nature reviews methods primers https://doi.org/10.

1038/s43586-023-00257-4

Primer Check for updates

Finite-difference
time-domain methods
F. L. Teixeira 1, C. Sarris 2, Y. Zhang 3
, D.-Y. Na 4
, J.-P. Berenger 5
, Y. Su 6
, M. Okoniewski , W. C. Chew8,
7

V. Backman6 & J. J. Simpson 3

Abstract Sections

The finite-difference time-domain (FDTD) method is a widespread Introduction

numerical tool for full-wave analysis of electromagnetic fields in Experimentation

complex media and for detailed geometries. Applications of the Results
FDTD method cover a range of time and spatial scales, extending
Applications
from subatomic to galactic lengths and from classical to quantum
physics. Technology areas that benefit from the FDTD method Reproducibility and data
deposition
include biomedicine — bioimaging, biophotonics, bioelectronics and
Limitations and optimizations
biosensors; geophysics — remote sensing, communications, space
weather hazards and geolocation; metamaterials — sub-wavelength Outlook

focusing lenses, electromagnetic cloaks and continuously scanning

leaky-wave antennas; optics — diffractive optical elements, photonic
bandgap structures, photonic crystal waveguides and ring-resonator
devices; plasmonics — plasmonic waveguides and antennas; and
quantum applications — quantum devices and quantum radar. This
Primer summarizes the main features of the FDTD method, along
with key extensions that enable accurate solutions to be obtained for
different research questions. Additionally, hardware considerations
are discussed, plus examples of how to extract magnitude and phase
data, Brillouin diagrams and scattering parameters from the output
of an FDTD model. The Primer ends with a discussion of ongoing
challenges and opportunities to further enhance the FDTD method
for current and future applications.

1
Department of Electrical and Computer Engineering, The Ohio State University, Columbus, OH, USA.
2
Department of Electrical and Computer Engineering, University of Toronto, Toronto, Ontario, Canada.
3
Department of Electrical and Computer Engineering, University of Utah, Salt Lake City, UT, USA. 4Department of
Electrical Engineering, Pohang University of Science and Technology, Pohang, Korea. 5Department of Electrical
and Electronic Engineering, The University of Manchester, Manchester, UK. 6Biomedical Engineering Department,
Northwestern University, Evanston, IL, USA. 7Department of Electrical and Software Engineering, University
of Calgary, Calgary, Canada. 8Department of Electrical and Computer Engineering, Purdue University, West
Lafayette, IN, USA. e-mail: [email protected]

Nature Reviews Methods Primers | (2023) 3:75 1

0123456789();:
Primer

Introduction method continues to grow in popularity because it is robust, flexible

As one of the four fundamental forces of nature, electromagnetic fields and relatively easy to implement.
are ubiquitous in the natural environment and modern technologies. This Primer provides an overview of FDTD modelling capabilities
The field of electromagnetics began with discoveries into static elec- applied to various applications across the electromagnetic spectrum.
tricity and the magnetic field of the Earth and has expanded to include It concludes with an outlook on future possibilities.
a range of scientific disciplines and engineering applications, such as
wireless radio communications, biomedical imaging and quantum Experimentation
technologies. Fundamental to these advances was the formation of The FDTD method solves the time-domain form of Faraday’s law
Maxwell’s equations in 1865 (ref. 1) as well as the development of robust and Ampere–Maxwell’s law to obtain the time evolution of elec-
and powerful solutions to those equations. The reason Maxwell’s equa- tric and magnetic fields across a spatial grid. Originally, the FDTD
tions and their solutions endure is that they are valid from subatomic to method was proposed using the differential forms of Faraday’s and
galactic length scales and from classical to quantum physics2. Ampere–Maxwell’s laws:
Initial efforts focused on obtaining exact solutions to Maxwell’s
equations for canonical geometries. During this time, the Sommerfeld ∂H
Faraday’s law : µ = −∇ × E − M , (1)
integral, Rayleigh scattering, Mie scattering3 and the Debye model4 ∂t
were pioneered. This was followed by an era of approximate solutions4,
including geometrical optics, asymptotic theory, the geometrical ∂E
Ampere − Maxwell’s law : ε = ∇ × H − Js − σ E , (2)
theory of diffraction and perturbational techniques. With the advent ∂t
of the modern computer, numerical solutions started to appear.
Compared with exact or approximate analytical approaches, in which E is the electric field, H is the magnetic field, Js represents an
numerical techniques provide greater flexibility and can consider electric current density source and M is the equivalent magnetic cur-
more complex scenarios. The early focus on the frequency domain rent density. Three primary constitutive parameters are used to
meant that, initially, numerical solutions to Maxwell’s equations were describe the electromagnetic properties of materials: ε, electric per-
only obtained in the frequency domain. mittivity; µ, magnetic permeability and σ , conductivity. Materials and
A robust numerical method for solving Maxwell’s equations objects of interest are generally placed into the grid by assigning
was finally formulated in the time domain with the introduction of constitutive material parameters on the basis of the cell location.
the finite-difference time-domain (FDTD) method in 1966 (ref. 5). The FDTD grids are typically generated in Cartesian coordinates using
approach decomposes the spatial domain of interest into grid cells, non-collocated, staggered electric and magnetic field components, as
in which each grid cell has staggered electric and magnetic field shown in Fig. 1a. Central differencing is used to approximate the time
components, as shown for Cartesian coordinates in Fig. 1a. and spatial partial derivatives in equations (1) and (2). Truncating the
The FDTD method has several advantages. It evolves in time, just Taylor series expansions of the field components yields a second-order
as nature evolves over time. As a result, FDTD results can give a direct accurate scheme in both space and time.
insight into the time evolution of electromagnetic fields. The models There are numerous source options for an FDTD model. These
are matrix-free and may be efficiently parallelized onto supercomput- include hard, soft, resistive voltage and plane wave sources. Hard
ers in a straightforward manner. Additionally, the FDTD method treats sources have field components set at specific locations to specific time
impulsive and nonlinear behaviour naturally, although sources of error waveforms and are the simplest source option. However, hard sources
are well understood and may be bounded to certain tolerance levels. are non-physical and scatter incoming electromagnetic waves. Current
Finally, FDTD models do not automatically need to be reformulated or soft sources that are implemented by assigning values to the Js term
for different problems. Instead, changes are typically only made to in Ampere–Maxwell’s law (equation (2)) are used to account for current
account for additional physics as needed. flow in the model. They are useful for modelling impressed sources
Before choosing the FDTD method, it is important to note its that do not perturb the underlying electromagnetic response (Green’s
disadvantages. FDTD simulations are relatively slow because the function) of the problem. Resistive voltage sources7 can simulate sources
Courant–Friedrichs–Lewy (CFL) condition6 must be satisfied to main- with prescribed impedance profiles, such as at the feeding point of an
tain stability. Typically, the grid cell dimensions — ∆x ∆y and ∆z in antenna. Plane waves from a distant source may be efficiently modelled
Cartesian coordinates — are chosen to satisfy the geometrical details using a perfectly matched total-field scattered-field formulation8. Using
of interest and have at least 10–20 grid cells per shortest electromag- the perfectly matched total-field scattered-field formulation, the numer-
netic wavelength for sufficient accuracy7. On the basis of these grid ical dispersion introduced by the FDTD grid is accounted for in the
cell sizes, the time-step increment is chosen that satisfies the CFL incident wave calculation. As a result, the addition or subtraction of
condition. For cubic grid cells of dimensions ∆x = ∆y = ∆z = ∆ , the the incident wave in the FDTD model is accurate down to machine
∆
CFL condition reads ∆t ≤ , in which c is the speed of light. Further- precision.
c 3
more, the FDTD method is sensitive to the shape and size of grid cells. Alongside an appropriate source, all FDTD models must have
As a result, FDTD models usually have uniformly shaped grid cells suitable boundary conditions. Candidate boundary conditions include
that may cause defeaturing of certain geometries. However, there are perfect electric conductor (PEC); perfect magnetic conductor (PMC);
techniques to model material boundaries inside a grid cell and fea- periodic boundary conditions (PBCs); surface impedance boundary
tures smaller than a grid cell, such as thin wires. Additionally, not all conditions (SIBCs); and absorbing or radiation boundary conditions.
electromagnetic field component values — especially those required PEC and PMC facets reflect electromagnetic waves and are the simplest
for implementing complex materials — are directly solved for by the boundary conditions. They are only appropriate for modelling (high
algorithm at the required locations and timepoints. This can some- conductivity) metallic structures at or below microwave frequen-
times lead to stability issues. Despite these disadvantages, the FDTD cies (PEC) or to take advantage of certain types of symmetries in the

Nature Reviews Methods Primers | (2023) 3:75 2

0123456789();:
Primer

a b
z
y Medium
x Ex boundary
Hz(i + ½,j + ½,k + 1)

Medium 1 Etan
Ey d2
Hz d3
Ez(i,j,k + ½) Az
d1

d4

Hy(i + ½,j,k + ½)

Hx(i + 1,j + ½,k + ½)

Medium 2

(i,j,k)

Ex(i + ½,j,k) Ey(i,j + ½,k)

c d e

1m

z Ø = 1 cm
y 1.2 m Ø = 2 mm
x
PEC
50 Ω 50 Ω
PEC 2m
2m

4m

Fig. 1 | Details for common finite-difference time-domain grid arrangements. edges of the metallic strip. d, Same as part c, but with a subgrid instead of
a, A 3D finite-difference time-domain (FDTD) (Yee) grid cell in Cartesian spatial non-uniform gridding. e, Example of thin-wire antennas modelled in an FDTD grid.
coordinates. The grid cell has dimensions ∆x × ∆y × ∆z in space, with staggered A monopole antenna and conical antenna on a 400 × 200 × 10-cm perfect electric
electric and magnetic field components oriented in each Cartesian direction. i, j, k conductor (PEC) plate. The structure is illuminated by a plane wave. The FDTD grid
are grid cell indices for the x-direction, y-direction and z-direction, respectively. resolution is ∆ x = ∆y = ∆ z = 5 cm and the time-step increment ∆t = 0.075 ns. The
b, Staircasing in an FDTD model comprised of regular, square Cartesian grid monopole is modelled as a thin wire (diameter φ = 1 cm) parallel to the z-direction
cells arising from the grid electromagnetic components not lining up with the (120-cm long, spanning 24 cells). The conical antenna is modelled using nine
material boundary. A locally conformal FDTD algorithm is applied to one grid cell straight thin wires (φ = 2 mm) connected to form the wire structure: one short
containing the curved object. c, A non-uniform gridding approach for an FDTD vertical wire (5-cm long, spanning one grid cell) and eight oblique wires (four wires
grid refinement applied to a thin microstrip line printed on a dielectric substrate. of 21 cells 5.11-cm long and four wires of 11 cells 4.90-cm long). Both antennas are
Higher resolution is needed to accurately model the fringing fields around the connected to the PEC plate through a 50-Ω resistor.

problem (PEC and PMC). PBCs9,10 connect opposite edges of the grid radiation boundary conditions are suitable for open-region problems,
to each other. This boundary condition is suitable when a structure in which surrounding objects may be ignored in the simulation because
is periodic — such as phased antenna arrays, photonic crystals and they are sufficiently far away from the FDTD grid or not of direct inter-
metamaterials — and to approximate a geometry that should appear to est. Radiation boundary conditions are used infrequently, for example,
continue beyond the edges of the computational domain, for instance, when absorbing boundary conditions exhibit stability issues, such as
when approximating the ionosphere as extending far beyond the for magnetized plasma.
FDTD grid11. SIBC avoids the need to extend a model into a material,
which usually requires smaller grid cell sizes owing to the reduced The perfectly matched layer
electromagnetic wavelength12,13. Both frequency-independent and The main goal of any absorbing boundary condition is to minimize spu-
frequency-dependent SIBC formulations are available. Absorbing or rious reflections along the outer boundaries of the grid. The perfectly

