Glass-Substrate Photonic Interposer Overview

• Glass-substrate photonic interposers are glass-supported optical redistribution layers that enable routing, fanout, pitch conversion, and wavelength handling between fibers and chiplets.
• They encompass diverse architectures including fused-silica optical interposers with buried waveguides, glass-supported nanocavities, and deposited-glass PIC platforms for WDM applications.
• They address critical challenges such as managing vertical asymmetry, achieving sub-micron packaging alignment, and implementing multilayer redistribution for high-density, low-loss performance.

A glass-substrate photonic interposer can be understood as a glass-supported optical redistribution layer that performs routing, fanout, pitch conversion, wavelength handling, vertical redirection, or chiplet interfacing between fibers, photonic chips, and heterogeneous electronic-photonic assemblies. In the present literature, the term spans direct fused-silica optical interposers with buried waveguides, embedded mirrors, and self-alignment sockets; glass-supported resonant nanophotonic blocks whose cladding geometry is part of the optical design; deposited-glass waveguide platforms for low-loss routing and resonant filtering; glass-carried printed chiplet couplers; and panel-scale proposals in which a single glass package substrate hosts a reconfigurable WDM optical network (Djogo et al., 1 Dec 2025, Kawata et al., 2023, Moss et al., 2014, Huang et al., 2024, Hsueh et al., 8 Aug 2025).

1. Architectural scope and typology

The most direct realization of a glass-substrate photonic interposer is a monolithic fused-silica optical routing chip inserted between external fibers and a silicon photonic die. In “Laser Structured Optical Interposer for Ultra-dense Vertical Coupling of Multi-core Fibers to Silicon Photonic Chip” the entire optical redistribution function is implemented inside a 1 mm thick polished fused-silica wafer: six 7-core MCFs are converted into a 4 × 10 silicon-photonic grating-coupler region by buried waveguides, embedded TIR micro-mirrors, and FLICE-opened fiber sockets, yielding 40 optical channels with an average single-pass insertion loss of 5.0 dB for the fully packaged system, and a minimum loss of 3.2 dB (Djogo et al., 1 Dec 2025).

A second architectural class uses glass primarily as a host or carrier for integrated photonic functions rather than as a fully monolithic routing medium. The clearest examples are the GaAs 2D photonic crystal slab nanocavity on glass with an upper glass thin film, where the glass environment determines cavity loss channels, and the 3D-nanoprinted interposer on a silica/glass carrier, where printed free-space reflective couplers, chip-coupling frames, and fiber-guiding funnels bridge silicon and InP chiplets without changing the foundry process flow of the constituent dies (Kawata et al., 2023, Huang et al., 2024).

A third class is architectural and prospective: glass is treated as a package-scale substrate that can host a photonic backplane. The panel-scale proposal defines two forms, a PIC-embedded glass interposer and a monolithic-photonics-integrated glass interposer, both aimed at a single large glass package/interposer substrate carrying chiplets, HBM stacks, WDM waveguides, optical crossbar switches, and comb sources over areas up to 500 mm × 500 mm or larger (Hsueh et al., 8 Aug 2025).

Adjacent work broadens, but also sharpens, the definition. The polymer-on-glass photonic-crystal spectrometer is best treated as substrate-guided in-plane light distribution with lithographically defined, spatially localized wavelength-selective outcouplers, not as a core interposer; the Hydex papers present a high-index doped silica glass CMOS compatible platform with low-loss routing and high-Q resonators, but not a complete interposer; the multilayer SiN microcomb interposer and the dual-layer SiN 3D interposer are not glass-native, yet they formalize the passive-backplane and multilayer-routing roles that a glass substrate may assume (Pervez et al., 2010, Duchesne et al., 2015, Rao et al., 2020, Xia et al., 14 Apr 2026).

2. Materials systems and structural implementations

Several distinct material stacks recur in the literature, and they imply different interposer roles.

