All-Chalcogenide Programmable All-Optical Deep Neural Networks

All-Chalcogenide Programmable All-Optical Deep Neural Networks | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article All-Chalcogenide Programmable All-Optical Deep Neural Networks Ting Yu, Xiaoxuan Ma, Ernest Pastor, Jonathan George, Simon Wall, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-259851/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Deeplearning algorithms are revolutionising many aspects of modern life. Typically, they are implemented in CMOS-based hardware with severely limited memory access times and inefficient data-routing. All-optical neural networks without any electro-optic conversions could alleviate these shortcomings. However, an all-optical nonlinear activation function, which is a vital building block for optical neural networks, needs to be developed efficiently on-chip. Here, we introduce and demonstrate both optical synapse weighting and all-optical nonlinear thresholding using two different effects in one single chalcogenide material. We show how the structural phase transitions in a wide-bandgap phase-change material enables storing the neural network weights via non-volatile photonic memory, whilst resonant bond destabilisation is used as a nonlinear activation threshold without changing the material. These two different transitions within chalcogenides enable programmable neural networks with near-zero static power consumption once trained, in addition to picosecond delays performing inference tasks not limited by wire charging that limit electrical circuits; for instance, we show that nanosecond-order weight programming and near-instantaneous weight updates enable accurate inference tasks within 20 picoseconds in a 3-layer all-optical neural network. Optical neural networks that bypass electro-optic conversion altogether hold promise for network-edge machine learning applications where decision-making in real-time are critical, such as for autonomous vehicles or navigation systems such as signal pre-processing of LIDAR systems. Photonics/optics Electrical Engineering Deep-learning Algorithms Nonlinear Activation Function Optical Synapse Weighting Photonic Memory Resonant Bond Destabilization Figures Figure 1 Figure 2 Figure 3 Figure 4 Full Text Due to technical limitations, full-text HTML conversion of this manuscript could not be completed. However, the latest manuscript can be downloaded and accessed as a PDF. Additional Declarations There is NO Competing Interest. Supplementary Files SOMAONN221v3.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-259851","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":14884128,"identity":"0fe4dfd8-5698-4d97-bf7a-d3df3af03c08","order_by":0,"name":"Ting Yu","email":"","orcid":"","institution":"Singapore University of Technology and Design","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Ting","middleName":"","lastName":"Yu","suffix":""},{"id":14884129,"identity":"5a5f5126-da8d-450e-baad-512f8c2af083","order_by":1,"name":"Xiaoxuan Ma","email":"","orcid":"","institution":"George Washington University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xiaoxuan","middleName":"","lastName":"Ma","suffix":""},{"id":14884130,"identity":"d6517c51-a878-41d0-ad38-4b73ebf1c5e7","order_by":2,"name":"Ernest Pastor","email":"","orcid":"https://orcid.org/0000-0001-5334-6855","institution":"ICFO-The Institute of Photonic Sciences","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Ernest","middleName":"","lastName":"Pastor","suffix":""},{"id":14884131,"identity":"6dea9a76-4719-4fe8-96b2-08d0edf1b570","order_by":3,"name":"Jonathan George","email":"","orcid":"","institution":"George Washington University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Jonathan","middleName":"","lastName":"George","suffix":""},{"id":14884132,"identity":"4d6db3e6-77c9-4dad-a3b6-c1a59648271f","order_by":4,"name":"Simon Wall","email":"","orcid":"https://orcid.org/0000-0002-6136-0224","institution":"Aarhus University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Simon","middleName":"","lastName":"Wall","suffix":""},{"id":14884133,"identity":"8e010c01-d69c-4148-9c1d-b808ae2fb454","order_by":5,"name":"Mario Miscuglio","email":"","orcid":"","institution":"George Washington University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Mario","middleName":"","lastName":"Miscuglio","suffix":""},{"id":14884134,"identity":"f9b5095e-b3ac-4721-a8e0-fc52db7b84ec","order_by":6,"name":"Robert Simpson","email":"","orcid":"https://orcid.org/0000-0002-3499-4950","institution":"Singapore University of Technology and Design","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Robert","middleName":"","lastName":"Simpson","suffix":""},{"id":14884135,"identity":"9e1f4eb5-1863-4831-9e1e-76ff31a83ae4","order_by":7,"name":"Volker Sorger","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAs0lEQVRIiWNgGAWjYPACGwjFQ4KWNNK1HCZBi3x77+MXH/ect9twI4Hxwds2IrQYnDluZjnj2e1koBZmw7lEaZFIYzPmOXA72exGAps0LzFa5GcAtfw5cA6khf03UVoYbqQxP2Y4cMAOZAszUVoMzhxjY+w5kJxgf+Zhs+Scc8Q4rL2N+cOPA3b2ku3JBz+8KSPGYQwMbBJAIrGBgbGBOPVAwPwBSNgTrXwUjIJRMApGHgAA5Sg41oKQHOEAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-5152-4766","institution":"George Washington University","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Volker","middleName":"","lastName":"Sorger","suffix":""}],"badges":[],"createdAt":"2021-02-20 18:00:31","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-259851/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-259851/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":6800454,"identity":"ac177b73-d714-4605-8ebc-e48ecad19d5e","added_by":"auto","created_at":"2021-03-10 15:07:53","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":298403,"visible":true,"origin":"","legend":"a, An