Nature Reviews Methods Primers | (2023) 3:75 3

0123456789();:
Primer

matched layer (PML)14,15 is the most popular absorbing boundary condi- grid cells that are rectangular or cubic, meaning that complex geom-
tion for the FDTD method owing to its overall efficiency, low reflection etries may not always be replicated, producing a staircase error. This
levels, robustness, relative ease of implementation and flexibility. may be seen in Fig. 1b, in which the material boundary does not line up
After the PML was originally introduced, it was reformulated as with the grid cell edges.
a complex-valued mapping applied to the spatial coordinates16–18. To mitigate staircasing errors, locally conformal FDTD grid imple-
This led to the PML version most commonly used today, which uses mentations may be used, in which the FDTD grid cells are deformed next
transformation optics principles applied to Maxwell’s equations19. to curved or sloped boundaries25,44–46. Although locally conformal FDTD
The coordinate mapping is fully recast as modified, permeability and algorithms produce more accurate results for problems with curved
permittivity tensors into the original, real-valued, coordinate domain. and sloped geometries, their implementation requires additional book-
The form of Maxwell’s equations is preserved via an anisotropic and keeping to handle the local geometric features. Extra care is needed to
dispersive medium — or an artificial PML medium — also known as avoid numerical instabilities47.
uniaxial PML17,20. Consequently, algorithms for implementing the FDTD For problems involving highly complex 3D structures, manually
method in anisotropic and dispersive media may be retooled for PML creating the FDTD grid according to the object shape may be costly
media in an FDTD model and vice versa21–23. and error prone. To avoid these problems, automatic grid generation
techniques based on input computer-aided design files and computer
Complex media graphics principles48 may be used.
Beyond simple linear materials with scalar constitutive relations, the
FDTD method has been successfully applied to a large class of disper- Subcell models and subgridding. Many geometries contain localized
sive, anisotropic and nonlinear materials. These include dispersive, geometric details and strong, localized fields. Consider, for example, an
frequency-dependent materials, such as Drude, Lorentz, Debye and array of plasmonic nanospheres embedded in a dielectric host medium.
gyromagnetic materials, for example, magnetized ferrites and plasma. Plasmonic modes may be strongly localized on the surface of the nano-
Two main approaches have been applied. The first approach involves spheres, in which the geometry does not conform to a grid of rectangular
numerical integration of the time convolution integrals. The exponential Yee grid cells. Therefore, grid refinement is necessary to accurately
nature of the susceptibility functions is exploited to avoid storing the simulate these structures and capture surface field resonances.
entire time history of the fields. In its place, recursive accumulators are There are four possibilities for implementing a grid refinement.
used so that only a few past time-step field values are required in the FDTD The first, brute-force approach involves uniformly applying the mini-
update equations24,25. The second approach involves auxiliary differen- mum cell size required anywhere in the computational domain. This
tial equations. Here, additional differential equations, which represent approach is simple to implement. However, it results in a large com-
the physics of the material, are solved simultaneously with Maxwell’s putational overhead. A second option is to use non-uniform gridding,
equations24,26. Examples include magnetized ferrites27,28 and magnetized as shown in Fig. 1c, in which only a fraction of the cells are refined49,50.
plasma29. The constitutive parameters of the dispersive media should A third option is a subcell method, in which the cell size is main-
be obtained30,31 before the FDTD implementation. It is also possible to tained, but effective material parameters are assigned to the cell to
implement dispersive media using a Z-transform approach32,33. account for material boundaries within the cells51. These parameters
In anisotropic media, the constitutive parameters are represented may be scalar or tensorial52,53 and may also account for the dispersive
as tensors. Several FDTD extensions have been successfully developed properties of materials46,54,55.
to handle anisotropic media34–37. An important example of anisotropic Subgridding, shown in Fig. 1d, is the most advanced technique.
media is the PML media18,19. In nonlinear media, the constitutive param- Grids of different resolutions are coupled together via a set of update
eters depend on the electric or magnetic field strengths. Constitutive equations, independent of the regular finite differences for Max-
relations for a nonlinear medium are expressed in terms of nonlinear well’s equations56–58. These interface update equations may result in a
susceptibilities, in addition to linear susceptibilities38,39. Special cases late-time instability59 but consistent formulations have been presented
of nonlinear susceptibilities include isotropic models, such as the Kerr to ensure that subgrids are coupled in a stable fashion60–62.
model with a quadratic dependency of the permittivity on the electric FDTD subgridding methods may be categorized on the basis
field strength, and the cubic-quintic Kerr model with both quadratic of how they treat time-stepping, as subgrids reduce the required
and quartic dependencies24. The implementation of nonlinear media time-stepping increment. To meet this stability constraint, two meth-
in FDTD is accomplished by either an iterative loop or a root-finding ods have been developed. The first uses a uniformly reduced time step
algorithm, such as Newton–Raphson’s algorithm applied within each on the basis of the CFL condition for the densest subgrid. The second
time step24. These approaches may be adapted to integrate nonlinear uses adaptive time-stepping, in which subgrids are assigned time-step
lumped circuit elements — diodes, transistors and other semiconductor increments according to their own CFL condition and the subgrids and
devices — into FDTD simulations34. In several multiphysics problems, coarse mesh are synchronized through temporal interpolations
Maxwell’s equations do not constitute a closed system and the consti- and extrapolations59,62.
tutive relations will not fully capture all the dynamic effects. In such
cases, the FDTD algorithm should be integrated as a field solver, with Thin-wire models. The thin-wire formalism is an important subcell
the set of equations discretizing the coupled dynamics40,41. Special care model that allows wires thinner than the grid cell dimensions to be
must be taken to minimize numerical FDTD artefacts that might subtly accounted for in an FDTD model. Straight wires of arbitrary radius,
affect the behaviour of the coupled dynamics42,43. length and orientation may be implemented63. Furthermore, wires
may be placed anywhere in the domain, even independent of the FDTD
Spatial gridding grid cell features. Sources and loads may be connected to wires to form
Complex geometries. Each grid cell of an FDTD model must be circuits. FDTD grid cells may be comprised of both thin wires and any
assigned material constitutive properties. It is common to use uniform other objects.

Nature Reviews Methods Primers | (2023) 3:75 4

0123456789();:
Primer

The basic methodology of the thin-wire method involves solving a Dimensional splitting unconditionally stable algorithms, such
set of partial differential equations for each straight wire using the 1D as the alternating direction implicit and the locally one-dimensional
FDTD method in parallel with conventional FDTD solution for Maxwell algorithms, involve only tridiagonal matrices and require iterative
equations63,64. Each wire is discretized independently from the rest of solvers at each time step. Alternating direction implicit-FDTD80–82 and
the model. The chosen spatial discretization of any wire should not be locally one dimensional-FDTD83,84 introduce additional splitting errors.
too large to obtain accurate sampling of the current and charge distri- Care must be exercised when choosing the simulation parameters to
bution of the wire. It also must not be too small, as the 1D solution on ensure that the numerical error remains within acceptable bounds.
the wire is also subject to the CFL stability condition.
Figure 1e illustrates some possibilities of the thin-wire formalism Parallel computing and hardware acceleration
through a canonical example that includes a two-cell thick PEC plate Conventional FDTD is formally an Iterative Stencil Loops simulator
along with two thin-wire antennas in the same FDTD domain. or a time marching, finely grained, nearest neighbour algorithm. It is
an explicit algorithm, in which new field values are computed using
Cylindrical and spherical FDTD models. For problems with cylindrical only previously computed field components. These characteristics
or spherical symmetries, it is desirable to use cylindrical and spherical make FDTD models highly parallel and well suited to various parallel
FDTD grids that are conformal to the geometries of interest. These computational hardware components, from mainstream multicore
grids may mitigate staircasing errors and remove the need for high grid central processing unit (CPU) chips and graphical processing unit (GPU)
resolutions or locally conformal implementations35,65. In cylindrical and chips85, to special-purpose hardware based on field programmable
spherical coordinates, the finite-difference approximations for the gate arrays. As of 2023, multicore CPUs and GPUs are the two main
spatial derivatives must properly account for the changing metric fac- computational hardware platforms available to an FDTD developer.
tors. They also require special treatment of the coordinate singularities Contemporary CPUs have up to 24 or 64 cores. From the per-
along the axis of symmetry or on the spherical poles. Cylindrical and spective of a programmer, each core is a fully functioning processor
spherical FDTD algorithms are routinely used for diverse applications that splits each core into two virtual cores, called threads. A single
such as optics66, geophysics67,68 and ionospheric propagation69,70. 18-core CPU may run 36 threads or independent processes. For the
FDTD method, developers may achieve effective parallelization by
Non-orthogonal and unstructured FDTD models. The FDTD method using OpenMP API with a simple parallelization pragma added in front
can be extended to non-orthogonal curvilinear grids, increasing the of the nested loops with field updates. If more complex field updates
geometric flexibility beyond regular Cartesian, cylindrical and spher- are required in a small area — for example, complex materials or sub-
ical grids. In this case, the FDTD grids are comprised of deformed cell models — it is beneficial not to break the parallel loops. Instead,
quasi-hexahedral cells65–68. A field representation in terms of covariant a backward time step can be performed and the field recomputed using
and contravariant field components may be used to account for grid updates, reflecting the local field interaction complexity.
cell deformation and non-orthogonality68. By contrast, GPUs are massively parallel computational platforms.
Alternatively, FDTD extensions can be obtained on the basis The highly parallel sections of the code are executed on the GPU device
of a finite-element framework. The conventional FDTD algorithm as kernels. A kernel executes in parallel across a set of threads, organ-
is equivalent to a Galerkin-type finite-element time domain (FETD) ized into thread blocks and arranged into a grid. To execute kernels,
algorithm on a Cartesian grid using mixed face or edge basis functions thread blocks are distributed across streaming multiprocessors on the
and low-order quadrature rules71. Although this approach is suitable for basis of resource availability. Multiple instructions, multiple data may
handling arbitrarily shaped geometries, conventional FETD requires a be executed across a thread processing cluster, with single instruction
linear system of equations to be solved in each time step, which is less multiple data on each streaming multiprocessor.
efficient than the FDTD method. Nevertheless, there are advantages A GPU FDTD programmer does not use loops to update fields.
of constructing FDTD schemes on more general grids on the basis of Instead, the simulation domain is divided into blocks of FDTD cells to
finite-element principles72–74. be distributed across thread processors. Computing fields at the edge
of the block requires data from neighbouring blocks at every time step.
Unconditionally stable FDTD methods At the time of writing, a single NVIDIA Hopper GPU has 80 GB
The conventional FDTD algorithm is conditionally stable because the of high-speed memory, and 18,176 CUDA cores, each with 2 floating
time-step increment is constrained by the CFL condition. It is possible point units. This resource enables large FDTD simulations with more
to construct unconditionally stable FDTD methods on the basis of than 3 billion grid cells, for example, a photonics problem with grid
modifying the time-stepping algorithm used for field updates75,76. The dimensions of 6,000 × 6,000 × 90, or an antenna problem with
Newmark-β FDTD algorithm77 and the Crank–Nicolson FDTD algorithm78 grid dimensions of 1,500 × 1,500 × 1,400. High-end PC machines can
are two popular methods. Unconditionally stable methods are particu- support multiple GPU cards. Message passing interface may be used to
larly useful for problems that require very fine grid resolutions of two or extend parallelization to multiple GPUs and machine clusters, which
more orders of magnitude less than the wavelength40. each have multiple GPUs. Most commercial FDTD codes can use a single
Unconditionally stable methods have drawbacks. The field values GPU at minimum.
to be solved at any given time step depend on known past field values and
unknown present field values at other grid points. The result is an Higher-order FDTD modelling
implicit time update that requires solution of a sparse matrix equa- The accuracy of the FDTD method for electrically large problems, which
tion at each time step to obtain the next field value. Additionally, the span many tens of wavelengths or more, is primarily limited by numeri-
condition number of the underlying system matrix tends to grow with cal or grid dispersion. This causes the phase velocity of the waves in
time-step size79. This leads to a slower field update for schemes that use the FDTD grid to exhibit spurious variations versus both frequency
iterative solvers for the sparse matrix equation. (dispersion) and propagation (anisotropy) angle.

Nature Reviews Methods Primers | (2023) 3:75 5

0123456789();:
Primer

Glossary wave propagation in inhomogeneous and nonlinear media, replicat-

ing the FDTD solution with excellent accuracy. Alternatively, the loss
function may be evaluated on an FDTD grid using finite differences to
Boundary conditions Maxwell’s equations approximate the derivatives. This approach was followed by ref. 98 for
Description of the behaviour of Two coupled partial differential coupled electromagnetic-thermal simulations.
the solutions at certain points in equations that govern the propagation A supervised strategy, using a convolutional neural network (CNN)
space, usually along the outer of electromagnetic waves. and a long short-term memory (LSTM) network trained with FDTD data,
edges of the finite-difference was used in ref. 99 to rapidly extract the scattering parameters of vari-
time-domain grid. Monte Carlo method ous planar microwave structures. The CNN processed the geometry of
When applied to the finite-difference the simulated structure, whereas the LSTM processed the scattering
Courant–Friedrichs–Lewy time-domain method, the Monte parameters computed by a coarse-mesh FDTD simulation. The hybrid
(CFL) condition Carlo method involves rerunning a CNN–LSTM network rapidly and accurately computed the scattering
A constraint that must be satisfied finite-difference time-domain simulation parameters of new structures not previously seen by the networks, in a
to achieve convergence and numerous times, often thousands or fraction of the simulated time needed for a dense grid FDTD simulation.
maintain numerical stability in a millions of times, to obtain a range A flowchart of this approach is shown in Fig. 2c.
simulation. of possible electric or magnetic field The current state of the art still requires significant computa-
outcomes for an uncertain modelling tional resources for off-line training of neural network models. This
Iterative Stencil Loops scenario. is achieved either by using ground-truth FDTD data99 or via the PINN
simulator approach of minimizing a cost function on the basis of the equations
A numerical data processing solution Scattering parameters to be solved97,98. Such models are efficient surrogates of FDTD in sce-
where an array of elements is updated S-parameters provide a relationship narios that need repetitive simulations, for instance, design optimiza-
according to a fixed pattern called between the input and output of an tion, uncertainty quantification and yield analysis. They should not
a stencil. electrical network. be seen as replacements of FDTD models or standalone simulators.