Platform	Stack	Reported figures
Fused-silica optical interposer	Corning 7980 polished fused silica, 1 mm thick, with laser-written waveguides, FLICE mirrors, and sockets	40 optical channels, six 7-core MCFs, 5.0 dB average single-pass packaged loss
Glass-supported PhC resonator	GaAs slab, $h = 130\ \mathrm{nm}$, on glass $n=1.45$ with upper glass film	$Q \sim 10^4$ bare on glass; $Q = 4.2\times10^5$ at $t \approx 260\ \mathrm{nm}$
Hydex deposited-glass PIC platform	High-index doped silica glass core $n = 1.7$ in SiO$_2$	17% contrast; $0.04\ \mathrm{dB/cm}$ propagation loss; $Q = 1.2\times10^6$ ring
Polymer-on-glass PhC slab guide	396 nm PMMA on 22 mm × 22 mm, 0.96 mm glass coverslip	450–700 nm operation; $3\times3$ array of PhC patterns

The fused-silica interposer uses femtosecond-laser-induced refractive-index modification for buried waveguides and femtosecond laser irradiation followed by chemical etching (FLICE) for hollow micro-optical structures. Its mirrors are elliptical disks, about 30 $n=1.45$0m tall and $n=1.45$1m wide, tilted 50° with respect to the waveguides to induce TIR, while the waveguides occupy four depth layers centered nominally at 50, 140, 230, and 320 $n=1.45$2m (Djogo et al., 1 Dec 2025).

The glass-supported nanocavity is structurally much smaller and much more sensitive to vertical refractive-index symmetry. It consists of a GaAs 2D photonic crystal slab nanocavity on glass with an additional upper glass thin film, using a W0.9 line-defect waveguide and a multistep heterostructure created by local elongation of the lattice constant in the $n=1.45$3-direction by +2.5% and +1.25% for designated holes. In air-clad conditions the same cavity supports $n=1.45$4 and mode volume $n=1.45$5 near $n=1.45$6 (Kawata et al., 2023).

Hydex represents a different strategy: the interposer-like function is realized in a deposited-glass waveguide circuit rather than in bulk glass machining. One paper reports a core refractive index $n=1.45$7, 17% core-cladding contrast, $n=1.45$8 propagation loss, $n=1.45$9 fiber coupling, and ring resonators with $Q \sim 10^4$0 and 200 GHz FSR (Moss et al., 2014). A related Hydex review reports propagation loss as low as $Q \sim 10^4$1, fiber pigtail / coupling loss about $Q \sim 10^4$2, and negligible bending losses for radii down to $Q \sim 10^4$3, emphasizing long routing paths and compact resonant functions in a silica-derived platform (Duchesne et al., 2015).

Glass can also act as a low-index slab guide supporting local patterned functions. In the polymer-on-glass spectrometer, the main waveguiding body is a glass coverslip with $Q \sim 10^4$4, while a PMMA overlayer with $Q \sim 10^4$5 is patterned into square-lattice PhC patches whose lattice constants span 300 to 420 nm in the final device. This is a weak-index-contrast substrate-guided system rather than a channelized interconnect, but it demonstrates that local wavelength-selective extraction can be integrated directly on glass (Pervez et al., 2010).

3. Optical physics specific to glass-supported interposers

A central optical problem for high-confinement devices on glass is vertical asymmetry. The GaAs PhC study makes this explicit: on glass with air above, the refractive-index environment is asymmetric, so the intended TE-like cavity mode hybridizes with TM-like slab modes. The paper identifies the loss chain as TE–TM mode conversion, followed by lateral leakage because the TM-polarized light does not feel any photonic bandgap effect. In the multistep heterostructure cavity this collapses the cavity from $Q \sim 10^4$6 in air to $Q \sim 10^4$7 for the bare on-glass case. Adding a top glass film suppresses this asymmetry-driven conversion, and the optimized thickness $Q \sim 10^4$8 yields $Q \sim 10^4$9, which is 1.7× higher than the thick symmetric-glass case that saturates near $Q = 4.2\times10^5$0 (Kawata et al., 2023).

