Electromagnetic Theory

CPML (Convolutional Perfectly Matched Layer)

Q: What is the difference between CPML and the original Berenger split-field PML?

Berenger's 1994 split-field PML divides each field component into two subcomponents and assigns separate conductivities to each, doubling the number of stored field variables in the boundary region. CPML instead uses an unsplit, stretched-coordinate formulation in which the spatial derivative is modified by a complex frequency-shifted stretching factor, and the resulting convolution is evaluated efficiently with a recursive auxiliary variable. The practical result is that CPML uses roughly 30 percent less memory than split-field PML and is far easier to implement because it touches only the field update equations through two extra arrays per boundary. CPML also absorbs evanescent waves and low-frequency or near-grazing-incidence energy that the original PML reflects, thanks to the complex frequency shift parameter alpha.

Q: How thick should the CPML region be and how should sigma be graded?

A CPML region is typically 8 to 16 cells thick. The conductivity profile is polynomially graded from zero at the inner boundary to a maximum at the outer wall using sigma(x) = sigma_max times (x/d) raised to the power m, with the polynomial order m between 3 and 4 for most problems. The optimal maximum conductivity is sigma_max = 0.8 times (m+1) divided by (delta_x times eta), where delta_x is the cell size and eta is the intrinsic impedance of the medium. Eight cells with m equal to 3 and properly scaled sigma_max give reflections near minus 40 dB; 12 to 16 cells push reflection below minus 60 to minus 80 dB at the cost of more memory and run time. Grading is essential because an abrupt conductivity jump would itself reflect energy back into the grid.

Q: What does the CFS alpha parameter do in CPML?

The complex frequency-shifted (CFS) stretching introduces an extra parameter alpha alongside the conductivity sigma and the real stretch kappa. The alpha term shifts the PML pole off the real frequency axis, which dramatically improves absorption of low-frequency and evanescent fields and of waves arriving at near-grazing angles. These are exactly the cases where the classic PML degrades and produces late-time reflections that contaminate broadband or transient simulations. A typical recipe grades alpha from a maximum of about 0.05 to 0.24 at the inner edge down to zero at the outer wall, while kappa is graded from 1 up to about 5 to 15. Tuning alpha and kappa together lets CPML remain stable and low-reflection for dispersive, lossy, and open-region antenna problems.

/kon-voh-LOO-shuh-nuhl PER-fik-tlee matcht LAY-er/

Used to terminate the edges of a finite-difference time-domain grid, the convolutional perfectly matched layer is an absorbing boundary that lets outgoing waves leave the computational domain as if it extended to infinity. Where the original split-field PML stores two subcomponents per field, CPML applies a complex frequency-shifted stretched-coordinate transform and evaluates it through recursive convolution, so only two extra auxiliary arrays per boundary are needed. The complex frequency shift (the alpha parameter) makes CPML absorb evanescent, low-frequency, and near-grazing-incidence energy that classic PML reflects, which is why it is the default boundary in modern computational electromagnetics solvers. A well-graded 8 to 16 cell CPML reaches reflection levels of minus 40 dB to minus 80 dB while using roughly 30 percent less memory than the Berenger formulation.

Category: Electromagnetic Theory

Layer Thickness: 8 to 16 cells

Reflection: −40 to −80 dB

How CPML Truncates an FDTD Grid

Every time-domain field solver works on a finite mesh, yet most antenna, scattering, and waveguide problems are open-region: the fields radiate outward toward infinity. A perfectly matched layer fixes this by surrounding the grid with a fictitious anisotropic absorber whose wave impedance is matched to the interior medium at every angle of incidence, so an outgoing wave passes the inner interface without reflection and is then attenuated inside the layer before it can return. The convolutional PML, introduced by Roden and Gedney in 2000, recasts that absorber in stretched coordinates and applies a complex frequency-shifted (CFS) stretching variable to each spatial derivative. Because the stretching is frequency dependent, the inverse transform back to the time domain becomes a convolution, which CPML evaluates with a single recursive auxiliary variable rather than storing the full field history.