Results
Higher-order FDTD algorithms may be used to reduce the numeri- The electromagnetic fields predicted in space and time by FDTD models
cal dispersion. Conventional higher-order FDTD algorithms are are ideal for generating animations of the field behaviour. When dealing
based on extended finite-difference stencils to approximate the field with large or long models running on supercomputers, parallel input
derivatives86. This is often denoted as (n,p) FDTD, in which n and p refer and output — such as HDF5100 — is highly beneficial to save the desired
to the order of the truncation error in time and space, respectively. output data. When sampling a field component in space, the field values
Using this notation, the conventional second-order FDTD algorithm, are averages over the full dimensions of a grid cell, even though they are
Yee’s scheme, is (2,2) FDTD. The (2,4) FDTD algorithm strikes a good typically drawn (as in Fig. 1a) or stored (in computer memory) in a manner
balance between accuracy and computational cost87. that implies that the field components are located at just one specific
The main limitation of higher-order schemes comes from the behav- position in space.
iour of larger spatial stencils across material interfaces, in which the higher When the spectrum of a field component is desired, a post-run
order of the finite-difference approximation is not retained. Although discrete Fourier transform (DFT) may be applied to the data. For simula-
approaches are available to maintain the accuracy order of higher-order tions involving a large number of time steps, an on-the-fly DFT may be
FDTD schemes across interfaces88,89, they are characterized by a decrease performed within the time-stepping loop7. To avoid any artificial high
in generality and a substantial increase in the implementation complexity. frequency components, the DFT should be performed either on results
In broadband problems, grid dispersion effects are exacerbated at that decay to zero or until the data reach a zero-crossing.
the high end of the frequency band. The phase error may be more uni- Overall, it is good practice to validate new FDTD models or new
formly minimized across a wide frequency band by modifying the coef- algorithmic approaches against available analytical results, measure-
ficients of the enlarged finite-difference stencils90–92. Figure 2a provides ment data, other computational approaches or previous FDTD models
a notional view of the most suitable FDTD algorithms according to such as convergence studies. This helps to confirm the validity of the
the electrical size of the problem domain and chosen grid resolution. model before applying it to new problems. For example, if the grid reso-
lution is doubled and the model predicts analogous electromagnetic
Scientific machine learning fields, it is assumed that the results have converged.
Deep learning methods have been developed as alternatives to conven-
tional numerical techniques for solving partial differential equations Uncertainty quantification
and inverse problems. Developments include physics-informed Many devices and systems of interest have uncertainties owing to toler-
neural networks (PINNs)93,94, the deep Galerkin method95 and deep ances in the manufacturing process, statistical fluctuations of material
Ritz method96. These methods do not explicitly use any spatial or tem- properties or changes in environmental conditions. These uncertainties
poral discretization. As a result, they are not subject to the numerical mean that the output response of a model is also statistically variable.
dispersion and stability constraints of the FDTD method. Instead, the Quantifying this statistical variability is important to establish reli-
PINN-based solution of Maxwell’s equations relies on minimization of able models with realistic design margins for robust optimization.
a loss function that represents how well the generated data satisfy the Uncertainty quantification addresses this challenge.
equation, along with its initial and boundary conditions. Various uncertainty quantification methods have been integrated
In ref. 97, a fully connected artificial neural network, with a loss with FDTD, across a range of applications. For example, the Monte Carlo
function evaluated by automatic differentiation, was used to simulate method was used to evaluate scattering from rough surfaces in ref. 101.

Nature Reviews Methods Primers | (2023) 3:75 6

0123456789();:
Primer

a b
L/λ
Pre-asymptotic High-order FDTD
high-order FDTD
Electrical size

Spectral
methods
Yee’s (2,2) FDTD Unconditionally
stable FDTD
(ADI, LOD)

λ/α
Nyquist Mesh resolution
limit Uncertainties Distorted mesh:
realization of a
random geometry

Layout of circuit and Predicted/compensated S-parameters

substrate information
0

–10

|S21| (dB)
–20

–30

Dense mesh
–40
Prediction
–50
0 5 10 15 20
Frequency (GHz)

+
CNN-based prediction
0 Numerical error compensation

–10 Dense mesh results

(ground-truth data)
|S21| (dB)

–20

–30

–40
Coarse mesh
–50
0 5 10 15 20
Coarse-mesh Full-wave electromagnetic
Frequency (GHz)
results solver (FDTD)

Fig. 2 | Advanced finite-difference time-domain modelling approaches. convolutional neural network (CNN) network uses the information from the
a, Suitable finite-difference time-domain (FDTD) algorithms, depending geometry image and a long short-term memory network uses the coarse FDTD
on the electrical size of the problem, the FDTD grid and the chosen grid data. This method is used to rapidly produce the scattering parameters of planar
resolution. Spectral methods use a Fourier transform to represent spatial microwave structures over a broad bandwidth (0–20 GHz). α, FDTD cell size; λ,
derivatives. b, Example of a stochastically deformable, generalized curvilinear wavelength; ADI, alternating direction implicit; L, domain size; LOD, locally one
computational grid applied at a metallic edge having a statistically variable dimensional. Part b adapted with permission from ref. 105, IEEE. Part c reprinted
geometry. c, Example flowchart for a supervised training method, in which a with permission from ref. 99, IEEE.

The Monte Carlo method is simple to implement, versatile and Methods that use polynomial chaos expansion (PCE) have been
non-intrusive as it is based on post-processing iterative simulations. formulated as robust alternatives to Monte Carlo. PCE is based on
However, the statistics — mean and variance — converge at a rate expanding the output function of interest, a second-order random
proportional to 1/ N , where N is the number of simulations102. This process, as orthogonal polynomials of the random input parameters103.
limits the Monte Carlo method when applied to computationally A major challenge is that the number of polynomials grows with the
large problems. number of random input parameters and polynomial order.

Nature Reviews Methods Primers | (2023) 3:75 7

0123456789();:
Primer

a b
Γ X M Γ
dx
y

0.8
dy

PBC
PBC y 0.6
x

Frequency, wd/2πc
dy
x
dx

c 0.4

0.25 mm

2.375 mm 0.25 mm
0.355 mm X
2.375 mm 0.2

0.3 mm
Γ M

1.19 mm 0
0 0.5 1
0.46 mm
3.33 mm |E| (dB)
2.62 mm
0.25 mm

2.62 mm

5 mm
5 mm

d
1 10

0
0.5

–10
Electric field amplitude

S parameters (dB)

0
–20

–30
–0.5

–40

–1
Incident field –50
S11
Transmitted field
Reflected field S21
–1.5 –60
0 0.5 1 1.5 6 8 10 12 14 16 18 20 22
Time (ns) Frequency (GHz)

Source Probe

Ex

PML Hy k PML

The expansion coefficients in PCE may be determined with a implementation starts by expanding all field components in terms
non-intrusive procedure by post-processing multiple FDTD simula- of polynomials of the random inputs and recasting the entire system of
tions at a set of input parameter vectors104. Alternatively, intrusive FDTD updates as a set of equations with respect to the field expansion

Nature Reviews Methods Primers | (2023) 3:75 8

0123456789();:
Primer

Fig. 3 | Example approaches for obtaining results from finite-difference and each PBC-based simulation runs for 8,192 time steps. b, Right: Brillouin
time-domain simulations. A band diagram (panels a and b) and S-parameters (band) diagram for the square lattice of aluminium cylinders with period
(panels c and d). a, A 2D lattice of cylinders. The periodic array is reduced to d = 2 cm computed by the FDTD method (crosses) and data from ref. 207
a single unit cell modelled in an finite-difference time-domain (FDTD) grid (thick yellow lines). Left: the magnitude of the Fourier transform of the electric
via periodic boundary conditions. Within the unit cell, a broadband source field sampled within the unit cell for two wavevectors within the irreducible
excitation — such a Gaussian source — excites electromagnetic waves, whose Brillouin zone. The detected peaks are transferred to the band diagram and
frequency spectra consist of peaks at the resonant frequencies corresponding shown with green and red dots, respectively. c, The geometry of a split ring
to the wavevector k enforced by the periodic boundary conditions (PBCs). resonator/strip wire unit cell. d, Computational domain of the FDTD simulation,
The maximum simulated frequency of interest is fmax = 15 GHz and the spatial incident/reflected/transmitted electric field time-domain waveforms and
resolution corresponds to 20 cells per wavelength in each direction at fmax. scattering parameters of the split ring resonator/strip wire unit cell. PML,
Twenty samples of the wavevector are taken between each edge of the perfectly matched layer. Parts c and d reprinted with permission from
Brillouin zone. The time step is 0.9 of the Courant–Friedrichs–Lewy limit ref. 119, IEEE.

coefficients. In this case, a curse of dimensionality translates to more obtained by subtracting the total field from the incident. The total
update equations. Intrusive PCE methods converge significantly faster transmitted electric field is sampled on a similar probe plane behind
than the non-intrusive PCE and Monte Carlo103. the unit cell. The waveforms of the incident, reflected and transmitted
Applications of intrusive PCE in FDTD are presented in refs. 105,106. fields are Fourier-transformed. Then, the Fourier transform of the
Fabrication tolerances for the metallic parts of microwave circuits reflected and transmitted electric fields is divided by the Fourier trans-
were taken into account105, and FDTD was formulated on a statistically form of the incident field to compute S11 and S21, respectively.
deformable, generalized curvilinear computational mesh (Fig. 2b). The magnitude of these coefficients is shown in Fig. 3d.

Brillouin diagrams for periodic structures Applications

Periodic structures are common in electromagnetic problems. Fields in This section highlights some technological areas where the FDTD
periodic structures are subject to the Floquet theorem107, which deter- method has provided important results and insights.
mines the relation between field components one lattice period apart.
This relation is translated to PBCs enabling infinite periodic structures Biomedical
to be modelled via the simulation of a single unit cell, terminated in FDTD is a valuable tool for simulation and optimization in bioimaging,
PBCs, as shown in Fig. 3a. biophotonics, bioelectronics and biosensors. It is versatile at modelling
PBC algorithms for FDTD models may be divided into two groups: different imaging modality setups, tissue structures and physical prop-
direct field methods, which use Maxwell’s equations and their standard erties. The FDTD method has been used to validate and improve opti-
FDTD implementation, and field transformation methods that intro- cal coherence tomography (OCT), partial wave spectroscopy (PWS),
duce auxiliary fields and update equations for these fields. A thorough surface-enhanced Raman scattering and MRI. For example, the FDTD
review is provided in ref. 108. method was used to design a silicon-based metalens for the sample
FDTD models equipped with PBCs have extracted Brillouin band arm of a conventional ultrahigh resolution OCT120, achieving a 30-fold
diagrams — angular frequency ω versus wavevector k — of periodic increase in depth of focus compared with traditional objectives with
structures109–112; band diagrams for microwave113 and optical114 metama- similar resolution. By determining the underlying relationship between
terials; scattering from periodic structures, such as frequency selective subdiffractional information in OCT images and biological sample sta-
surfaces115,116; and analysis of infinite phased arrays117,118. As an example, tistics,121 the high-resolution ultrastructure of biological features can
a 2D lattice of alumina cylinders modelled as one unit cell in an FDTD grid be investigated using a traditional diffraction-limited setup. The FDTD
is shown in Fig. 3a. Figure 3b shows the corresponding FDTD-calculated method had a significant role in validating interferometric spectros-
band diagram for the geometry of Fig. 3a. copy for optical statistical nanosensing, specifically the quantification
of subdiffractional refractive-index variations in biological media,
S-parameters for microwave circuits which led to the development of PWS microscopy122, as shown in Fig. 4a.
Scattering parameters (S-parameters) are the most common network PWS has been applied to measure genome structures at length scales
parameters to describe the performance of microwave circuits and below the resolution of conventional microscopy123. An FDTD-aided
devices in the frequency domain. Figure 3c shows a split ring resona- design of nanoparticle layers was used124 as a surface-enhanced Raman
tor or strip wire structure119. In Fig. 3d, the results of an FDTD simulation scattering substrate to enhance small molecule detection. To achieve
are presented, with the structure placed in a computational domain simultaneous improvement in transmission efficiency and reception
terminated in PMCs in the lateral direction, PECs in the vertical direc- sensitivity in MRI, FDTD was used125 to design a band-pass birdcage radio
tion and PMLs in the longitudinal direction. The normal incidence of frequency coil by combining a multichannel wireless radio frequency
a vertically polarized plane wave is simulated onto an infinite array element with a high permittivity material.
of split ring resonator/strip wires. FDTD is used to determine the reflec- Beyond bioimaging, the FDTD method has been used to study the
tion coefficient S11 and transmission coefficient S21. The source is a direct interaction of electromagnetic fields with biological materi-
modulated Gaussian pulse g(t) = exp(−(t−t0)2/Ts2) sin(2πfc t), in which als for the safety evaluation of electromagnetic absorption126,127 and
fc = 14 GHz, t0 = 3Ts and Ts = 7.5 ps to provide sufficient bandwidth to numerical dosimetry128, as has been shown in Fig. 4b. It has also been
extract the scattering parameters in the frequency range between 0 used to investigate cell behaviours and disease biomarkers related to
and 22 GHz. The incident and total electric fields are sampled on the retinal rods129 (Fig. 4c), cervical cells130, mitochondrial aggregation131
probe plane in front of the unit cell. The reflected electric field is and cancer cells132. The design of wearable devices133 and antennas134,

Nature Reviews Methods Primers | (2023) 3:75 9

0123456789();:
Primer

miniaturized biosensors36,135, nanolenses37 and microwave thermal 3D FDTD model may be required to accurately predict the propagation
ablation technologies127 has been enhanced with FDTD. characteristics144.
Both 3D localized and global models of the Earth–ionosphere
Geophysics waveguide have been generated68. Figure 5c provides example electro-
Vegetation, rocks, soil, snow and the ionosphere have complex con- magnetic propagation results for a pulse propagating globally around
stitutive properties often exhibiting inhomogeneity, dispersion and the Earth under ionospheric conditions predicted by the whole atmos-
anisotropy. Because the FDTD method is versatile at handling such phere community climate model with thermosphere and ionosphere
conditions, it has been applied to geophysical problems across a range extension.
of frequencies and spatial scales.
FDTD applications in geophysics include simulation of ground- Metamaterials
penetrating radar for the detection and localization of objects buried Artificial dielectrics with properties beyond those encountered in
underground or geological anomalies136, borehole sensors in geophysical natural media — referred to as metamaterials — have led to the discovery
exploration137, remote sensing of ionosphere anomalies and hydrocar- of novel structures that support unconventional wave phenomena.
bon deposits68, hypothesized electromagnetic earthquake precursors68, Examples include negative refraction and inverted Doppler shift, along
space weather hazards138, long-range radio communications139–144, with related applications such as planar, sub-wavelength focusing
propagation from lightning145 and high-frequency propagation through lenses, electromagnetic cloaks and continuously scanning leaky-wave
the magnetized ionosphere146,147. antennas. The transient behaviour of metamaterials may be naturally
Figure 5a shows an efficient strategy for modelling electromag- resolved by FDTD simulations. For example, the causal evolution of
netic wave propagation over long propagation paths in the Earth– negative refraction was initially disputed and was verified in several
ionosphere waveguide by taking advantage of the symmetry and signal FDTD papers, which illustrated the transient development of negatively
movement. Figure 5b provides results for an 8-Mm radio path using the refracted wavefronts at the interface between a positive-index and
model of Fig. 5a. For some propagation geometries and ranges, a fully negative-index medium148–150.