artificial photonic neuron based on structural and electronic transitions in phase change materials. Ultrafast photoexcitation forms a transient state by removing the resonant bonds and changing the optical properties. Rapid energy extraction (using silicon slab waveguide as heat-sink) prevents amorphization allowing for a volatile transition. Successive heating, after several picosecond, causes lattice heating, which thermally melts the GST. b, Schematic of the artificial photonic neuron which comprises an photonic programmable phase-change array (P3A), programmed according to quantized synaptic weights, and all-optical nonlinear activation function (NLAF) based on phase change materials (PCM). Inputs at 800 nm laser source are linearly combined according to a cascade of programmable switches (SiN-based photonics platform), the resulting optical power is amplified, e.g. using erbium doped fiber amplifier (EDFA), and used as pump signal for the optical NLAF module which uses a 1550 nm probe signal. The output of the NLAF is injected in the Si-based photonic layer. c, Weighting mechanism and summation rely on a cascade of Sb2S3-SiN hybrid photonic switches which performs the same operation as an optical equivalent of an FPGA (Field Programmable Gate Array). The PCM is placed on one of side of the directional coupler (DC), and according to the portion of material written (crystalline to amorphous) the light is partially configured to the cross state. The phase of the film is thermally set, with a resolution of ~1 µm yielding to a total of 16 distinguishable states (4-bit). Fully amorphous PCM film configure light in a cross state of the DC switch, while fully crystalline results in a bar state. The hybrid configurations in between bard- and cross states represent the intermediate states. d, The NLAF module consists of a single mode hybrid silicon waveguide covered for 1 µm by a 30 nm thick Ge2Sb2Te5 layer. A 1550 nm TM polarized light is used for sensing the nonlinear variation of the effective refractive index induced by the fs-laser coupling into the PIC e.g., from free space. Plasmonic antennas could be used for enhancing light matter interaction. After the NLAF the signal could be either detected using integrated photodetector or passed to a second stage photonic waveguide. ","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-259851/v1/a61aa010922d58118d36b293.png"},{"id":6800109,"identity":"6ac60f7f-bc58-4019-8bad-d88e20ac39b8","added_by":"auto","created_at":"2021-03-10 15:04:53","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":370335,"visible":true,"origin":"","legend":"Optical-equivalent of an electro-thermally controlled Programmable Photonic Phase-change Array (P3A) based on photonic nonvolatile memory on-chip. a) Schematic representation of the silicon nitride asymmetrical directional coupler (DC) operating at 800 nm. The DC is composed of two waveguides, a bare silicon nitride waveguide (SW, 500 nm width) and a hybrid Sb2S3-on-silicon nitride hybrid waveguide (HW, 420 nm width), which are separated by a 120 nm gap. The thickness of the Sb2S3 is 20 nm. A free space laser source can be used for writing portion of the Sb2S3 film deposited on top of the hybrid waveguide with a writing resolution of 1 µm. According to the percentage of the portion of the Sb2S3 layer that is amorphized, the percentage of light which switches to the HW changes proportionally. Numerical approach used Comsol Multiphysics. ","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-259851/v1/6116ca6af1bd71aec7c2c107.png"},{"id":6800456,"identity":"dadec6ab-71b1-40c7-b34b-4771e70d58d4","added_by":"auto","created_at":"2021-03-10 15:07:53","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":493709,"visible":true,"origin":"","legend":"Nonlinear activation function (NLAF) of photonic neurons. a, Pump-probe physics and temporal\nresponse of the all-optical nonlinear activation function based on resonant bond-state depopulation in GST thin\nfilms. Schematic of the pump probe setup. The 800 nm pump bears the weighted-addition (MAC) signal of 4-bit\nresolution from the upstream photonic perceptron (Fig. 1\u00262). Pump: fs-laser 35 fs pulses, λ = 800 nm. Probe:\nλ = 1550 nm. b, Peak change in dielectric function after photoexcitation for different pump fluences. Blue\ntriangles represent repeat-measurements after raising the fluence to over 17 mJcm-2, demonstrating no permanent\nchange is induced in the sample. c, Time traces of the dielectric function showing that the majority of the changes\nrecover in the first few picoseconds as the resonant bonding state reforms, enabling for rapid volatile NLAF\nfunctionality on-chip at near-zero delay when performing inference tasks. The shaded area indicates the change\nin extracted dielectric function assuming a range of initial thicknesses for the GST layer, from 24-30 nm. d,\nNonlinear activation function (NLAF) based on an electro-absorptive pump and probe scheme The electronic\ntransition in the PCM, heterogeneously integrated on top a SiN provides the all-optical nonlinearity. The pump\nsignal, i.e. the P3A outputs modulated according to the previous weights, is coupled from free-space, amplified\nand focused onto each GST segments of the NLAF modules. e, Transmittance of the 1550 nm probe signal as\nfunction of the pump-signal fluence (result of the weighted addition of the optical interference unit). Insets show\nthe normalized electric field distribution for a propagating hybrid TM mode in the xz and yz planes when the\ndielectric constant of the GST film is (i.) in the crystalline phase and (ii.) has electronically transitioned due to the\ndepletion of resonant bonds induced by the