Comparison of Lumerical FDTD and Tidy3D for three-dimensional FDTD simulations of passive silicon photonic components Journal: opticajournal Publication type: Research Article
Zuyang Liu 1* and Joyce K. S. Poon 1 · 2025
Read Comparison of Lumerical FDTD and Tidy3D for three-dimensional FDTD simulations of passive silicon photonic components Journal: opticajournal Publication type: Research Article by Zuyang Liu 1* and Joyce K. S. Poon 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-16665.txt.
Introduction
Zuyang Liu and Joyce Poon compare two leading simulators in 2025 for light-based chips. These tools model how light pulses through tiny silicon waveguides, a step called FDTD simulation. Designers rely on them to test parts before building, but the tools differ in speed and cost.
The team ran both tools on six standard parts, including a splitter, a crossing, and a ring resonator, under matched settings. They compared accuracy, run time, and behavior across colors of light and grid fineness. Both tools proved reliable, with only small gaps between their results. The study gives chip designers a grounded basis for picking a tool for three-dimensional work.
Catalog record
| CREATOR | Zuyang Liu 1* and Joyce K. S. Poon 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-16665
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Comparison of Lumerical FDTD and Tidy3D for three-dimensional FDTD simulations of passive silicon photonic components Journal: opticajournal Publication type: Research Article Zuyang Liu 1* and Joyce K. S. Poon 1 Abstract We benchmark Lumerical FDTD and Tidy3D for 3D simulations of passive silicon photonic components on the SOI platform. Five devices – including an MMI, directional coupler, waveguide crossing, mode converter, and polarization splitter rotator – are simulated under matched conditions using geometries from the generic PDK from GDSFactory. We compare simulation accuracy, runtime, and transmission spectra across varying resolutions and bandwidths. Results reinforce that both solver are reliable with minimal discrepancies. Overall, Tidy3D is well suited for rapid prototyping while Lumerical FDTD remains a robust option for complex simulations. 1 Introduction The finite-difference time-domain (FDTD) method is a widely used full-wave solution for modeling eletromagnetic fields in complex media and geometries [ 1 , 2 , 3 ] . It discretizes the simulation domain into a mesh and solves Maxwell’s equations at each point over time. As a time-domain method, FDTD naturally supports broadband analysis and allows accurate modeling of arbitrarily shaped, dispersive, and subwavelength structures. These capabilities make it especially valuable for simulating integrated photonic devices with high index contrast and complex geometries. Due to its versatility, FDTD is employed in the simulation of photonic components including waveguides, resonators, and grating couplers [ 4 , 5 , 6 , 7 , 8 , 9 ] . However, many devices require full three-dimensional (3D) simulations to capture asymmetric vertical structure and polarization effects, which significantly increase computational demands. Simulation time and memory usage scale rapidly with discretization. As a result, a careful balance between accuracy, resolution, and runtime is essential. Among these tools, Lumerical FDTD (Ansys) is a widely used commercial solver known for its robust performance and various built-in features with a rich desktop graphical user interface (GUI) [ 10 ] . Meanwhile, Tidy3D (Flexcompute) is an emerging cloud-native FDTD solver that offers graphics processing unit (GPU) acceleration, a Python interface for streamlined workflow, and a web-based GUI [ 11 ] . They differ significantly in computational performance, runtime, and user experience. However, to our knowledge, there have not been a study that systematically compares these electromagnetic solvers under matched simulation conditions. In this study, we benchmark Lumerical FDTD and Tidy3D for 3D FDTD simulations of a series of passive silicon photonic components on the silicon-on-insulator (SOI) platform in terms of simulation accuracy, broadband simulation reliability, runtime, and usability. We consider a representative set of devices including a 2 \times 2 multimode interferometer (MMI), a directional coupler, a waveguide crossing, a mode converter, and a polarization rotator splitter (PRS). All geometries are picked from the generic process design kit (PDK) of GDSFactory, an open source electronics design automation (EDA) tool for integrated circuits. [ 12 ] . Each device is simulated using the same material models, boundary conditions, and source configurations across both solvers. The simulations are centered at 1550 nm. The remainder of this manuscript is structured as follows. Section 2 describes the simulation setup and configurations. Section 3 presents the simulation results and comparative analysis. Section 4 discusses the implications of our findings in practical photonic design workflows. Finally, Section 5 concludes the paper. 