a c

× 10–4

× 10–5
10–1 ~150 µm

0.04

0.02

0.5
0.2

0.1

1.5
0

0
2

1
FDTD
0.6 NA
0.3 NA
∑/Γ2

Quadrature-form
~

0.6 NA
0.3 NA 7 7 7 7
Closed-form
L = 2.0 µm 0.6 NA
10–2 0.3 NA
102 lc (nm) 6 6 6 6

b
5 5 5 5

4 4 4 4

x (µm)
3 3 3 3

2 2 2 2

1 1 1 1

0 2.8 × 10–14 –2 –1 0 1 2 –2 –1 0 1 2 –2 –1 0 1 2 –2 –1 0 1 2
–1
SA (J kg ) z (µm) z (µm) z (µm) z (µm)

Fig. 4 | The finite-difference time-domain method applied to the biomedical numerical dosimetry study. c, Propagation of |Ey| inside the outer segment of the
field. a, Left: theoretical derivation of the partial wave spectroscopy imaging photoreceptor cell for four different snapshots in time calculated by the FDTD
modality, supported by the finite-difference time-domain (FDTD) method. Right: method. NA, numerical aperture. Part a reprinted with permission from ref. 122,
experimental partial wave spectroscopy image (scale bar, 20 µm), revealing APS, and ref. 123, AAAS. Part b reprinted with permission from ref. 128, IEEE. Part c
chromatin activities in live cells. b, Distribution of specific energy absorption reprinted with permission from ref. 129, IEEE.
(SA) on the surface (left) and inside the human head (right) of an FDTD-based

Nature Reviews Methods Primers | (2023) 3:75 10

0123456789();:
Primer

a Fig. 5 | The finite-difference time-domain method

Axis of symmetry applied to electromagnetic propagation in the
Earth–ionosphere waveguide. a, An efficient axially
symmetric (2.5-D) finite-difference time-domain (FDTD)
Perfect conductor
computational domain model. b, Electromagnetic
propagation results calculated by the model of part a
Ionosphere
Ionosphere over an 8-Mm radio path originating from a ground-level
1-kW vertical transmitter operating at 30 kHz. Ground
Vacuum ( εr = 15, σ = 0.01 ohmm) is assumed from the transmitter
Vacuum
to a distance of 2 Mm and seawater ( εr = 80, σ = 4ohmm)
is assumed thereafter. The geomagnetic field is
Moving FDTD Ground
oriented 45° from the ground and 45° from the plane
Transmitter domain of propagation. The vertical electric field strength is
plotted for typical day ( β = 0.3, h’ = 72km) and night
( β = 0.5, h’ = 87 km) exponential ionospheres over the
b
whole path. For the day + night case, day is assumed
80 Radio path in daytime up to 3 Mm and night beyond 4 Mm, with a continuous
Radio path in night-time terminator from 3 Mm to 4 Mm. The reverse is assumed
Radio path in daytime + night-time for the night + day scenario. Compared with the uniform
70
30 kHz Radio path in night-time + daytime
Vertical E field on the ground (dB (µV m–1))

ionosphere scenarios, the effect of the terminator on the

60 signal is visible from about 3.5 Mm from the transmitter.
The calculations were performed using horizontal
50
and vertical grid cell dimensions of 1 km and 1.33 km,
respectively, and time-step increment of 2.53 ms.
c, Example of electromagnetic propagation results from
40
a 3D global FDTD model69. A horizontal plane of radial
electric field components (Er) sampled just above the
30
surface of the Earth at 3.75 ms for a 300-Hz pulse
occurring from Salt Lake City, UT, USA at 10 pm UTC.
20
The 3D ionospheric conditions are defined by the
whole atmosphere community climate model with
10
thermosphere and ionosphere extension.

0
0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000
Range (km)

c
0
0

–1
-1
-1

--22
–2
log10(abs(Er))
Latitude

00
–3
--

–4
- 44
-

–5
-5
-5
Longitude

Figure 6 provides an example of FDTD simulation results for at the input owing to growing evanescent waves within the negative
metamaterials. As shown in the inset of Fig. 6, a negative index lens is index slab.
modelled consisting of a dispersive slab in between two semi-infinite
free space regions151. Figure 6 shows the transient evolution of the Optics and plasmonics
electric field on the left and right interfaces of the lens at the design The FDTD method is one of the main workhorse simulation tech-
frequency (15.98 GHz). As predicted by the theory of the perfect niques in optics153. FDTD is a flexible tool for accurately modelling
lens152, the electric field on the exit face is larger than the electric field optical devices with fine geometrical details and a high degree of field

Nature Reviews Methods Primers | (2023) 3:75 11

0123456789();:
Primer

0.3 quantization method is the standard quantization approach170–174.

First interface
Second interface The main idea is to find eigenmodes of Helmholtz or vector wave
0.2 equations, which may be viewed as electromagnetic uncoupled
harmonic oscillators.
Finding numerical eigenmodes may be computationally expensive
0.1
when photons carry broadband quantum information. In this case, the
Ez (V m–1)

FDTD method is advantageous. In recent work175, the original second

0 quantization formulation was extended to use the FDTD method. By
applying the unitary transformation to mode-ladder operators, electric
x field operators are rewritten in terms of the convolution between a new
–0.1
propagator and coordinate-ladder operators. The new propagator
n = +1 n = –1 n = +1
may be found using typical FDTD simulations. Once the new propaga-
–0.2 tor is found, the expectation value of arbitrary observables — such
as coincidence count — with respect to an initial quantum state are
y
–0.3 found. Note that initial quantum states are solutions to the quantum
0 500 1,000 1,500 2,000 state equation or Schrödinger equation, describing the initial state
Time steps of a system170,171,174–177. The Hamiltonian is time-independent in this
Fig. 6 | The finite-difference time-domain method applied to a metamaterial: system. See, for example, ref. 175, which provides a 1D FDTD solution
a negative index lens. Electric field as a function of time at the first and second for the new propagator.
interfaces of the negative index lens ( n = − 1 + j 0.01), with geometry shown in the Figure 8 depicts simulation results of the Hong–Ou–Mandel
inset. The source and image planes are indicated with dashed lines. The negative (HOM) effect, a well-known quantum physics phenomenon to meas-
index medium is implemented as a Drude dispersive medium. Adapted with ure the indistinguishability of two input photons. If the photons are
permission from ref. 151, IEEE. indistinguishable, the destructive interference between them in a
50/50 beam splitter produces a bunching effect, causing the output
photons to be measured at the same output port randomly. The degree
of indistinguishability is often quantified using the second-order cor-
confinement. Owing to this flexibility, it has been used to make new relation function or normalized coincidence of two photodetections
discoveries, such as the existence of the photonic nanojet154,155. Other at both output ports, represented by g (2), which encodes the intensity
examples include diffractive optical elements156, photonic bandgap fluctuations. A value of g(2) below 0.5 indicates the quantum nature of
structures157, grating couplers158, photonic crystal waveguides159, electromagnetic fields. The HOM effect has been successfully modelled
waveguide bends160 and ring-resonator devices161. using the numerical mode decomposition framework175. The aim is
The FDTD method is also suitable for simulating plasmonic struc- to achieve the same HOM effect using the quantum FDTD scheme.
tures that explore surface plasmons — collective electron excitations — As shown in Fig. 8, both results are in strong agreement, showing the
at metal–dielectric interfaces162–164, or for confining and manipulating validity of the quantum FDTD method.
light below the diffraction limit. For example, the FDTD method has
been applied to the design of a plasmonic super-directive antenna165. Multiphysics modelling
As another example, Fig. 7 shows cross-sectional snapshots of a 3D The FDTD method has been coupled to several other solvers to model
FDTD simulation of an electric field guided at λ = 516.67 nm along a multiphysics phenomena. For example, electromagnetic–thermal
T-shaped chain of gold nanospheres with 25-nm diameter and 75-nm coupling was shown in ref. 178, in which the FDTD method was used to
intercentre spacing. The FDTD mesh is terminated by a PML and is com- determine the steady-state field distribution on the conducting parts
posed of 448 × 544 × 256 cells with cubic grid cells of size 1.5625 nm. of a microelectromechanical switch. Additionally, FDTD algorithms
As seen in Fig. 7, the coupling of plasmon resonances between adja- have been developed for the Schrödinger equation to model quantum
cent nanospheres produces a guided field with deep-sub-wavelength phenomena and for the coupled system of Maxwell and Schrödinger
confinement along the chain. equations for coupled electromagnetic–quantum effects. Further-
more, the FDTD method has been coupled to charge transport mod-
Quantum FDTD els to replace quasistatic models for semiconductor devices, such
There are two fundamental classes of subatomic particles: fermions — as microwave field effect transistors and power amplifiers179,180. The
electrons, protons and neutrons — which make up matter; and bosons — FDTD-computed electromagnetic fields are linked to hydrodynamic
photons, gluons and bosons — which carry forces. Density-functional equations — current continuity, energy and momentum conserva-
theory166–169 is an effective quantum-mechanical computational method tion equations — that were used to determine the charge distribution
for solving the approximate electronic structures of many-body, fermi- and current density in the device. As a last example, FDTD time step-
onic systems. It is particularly useful for designing (artificial) materi- ping may be linked to the Simulation Program with Integrated Circuit
als in solid-state physics. However, single photons, bosons, are ideal Emphasis (SPICE) models of linear and nonlinear electronic circuits
for carrying quantum information as they barely interact with the and systems181,182 to model wave–circuit interactions in structures such
environment or with each other. as active antennas, diode-reconfigurable microwave components and
Superposition and entanglement are the two major factors in high-speed interconnects.
developing quantum information science technology. First, to cor- FDTD is readily coupled to ordinary and partial differen-
rectly model and design quantum electromagnetic systems, the quan- tial equation solvers in the time domain, facilitating modelling of
tization of classical electromagnetic systems is needed. The second dynamic phenomena. For example, coupled electromagnetic-circuit

Nature Reviews Methods Primers | (2023) 3:75 12

0123456789();:
Primer

simulations may be run by linking FDTD to SPICE-type solvers and results from individual research laboratories and publications.
using the same time step to synchronously model wave–circuit Furthermore, many commercial software companies offer data sets
interactions181,182. Similarly, FDTD may be coupled to heat equation that may be used as inputs to FDTD models.
solvers to model the transient evolution of temperature in lossy
media in the presence of electromagnetic fields. Such simulations Limitations and optimizations
can include temperature-dependent media, whose constitutive To avoid poor FDTD modelling results and unexpected outcomes, it is
properties may be adjusted as the temperature changes, taking important to understand the limitations of the FDTD method. Several
advantage of the dynamic nature of FDTD modelling. Finally, mul- limitations — numerical dispersion, staircasing error and stability limit —
tiphysics phenomena can also involve extreme multiscale model- were discussed in the preceding sections and are not considered again
ling. A recent example is the quantum–electromagnetic modelling of here. Rather, some additional limitations of the FDTD algorithm are
nanowires, in which an alternating-direction hybrid implicit–explicit briefly discussed, along with workarounds and optimizations.
FDTD technique was used to circumvent the extreme stability limit
of conventional FDTD imposed by the presence of nanowires in the Speeding up FDTD simulations
computational domain183. The maximum time-step increment for the conventional FDTD method
is limiting. Alongside unconditionally stable FDTD methods, other
Reproducibility and data deposition workarounds are available to improve the performance of the FDTD
Reproducing FDTD data using alternative implementations of the FDTD method for long simulations. For example, in lossy media at sufficiently
algorithm requires information about the geometry of the problem — low frequencies, the conduction current may be stronger than the
including specification of sources, material and boundary conditions — displacement current by several orders of magnitude. It is possible to
and information on the parameters of the FDTD algorithm, such as scale up the permittivity and time-step increment ∆t while minimally
grid cell size and time step. If all simulation parameters are aligned, affecting the accuracy of the results184. Other types of parameter scaling
minor discrepancies between the results, for instance, deviations in the may also be used185.
computed time-domain waveforms, may be attributed to differences Other options for speeding up an FDTD simulation include relax-
in the absorbing boundary conditions, including implementation of ing the stability limit by applying a spatial filter to the simulated
PMLs, or differences in the treatment of material boundaries using field components to suppress unstable spatial harmonics186; apply-
staircasing, subcell or other techniques. ing signal-processing techniques to extrapolate the results from a
Discrepancies between FDTD simulated and measured data may short FDTD simulation to later time points187,188 and approximating
result from minor discrepancies between material parameters owing the continuation of geometries via PBCs. The final approach was used
to random fluctuations of the dielectric permittivity and conductiv- to simulate high-frequency electromagnetic propagation through
ity of materials, or discrepancies between the physical and simulated the ionosphere189.
geometry owing to discretization, for instance, staircasing of slanted
boundaries. Additive errors from numerical dispersion
FDTD results may be reproduced by other simulation methods, In addition to phase error, numerical dispersion effects may also intro-
both differential and integral equation-based. Comparison between duce additive errors into FDTD-calculated data. Numerical dispersion
such data sets must consider the different ways each method accumu- effects establish a dependence of the numerical wave impedance on the
lates numerical errors. As a result, comparisons cannot rely on aligning FDTD grid cell size. This dependence causes an impedance mismatch
the simulation parameters of each method. However, comparisons may at the subgridding interfaces where the cell size abruptly changes.
be made between the convergent results of FDTD and other methods, Such impedance mismatches produce spurious reflections, which
determined by varying the simulation parameters, such as cell size. are additive errors. The application of spatial filters to the modified
No data repository specific for FDTD results exists at present. FDTD update equations at the subgridding interfaces may mitigate
However, many university and journal-based repositories provide additive errors190.