That result also corrects a common simplification: a perfectly thick symmetric top cladding is not always the optimum on glass. The same paper attributes the improvement at intermediate $Q = 4.2\times10^5$1 to mode mismatch between the cavity’s radiation and the optical modes supported by the upper glass film. In the authors’ decomposition, lateral leakage dominates for $Q = 4.2\times10^5$2, bottom leakage is relatively insensitive to $Q = 4.2\times10^5$3, and top leakage exhibits a distinct dip around $Q = 4.2\times10^5$4. This establishes that in glass-supported nanophotonics the overglass is part of the optical design, not merely encapsulation (Kawata et al., 2023).

A second, very different, glass-specific optical regime is multimode slab transport with local extraction. In the photonic-crystal spectrometer, light is edge-coupled into a 0.96 mm thick glass slab and scattered upward by reciprocal lattice vectors $Q = 4.2\times10^5$5 from small PhC patches. The strongest extraction occurs when $Q = 4.2\times10^5$6, giving the design rule $Q = 4.2\times10^5$7. The architecture therefore uses different lattice constants to realize different spectral response functions and reconstructs a spectrum through the matrix equation $Q = 4.2\times10^5$8 and a Moore–Penrose pseudoinverse (Pervez et al., 2010). This suggests that a glass interposer need not be restricted to point-to-point channels; it can also host spatially distributed monitor taps or wavelength-selective pickoff sites.

A third optical regime is low-loss, moderate-contrast deposited-glass routing. Hydex is explicitly framed as promising for telecommunications and on-chip WDM optical interconnects for computing, with negligible nonlinear absorption in the telecom band, $Q = 4.2\times10^5$9 in one report, and $t \approx 260\ \mathrm{nm}$0 propagation loss (Moss et al., 2014). The related review gives $t \approx 260\ \mathrm{nm}$1, $t \approx 260\ \mathrm{nm}$2, $t \approx 260\ \mathrm{nm}$3, and operation with no detectable multiphoton absorption up to $t \approx 260\ \mathrm{nm}$4 (Duchesne et al., 2015). For interposer use, a plausible implication is that the dominant value of such glass platforms lies less in their nonlinear demonstrations than in their passive routing, WDM filtering, and fiber-compatible I/O.

4. Routing, fanout, multilayer redistribution, and wavelength handling

At the package level, the fused-silica optical interposer is the clearest demonstration of what a glass-substrate photonic interposer does operationally. It accepts six 7-core MCFs, routes 40 channels through a 4 to 6 mm fanout, uses four optical depth layers, redirects light with embedded TIR mirrors, and launches toward SiP gratings at 10° off vertical. The design space includes three routing topologies, with Design C emerging as the best compactness-loss compromise because crossings at 80° contribute < 0.01 dB/intersect and produced no measurable crosstalk above a 40 dB noise floor, even with up to 16 crossings. The final socketed interposer showed -2.75 dB average channel loss with $t \approx 260\ \mathrm{nm}$5 dB, and the packaged loop response had 48 nm 3 dB bandwidth with a peak near 1504 nm, the bandwidth limit being dominated by the SiP grating couplers rather than the glass interposer itself (Djogo et al., 1 Dec 2025).

Broadband passive-backplane logic appears in the microcomb interposer architecture, even though that work is implemented in stoichiometric silicon nitride rather than on glass. Its interposer collects broadband light from chiplets, routes it, splits it into the $t \approx 260\ \mathrm{nm}$6, $t \approx 260\ \mathrm{nm}$7, and $t \approx 260\ \mathrm{nm}$8 bands with two cascaded dichroic directional couplers, suppresses the pump with thermally tunable microrings, and interfaces to SHG, detectors, and photonic chiplets. The reported extinction ratios are typically in the 15–25 dB class, the MMIs have < 0.5 dB excess loss, and the thick-thin SiN bilayer taper is simulated to provide < 1 dB across an octave (Rao et al., 2020). This suggests that the passive-backplane role of an interposer is largely substrate-agnostic even when its compact implementations are not.

