Rough interfaces scatter radiation out of the specular direction. Two options control how BornAgain computes this diffuse scattering:
options = simulation.options()
options.setRoughnessApproximation("FullDWBA", 64)
options.setDwbaReferenceWaves("Sharp") # optional, default "Transient"
setRoughnessApproximation(approximation, terms) chooses the
approximation for the scattering from roughness, and the number of
terms of the sum that this approximation computes.setDwbaReferenceWaves(waves) chooses the waves from which roughness
and particles scatter.The options apply to ScatteringSimulation and OffspecSimulation.
Without these calls, BornAgain uses the defaults:
options.setRoughnessApproximation("FullDWBA", 1)
options.setDwbaReferenceWaves("Transient")
Versions before 25.1 had no such options and always computed this.
A rough interface deviates from its mean plane. Where interface $i$ lies at a height $h>0$ above its mean position, the material of the lower layer fills the space between $0$ and $h$ that the upper layer would fill at a flat interface. Where $h$ is negative, the two materials change roles. This displaced material scatters with the amplitude
$$ \int_0^h K_i(z)\thinspace dz. $$
For scalar waves at a nonmagnetic interface, the height kernel $K_i(z)$ is the difference between the scattering potentials of the two layers, multiplied by two reference waves at the height $z$: the wave that comes from the source, and the wave calculated for illumination from the direction of the detector.
The sum below assumes that the heights of all interfaces are jointly Gaussian random variables. Interface $i$ has the RMS height $\sigma_i$. The covariance $C_{ij}(r)$ says how the heights of the interfaces $i$ and $j$ are correlated at two points with the lateral distance $r$. The diffuse cross section per illuminated area is then a sum of terms $n=1,2,\ldots,N$:
$$ \frac{1}{A}\frac{d\sigma}{d\Omega} =\frac{k_0^4}{16\pi^2}\sum_{n=1}^{N}\sum_{i,j} b_{n,i}\thinspace S_{n,ij}(\nu)\thinspace b_{n,j}^{\ast}. $$
Here $A$ is the illuminated area, $k_0=2\pi/\lambda$ with the wavelength $\lambda$, and $\nu=|\mathbf q_\parallel|/2\pi$ is the spatial frequency, in cycles/nm, that belongs to the lateral wavevector transfer $\mathbf q_\parallel$. Each term is a product of two kinds of factors.
The coefficients
$$ b_{n,i}=\frac{1}{\sqrt{n!}} \left\langle K_i(h)\thinspace\mathrm{He}_{n-1}(h/\sigma_i)\right\rangle $$
describe the waves at interface $i$. The angle brackets denote the average over the Gaussian height $h$, and $\mathrm{He}_{n-1}$ is the probabilists’ Hermite polynomial of degree $n-1$.
The spectra
$$ S_{n,ij}(\nu)=\frac{2\pi}{(\sigma_i\sigma_j)^{n-1}} \int_0^\infty r\thinspace C_{ij}^n(r)\thinspace J_0(2\pi\nu r)\thinspace dr $$
describe the roughness alone; $J_0$ is the Bessel function of order zero. $S_1$ is the two-dimensional PSD of the heights; $S_n$ is the same kind of spectrum for the $n$-th power of the covariance.
Every term of the sum is nonnegative. For a given approximation and given reference waves, a sum that stops after $N$ terms therefore never exceeds the complete sum.
The options on this page choose the reference waves, the waves that enter the height kernel, and the number $N$ of terms.
With polarized neutrons, $K_i$ is a 2×2 matrix in spin space; the polarizer and the analyzer determine which of its components are measured. With polarized x rays, $K_i$ contains the products of the electric-field vectors of the two waves. The spectra $S_{n,ij}$ do not depend on polarization.
In the distorted-wave Born approximation (DWBA), particles and rough
interfaces scatter the waves of a simpler reference sample, which
consists of laterally uniform layers. setDwbaReferenceWaves chooses
how the interfaces of this reference sample are described:
"Transient" (default): the wave amplitudes in the layers are
computed as for specular reflection, with the factors of the transient
models (ErfTransient, TanhTransient). These factors approximate
the effect of a smooth transition between the two materials. The
roughness of one interface thus also changes the waves that reach
particles and other rough interfaces."Sharp": every interface is perfectly sharp and lies at its mean
position. The waves are those of the sample without roughness. The
rough interfaces still scatter diffusely, with the RMS heights and
spectra given in the sample.The same choice is used for particles and for roughness. It does not affect the specular beam: the reflectivity of a specular scan and the specular peak in a GISAS image are always computed with the transient models.
The transient waves are an approximation. In each layer they are plane waves, as for sharp interfaces, only with other amplitudes. They do not describe how the waves vary inside the transition region of a rough interface. This can become important at larger RMS heights.
Above and below the mean interface, the two layers have different waves.
setRoughnessApproximation(approximation, terms) chooses which of them
enter the height kernel $K_i(z)$:
When the RMS heights are small, all four approximations give the same result.
Rules for the arguments:
setRoughnessApproximation("HeightPerturbation").ErfTransient, the transient model of
Gaussian heights.TanhTransient and transient
reference waves, the tanh factors change the wave amplitudes, but the
height distribution of this model is not used. This is what earlier
versions computed.All approximations work with polarized x rays and neutrons, with analyzers, and with magnetic layers.