a b c Fig. 7 | The finite-difference time-domain method

Ex @ 27.5 ps Ey @ 27.5 ps applied to plasmonics. a, A diagram of a T-junction
chain of gold nanospheres. b,c, Finite-difference
800 800 time-domain cross-sectional snapshots of the
700 700
amplitude of the electric field components Ex
T-junction (panel b) and Ey (panel c) along the T-junction chain
600 600 of gold nanospheres. The chain is excited by an
500 500 x-oriented dipole to the left (indicated by the arrow),
y (nm)
y (nm)

with wavelength λ = 516.67 nm. The nanospheres

400 400
have 25-nm diameter and 75-nm intercentre
300 300 spacing. The gold permittivity is represented by a
200 200 Drude model with a relaxation time τ = 4 fs and bulk
plasmon frequency ω = 6.79 × 1015 rad s−1. Coupling
100 100
of the plasmon resonances between adjacent
0 0 nanospheres produces the guided field propagation
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700 with deep-sub-wavelength confinement.
x (nm) x (nm)

Nature Reviews Methods Primers | (2023) 3:75 13

0123456789();:
Primer

a b
бx = cτ
Delay

50/50 Beam splitter

1
At t = 0

x
0.8

0.6

g(2)(τ)
Single photon Single photon

0.4
50/50 Beam splitter

After
interference
0.2
x
Quantum FDTD method
Numerical mode decomposition
2 1 0
–0.5 0 0.5
τ (ns)

Coincidence

Fig. 8 | Results from a quantum finite-difference time-domain model. a, Schematic of the simulated quantum beam splitter to observe the Hong–Ou–Mandel effect.
b, The Hong–Ou–Mandel effects are numerically evaluated by using numerical canonical quantization (solid line) and Q-finite-difference time-domain (FDTD) method
(square markers).

Source implementation artefacts by the FDTD method only on high-performance computing platforms.
FDTD is a charge-conserving algorithm. As a result, current source Unlike integral equation methods, which connect source and observa-
distributions must either be divergence-free or have no direct cur- tion points via Green’s functions, the FDTD method requires global
rent offset to avoid charge deposition on the FDTD grid nodes. Such discretization of the entire computational domain. To accelerate the
node-deposited charges are stationary and equivalent to infinitely modelling of thin sheets, including metasurfaces, stable FDTD imple-
massive-charged particles that produce spurious static fields191. mentations of surface impedance and generalized sheet transition
A sudden initiation of a source at t = 0 may lead to the excitation of conditions are needed. PMLs add to the computational cost of a 3D
high-frequency field transients that are poorly resolved in the FDTD simulation. Boundary integral approaches destroy the matrix-free
grid that alter the results. Smoothing ramp functions may be applied nature of FDTD. A challenge remains to find a thinner, but still effective,
to the early source signals to reduce such high-frequency components184. absorbing boundary condition while preserving the matrix-free nature
of FDTD. Finally, subgridding approaches have been developed but
Unconditional instabilities more stable algorithms that permit larger time-stepping increments
Special care must be taken when implementing nontrivial modifications are needed.
to the FDTD method, such as adding subgrids or using non-orthogonal
FDTD algorithms, to avoid unconditional instabilities192–194. Such insta- Opportunities for FDTD
bilities may arise from subtle inconsistencies in the spatial discretiza- Although there are challenges, there are also numerous opportunities
tion. This may lead to a violation of the mathematical properties (div for the FDTD method.
curl = 0 or curl grad = 0) or physical principles, such as positive defi-
niteness of the electromagnetic energy density or reciprocity, in the Scientific machine learning. Scientific machine learning is driven by
discrete setting. Discrete exterior calculus tools are useful for avoiding three fundamental questions: how can training data sets be generated
inconsistencies during the spatial discretization process195–197. by sampling a set of model input parameters? How should efficient
learning structures be chosen among the many existing and emerg-
Outlook ing forms of neural networks? And what types of objectives — field
This section highlights current challenges and opportunities for the computation, scattering parameters, optimization and uncertainty
FDTD method, as well as ongoing and future priorities and applications. quantification — should be targeted by developing such models?
Each of these questions may be considered in the context of FDTD
Challenges modelling.
Despite the rapid growth and widespread use of the FDTD method, chal- Standard PINNs use automatic differentiation to compute any
lenges remain. For example, the FDTD method is limited by numerical derivatives included in their loss functions. Alternatively, decades of
dispersion and stability. Large-scale problems are generally solvable FDTD research may be applied to approximate these derivatives on

Nature Reviews Methods Primers | (2023) 3:75 14

0123456789();:
Primer

uniform, non-uniform or subgridded FDTD meshes. A recent study has Future priorities and applications
shown that such an approach results in PINNs that may use time steps Priorities for FDTD research include exploiting new high-performance
beyond the CFL constraint for the corresponding FDTD grid98. This is computing resources and hardware to expand FDTD to larger-scale
a promising direction for further research. problems; integrating machine-learning algorithms in FDTD computa-
tions; developing FDTD algorithms on quantum computers; using the
FDTD as a design optimization tool. The FDTD method has three FDTD method for inverse problems and design optimization by inte-
distinct advantages as a design tool: accuracy, versatility and the ability grating with other methods and topology optimizations; developing
to provide broadband characterization of geometries in a single simula- stable, highly-accurate space–time subgrids for extreme multiscale
tion. Computational costs may be overcome by using efficient tech- problems; and exploring multiphysics FDTD algorithms, including
niques to compute the output function derivatives with respect to the quantum, chemical and mechanical effects.
input parameters, such as the adjoint variable method198. Surrogate The impact of the FDTD method is likely to see continual growth
models can be developed via a reduced number of FDTD simulations199. for a diverse range of applications. This includes inverse designs and
Furthermore, high-performance computing resources and hardware optimizations of meta-structures in microwaves and optics; modelling
acceleration can be used, as shown in a recent 3D simulation of a and optimization of time-varying metasurfaces; and analysis of wire-
100λ × 100λ area metalens simulation in less than 5 min using less communication links, such as radiation exposure studies. Another
hardware-accelerated FDTD200. area where growth is expected is in simulating biological phenomena,
including multiphysics interactions in molecular communications,
Quantum FDTD prospects. Evolving from scientific proof-of-principle for example, protein folding stimulated by terahertz waves and thera-
experiments, quantum technologies are poised to be distributed in peutic applications of electromagnetic fields. One last example is
chips on the basis of nano-optics technology, photonic integrated modelling extreme multiscale structures, including quantum effects.
circuits and microwave engineering technology201. This is a crucial
step towards scalable quantum systems. It also necessitates an effec- Published online: xx xx xxxx
tive full-wave physics numerical framework that can simulate the time References
evolution of quantum information carried by single or entangled 1. Maxwell, J. C. VIII. A dynamical theory of the electromagnetic field. Philos. Trans. R.
photons in a complex environment. Extending the 1D quantum FDTD Soc. Lond. 155, 459–512 (1865).
2. Chew, W. C. et al. Quantum Maxwell’s equations made simple: employing scalar and
scheme into higher-dimensional space involving a lossy environment vector potential formulation. IEEE Antennas Propag. Mag. 63, 14–26 (2020).
will be important to understand the full physics of quantum informa- 3. Kong, J. A. Electromagnetic Wave Theory (Wiley, 1990).
tion propagation at the on-chip level. It will also benefit a range of other 4. Balanis, C. A. Advanced Engineering Electromagnetics 2nd edn (Wiley, 2012).
5. Yee, K. Numerical solution of initial boundary value problems involving Maxwell’s
quantum technologies, such as quantum communications, quantum equations in isotropic media. IEEE Trans. Antennas Propag. 14, 302–307 (1966).
computing and quantum sensing. Paper in which Yee introduced the FDTD method.
6. Courant, R., Friedrichs, K. & Lewy, H. Über die partiellen differenzengleichungen der
mathematischen Physik. Math. Ann. 100, 32–74 (1928).
Other novel approaches in development. Research developments 7. Taflove, A. & Hagness, S. C. Computational Electromagnetics: The Finite-Difference
include higher-order methods for high accuracy at coarse discretiza- Time-Domain Method 3rd edn (Artech House, Inc., 2005).
A thorough overview of many key algorithms and applications of the method.
tion rates, using concepts from FETD and the discontinuous Galer-
8. Tan, T. & Potter, M. FDTD discrete planewave (FDTD-DPW) formulation for a perfectly
kin time-domain method202. The stable and robust hybridization of matched source in TFSF simulations. IEEE Trans. Antennas Propag. 58, 2641–2648
FDTD with other methods — such as the method of moments, the (2010).
9. Harms, P., Mittra, R. & Ko, W. Implementation of the periodic boundary condition in the
finite-element method, ray tracing and the vector parabolic equation
finite-difference time-domain algorithm for FSS structures. IEEE Trans. Antennas Propag.
method — continues to attract attention. 42, 1317–1324 (1994).
The FDTD method registers the complete history of pulses propa- 10. Kesler, M. P., Maloney, J. G., Shirley, B. L. & Smith, G. S. Antenna design with the use of
photonic band-gap materials as all-dielectric planar reflectors. Microw. Opt. Technol. Lett.
gating through a computational domain, along with its multiple reflec-
11, 169–174 (1996).
tions from objects and discontinuities. However, the sharp edge of a 11. Smith, D. R., Burns, S., Simpson, J. J. & Ferrone, S. M. FDTD modeling of scattered
microstrip structure may not be continuously illuminated by the initial ultra-low frequency electromagnetic waves from objects submerged in the ocean.
IEEE Trans. Antennas Propag. 67, 2534–2541 (2019).
and reflected pulses. Therefore, static mesh refinement is a suboptimal 12. Oh, K. S. & Schutt-Aine, J. E. An efficient implementation of surface impedance boundary
solution to the grid refinement problem. To overcome this, adaptive conditions for the finite-difference time-domain method. IEEE Trans. Antennas Propag.
mesh refinement strategies based on the intelligent mesh regeneration 43, 660–666 (1995).
13. Beggs, J. H., Luebbers, R. J., Yee, K. S. & Kunz, K. S. Finite-difference time-domain
and dynamic adaptivity during simulation time203 may be used. implementation of surface impedance boundary conditions. IEEE Trans. Antennas
Attempts have been made to hybridize FDTD with other Propag. 40, 49–56 (1992).
frequency-domain and time-domain numerical techniques. For exam- 14. Berenger, J.-P. A perfectly matched layer for the absorption of electromagnetic waves.
J. Comput. Phys. 114, 185–200 (1994).
ple, ref. 204 hybridized FDTD with the method of moments to effi- 15. Katz, D. S., Thiele, E. T. & Taflove, A. Validation and extension to three dimensions of the
ciently embed thin-wire antennas in an FDTD mesh without invoking Berenger PML absorbing boundary condition for FDTD meshes. IEEE Microw. Guided
thin-wire or subgridding techniques. In ref. 205, FDTD was combined Wave Lett. 4, 268–270 (1994).
16. Chew, W. C. & Weedon, W. H. A 3D perfectly matched medium from modified
with ray tracing to solve indoor wireless propagation problems. Ray Maxwell’s equations with stretched coordinates. Microw. Opt. Technol. Lett. 7, 599–604
tracing modelled propagation over electrically large distances, whereas (1994).
FDTD was applied to compute the interaction of wireless signals with 17. Teixeira, F. L. & Chew, W. C. Complex space approach to perfectly matched layers:
a review and some new developments. Int. J. Numer. Model. Electron. Netw. Devices
walls and corners. Generally, hybrid methods follow a similar pattern Fields 13, 441–455 (2000).
of exploiting the versatility of FDTD, while improving its efficiency for a Describes the PML concept and its various extensions with an emphasis on the
particular class of problems by using a different method. These hybrid- complex coordinate mapping approach.
18. Kuzuoglu, M. & Mittra, R. Frequency dependence of the constitutive parameters of causal
izations take advantage of the fundamental connections between perfectly matched anisotropic absorbers. IEEE Microw. Guided Wave Lett. 6, 447–449
numerical techniques206. (1996).