fs pump laser. f, Crystallisation of GST. The change in optical\nreflectivity with temperature shows a logistic function is differentiated to determine the crystallisation temperature\n(see Methods and SOM). Inset: full temperature range, from room temperature to 250 ℃.","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-259851/v1/3ea08a14eaad0ee2a692386c.png"},{"id":6800455,"identity":"6977406b-bc3e-49ef-bbdc-9aba226e394d","added_by":"auto","created_at":"2021-03-10 15:07:53","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":225695,"visible":true,"origin":"","legend":"All-Photonic Deep Neural Network Performance. a, Schematic of the fully connected network\ncomposed by 2 layers of 100 neurons. b, Dataflow graph of the NN which comprises 100 neuros and 2 fully\nconnected layers. c, Accuracy results for the inference on unseen data for NN trained with 2% of Gaussian noise.\nThe evaluation of the effect of NLAF module and synaptic quantized weights on inference accuracy. Soft max\noperation is considered to be performed electronically. d, The speed limit of this all-photonic deep neural network\ndepends on the thermal transients of GST films providing the neuron’s nonlinearity, (assuming a fixed and\nprogrammed kernel. The temperature values are extracted from literature23 and the cooling behavior was simulated\nusing the finite element method on COMSOL Multiphysics.","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-259851/v1/b5f0f9d94ed7d91b3953baf4.png"},{"id":13602681,"identity":"022cb008-ab1c-40bf-810c-30441a5e0c93","added_by":"auto","created_at":"2021-09-17 05:52:27","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2701117,"visible":true,"origin":"","legend":"Article File","description":"","filename":"AONN221v4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-259851/v1_covered.pdf"},{"id":6800698,"identity":"22bc4bf0-c452-425e-8051-252b0a345b38","added_by":"auto","created_at":"2021-03-10 15:13:58","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4240078,"visible":true,"origin":"","legend":"Article File","description":"","filename":"AONN221v4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-259851/v1_stamped.pdf"},{"id":6800551,"identity":"3fa9f573-d3bd-4c3d-aaa9-0f4dbe543bbe","added_by":"auto","created_at":"2021-03-10 15:10:53","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":1181113,"visible":true,"origin":"","legend":"","description":"","filename":"SOMAONN221v3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-259851/v1/dcba9a791be33845583d2a21.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"All-Chalcogenide Programmable All-Optical Deep Neural Networks","fulltext":[{"header":"Full Text","content":"Due to technical limitations, full-text HTML conversion of this manuscript could not be completed. However, the latest manuscript can be downloaded and \u003ca href='/article/rs-259851/latest.pdf' target='_blank'\u003e accessed as a PDF.\u003c/a\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Deep-learning Algorithms, Nonlinear Activation Function, Optical Synapse Weighting, Photonic Memory, Resonant Bond Destabilization","lastPublishedDoi":"10.21203/rs.3.rs-259851/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-259851/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Deeplearning algorithms are revolutionising many aspects of modern life. Typically, they are implemented in CMOS-based hardware with severely limited memory access times and inefficient data-routing. All-optical neural networks without any electro-optic conversions could alleviate these shortcomings. However, an all-optical nonlinear activation function, which is a vital building block for optical neural networks, needs to be developed efficiently on-chip. Here, we introduce and demonstrate both optical synapse weighting and all-optical nonlinear thresholding using two different effects in one single chalcogenide material. We show how the structural phase transitions in a wide-bandgap phase-change material enables storing the neural network weights via non-volatile photonic memory, whilst resonant bond destabilisation is used as a nonlinear activation threshold without changing the material. These two different transitions within chalcogenides enable programmable neural networks with near-zero static power consumption once trained, in addition to picosecond delays performing inference tasks not limited by wire charging that limit electrical circuits; for instance, we show that nanosecond-order weight programming and near-instantaneous weight updates enable accurate inference tasks within 20 picoseconds in a 3-layer all-optical neural network. Optical neural networks that bypass electro-optic conversion altogether hold promise for network-edge machine learning applications where decision-making in real-time are critical, such as for autonomous vehicles or navigation systems such as signal pre-processing of LIDAR systems.","manuscriptTitle":"All-Chalcogenide Programmable All-Optical Deep Neural Networks","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2021-03-10 15:04:51","doi":"10.21203/rs.3.rs-259851/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"30f78c5f-6bed-42d9-8afd-8c9734e34e44","owner":[],"postedDate":"March 10th, 2021","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":2872404,"name":"Photonics/optics"},{"id":2872405,"name":"Electrical Engineering"}],"tags":[],"updatedAt":"2021-04-23T14:55:44+00:00","versionOfRecord":[],"versionCreatedAt":"2021-03-10 15:04:51","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-259851","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-259851","identity":"rs-259851","version":["v1"]},"buildId":"WrCJVZZCHTDjtuVLN7oU0","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}