile.

For the microwave device (Fig. 3), the design is furtherconstrainedsothateachlayerisfullyconnectedwithnofloatingpieces. We directly impose connectivity by periodically addingbridges between disconnected islands within each layer. Thisinterventiondoesnot takesensitivity intoaccount, and typicallycausesasmalldecreaseindeviceperformance.Therefore,weonlyapplytheconnectivityconstraintonceper40iterations,allowingtheperformancetorecoverthereafter.

III. Sorting efficiency of 3D printed designThefreeformscatteringelementshowninFig.1isdesignedfor

high-resolution 3D printing. It consists of amonolithic polymercube(index=1.56),infiltratedbyaseriesofholes(Fig.S3a).Aswithall the designs presented in this manuscript, the device sortsincidentlighttodifferentlocationsdependingonthewavelengthandpolarization.Hereweshowthesortingefficiencyunderlinearlypolarizedillumination.Thevisiblespectrumisdividedintothreespectralbands(400nm–500nm,500nm–600nm,600nm–700nm),eachfocusedonadifferentquadrantinthefocalplane(Fig.S3b).Weimposemirrorsymmetryalongadiagonalplanebisectingtheredandbluequadrants.Thisguaranteesthatthegreenspectralbandwill focus orthogonal polarizations to opposite quadrants,while the red and blue spectral bands will be polarizationindependent.Fig.S3cshowsthesortingefficiencyforeachquadrantacross the visible spectrum, defined as the fraction of incidentoptical power that reaches the target quadrant. This designachievesgreaterefficiencythansimilar layereddevices(Fig.2d),whichimposestrongerrestrictionsonthedevicetopology.

IV. Dependence on device thicknessEfficientmultifunctional devices require a large number of

opticaldegreesof freedom,whichscaleswith theoveralldevicethickness.Inthecaseofspectral-andpolarizationsortingelements,thisrestrictstheperformanceofultrathindevices.Hereweshowhow complex sorting behavior emerges with increased devicethickness.Fig. S4 shows thesortingefficiencyof several layereddevices (similar to Fig. 2), optimized with overall thicknessesranging from 400 nm – 2000 nm. Thin devices are unable toeffectivelysortincidentlight.Asthicknessincreases,theefficiencyimproveswhile crosstalk isminimized. Webelieve that addingevenmorelayersthanshownherewillincreasetheperformanceoftheoptimizeddevicealthoughnot indefinitely. Theeffectof thenumberof layersusedaswellas the thicknessof the individuallayersisaninterestingtopicforfutureinvestigation.

V. Dependence on incident angleThedevicespresented in thismanuscriptweredesigned to

sort light at normal incidence. In realistic imaging systems, theillumination will span a range of angles that depends on thenumericalapertureof the imagingoptics.Forexample,a typicalsmartphone camera[5] has a numerical aperture of NA = 0.21,corresponding to an acceptance cone spanning ±7.8° in SiO2. In

Fig. S4. Sorting efficiency for increasing device thickness.EfficiencyspectraforthemultilayerdesignsshowninFig2,optimizedfor1–5devicelayers.Eachplotcorrespondstoadifferentfocalregion,includingthe cross-polarized sub-pixel shown in yellow.Darker lines indicateincreasingdevicethickness.

Fig.S3.Sortingefficiencyof3Dprintedscatteringelement.(a)Deviceconsistsofasinglepieceofpolymer,asshowninFig.1.Lightincidentfrom above is focused to different quadrants depending on thepolarizationandfrequency.(b)Intensitydistributionatthefocalplaneunderbroadband,linearlypolarizedillumination.Colorscorrespondtotheobservedhueforvisiblelight.Theupperrightquadrantisassignedtocross-polarizedillumination,andthereforeremainsdark.(c)Sortingefficiencyforeachofthefourfocalregions,definedasthefractionofincidentlightfocusedintothetargetquadrant.

3

ordertoevaluatethedeviceperformanceunderrealisticconditions,wehavecharacterizedthesortingefficiencyofthelayereddesign(shown in Fig. 2) for different illumination angles. As the angledeviatesfromthesurfacenormaltheefficiencydecreases(Fig.S5).Thisisduetoanincreaseinscatteringintonon-targetedquadrantswithinthefocalplane,aswellasreflectionfromthesurface.Thedrop-off inperformancewith incidentangle isaconsequenceofdecorrelation in the angular scattering through disorderedmedia[6],whichscaleswiththethicknessofthescatteringmedium.Since the device maintains high performance near normalincidence, we define the steepest functional angle where theperformancedropstohalfofthemaximumefficiency,roughly±8°.Therefore,thisdesignmaybedeployedincolorandpolarizationfiltersinmobileimagingsystems.Therangeofincidenceanglescanbeextended to steeperanglesby reducing thedevice thickness.Furthermore,wemaydirectlyincludenon-normalincidenceintothedesignalgorithm[2],whichhassofaronlyconsiderednormalincidence.