The CFS stretching factor along a coordinate, written s = kappa + sigma / (alpha + jωε₀), carries three graded profiles: kappa, the real coordinate stretch; sigma, the loss term that does the actual absorbing; and alpha, the frequency shift that moves the absorber pole off the real axis. Setting alpha greater than zero is what separates CPML from the earlier uniaxial and split-field layers. It restores low-reflection performance for evanescent fields near a structure and for waves striking the boundary at glancing angles, the two regimes where standard PML develops slow, late-time reflections that corrupt broadband and transient results.

Compared with analytical boundaries such as the first or second-order Mur condition, CPML is dramatically more absorptive (Mur typically reflects at the minus 25 dB to minus 35 dB level and only near normal incidence), at the cost of the extra memory for the auxiliary arrays inside the boundary cells. That trade is almost always worthwhile for radar cross-section, antenna pattern, and full-wave electromagnetic work where late-time accuracy matters.

Governing Equations and Grading

CFS stretched-coordinate factor:
s_x = κ_x + σ_x / (α_x + jωε₀)

Polynomial conductivity grading:
σ(x) = σ_max × (x / d)^m, m = 3 to 4

Optimal maximum conductivity:
σ_max ≈ 0.8 × (m + 1) / (Δx × η)

Theoretical normal-incidence reflection:
R(0) = e^{−2σ_maxd / [(m+1)ε₀c]} → target R < −60 dB

Where d = layer thickness, Δx = cell size, η = intrinsic impedance, c = speed of light, κ grades from 1 to 5–15, and α grades from ~0.05–0.24 down to 0 at the outer wall.

CPML vs Other FDTD Boundaries

Boundary	Type	Typical Reflection	Memory Overhead	Grazing / Evanescent	Best Use
CPML (CFS)	Convolutional, unsplit	−40 to −80 dB	2 arrays / boundary	Excellent	Antennas, RCS, dispersive media
Berenger split-field PML	Split field	−40 to −80 dB	~30% more than CPML	Poor at grazing	Legacy / reference solvers
UPML	Uniaxial anisotropic	−40 to −70 dB	Moderate	Poor at grazing	Lossy media, simple layouts
Mur (2nd order)	Analytical ABC	−25 to −35 dB	Negligible	Poor off-normal	Quick checks, tight memory

Common Questions

Frequently Asked Questions

What is the difference between CPML and the original Berenger split-field PML?

Berenger's 1994 split-field PML divides each field component into two subcomponents with separate conductivities, doubling the stored variables in the boundary. CPML uses an unsplit, stretched-coordinate form whose convolution is evaluated by a recursive auxiliary variable, cutting memory by roughly 30 percent and simplifying the code to two extra arrays per boundary. The CFS shift also lets CPML absorb evanescent and near-grazing energy that split-field PML reflects.

How thick should the CPML region be and how should sigma be graded?

Use 8 to 16 cells. Grade conductivity polynomially as σ(x) = σ_max × (x/d)^m with m = 3 to 4, and set σ_max ≈ 0.8(m+1)/(Δx×η). Eight cells give about minus 40 dB; 12 to 16 cells reach minus 60 to minus 80 dB. Smooth grading is essential, since an abrupt conductivity step would itself reflect energy back into the grid.

What does the CFS alpha parameter do in CPML?

The complex frequency-shift parameter α moves the PML pole off the real frequency axis, which sharply improves absorption of low-frequency, evanescent, and near-grazing-incidence fields, the cases where classic PML develops late-time reflections. A common recipe grades α from about 0.05 to 0.24 at the inner edge to zero at the outer wall, while κ grades from 1 up to 5 to 15. Tuning α and κ together keeps CPML stable and low-reflection for dispersive and open-region antenna problems.

Related Terms

← Cpk CPR Flange →