The scaling problem of dense optical redistribution is formalized in the 3D SiN interposer. For a fully connected planar network of $t \approx 260\ \mathrm{nm}$9 nodes, the paper gives $n = 1.7$0, and for the demonstrated 12-node case the all-planar routing baseline contains 495 intralayer crossings. By moving to a dual-layer 3D routing scheme, the fabricated interposer reduces this to 150, below the cited all-planar lower bound of 153, and achieves a 45.8% reduction experimentally in the average loss per waveguide. The measured component losses are $n = 1.7$1 and $n = 1.7$2 per intralayer crossing for the two layers, $n = 1.7$3 per interlayer crossing, $n = 1.7$4 per interlayer taper, and $n = 1.7$5 propagation loss (Xia et al., 14 Apr 2026). Although not a glass implementation, this is directly relevant to any glass interposer that uses multilayer dielectric routing.

At the largest proposed scale, the panel-scale glass interposer defines a WDM optical network-on-package. Its per-waveguide bandwidth is given by $n = 1.7$6 with $n = 1.7$7 and $n = 1.7$8, producing $n = 1.7$9 per waveguide. In the unit-interposer comparison, the photonic system supports 832 links and 26.624 Tb/s per XPU using 26 waveguides, whereas the electrical silicon comparison uses 736 transmission lines for 13.312 Tb/s. The path-loss model assigns 0.25 dB to each straight bypass and 0.5 dB to each 90° turn, and for a worst-case 8 × 8 tiled route the paper budgets twelve 0° bypass switches, two 90° turn switches, and an additional 3 dB for edge couplers and waveguides, leading to the stated $_2$0 per-carrier launch requirement for $_2$1 receiver sensitivity and the associated >500 mm repeater-free reach claim (Hsueh et al., 8 Aug 2025).

5. Heterogeneous integration, alignment, and packaging

The strongest packaging result on a glass carrier is the 3D-nanoprinted interposer. Its chip-coupling frame uses printed fixers and clamps, with stated positioning accuracy of ±1 µm from fixers and sub-micrometer accuracy with clamps. Optical coupling is performed by pairs of off-chip parabolic reflectors at polymer/air interfaces, governed by the mode-conversion relation $_2$2 and the ratio $_2$3. The demonstrated mode-field-dimension conversion ratio of 5:2 from fiber to chip is described as the largest mode size conversion using non-waveguided components, with 1.7 dB inherent coupling loss and measured fiber-to-chip loss around 2.4 dB/facet I/O over 1480–1620 nm. The system further demonstrates 2.5 dB die-to-die coupling loss between silicon and InP chips over the same 140 nm wavelength range (Huang et al., 2024).

The fused-silica interposer emphasizes another packaging mode: passive fiber registration combined with active chip alignment. On the fiber side it uses FLICE-opened self-aligning sockets with length increased to 250 $_2$4m, and assembly occurs by a plug-and-twist procedure with Norland NOA81 UV-curable optical adhesive. On the chip side the interposer-to-SiP assembly uses active alignment and a thin index-matched epoxy layer. The reported packaging-induced penalty is about 0.3 dB, of which 0.2 dB is linked to random $_2$5 waveguide-core position misalignment along the MCF length, and re-measurement after two weeks showed < 0.1 dB degradation (Djogo et al., 1 Dec 2025).

Mechanical stack design becomes more critical when the photonic layer is thin, laminated, or stressed. The flexible chalcogenide-glass platform shows that conventional multilayer bending theory fails when PI and SU-8 have moduli of $_2$6 and $_2$7 while the intermediate silicone adhesive is only $_2$8. The resulting multiple neutral axes permit placement of the neutral axis inside the SU-8 photonic region, and the paper derives $_2$9 with the condition $0.04\ \mathrm{dB/cm}$0 for the SU-8 neutral axis to remain inside SU-8. Optically, the same platform reports $0.04\ \mathrm{dB/cm}$1 waveguide loss, intrinsic resonator $0.04\ \mathrm{dB/cm}$2 up to 460,000, and repeated bending to 0.5 mm radius without measurable degradation (Li et al., 2013). A plausible implication is that laminated glass-photonic redistribution layers may require mechanics-aware co-design even when the primary interposer is nominally passive.

The GaAs-on-glass nanocavity paper adds a heterogeneous-integration detail of direct relevance: it explicitly notes that such structures can be implemented via pick-and-place integration techniques including transfer printing (Kawata et al., 2023). That statement is narrow, but important. It places glass-supported high-Q resonant blocks within a chiplet-assembly ecosystem rather than treating them as purely monolithic laboratory objects.

6. Limits, misconceptions, and research trajectory

One recurring misconception is that a glass-substrate photonic interposer is necessarily a single technology. The literature instead spans at least five distinct modalities: bulk fused-silica optical redistribution, deposited-glass waveguide PICs, glass-supported nanophotonic resonators, glass-carried printed free-space couplers, and proposed panel-scale glass photonic fabrics. Another misconception is that glass only supports low-functionality passive fanout. The cited work includes high-Q nanocavities on glass, 200 GHz FSR Hydex rings, distributed spectral taps on glass slabs, octave-spanning passive interposer logic, and reconfigurable WDM crossbar concepts (Kawata et al., 2023, Moss et al., 2014, Pervez et al., 2010, Rao et al., 2020, Hsueh et al., 8 Aug 2025).

A second misconception is that a fully symmetric or fully thick dielectric cladding is always the best route to high performance on glass. The GaAs PhC study explicitly reports the opposite for its cavity family: $0.04\ \mathrm{dB/cm}$3 at $0.04\ \mathrm{dB/cm}$4 exceeds the thick symmetric-glass result of $0.04\ \mathrm{dB/cm}$5 because thin-film modal mismatch suppresses an additional radiation channel (Kawata et al., 2023). Similarly, a plausible implication of the rhombic top-glass analysis is that local overglass geometry can become a secondary loss-engineering parameter rather than a purely protective layer.

The limitations are equally clear. Several of the most informative papers are not complete interposer demonstrations. The PhC-on-glass work is entirely numerical, does not treat waveguide-cavity coupling, loaded versus unloaded $0.04\ \mathrm{dB/cm}$6, insertion loss, thermal resistance, packaging reliability, or fabrication tolerances in a systematic way (Kawata et al., 2023). The Hydex papers demonstrate an excellent passive and nonlinear platform, but not an interposer with active devices, package-qualified integration, or electrical redistribution (Moss et al., 2014, Duchesne et al., 2015). The spectrometer is explicitly useful background, but not a core interposer paper (Pervez et al., 2010). The microcomb interposer and the 3D SiN interposer are architecturally relevant but not glass-native (Rao et al., 2020, Xia et al., 14 Apr 2026). The panel-scale glass interposer remains a proposal whose >500 mm repeater-free claim is supported by link-budget accounting rather than a measured panel-scale system, and whose open issues include thermal control, route contention, multilayer process maturity, large-area packaging, and defect tolerance (Hsueh et al., 8 Aug 2025).

The present state of the field therefore supports a precise conclusion. Glass-substrate photonic interposers are no longer limited to the idea of a transparent carrier with a few passive waveguides. The literature already contains a direct fused-silica interposer with dense vertical coupling and packaged multichannel performance, glass-supported resonant blocks whose loss physics is now quantitatively understood, deposited-glass platforms with low-loss WDM-relevant circuitry, glass-carried chiplet couplers with passive assembly, and a credible panel-scale architectural proposal. What remains unresolved is their unification into a single package technology that simultaneously delivers low-loss optical redistribution, high-density multilayer routing, robust heterogeneous assembly, electrical co-integration, manufacturable panel-scale process control, and experimentally validated system-level operation on glass.