Nature Reviews Methods Primers | (2023) 3:75 15

0123456789();:
Primer

19. Teixeira, F. L. & Chew, W. C. Differential forms, metrics, and the reflectionless absorption 51. Maloney, J. G. & Smith, G. S. The efficient modeling of thin material sheets in the
of electromagnetic waves. J. Electromagn. Waves Appl. 13, 665–686 (1999). finite-difference time-domain (FDTD) method. IEEE Trans. Antennas Propag. 40, 323–330
20. Sacks, Z. S., Kingsland, D. M., Lee, R. & Lee, J.-F. A perfectly matched anisotropic (1992).
absorber for use as an absorbing boundary condition. IEEE Trans. Antennas Propag. 43, 52. Farjadpour, A. et al. Improving accuracy by subpixel smoothing in the finite-difference
1460–1463 (1995). time domain. Opt. Lett. 31, 2972–2974 (2006).
21. Roden, J. A. & Gedney, S. D. Convolution PML (CPML): an efficient FDTD implementation 53. Nadobny, J., Sullivan, D., Wlodarczyk, W., Deuflhard, P. & Wust, P. A 3-D tensor
of the CFS–PML for arbitrary media. Microw. Opt. Technol. Lett. 27, 334–339 (2000). FDTD-formulation for treatment of sloped interfaces in electrically inhomogeneous
22. Moreno, E. et al. Implementation of open boundary problems in photo-conductive media. IEEE Trans. Antennas Propag. 51, 1760–1770 (2003).
antennas by using convolutional perfectly matched layers. IEEE Trans. Antennas Propag. Introduces a tensor FDTD formulation for the treatment of non-rectangular boundaries
64, 4919–4922 (2016). crossing a Yee cell.
23. Wang, J.-F., Li, G. & Chen, Z. Convolutional implementation and analysis of the CFS-PML 54. Karkkainen, M. K. Subcell FDTD modeling of electrically thin dispersive layers. IEEE Trans.
ABC for the FDTD method based on wave equation. IEEE Microw. Wirel. Compon. Lett. 32, Microw. Theory Tech. 51, 1774–1780 (2003).
811–814 (2022). 55. Zhao, Y. & Hao, Y. Finite-difference time-domain study of guided modes in
24. Teixeira, F. L. Time-domain finite-difference and finite-element methods for nano-plasmonic waveguides. IEEE Trans. Antennas Propag. 55, 3070–3077 (2007).
Maxwell equations in complex media. IEEE Trans. Antennas Propag. 56, 2150–2166 56. Chevalier, M. W., Luebbers, R. J. & Cable, V. P. FDTD local grid with material traverse.
(2008). IEEE Trans. Antennas Propag. 45, 411–421 (1997).
Provides a thorough review of the FDTD algorithm applied to complex media. 57. Kim, I. S. & Hoefer, W. J. R. A local mesh refinement algorithm for the time domain-finite
25. Shibayama, J., Suzuki, K., Yamauchi, J. & Nakano, H. Trapezoidal recursive convolution‐ difference method using Maxwell’s curl equations. IEEE Trans. Microw. Theory Tech. 38,
based FDTD method for arbitrary‐shaped dispersive materials. Electron. Lett. 54, 812–815 (1990).
1429–1430 (2018). 58. Zivanovic, S. S., Yee, K. S. & Mei, K. K. A subgridding method for the time-domain
26. Okoniewski, M., Mrozowski, M. & Stuchly, M. A. Computationally efficient algorithms finite-difference method to solve Maxwell’s equations. IEEE Trans. Microw. Theory Tech.
for multi-term dielectric dispersion in FDTD. in IEEE Antennas and Propagation Society 39, 471–479 (1991).
International Symposium 1997 364–367 (IEEE, 1997). 59. Okoniewski, M., Okoniewska, E. & Stuchly, M. A. Three-dimensional subgridding
27. Pereda, J. A., Vielva, L. A., Solano, M. A., Vegas, A. & Prieto, A. FDTD analysis of algorithm for FDTD. IEEE Trans. Antennas Propag. 45, 422–429 (1997).
magnetized ferrites: application to the calculation of dispersion characteristics 60. Kulas, L. & Mrozowski, M. Low-reflection subgridding. IEEE Trans. Microw. Theory Tech.
of ferrite-loaded waveguides. IEEE Trans. Microw. Theory Tech. 43, 350–357 (1995). 53, 1587–1592 (2005).
28. Okoniewski, M. & Okoniewska, E. FDTD analysis of magnetized ferrites: a more efficient 61. Thoma, P. & Weiland, T. A consistent subgridding scheme for the finite difference time
algorithm. IEEE Microw. Guided Wave Lett. 4, 169–171 (1994). domain method. Int. J. Numer. Model. Electron. Netw. Devices Fields 9, 359–374 (1996).
29. Yu, Y., Niu, J. & Simpson, J. J. A 3-D global earth–ionosphere FDTD model including Presents consistent and provably stable formulations for FDTD subgridding.
an anisotropic magnetized plasma ionosphere. IEEE Trans. Antennas Propag. 60, 62. Xiao, K., Pommerenke, D. J. & Drewniak, J. L. A three-dimensional FDTD subgridding
3246–3256 (2012). algorithm with separated temporal and spatial interfaces and related stability analysis.
30. Ha, S.-G., Cho, J., Choi, J., Kim, H. & Jung, K.-Y. FDTD dispersive modeling of human IEEE Trans. Antennas Propag. 55, 1981–1990 (2007).
tissues based on quadratic complex rational function. IEEE Trans. Antennas Propag. 61, Presents consistent and provably stable formulations for FDTD subgridding.
996–999 (2013). 63. Berenger, J.-P. Origin of parasitic solutions with Holland and Simpson thin wires in the
31. Choi, H., Baek, J.-W. & Jung, K.-Y. Comprehensive study on numerical aspects of modified FDTD grid. IEEE Trans. Electromagn. Compat. 61, 487–494 (2019).
Lorentz model-based dispersive FDTD formulations. IEEE Trans. Antennas Propag. 67, Describes the origin of the parasitic solutions for the thin-wire formulation.
7643–7648 (2019). 64. Holland, R. & Simpson, L. Finite-difference analysis of EMP coupling to thin struts and
32. Sullivan, D. M. Frequency-dependent FDTD methods using Z transforms. IEEE Trans. wires. IEEE Trans. Electromagn. Compat. EMC-23, 88–97 (1981).
Antennas Propag. 40, 1223–1230 (1992). 65. Teixeira, F. L. & Chew, W. C. Finite-difference computation of transient electromagnetic
33. Kosmas, P., Rappaport, C. M. & Bishop, E. Modeling with the FDTD method for breast waves for cylindrical geometries in complex media. IEEE Trans. Geosci. Remote Sens. 38,
cancer detection. IEEE Trans. Microw. Theory Tech. 52, 1890–1897 (2004). 1530–1543 (2000).
34. Kast, M. J. & Elsherbeni, A. Z. Integration of nonlinear circuit elements into FDTD method 66. Pérez-Ocón, F., Pozo, A. M., Jiménez, J. R. & Hita, E. Fast single-mode characterization
formulation. in Advances in Time‐Domain Computational Electromagnetic Methods of optical fiber by finite-difference time-domain method. J. Lightwave Technol. 24, 3129
(ed. Rodolfo, A.) 1–31 (SciTech Publishing, 2022). (2006).
35. Holland, R. THREDS: a finite-difference time-domain EMP code in 3d spherical 67. Lee, H. O. & Teixeira, F. L. Cylindrical FDTD analysis of LWD tools through anisotropic
coordinates. IEEE Trans. Nucl. Sci. 30, 4592–4595 (1983). dipping-layered earth media. IEEE Trans. Geosci. Remote Sens. 45, 383–388 (2007).
36. Said, F. A. et al. FDTD analysis of structured metallic nanohole films for LSPR-based 68. Simpson, J. J. Current and future applications of 3-D global earth–ionosphere models
biosensor. in IEEE Regional Symposium on Micro and Nanoelectronics (RSM) (IEEE, 2015). based on the full-vector Maxwell’s equations FDTD method. Surv. Geophys. 30, 105–130
37. Mudanyali, O. et al. Wide-field optical detection of nanoparticles using on-chip (2009).
microscopy and self-assembled nanolenses. Nat. Photon. 7, 247–254 (2013). Summarizes developments and applications of FDTD modelling of the Earth–ionosphere
38. Joseph, R. M. & Taflove, A. FDTD Maxwell’s equations models for nonlinear waveguide.
electrodynamics and optics. IEEE Trans. Antennas Propag. 45, 364–374 (1997). 69. Simpson, J. J. & Taflove, A. Three-dimensional FDTD modeling of impulsive ELF
39. Boyd, R. Nonlinear Optics 3rd edn (Academic Press, 2018). propagation about the earth-sphere. IEEE Trans. Antennas Propag. 52, 443–451 (2004).
40. Bhardwaj, S., Teixeira, F. L. & Volakis, J. L. Fast modeling of terahertz plasma-wave devices 70. Samimi, A. & Simpson, J. J. Parallelization of 3-D global FDTD earth–ionosphere
using unconditionally stable FDTD methods. IEEE J. Multisc. Multiphys. Comput. Tech. 3, waveguide models at resolutions on the order of ∼1 km and higher. IEEE Antennas Wirel.
29–36 (2018). Propag. Lett. 15, 1959–1962 (2016).
41. Biabani, S. & Foroutan, G. Self consistent multi-fluid FDTD simulations of a nanosecond 71. Lee, R. A note on mass lumping in the finite element time domain method. IEEE Trans.
high power microwave discharge in air. Phys. Lett. A 382, 2720–2731 (2018). Antennas Propag. 54, 760–762 (2006).
42. Godfrey, B. B. & Vay, J.-L. Suppressing the numerical Cherenkov instability in FDTD PIC 72. Teixeira, F. L. A summary review on 25 years of progress and future challenges in FDTD
codes. J. Comput. Phys. 267, 1–6 (2014). and FETD techniques. Appl. Comput. Electromagn. Soc. J. 25, 1–14 (2022).
43. Na, D.-Y. et al. Diagnosing numerical Cherenkov instabilities in relativistic plasma 73. Rylander, T. & Bondeson, A. Stable FEM–FDTD hybrid method for Maxwell’s equations.
simulations based on general meshes. J. Comput. Phys. 402, 108880 (2020). Computer Phys. Commun. 125, 75–82 (2000).
44. Nicolaides, R. A method for complex geometries in finite-difference solutions of 74. Chilton, R. A. & Lee, R. Conservative and provably stable FDTD subgridding. IEEE Trans.
Maxwell’s equations. Comput. Math. Appl. 48, 1111–1119 (2004). Antennas Propag. 55, 2537–2549 (2007).
45. Lee, H. O. & Teixeira, F. L. Locally-conformal FDTD for anisotropic conductive interfaces. 75. Jung, K.-Y. & Teixeira, F. L. in Advanced Time Domain Modeling for Electrical Engineering
IEEE Trans. Antennas Propag. 58, 3658–3665 (2010). (ed. Rodolfo, A.) 217–252 (SciTech Publishing, IET Press, 2022).
46. Zhao, Q. & Sarris, C. D. Generalized tensor FDTD method for sloped dispersive interfaces 76. Tan, E. L. Fundamental schemes for efficient unconditionally stable implicit
and thin sheets. Opt. Express 27, 15812–15826 (2019). finite-difference time-domain methods. IEEE Trans. Antennas Propag. 56, 170–177 (2008).
47. Cabello, M. R. et al. A new efficient and stable 3D conformal FDTD. IEEE Microw. Wirel. 77. Shi, S.-B., Shao, W., Wei, X.-K., Yang, X.-S. & Wang, B.-Z. A new unconditionally stable
Compon. Lett. 26, 553–555 (2016). FDTD method based on the Newmark-Beta algorithm. IEEE Trans. Microw. Theory Tech.
48. Srisukh, Y., Nehrbass, J., Teixeira, F. L., Lee, J.-F. & Lee, R. An approach for automatic grid 64, 4082–4090 (2016).
generation in three-dimensional FDTD simulations of complex geometries. IEEE Antennas 78. Sun, G. & Trueman, C. W. Efficient implementations of the Crank–Nicolson scheme
Propag. Mag. 44, 75–80 (2002). for the finite-difference time-domain method. IEEE Trans. Microw. Theory Tech. 54,
Describes an automatic 3D FDTD grid generation for complex geometries. 2275–2284 (2006).
49. Heinrich, W., Beilenhoff, K., Mezzanotte, P. & Roselli, L. Optimum mesh grading for 79. Moon, H., Teixeira, F. L., Kim, J. & Omelchenko, Y. A. Trade-offs for unconditional stability
finite-difference method. IEEE Trans. Microw. Theory Tech. 44, 1569–1574 (1996). in the finite-element time-domain method. IEEE Microw. Wirel. Compon. Lett. 24,
50. Mezzanotte, P., Roselli, L., Huber, C., Zscheile, H. & Heinrich, W. On the accuracy of 361–363 (2014).
the finite-difference method using mesh grading. in Proceedings of 1995 IEEE MTT-S 80. Namiki, T. A new FDTD algorithm based on alternating-direction implicit method.
International Microwave Symposium Vol. 2, 781–784 (IEEE, 1995). IEEE Trans. Microw. Theory Tech. 47, 2003–2007 (1999).

