A Computational Milestone in Full-Wave Kerr Comb Simulation Using GPU-Accelerated FDTD in Tidy3D

Source:

Chenchen Wang, Qingyi Zhou, & Zongfu Yu, "Full-Wave Simulation of Kerr Comb Generation Using FDTD",  IEEE Journal of Selected Topics in Quantum Electronics, 32(3), 2100107 (2026) https://doi.org/10.1109/JSTQE.2026.3659816 

Tidy3D Simulation Projet:

Full-wave simulation of Kerr comb generation using FDTD

FDTD simulation of dissipative Kerr solitons in a Kerr ring resonator

Overview

The generation of Kerr frequency combs in high-quality (high-Q) microresonators is a cornerstone of modern photonics, enabling advancements in spectroscopy, metrology, and telecommunications. While mean-field models like the Lugiato-Lefever Equation (LLE) have historically dominated the field, they rely on approximations that can miss critical spatiotemporal details. This study presents a computational milestone: the first-principles simulation of Kerr comb dynamics using the Finite-Difference Time-Domain (FDTD) method, capturing the exact spatial and temporal evolution without relying on traditional physical assumptions.

The Challenge: Overcoming the "Impossible" Simulation 

Directly solving Maxwell’s equations for a frequency comb presents three primary technical hurdles: 

• Nonlinear Dynamics: The simulation must capture complex interactions like self-steepening and Raman scattering without the "slowly varying envelope" approximation.

• Massive Problem Size: To achieve high-fidelity 3D geometry for a 50 µm-radius ring, the researchers used over 1.3 billion grid points.

• Temporal Duration: The simulation required approximately 6 million time steps to cover over 400 ps of real-time evolution, bridging the gap between femtosecond optical oscillations and steady-state soliton formation.

The Solution: GPU-Accelerated First-Principles Modeling

The researchers leveraged GPU-accelerated computing to overcome these immense computational demands.

Key Methodology:

• High Spatial Resolution: The minimum spatial step was set to 1/30th of the target wavelength.

• Direct Dispersion Integration: FDTD inherently incorporates all geometric and higher-order dispersion terms directly from the Maxwell equations, removing the need for additional modeling corrections.

• Validation through Cross-Modeling: To verify the framework, the authors translated FDTD parameters into equivalent LLE models for direct comparison.

"GPU acceleration was the key enabler, turning an otherwise prohibitive computation into a tractable workflow. It allowed the simulation to sustain the resolution and runtime needed to bridge from initial transients to steady comb states." - Chenchen Wang, the first author of the source paper

(a) Schematic diagram of the simulation and experiment setup. (b) Illustrated evolution stage of the Kerr comb

Comparison of Operating Regimes

The FDTD framework successfully reproduced three distinct comb states, showing clear agreement with traditional theoretical models while capturing more complex dynamics.

To ensure the reliability of the comparison across these regimes, the following system parameters were kept constant in the 2D setup:

• Loaded Quality Factor (Q_loaded): 1.8 χ 10⁴.

• Second-Order Dispersion (ζ₂): 2π x 0.55 GHz.

• Dimensionless Dispersion (β): Approximately -0.2.

• Pump Wavelength: 1550 nm.

• Initial Condition: A short pulse injected at t = 0 to facilitate efficient access to the steady state.

The full-wave simulation cost for each case was approximately 2 A100 GPU-hours.

Unique Insight: Subcomb Mismatch

A significant finding of this full-wave approach is the accurate capture of subcomb mismatch. Traditional models use a "rigid" frequency grid that assumes all modes are perfectly commensurate. However,  FDTD revealed that:

• Secondary combs originating from different primary combs have distinct offset frequencies.

• This physical reality results in a frequency mismatch that is critical to understanding soliton formation and comb noise.

  
  

Conclusion

This work demonstrates that full-wave modeling has matured into a practical, capable design tool for next-generation microresonator sources. Because FDTD captures transient and spatial dynamics that mean-field models tend to average away, it offers a fuller view of how the comb actually forms.

This scale of simulation is only feasible because of Tidy3D's GPU-native FDTD solver, which handles over a billion grid points and millions of time steps within a practical compute budget – without resorting to the reduced-order approximations that mean-field methods require.