The number of terms $N$ says after which term the sum over $n$ stops. It is a numerical parameter: the physical approximation is the same for every $N$, and a larger $N$ only brings the sum closer to its limit. A larger $N$ does not make the DWBA or the reference waves more accurate.
How many terms are needed depends on the RMS height $\sigma$ and on the vertical wavevector transfer $q_z$. A useful scale is $|q_z\sigma|^2$. With LowerField or UpperField and a single pair of plane waves, the coefficients $|b_n|^2$ follow a Poisson distribution with this mean: they are largest for $n$ near $|q_z\sigma|^2$ and decay beyond it over a few times its square root. A sum that stops earlier can miss a large part of the intensity.
The table gives rough estimates of the relative error of the intensity that results from stopping after $N$ terms. Here $x=|q_z\sigma|^2$, where $\sigma$ is the largest RMS height in the sample and $q_z\approx k_0(\sin\alpha_i+\sin\alpha_f)$ is the largest vertical wavevector transfer on the detector. The estimates are for self-affine (fractal) roughness with Hurst exponent below 1.
| Number of terms $N$ | Error of the intensity |
|---|---|
| $x+2\sqrt{x}+2$ | 2 % |
| $x+3\sqrt{x}+3$ | 0.1 % |
| $x+4\sqrt{x}+5$ | 0.01 % |
The first row is usually enough for comparisons with measured data.
In some cases, more terms are needed than the table says:
To check convergence, repeat the simulation with $2N$ terms and compare the detector regions of interest.
The computing time grows with $N$, most for FullDWBA when $q_z\sigma$ is large.
For LowerField and UpperField, the sum over all $n$ is known in closed form. In each layer, the reference waves are plane waves; an incident and an outgoing plane wave together have a vertical wavevector transfer. Take one such pair with the transfer $q_a$ at interface $i$, and one with $q_b$ at interface $j$. Their contribution to the complete sum is proportional to $2\pi\int_0^\infty r\thinspace f(r)\thinspace J_0(2\pi\nu r)\thinspace dr$ with
$$ f(r)=e^{-\frac12(\sigma_i^2q_a^2+\sigma_j^2q_b^{\ast 2})} \frac{e^{q_a q_b^{\ast}C_{ij}(r)}-1}{q_a q_b^{\ast}}. $$
This is the form in which the result of Sinha et al. is usually written. BornAgain does not evaluate this integral; it adds the terms up to $N$.
If two adjacent layers have different magnetizations, a displaced interface adds or removes a thin sheet of magnetized material. The magnetic field of this sheet has two parts: the field inside the sheet, and a stray field that extends into the layers above and below and decays with the distance from the interface. Neutrons are scattered by both.
Consider one lateral Fourier component of the heights, with the wavevector $\mathbf q_\parallel$, $Q=|\mathbf q_\parallel|>0$ and the unit vector $\mathbf u=\mathbf q_\parallel/Q$. Let $\Delta\mathbf M$ be the magnetization below the interface minus that above, $\Delta\mathbf M_\parallel$ its part in the plane of the interface, $m_u=\mathbf u\cdot\Delta\mathbf M$, and $m_z=\Delta M_z$. Let $\widehat{\mathbf z}$ be the unit vector along the surface normal. For an interface at the height $z_j$, the field per unit upward displacement, divided by $\mu_0$, is
$$ \Delta\mathbf M_\parallel\delta(z-z_j) +\frac{Q}{2}e^{-Q|z-z_j|} (-m_u+i s m_z)(\mathbf u-i s\widehat{\mathbf z}), \qquad s=\operatorname{sign}(z-z_j). $$
The first term is the field of the sheet, the second the stray field.
Because the stray field reaches into the neighboring layers, the height kernel of a magnetic interface collects contributions from the reference waves in all layers, not only at the interface. This holds for every approximation. With HeightPerturbation, only the field of the sheet is taken at the mean interface; the stray field still enters with the waves of all layers.
LowerField and UpperField take the dependence of the kernel on the height from the lower or the upper side of the interface and use it on both sides, as for nonmagnetic interfaces. The field of the sheet is treated in the same way. This defines the two approximations for magnetic interfaces.
If the source or the detector lies below the sample, the corresponding reference wave passes through the substrate. It is normalized at the sample; in an absorbing substrate it therefore grows exponentially with the distance from the sample. If it grows faster than the stray field decays, the kernel is infinite, and the simulation stops with an error.
For scalar waves at a single nonmagnetic interface between vacuum and a medium with the refractive index $n$, HeightPerturbation gives
$$ \frac{1}{A}\frac{d\sigma}{d\Omega} =\frac{k_0^4}{16\pi^2} \left|(1-n^2)t_i t_f\right|^2 S_1(\nu). $$
Here $t_i$ and $t_f$ are the transmission amplitudes of the incident and
of the outgoing reference wave, and $S_1$ is the two-dimensional PSD of
the heights, in nm⁴. With Sharp waves, $t_i$ and $t_f$ are the Fresnel
amplitudes. With Transient waves, they are the Fresnel amplitudes
times the factors of the transient model.
HeightPerturbation is accurate if the RMS height is small compared with the vertical wavelengths: at each interface, $\sigma|k_z|\ll1$ must hold for the incident and the outgoing wave in both adjacent layers, where $k_z$ is the vertical wavenumber. Below the critical angle, $k_z$ is mostly imaginary; the condition then concerns the decay rate of the evanescent wave. In a magnetic layer, polarized neutrons have two vertical wavenumbers, and the condition must hold for both.