Nature Reviews Methods Primers | (2023) 3:75 16

0123456789();:
Primer

81. Zheng, F., Chen, Z. & Zhang, J. A finite-difference time-domain method without the 115. Aminian, A., Yang, F. & Rahmat-Samii, Y. Bandwidth determination for soft and hard
Courant stability conditions. IEEE Microw. Guided Wave Lett. 9, 441–443 (1999). ground planes by spectral FDTD: a unified approach in visible and surface wave regions.
82. Yuan, C. & Chen, Z. D. On the modeling of conducting media with the unconditionally IEEE Trans. Antennas Propag. 53, 18–28 (2005).
stable ADI-FDTD method. IEEE Trans. Microw. Theory Tech. 51, 1929–1938 (2003). 116. Yang, F., Chen, J., Qiang, R. & Elsherbeni, A. A simple and efficient FDTD/PBC algorithm
83. Shibayama, J., Muraki, M., Yamauchi, J. & Nakano, H. Efficient implicit FDTD algorithm for scattering analysis of periodic structures. Radio Sci. https://doi.org/10.1029/
based on locally one-dimensional scheme. Electron. Lett. 41, 1–2 (2005). 2006rs003526 (2007).
84. Nascimento, V. E., Borges, B.-H. V. & Teixeira, F. L. Split-field PML implementations for 117. Holter, H. & Steyskal, H. Infinite phased-array analysis using FDTD periodic boundary
the unconditionally stable LOD-FDTD method. IEEE Microw. Wirel. Compon. Lett. 16, conditions-pulse scanning in oblique directions. IEEE Trans. Antennas Propag. 47,
398–400 (2006). 1508–1514 (1999).
85. Ong, C. et al. Speed it up. IEEE Microw. Mag. 11, 70–78 (2010). 118. Turner, G. M. & Christodoulou, C. FDTD analysis of phased array antennas. IEEE Trans.
86. Shlager, K. L. & Schneider, J. B. Comparison of the dispersion properties of several Antennas Propag. 47, 661–667 (1999).
low-dispersion finite-difference time-domain algorithms. IEEE Trans. Antennas Propag. 119. Zhao, Q. & Sarris, C. D. Discontinuous Galerkin time domain modeling of metasurface
51, 642–653 (2003). geometries with multi-rate time stepping. in 2021 IEEE MTT-S International Microwave
87. Hadi, M. F. & Piket-May, M. A modified FDTD (2, 4) scheme for modeling electrically large Symposium (IMS) (IEEE, 2021).
structures with high-phase accuracy. IEEE Trans. Antennas Propag. 45, 254–264 (1997). 120. Guo, S. et al. Metalens for improving optical coherence tomography. J. Korean Phys. Soc.
88. Zhao, S. & Wei, G. W. High-order FDTD methods via derivative matching for Maxwell’s 81, 32–37 (2022).
equations with material interfaces. J. Comput. Phys. 200, 60–103 (2004). 121. Eid, A., Winkelmann, J. A., Eshein, A., Taflove, A. & Backman, V. Origins of subdiffractional
89. Law, Y.-M. & Nave, J.-C. High-order FDTD schemes for Maxwell’s interface problems with contrast in optical coherence tomography. Biomed. Opt. Express 12, 3630–3642 (2021).
discontinuous coefficients and complex interfaces based on the correction function 122. Cherkezyan, L. et al. Interferometric spectroscopy of scattered light can quantify the
method. J. Sci. Comput. https://doi.org/10.1007/s10915-022-01797-9 (2022). statistics of subdiffractional refractive-index fluctuations. Phys. Rev. Lett. https://doi.org/
90. Wang, S. & Teixeira, F. L. Lattice models for large-scale simulations of coherent wave 10.1103/physrevlett.111.033903 (2013).
scattering. Phys. Rev. E https://doi.org/10.1103/physreve.69.016701 (2004). 123. Li, Y. et al. Nanoscale chromatin imaging and analysis platform bridges 4D chromatin
91. Finkelstein, B. & Kastner, R. A comprehensive new methodology for formulating FDTD organization with molecular function. Sci. Adv. 7, eabe4310 (2021).
schemes with controlled order of accuracy and dispersion. IEEE Trans. Antennas Propag. 124. Sun, G., Fu, C., Dong, M., Jin, G. & Song, Q. The finite-difference time-domain (FDTD)
56, 3516–3525 (2008). guided preparation of Ag nanostructures on Ti substrate for sensitive SERS detection of
92. Zygiridis, T. T., Papadopoulos, A. D. & Kantartzis, N. V. Error-optimized finite-difference small molecules. Spectrochim. Acta Pt A: Mol. Biomolecular Spectrosc. 269, 120743 (2022).
modeling of wave propagation problems with Lorentz material dispersion. J. Comput. 125. Seo, J.-H., Han, Y. & Chung, J.-Y. A comparative study of birdcage RF coil configurations
Phys. 452, 110916 (2022). for ultra-high field magnetic resonance imaging. Sensors 22, 1741 (2022).
93. Karniadakis, G. E. et al. Physics-informed machine learning. Nat. Rev. Phys. 3, 422–440 126. Taflove, A. & Brodwin, M. E. Computation of the electromagnetic fields and induced
(2021). temperatures within a model of the microwave-irradiated human eye. IEEE Trans.
94. Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: a deep Microw. Theory Tech. 23, 888–896 (1975).
learning framework for solving forward and inverse problems involving nonlinear partial 127. Chiang, J., Wang, P. & Brace, C. L. Computational modelling of microwave tumour
differential equations. J. Comput. Phys. 378, 686–707 (2019). ablations. Int. J. Hyperth. 29, 308–317 (2013).
95. Sirignano, J. & Spiliopoulos, K. DGM: a deep learning algorithm for solving partial 128. Chakarothai, J., Watanabe, S. & Wake, K. Numerical dosimetry of electromagnetic pulse
differential equations. J. Comput. Phys. 375, 1339–1364 (2018). exposures using FDTD method. IEEE Trans. Antennas Propag. 66, 5397–5408 (2018).
96. E, W. & Yu, B. The deep Ritz method: a deep learning-based numerical algorithm for 129. Hajiaboli, A. & Popovie, M. FDTD analysis of light propagation in the human
solving variational problems. Commun. Math. Stat. 6, 1–12 (2018). photoreceptor cells. IEEE Trans. Magn. 44, 1430–1433 (2008).
97. Zhang, P. et al. A Maxwell’s equations based deep learning method for time domain 130. Arifler, D. et al. Light scattering from normal and dysplastic cervical cells at different
electromagnetic simulations. IEEE J. Multisc. Multiphys. Comput. Tech. 6, 35–40 (2021). epithelial depths: finite-difference time-domain modeling with a perfectly matched layer
98. Qi, S. & Sarris, C. D. Electromagnetic-thermal analysis with FDTD and physics-informed boundary condition. J. Biomed. Opt. 8, 484–494 (2003).
neural networks. IEEE J. Multisc. Multiphys. Comput. Tech. 8, 49–59 (2023). 131. Su, X.-T., Singh, K., Rozmus, W., Backhouse, C. & Capjack, C. Light scattering
99. Qi, S. & Sarris, C. D. Deep neural networks for rapid simulation of planar microwave characterization of mitochondrial aggregation in single cells. Opt. Express 17,
circuits based on their layouts. IEEE Trans. Microw. Theory Tech. 70, 4805–4815 (2022). 13381–13388 (2009).
100. Folk, M., Heber, G., Koziol, Q., Pourmal, E. & Robinson, D. An overview of the HDF5 132. Simpson, J. J., Capoglu, I. R. & Backman, V. Using FDTD to improve our understanding
technology suite and its applications. in AD ‘11 Proc. EDBT/ICDT 2011 Workshop on Array of partial wave spectroscopy for advancing ultra early-stage cancer detection
Databases 36–47 (ACM, 2011). techniques. in 2009 13th International Symposium on Antenna Technology and Applied
101. Hastings, F. D., Schneider, J. B. & Broschat, S. L. A Monte-Carlo FDTD technique for rough Electromagnetics and the Canadian Radio Science Meeting (IEEE, 2009).
surface scattering. IEEE Trans. Antennas Propag. 43, 1183–1191 (1995). 133. Silla, G., Bastianelli, L., Colella, E., Moglie, F. & Primiani, V. M. SAR computation due to
102. Le Maître, O. & Knio, O. M. Spectral Methods for Uncertainty Quantification: With wearable devices by using high-resolution body models and FDTD numerical code. in
Applications to Computational Fluid Dynamics (Springer-Verlag, 2010). 2022 International Symposium on Electromagnetic Compatibility — EMC Europe. (IEEE).
103. Xiu, D. & Karniadakis, G. E. The Wiener–Askey polynomial chaos for stochastic differential 134. Al-Sehemi, A., Al-Ghamdi, A., Dishovsky, N., Atanasov, N. & Atanasova, G. Miniaturized
equations. SIAM J. Sci. Comput. 24, 619–644 (2002). wearable antennas with improved radiation efficiency using magneto-dielectric
104. Xiu, D. Numerical Methods for Stochastic Computations: A Spectral Method Approach composites. IETE J. Res. 68, 1157–1167 (2022).
(Princeton Univ. Press, 2010). 135. Arunkumar, R., Suaganya, T. & Robinson, S. Design and analysis of 2D photonic crystal
105. Austin, A. C. M. & Sarris, C. D. Efficient analysis of geometrical uncertainty in the FDTD based biosensor to detect different blood components. Photonic Sens. 9, 69–77 (2019).
method using polynomial chaos with application to microwave circuits. IEEE Trans. 136. Warren, C., Giannopoulos, A. & Giannakis, I. gprMax: open source software to simulate
Microw. Theory Tech. 61, 4293–4301 (2013). electromagnetic wave propagation for ground penetrating radar. Computer Phys.
Provides an FDTD-based method for the modelling of geometric uncertainties. Commun. 209, 163–170 (2016).
106. Edwards, R. S., Marvin, A. C. & Porter, S. J. Uncertainty analyses in the finite-difference 137. Lee, H. O., Teixeira, F. L., San Martin, L. E. & Bittar, M. S. Numerical modeling of
time-domain method. IEEE Trans. Electromagn. Compat. 52, 155–163 (2010). eccentered LWD borehole sensors in dipping and fully anisotropic earth formations.
107. Ishimaru, A. Electromagnetic Wave Propagation, Radiation, and Scattering: From IEEE Trans. Geosci. Remote Sens. 50, 727–735 (2012).
Fundamentals to Applications 2nd edn (John Wiley & Sons, 2017). 138. Pokhrel, S., Nguyen, B., Rodriguez, M., Bernabeu, E. & Simpson, J. J. A finite difference
108. Taflove, A. & Hagness, S. C. Periodic structures. in Computational Electromagnetics: time domain investigation of electric field enhancements along ocean–continent
The Finite-Difference Time-Domain Method Ch. 13 (Artech House, Inc., 2005). boundaries during space weather events. J. Geophys. Res. Space Phys. 123, 5033–5046
109. Chan, C. T., Yu, Q. L. & Ho, K. M. Order-N spectral method for electromagnetic waves. (2018).
Phys. Rev. B 51, 16635–16642 (1995). 139. Thevenot, M., Bérenger, J.-P., Monediere, T. & Jecko, F. A FDTD scheme for the
110. Luo, C., Johnson, S. G., Joannopoulos, J. D. & Pendry, J. B. Negative refraction without computation of VLF-LF propagation in the anisotropic earth–ionosphere waveguide.
negative index in metallic photonic crystals. Opt. Express 11, 746–754 (2003). Ann. Telecommun. 54, 297–310 (1999).
111. Ward, A. J. & Pendry, J. B. Calculating photonic Green’s functions using a nonorthogonal 140. Bérenger, J.-P. FDTD computation of VLF–LF propagation in the Earth–ionosphere
finite-difference time-domain method. Phys. Rev. B 58, 7252–7259 (1998). waveguide. Ann. Telecommun. 57, 1059–1090 (2002).
112. Ward, A. J. & Pendry, J. B. A program for calculating photonic band structures, Green’s 141. Bérenger, J.-P. Reduction of the angular dispersion of the FDTD method in the
functions and transmission/reflection coefficients using a non-orthogonal FDTD method. Earth–ionosphere waveguide. J. Electromagn. Waves Appl. 17, 1225–1235 (2003).
Computer Phys. Commun. 128, 590–621 (2000). 142. Bérenger, J.-P. An implicit FDTD scheme for the propagation of VLF–LF radio waves in the
113. Kokkinos, T., Sarris, C. D. & Eleftheriades, G. V. Periodic finite-difference time-domain Earth–ionosphere waveguide. Comptes Rendus Phys. 15, 393–402 (2014).
analysis of loaded transmission-line negative-refractive-index metamaterials. IEEE Trans. 143. Bérenger, J.-P. FDTD propagation of VLF–LF waves in the presence of ions in the
Microw. Theory Tech. 53, 1488–1495 (2005). Earth–ionosphere waveguide. Ann. Telecommun. 75, 437–446 (2020).
114. Liu, Y., Sarris, C. D. & Eleftheriades, G. V. Triangular-mesh-based FDTD analysis 144. Burns, S., Gasdia, F., Simpson, J. J. & Marshall, R. A. 3-D FDTD modeling of long-distance
of two-dimensional plasmonic structures supporting backward waves at optical VLF propagation in the Earth–ionosphere waveguide. IEEE Trans. Antennas Propag. 69,
frequencies. J. Lightwave Technol. 25, 938–945 (2007). 7743–7752 (2021).