2 Simulation configurations Simulation of the same set of devices are conducted using Lumerical FDTD and Tidy3D under matched settings for comparison. The Lumerical FDTD solver version is 2025 R1 (8.33.3999), running locally using CPU on a workstation with AMD 3960X 24-core processor, 64 GB RAM, and Windows 10 OS. Tidy3D simulations are executed using version 2.7.7 and run on the cloud. For both solvers, the geometry and simulation domain setup, boundary settings, and mesh configuration are all handled through their Python application programming interfaces (APIs). The Python codes are included in [ 13 ] . The auto-shutoff threshold is set to 1 \times 10 -5 for early termination. The simulation domain was defined to enclose the component with a distance of 2 \mu m from its top and bottom edges and 1 \mu m from edges on the device plane. Each device is extended with 10 \mu m long straight waveguides at every port before importing into the solvers. This preprocessing step is to ensure extrusion of input and output waveguides through the boundaries in the propagation direction and therefore a reliable result. Perfectly matched layer (PML) boundary condition is applied to all six sides of the 3D domain. No symmetry planes are used to maintain generality. We selected five passive silicon photonic components to span an range of mode coupling mechanisms and structure complexity. All components are designed for a wavelength around 1550 nm and sourced directly from the generic PDK of GDSFactory. To study the interactions of fundamental modes, we include a directional coupler and a waveguide crossing. A 2 \times 2 MMI is simulated to evaluate multimode propagation. And a mode converter is modeled for higher-order mode excitation. Finally, we select a polarization splitter rotator (PSR) to study higher-order mode propagation and polarization handling. We generate GDSII files of the components and import them to the solvers through Python APIs. The layer stack consists of a standard SOI platform with 220 nm silicon (Si) in top and bottom cladding of silicon oxide (SiO 2 ). One exception is the PSR, where the top cladding is silicon nitride (Si 3 N 4 ) and the bottom cladding is SiO 2 , as demonstrated by Dai et al. [ 14 ] The dispersion of Si and SiO 2 are obtained from Handbook of Optical Constants of Solids by E. Palik et al . [ 15 ] The refractive index of Si 3 N 4 is estimated as 2.0 around the center wavelength of 1550 nm. Lumerical FDTD models dispersive materials using multi-coefficient material (MCM) model, which can describe material dispersion with higher accuracy and lower computational need [ 16 ] . Meanwhile, Tidy3D fits the dispersion data using the pole-residue model that minimizes the fitting error [ 17 ] . To evaluate the resolution-dependent performance, we applied different discretization sizes in both solvers. The spatial resolution is determined by the number of cells per wavelength, which is the wavelength in the highest-index medium in the simulation domain. We consider a series of different resolutions in the comparison. No mesh override is manually applied. In Lumerical FDTD, we choose the custom non-uniform mesh and update the minimum mesh cells per wavelength with the selected spatial resolution. The meshing algorithm applies a smaller mesh in high index materials. Similarly, Tidy3D provides an automatic non-uniform meshing option where the resolution is specified by the minimum steps per wavelength in each material. Note that the automatic grid conforms to the structures in Tidy3D, which helps avoiding staircasing of permittivity at the edges. Lumerical FDTD offers conformal mesh technology (CMT) that can account for subcell features near the edges [ 18 ] . Each device is excited using a mode source with bandwidth of 20 nm or 50 nm centered at 1550 nm. The input is fundamental TE mode, except in the PRS where fundamental TM mode is considered to evaluate polarization evolution. In Lumerical FDTD, we insert port objects at each input and output