A waveguide confines a field, not a current
From far away, silicon photonics looks like electronics: chips, layouts, foundries, design kits, components, circuits, packaging. Up close the physics is different. A wire confines a current; a waveguide confines an electromagnetic field. The designer is not only asking where a signal goes, but what shape the field takes while it travels. That is what makes electromagnetic simulation central to photonic design [Chrostowski & Hochberg 2015].
The first question: what can propagate?
The simplest photonic simulation is a mode solve. Take a cross-section of a waveguide, a silicon core in oxide cladding, and at a fixed optical frequency solve for the field patterns that can travel down it. These are the waveguide eigenmodes, each with a field profile and a propagation constant, summarized as effective index, group index, confinement, polarization, and loss. Those numbers set phase delay, dispersion, bend sensitivity, and how strongly a modulator or heater can steer the light. These mode solves are run constantly through photonic design, but usually as many small local problems rather than one solve for the whole chip.
The second question: where does the light go?
Once the supported modes are known, the next question is scattering. A directional coupler splits power between guides; a grating coupler turns a fiber mode into an on-chip mode; a ring stores light over many round trips; a taper changes mode size without reflecting. These are driven problems: a source launches a mode, and the simulator computes transmission, reflection, loss, and the coupling between modes, often as scattering parameters. Small or strongly scattering devices use full-wave methods, finite-difference time-domain or finite elements; long, slowly varying devices use eigenmode expansion, which slices the structure, solves local modes, and matches fields between sections. Eigenmodes return here, now as ports and basis functions.
A ring resonator shows the whole workflow
A ring resonator ties the steps together. A mode solver gives the waveguide's effective and group index, which set the ring's free spectral range and phase response. A coupling calculation or a full-wave run gives the coupling between the bus and the ring. Then a compact model turns the ring into a reusable element with a resonance wavelength, a quality factor, a coupling coefficient, a loss, and a tuning efficiency. Solve the modes, validate with full-wave electromagnetics, then abstract the result.
The whole chip is not solved at once
Photonic design borrows the hierarchy of electronic design automation. A foundry ships a process design kit of validated building blocks, each with a compact model built from measurement, analytic formulas, and electromagnetic simulation. At the circuit level the designer connects compact models instead of meshing the whole chip [Chrostowski & Hochberg 2015]. Device-level electromagnetics produces the scattering parameters and coefficients; circuit-level simulation propagates optical signal, phase, and noise across hundreds or thousands of elements. Full-wave simulation answers the component question; circuit simulation answers the system question.
Inverse design puts the simulator in the loop
Classical design starts from a parameterized shape: width, gap, bend radius, grating period, taper length. Inverse design flips it. The designer states an objective, such as routing one input mode to a chosen output with low loss and reflection, and an optimizer changes the geometry while the simulator scores each candidate. Adjoint methods make this practical, computing the gradient with respect to many shape variables from one forward and one adjoint solve, so the search reaches geometries a human would not draw [Molesky et al. 2018]. Simulation moves from checking a device to helping synthesize it.
Why this matters now
For years silicon photonics meant transceivers and telecom. The center of gravity has shifted to moving data inside AI systems, where copper interconnect is becoming expensive in power and reach and light offers another way to scale. Much of the recent activity is optical input/output and co-packaged optics rather than pure photonic computing: Ayar Labs raised a $155M round in 2024 backed by AMD, Intel, and NVIDIA; Lightmatter has raised over $800M for its photonic interconnect platform; Marvell moved to acquire Celestial AI. The common thread is that every modulator, coupler, grating, resonator, and switch begins as an electromagnetic field problem in a small structure, and more of those problems are entering commercial design flows.
Simulation is how photonics is designed
Electronics has a clean abstraction: wires, transistors, gates, timing models. Photonics has a harder one, because the primitive object is a mode, and modes are fields that bend, leak, interfere, couple, and resonate. That is why simulation sits so close to photonic design, from the first eigenmode to the final circuit.
Planck Labs is building an integrated design platform for high-frequency 3D devices, with electromagnetic simulation at its core. Silicon photonics shows why that matters: the final object may look like a circuit, but the component-level design problem is still Maxwell's equations.
References
- Chrostowski, L., & Hochberg, M. (2015). Silicon Photonics Design: From Devices to Systems. Cambridge University Press.
- Molesky, S., Lin, Z., Piggott, A. Y., et al. (2018). Inverse design in nanophotonics. Nature Photonics 12, 659–670.