Publication:

Learned Manifold Reparameterization for DRC-Compatible Inverse Design in Silicon Photonics

Date

 
cris.virtual.department#PLACEHOLDER_PARENT_METADATA_VALUE#
cris.virtual.orcid#PLACEHOLDER_PARENT_METADATA_VALUE#
cris.virtualsource.departmentd1dac5b2-325d-4956-a26b-9583d552bac4
cris.virtualsource.orcidd1dac5b2-325d-4956-a26b-9583d552bac4
dc.contributor.authorDanis, Bahrem Serhat
dc.contributor.authorVit, Aycan Deniz
dc.contributor.authorDasdemir, Ahmet Onur
dc.contributor.authorDesdemir, Demet Baldan
dc.contributor.authorAkcakoca, Enes
dc.contributor.authorYanmaz, Zeynep Ipek
dc.contributor.authorPolat, Gulzade
dc.contributor.authorAydogan, Aytug
dc.contributor.authorMagden, Abdullah
dc.contributor.authorMagden, Emir Salih
dc.date.accessioned2026-11-03T12:29:03Z
dc.date.available2026-11-03T12:29:03Z
dc.date.createdwos2026
dc.date.issued2026
dc.description.abstractWe recast fabrication-compatible photonic inverse design as an unconstrained optimization problem on a learned low-dimensional manifold, intrinsically eliminating design-rule violations. A deep generative network is used to simultaneously enforce minimum feature size, binary material composition in an SOI platform, and CMOS design-rule checks. Optimizations performed in this resulting latent space are mapped onto the full space of manufacturable device geometries. This procedure converts length-scale-constrained topology optimization into a standard stochastic gradient problem, removing the need for post-processing or device correction steps. We demonstrate separate generator models for electron-beam lithography (60 nm minimum feature) and photolithography (150 nm minimum feature), and show that the same framework scales across fabrication processes by adjusting kernel configurations and upsampling ratios. We validate this approach by designing three compact device classes using FDFD-based optimizations: broadband power splitters with 50/50, 30/70, and 10/90 splitting ratios (losses between 0.05–0.12 dB; 4.2 × 4.2 μm2), a wavelength demultiplexer (0.25 dB loss, 16 dB extinction; 7×7 μm2), and a TE0-to-TE1 mode converter (0.03 dB loss; 2.1×2.1 μm2). This approach establishes a scalable, process-agnostic backbone for fabrication-aware integrated photonic design.
dc.description.wosFundingTextThis work is supported by the Scientific and Technological Research Council of Turkey (TUBITAK) under grant number 122E214.
dc.identifier.doi10.1117/12.3099627
dc.identifier.isbn979-8-9023-2167-5
dc.identifier.issn0277-786X
dc.identifier.urihttps://imec-publications.be/handle/20.500.12860/60714
dc.language.isoeng
dc.provenance.editstepusergreet.vanhoof@imec.be
dc.publisherSPIE-INT SOC OPTICAL ENGINEERING
dc.source.beginpage1410407
dc.source.conferenceMACHINE LEARNING IN PHOTONICS II
dc.source.conferencedate2026-04-12
dc.source.conferencelocationStrasbourg
dc.source.journalMACHINE LEARNING IN PHOTONICS II
dc.source.numberofpages7
dc.title

Learned Manifold Reparameterization for DRC-Compatible Inverse Design in Silicon Photonics

dc.typeProceedings paper
dspace.entity.typePublication
imec.internal.crawledAt2026-09-07
imec.internal.sourcecrawler
imec.internal.wosCreatedAt2026-09-07
Files
Publication available in collections: