As traditional electronic integrated circuits approach the physical boundaries of Moore’s Law, data centers, high-performance computing clusters, and artificial intelligence accelerators face a critical bottleneck: interconnect bandwidth. Copper wires suffer from severe RC delays, high power dissipation, and strict physical routing limitations when scaling to meet modern data demands.
Optics offers an elegant alternative. Because photons do not readily interact with one another, light waves can pass through the same physical medium without interference. To fully exploit this inherent parallelism, researchers leverage Wavelength Division Multiplexing (WDM)—a technique that slices the optical spectrum into distinct colors (wavelengths), allowing each color to carry an independent data stream simultaneously through a single waveguide or optical fiber. As Yun Zhang et al. and Haoran Zhang et al. demonstrate, WDM drastically multiplies data throughput without expanding the physical footprint of the chip's routing infrastructure.
The Scaling Challenge: From Laser Arrays to Microcombs
While WDM theoretically provides massive scalability, traditional implementations suffer from a severe hardware bottleneck at the light source. To illuminate, say, 40 distinct wavelength channels, conventional systems require an array of 40 individual lasers. Each laser must be independently manufactured, powered, and precisely stabilized against thermal drift. This approach scales poorly, resulting in bulky, power-hungry, and complex hardware architectures (Zhou et al.; Haipeng Zhang et al.).
The modern paradigm shifts this burden from an array of hardware sources to a single, highly efficient component: the optical frequency comb. Instead of generating each wavelength from a separate laser cavity, an optical frequency comb produces a broad spectrum of discrete, equally spaced, and phase-locked frequency lines from a single continuous-wave (CW) laser source (Qihua Liu et al.).
Traditional WDM: [Laser λ1] [Laser λ2] [Laser λ3] [Laser λ4] ---> [Waveguide]
Microcomb WDM: [Single CW Laser] ---> [Microresonator] ---------> [Waveguide]
(Comb Generation: ||||)
This is achieved through soliton microcombs. As explained by Yuanbin Liu et al., a single CW laser pumps an integrated microresonator cavity. Due to the strong optical nonlinearity of the cavity material—such as silicon carbide (Haipeng Zhang et al.)—the pump light undergoes parametric frequency conversion. Under the right conditions, the intracavity light organizes into a stable, self-reinforcing solitary wave (a soliton).
The mathematical consequence of a circulating soliton in a microcavity is a perfectly uniform "rainbow" of light emitted into the output waveguide. This provides dozens of parallel communication channels with inherent frequency stability, eliminating the physical footprint and power overhead of discrete laser arrays (Zhou et al.).
On-Chip Routing and Spectral Sorting
Once a single waveguide is packed with dozens of modulated wavelengths, managing that light without inducing signal degradation or thermal instability becomes the next engineering hurdle.
On-chip nanophotonic interconnects must route these multi-wavelength signals through shared waveguide trunks. Because microresonators and silicon components are highly sensitive to temperature fluctuations, researchers have developed automated design frameworks like GLOW and O-Router to optimize channel assignments (Ding). As Yu et al. show, strategically balancing trunk sharing against thermal drift and timing constraints can reduce optical power loss by up to 50% compared to naive routing strategies.
┌───────────┐ λ1 ┌────────────┐
│ ├───────────>│ Receiver 1 │
[Multi-λ Signal]─>│ Demux │ λ2 ├────────────┐
│ (Sorting) ├───────────>│ Receiver 2 │
│ │ λ3 ├────────────┐
└───────────┘ ──────────>│ Receiver 3 │
At the destination, the combined wavelengths must be separated back into individual data streams via a demultiplexer (DEMUX). The physics of demultiplexing relies on spatially dispersing light based on its color:
- Grating-Based Structures: Devices like echelle gratings or Littrow structures use diffraction to angle different wavelengths toward distinct physical outputs. Early foundational work by Qiao et al. demonstrated that broadening the optical passband of an echelle grating relaxes the wavelength tolerances of other components, enabling low-loss, athermalized operation without active temperature control.
- Photonic Crystal Fibers: Modern integrated architectures also leverage multicore photonic crystal fibers to achieve efficient wavelength separation over incredibly short propagation lengths (e.g., separating 8 channels across the C-band within just 5 cm), bypassing the need for bulky cascaded filter units (Malka).
- Wavelength-Selective Switches: For dynamic networks, Chi demonstrated that integrating MEMS mirror arrays directly onto silicon-on-insulator chips allows individual wavelength channels to be mechanically switched and cross-connected without external alignment.
Multi-Dimensional Multiplexing and Future Applications
To push past the limits of spectral efficiency, modern research combines WDM with other orthogonal dimensions of light. For instance, Han et al. demonstrated the integration of mode-division multiplexing alongside WDM within thin-film lithium niobate circuits, allowing data to be transmitted simultaneously across multiple independent spatial modes within a single waveguide. Similarly, Karadimitrakis et al. explore optical MIMO techniques to resolve crosstalk in multi-core and few-mode fibers, unlocking spatial division multiplexing to bypass the capacity limits of single-mode fibers.
This scalable optical parallelism is driving transformations across several domains:
Photonic AI Accelerators
In deep learning hardware, matrix multiplication can be executed at the speed of light by mapping weights and inputs to different wavelengths. Photonic neuromorphic accelerators like LiteCON utilize dense WDM (DWDM) to stream massive parallel data vectors through a single waveguide, achieving massive throughput and energy efficiency gains (Dang et al.). To scale this to trillion-parameter models, Hsueh et al. propose panel-scale glass interposers driven by optical combs, replacing hundreds of electrical lines with a few high-density optical paths.
Next-Generation Telecom and Access Networks
Beyond the chip, comb-driven WDM scales long-haul transport networks. Researchers have achieved single-channel transmission rates exceeding 1 Tb/s per wavelength over record distances using hollow-core fibers (Ge et al.), supported by multi-layer switching architectures that combine digital sub-wavelength grooming with flexible optical grid switching (Roy et al.). In cost-sensitive passive optical networks (PONs) and indoor wireless links, WDM enables the overlay of multicast services onto unicast channels (Zhu et al.) and maximizes aggregate data rates using multi-color LED sources (Weng et al.) or digital orthogonal filtering on low-cost hardware (Deng et al.).
By consolidating the multi-wavelength source into a single microresonator and optimizing on-chip routing, integrated optical frequency combs have transformed WDM from a bulky telecommunications architecture into a highly scalable, energy-efficient fabric for next-generation computing and communication.
Key Foundational Work
Wavelength division multiplexers/demultiplexers for high-throughput optical links (2002) — contributed 6 concept elements to this direction.
<title>Optimal design of grating-based wavelength division (de)multiplexers for optical network</title> (2001) — contributed 1 concept elements to this direction.
Athermalized low-loss echelle-grating-based multimode dense wavelength division demultiplexer (2002) — contributed 1 concept elements to this direction.