VI. Microwave near field measurement WesamplethescatteredmicrowavenearfieldintheKaband

usingascanningantenna.AnincidentGaussianbeam(FWHM=25mm) isgeneratedbyavectornetworkanalyzercoupled to freespaceviaamicrowavehornantennaandfocusingmirror(Fig.S6).Theinputbeampassesthroughthecube,scatteringintothefarfield.Wesamplethelocalelectricfieldatameasurementplane62mmbeyondtheoutputapertureofthedeviceusingaWR-28waveguideflangeinordertorecoverthecomplexscatteringamplitudeS21.Byscanning thepositionof theprobeantenna in themeasurementplane,wecanmeasurethelocalelectricfieldprofile.Weapplyadeconvolutionfiltertoaccountfortheanisotropyoftheprobeandthe finite aperture of the device. The complex fields arecomputationallyback-propagatedtothefocalplaneofthedevice.This analysis is repeated for a range ofmicrowave frequencieswithin the Ka band (26–40 GHz), and for both orthogonalpolarizations of the input beam. To measure the scatteringparametersforanorthogonalpolarization,werotatethedevice90°.We also measure the cross-polarized fields by rotating thewaveguideflange90°.

VII. Power flow analysis Weexperimentallyvalidatethedesignofvolumetricscattering

elementsusingamicrowaveanalogbymeasuring the scatteredfields.Thisanalysisrevealscloseagreementbetweensimulatedandmeasured sorting efficiencies, roughly 40% across themeasurementspectrum.However,theperformanceislowerthanpreviouslydesigneddevices(Fig.1and2).Hereweaccountfortheremaining energy that is lost in the system. From full-wavesimulationswecanquantifyfoursourcesofloss:reflectionfromthedevice,obliquescatteringawayfromthefocalplane,andabsorptionwithinthematerial(Fig.S7a).Bothreflectionandobliquescatteringcontributeequallytotheoveralllossinthesystem,independentofwavelength (Fig. S7b). We believe that these losses can besignificantly mitigated by explicitly including them in theoptimization algorithm, or through the appropriate choice ofboundary conditions and anti-reflection coatings. Due to theagreementbetween simulatedandexperimental efficiencies,weconcludethatabsorptionplaysanegligiblerole,whichisconsistentwiththeextremelylowextinctioncoefficientofpolypropyleneintheKaband[7].

Fig. S5. Sorting performance at non-normal incidence. The layereddevice is optimized to sort light incident from above. The sortingperformancedegradesastheincidentangleincreases.Comparedtothedesignatnormalincidence,theefficiencyisreducedbyhalfatanglesbeyond±8°.

Fig. S6. Experimental characterization at microwave frequencies.Microwave device consists of 20 patterned polypropylene sheets,assembled into a cube. The cube is illuminated by a collimatedmicrowave source operating in theKa band (26.5 – 40GHz). Localelectricfieldsarecollectedbyascanningnear-fieldprobe,andback-propagatedtothefocalplane.

4

References 1. T. D. Gerke and R. Piestun, "Aperiodic volume optics," Nat.

Photonics 4, 188–193 (2010). 2. D. Sell, J. Yang, S. Doshay, R. Yang, and J. A. Fan, "Large-Angle,

Multifunctional Metagratings Based on Freeform Multimode Geometries," Nano Lett. 17, 3752–3757 (2017).

3. F. Wang, B. S. Lazarov, and O. Sigmund, "On projection methods, convergence and robust formulations in topology optimization," Struct. Multidiscip. Optim. 43, 767–784 (2011).

4. J. K. Guest, J. H. Prévost, and T. Belytschko, "Achieving minimum length scale in topology optimization using nodal design variables and projection functions," Int. J. Numer. Methods Eng. 61, 238–254 (2004).

5. N. A. Switz, M. V. D’Ambrosio, and D. A. Fletcher, "Low-Cost Mobile Phone Microscopy with a Reversed Mobile Phone Camera Lens," PLoS One 9, e95330 (2014).

6. I. Freund, M. Rosenbluh, and S. Feng, "Memory effects in propagation of optical waves through disordered media," Phys. Rev. Lett. 61, 2328–2331 (1988).

7. J. Krupka, "Measurements of the Complex Permittivity of Low Loss Polymers at Frequency Range from 5 GHz to 50 GHz," IEEE Microw. Wirel. Components Lett. 26, 464–466 (2016).

Fig.S7.Power flowanalysisofmicrowavedevices. (a)The incidentmicrowave beam is deflected into different regions, contributing tolower overall sorting efficiency. (b)Distribution of optical power insimulations,includingtransmission,reflection,andobliquescattering.Thepowerisnormalizedtotheinputapertureofthedevice.

5