Nature Reviews Methods Primers | (2023) 3:75 17

0123456789();:
Primer

145. Hu, W. & Cummer, S. A. An FDTD model for low and high altitude lightning-generated EM 175. Na, D.-Y. & Chew, W. C. Quantum electromagnetic finite-difference time-domain solver.
fields. IEEE Trans. Antennas Propag. 54, 1513–1522 (2006). Quantum Rep. 2, 253–265 (2020).
146. Smith, D. R., Huang, C. Y., Dao, E., Pokhrel, S. & Simpson, J. J. FDTD modeling of high‐ 176. Na, D.-Y., Zhu, J. & Chew, W. C. Diagonalization of the Hamiltonian for finite-sized
frequency waves through ionospheric plasma irregularities. J. Geophys. Res. Space Phys. dispersive media: canonical quantization with numerical mode decomposition.
125, e2019JA027499 (2020). Phys. Rev. A 103, 063707 (2021).
147. Niknam, K. & Simpson, J. A review of grid-based, time-domain modeling of 177. Na, D.-Y., Zhu, J., Chew, W. C. & Teixeira, F. L. Quantum information preserving
electromagnetic wave propagation involving the ionosphere. IEEE J. Multisc. Multiphys. computational electromagnetics. Phys. Rev. A 102, 013711 (2020).
Comput. Tech. 6, 214–228 (2021). 178. Thiel, W., Tornquist, K., Reano, R. & Katehi, L. P. B. A study of thermal effects in
Summarizes the FDTD method applied to propagation in the ionosphere. RF-MEM-switches using a time domain approach. in 2002 IEEE MTT-S International
148. Cummer, S. A. Dynamics of causal beam refraction in negative refractive index materials. Microwave Symposium Digest (Cat. No. 02CH37278) (IEEE, 2002).
Appl. Phys. Lett. 82, 2008–2010 (2003). 179. Alsunaidi, M. A., Imtiaz, S. M. S. & El-Ghazaly, S. M. Electromagnetic wave effects on
Uses FDTD to demonstrate the dynamic evolution of negative refraction and microwave transistors using a full-wave time-domain model. IEEE Trans. Microw. Theory
sub-wavelength focusing with planar negative index lenses. Tech. 44, 799–808 (1996).
149. Foteinopoulou, S., Economou, E. N. & Soukoulis, C. M. Refraction in media with a 180. Grondin, R. O., El-Ghazaly, S. M. & Goodnick, S. A review of global modeling of charge
negative refractive index. Phys. Rev. Lett. https://doi.org/10.1103/physrevlett.90.107402 transport in semiconductors and full-wave electromagnetics. IEEE Trans. Microw. Theory
(2003). Tech. 47, 817–829 (1999).
150. Kokkinos, T., Islam, R., Sarris, C. D. & Eleftheriades, G. V. Rigorous analysis of negative 181. Piket-May, M. et al. High-speed electronic circuits with active and nonlinear components.
refractive index metamaterials using FDTD with embedded lumped elements. in 2004 in Computational Electrodynamics: The Finite-Difference Time-Domain Method Ch. 15
IEEE MTT-S International Microwave Symposium Digest (IEEE Cat. No. 04CH37535) (Artech House, Inc., 2005).
(IEEE, 2004). 182. Sui, W., Christensen, D. A. & Durney, C. H. Extending the two-dimensional FDTD method
151. Li, D. & Sarris, C. D. A unified FDTD lattice truncation method for dispersive media based to hybrid electromagnetic systems with active and passive lumped elements. IEEE Trans.
on periodic boundary conditions. J. Lightwave Technol. 28, 1447–1454 (2010). Microw. Theory Tech. 40, 724–730 (1992).
152. Pendry, J. B. Negative refraction makes a perfect lens. Phys. Rev. Lett. 85, 3966–3969 183. Decleer, P. & Vande Ginste, D. A hybrid EM/QM framework based on the ADHIE-FDTD
(2000). method for the modeling of nanowires. IEEE J. Multisc. Multiphys. Comput. Tech. 7,
153. Taflove, A., Oskooi, A. & Johnson, S. G. Advances in FDTD Computational 236–251 (2022).
Electrodynamics: Photonics and Nanotechnology (Artech House, 2013). 184. Hue, Y.-K., Teixeira, F. L., Martin, L. S. & Bittar, M. S. Three-dimensional simulation of
A perspective on advanced FDTD algorithms in the context of photonics and eccentric LWD tool response in boreholes through dipping formations. IEEE Trans.
nanotechnology. Geosci. Remote Sens. 43, 257–268 (2005).
154. Chen, Z., Taflove, A. & Backman, V. Photonic nanojet enhancement of backscattering 185. Zhang, Y., Simpson, J. J., Welling, D. & Liemohn, M. Improving the efficiency of Maxwell’s
of light by nanoparticles: a potential novel visible-light ultramicroscopy technique. equations FDTD modeling for space weather applications by scaling the speed of light.
Opt. Express 12, 1214–1220 (2004). in 2020 IEEE USNC-CNC-URSI North American Radio Science Meeting (Joint with AP-S
This paper uses FDTD to discover the existence of photonic nanojets. Symposium) (IEEE, 2020).
155. Karamehmedović, M., Scheel, K., Listov-Saabye Pedersen, F., Villegas, A. & Hansen, P.-E. 186. Sarris, C. D. Extending the stability limit of the FDTD method with spatial filtering. IEEE
Steerable photonic jet for super-resolution microscopy. Opt. Express 30, 41757–41773 Microw. Wirel. Compon. Lett. 21, 176–178 (2011).
(2022). 187. Jandhyala, V., Michielssen, E. & Mittra, R. FDTD signal extrapolation using the
156. Prather, D. W. & Shi, S. Formulation and application of the finite-difference time-domain forward–backward autoregressive (AR) model. IEEE Microw. Guided Wave Lett. 4, 163–165
method for the analysis of axially symmetric diffractive optical elements. J. Opt. Soc. (1994).
Am. A 16, 1131–1142 (1999). 188. Nayak, I., Kumar, M. & Teixeira, F. L. Detection and prediction of equilibrium states in
157. Tong, M.-S. et al. Analysis of photonic band-gap (PBG) structures using the FDTD method. kinetic plasma simulations via mode tracking using reduced-order dynamic mode
Microw. Opt. Technol. Lett. 41, 173–177 (2004). decomposition. J. Comput. Phys. 447, 110671 (2021).
158. Sánchez‐Postigo, A. et al. Breaking the coupling efficiency–bandwidth trade‐off in 189. Smith, D. R., Tan, T., Dao, E., Huang, C. & Simpson, J. J. An FDTD investigation of
surface grating couplers using zero‐order radiation. Laser Photon. Rev. 15, 2000542 orthogonality and the backscattering of HF waves in the presence of ionospheric
(2021). irregularities. J. Geophys. Res. Space Phys. https://doi.org/10.1029/2020ja028201
159. Lavrinenko, A. V., Novitsky, A. & Zhilko, V. V. ARROW-based silicon-on-insulator photonic (2020).
crystal waveguides with reduced losses. Opt. Quantum Electron. 38, 815 (2007). 190. Donderici, B. & Teixeira, F. L. Domain-overriding and digital filtering for 3-D FDTD
160. Bahadori, M., Nikdast, M., Cheng, Q. & Bergman, K. Universal design of waveguide subgridded simulations. IEEE Microw. Wirel. Compon. Lett. 16, 10–12 (2006).
bends in silicon-on-insulator photonics platform. J. Lightwave Technol. 37, 3044–3054 191. Wagner, C. L. & Schneider, J. B. Divergent fields, charge, and capacitance in FDTD
(2019). simulations. IEEE Trans. Microw. Theory Tech. 46, 2131–2136 (1998).
161. Kedia, J. & Gupta, N. An FDTD analysis of serially coupled double ring resonator for 192. Schuhmann, R. & Weiland, T. Stability of the FDTD algorithm on nonorthogonal
DWDM. Optik 126, 5641–5644 (2015). grids related to the spatial interpolation scheme. IEEE Trans. Magn. 34, 2751–2754
162. Gray, S. K. & Kupka, T. Propagation of light in metallic nanowire arrays: finite-difference (1998).
time-domain studies of silver cylinders. Phys. Rev. B https://doi.org/10.1103/ 193. Wang, S. & Teixeira, F. L. Some remarks on the stability of time-domain electromagnetic
physrevb.68.045415 (2003). simulations. IEEE Trans. Antennas Propag. 52, 895–898 (2004).
163. Oubre, C. & Nordlander, P. Optical properties of metallodielectric nanostructures 194. Douvalis, V., Hao, Y. & Parini, C. Stable non-orthogonal FDTD method. Electron. Lett. 40,
calculated using the finite difference time domain method. J. Phys. Chem. B 108, 850–851 (2004).
17740–17747 (2004). 195. Teixeira, F. L. Lattice Maxwell’s equations. Prog. Electromagn. Res. 148, 113–128 (2014).
164. Zeng, Z., Venuthurumilli, P. K. & Xu, X. Inverse design of plasmonic structures with FDTD. 196. Glasser, A. S. & Qin, H. Generalizing Yee’s method: scalable geometric higher-order FEEC
ACS Photon. 8, 1489–1496 (2021). algorithms for Maxwell’s equations on an unstructured mesh. Preprint at https://doi.org/
165. Ludwig, A., Sarris, C. D. & Eleftheriades, G. V. Metascreen-based superdirective 10.48550/arXiv.2301.01753 (2023).
antenna in the optical frequency regime. Phys. Rev. Lett. https://doi.org/10.1103/ 197. Teixeira, F. L. & Chew, W. C. Lattice electromagnetic theory from a topological viewpoint.
physrevlett.109.223901 (2012). J. Math. Phys. 40, 169–187 (1999).
166. Gabay, D., Yilmaz, A., Boag, A. & Natan, A. Modeling electromagnetic wave phenomena 198. Nikolova, N. K., Tam, H. W. & Bakr, M. H. Sensitivity analysis with the FDTD method on
in large quantum systems: formulation and computational costs. IEEE Antennas structured grids. IEEE Trans. Microw. Theory Tech. 52, 1207–1216 (2004).
Propag. Mag. 63, 29–39 (2021). 199. Zhao, Q. & Sarris, C. D. Space–time adaptive modeling and shape optimization of
167. Martin, R. M. Electronic Structure: Basic Theory and Practical Methods (Cambridge microwave structures with applications to metasurface design. IEEE Trans. Microw.
Univ. Press, 2004). Theory Tech. 70, 5440–5453 (2022).
168. Sholl, D. S. & Steckel, J. A. Density Functional Theory (John Wiley & Sons, Ltd, 2009). 200. Hughes, T. W., Minkov, M., Liu, V., Yu, Z. & Fan, S. A perspective on the pathway toward
169. Payne, M. C., Teter, M. P., Allan, D. C., Arias, T. A. & Joannopoulos, J. D. Iterative full wave simulation of large area metalenses. Appl. Phys. Lett. 119, 150502 (2021).
minimization techniques for ab initio total-energy calculations: molecular dynamics 201. Zivari, A. et al. On-chip distribution of quantum information using traveling phonons.
and conjugate gradients. Rev. Mod. Phys. 64, 1045–1097 (1992). Sci. Adv. 8, eadd2811 (2022).
170. Chew, W. C., Liu, A. Y., Salazar-Lazaro, C. & Sha, W. E. Quantum electromagnetics: 202. Balsara, D. S. & Simpson, J. J. Making a synthesis of FDTD and DGTD schemes for
a new look — part I. IEEE J. Multisc. Multiphys. Comput. Tech. 1, 73–84 (2016). computational electromagnetics. IEEE J. Multisc. Multiphys. Comput. Tech. 5, 99–118
171. Chew, W. C., Liu, A. Y., Salazar-Lazaro, C. & Sha, W. E. Quantum electromagnetics: (2020).
a new look — part II. IEEE J. Multisc. Multiphys. Comput. Tech. 1, 85–97 (2016). 203. Sarris, C. D. Adaptive mesh refinement for time-domain numerical electromagnetics. in
172. Fox, A. M. & Fox, M. Quantum Optics: An Introduction Vol. 15 (Oxford Univ. Press, 2006). Synthesis Lectures on Computational Electromagnetics Vol. 1, 1–154 (Morgan & Claypool
173. Gerry, C., Knight, P. & Knight, P. L. Introductory Quantum Optics (Cambridge Univ. Publishers, 2007).
Press, 2005). 204. Huang, Z., Demarest, K. R. & Plumb, R. G. An FDTD/MoM hybrid technique for modeling
174. Miller, D. A. B. Quantum mechanics for scientists and engineers (Cambridge University complex antennas in the presence of heterogeneous grounds. IEEE Trans. Geosci.
Press, 2008). Remote Sens. 37, 2692–2698 (1999).

Nature Reviews Methods Primers | (2023) 3:75 18

0123456789();:
Primer

205. Wang, Y., Safavi-Naeini, S. & Chaudhuri, S. K. A hybrid technique based on combining ray Competing interests
tracing and FDTD methods for site-specific modeling of indoor radio wave propagation. The authors declare no competing interests.
IEEE Trans. Antennas Propag. 48, 743–754 (2000).
206. Chen, Z., Wang, C.-F. & Hoefer, W. J. R. A unified view of computational electromagnetics. Additional information
IEEE Trans. Microw. Theory Tech. 70, 955–969 (2022). Peer review information Nature Reviews Methods Primers thanks Theodoros Zygiridis, Jiefu
207. Joannopoulos, J. D., Johnson, S. G., Winn, J. N. & Meade, R. D. Photonic Crystals: Molding Chen, Jun Shibayama, Stephen Gedney and Zhizhang Chen for their contribution to the peer
the Flow of Light 2nd edn (Princeton Univ. Press, 2008). review of this work.

Acknowledgements Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
F.L.T. acknowledges support from the US Department of Energy Grant No. DE-SC0022982 published maps and institutional affiliations.
through the NSF/DOE Partnership in Basic Plasma Science and Engineering. Part of the
material from J.J.S. is based on work supported by the National Science Foundation Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this
under Grant No. 1662318. C.S. acknowledges support from the Natural Sciences and article under a publishing agreement with the author(s) or other rightsholder(s); author
Engineering Research Council of Canada (NSERC) through a Discovery Grant. The authors self-archiving of the accepted manuscript version of this article is solely governed by the
acknowledge the help of K. Niknam in the generation of Fig. 5e using input data provided terms of such publishing agreement and applicable law.
by T. Reichler.

Author contributions Related links

Introduction (F.L.T., Y.Z. and J.J.S.); Experimentation (F.L.T., C.S., J.-P.B., M.O. and J.J.S.); Results
Message passing interface: https://www.open-mpi.org/
(F.L.T., C.S. and J.J.S.); Applications (F.L.T., C.S., Y.Z., D.-Y.N., J.-P.B., Y.S., V.B., J.J.S. and W.C.C.);
OpenMP: https://www.openmp.org/resources/
Reproducibility and data deposition (C.S., Y.Z. and J.J.S.); Limitations and optimizations
(F.L.T., Y.Z. and J.J.S.); Outlook (C.S., Y.Z., D.-Y.N., J.J.S. and W.C.C.); Overview of the Primer
© Springer Nature Limited 2023
(all authors).

Nature Reviews Methods Primers | (2023) 3:75 19

0123456789();: