Rough interface models

Interface model

The mathematical model for one interface must be specified through one of

ba.SelfAffineFractalModel(...)
ba.PoweredExponentialModel(...)
ba.LinearGrowthModel(...)

These models define the behavior of autocorrelation function of roughness or, better to say, its Fourier spectrum in the full range of spatial frequencies from 0 to infinity.

Self-affine fractal model
ba.SelfAffineFractalModel(sigma, hurst, lateralCorrLength,
                          maxSpatFrequency=math.inf)

where

  • sigma, $\sigma$, is the RMS height before applying the frequency cutoff, in nm.
  • hurst, $H$, is a fractal exponent H with 0<H<=1. The smaller $H$ is, the more serrate the surface profile looks.
  • lateralCorrLength, $\xi$, is the positive lateral (in-plane) correlation length.
  • maxSpatFrequency, $\nu_{max}$, is a cut-off spatial frequency in nm⁻¹.

See the specular roughness example for a complete simulation using SelfAffineFractalModel.

This is the K-correlation model of Palasantzas 1993. The autocorrelation spectrum is

$$ S(\nu)=\dfrac{4 \pi H \sigma^2\xi^2}{(1+(2\pi\xi\nu)^2)^{1+H}} ,\quad for \quad \nu<\nu_{max} $$

In case there is no cut-off, the real-space roughness correlation function at the interface is expressed as: $$ < U(x, y) U(x’, y’)> = \sigma^2 \frac{2^{1-H}}{\Gamma(H)} \left( \frac{\tau}{ξ} \right)^H K_H \left( \frac{\tau}{ξ} \right), \quad \tau=[(x-x’)^2+(y-y’)^2]^{\frac{1}{2}} $$

where $U(x, y)$ is the height deviation at position $(x, y)$.

The main property is that it remains nearly constant at low $\nu$ and transitions into a straight line on a log-log scale at high $\nu$, exhibiting an inflection point between these two regions.

Powered-exponential model
ba.PoweredExponentialModel(sigma, hurst, lateral_corr_length,
                           max_spatial_frequency=math.inf)

where

  • sigma, $\sigma$, is the RMS height before applying the frequency cutoff, in nm.
  • hurst, $H$, is the Hurst exponent; the supported numerical range is 0.05 <= H <= 1.
  • lateral_corr_length, $\xi$, is the positive lateral correlation length, in nm.
  • max_spatial_frequency, $\nu_{max}$, is a cut-off spatial frequency in nm⁻¹.

Without a spatial-frequency cutoff, the height covariance is:

$$ C(r)=\sigma^2\exp\left[-(r/\xi)^{2H}\right], \qquad r=\sqrt{(x-x’)^2+(y-y’)^2}. $$

This covariance is called powered exponential (also generalized exponential); see Gneiting et al., Table 1. It is the covariance used in the rough-surface treatment of Sinha et al..

The two-dimensional PSD, in nm⁴, is

$$ S(\nu)=2\pi\int_0^\infty r C(r)J_0(2\pi\nu r)\thinspace dr. $$

For $H=0.5$, this is exactly the spectrum of SelfAffineFractalModel with the same parameters. For $H=1$,

$$ S(\nu)=\pi\sigma^2\xi^2\exp[-(\pi\xi\nu)^2]. $$

See the roughness spectra example for a comparison of the two models at several Hurst exponents.

Linear growth model
ba.LinearGrowthModel(particle_volume, damp1, damp2, damp3, damp4,
                     maxSpatFrequency=0.5)

where

  • particle_volume, $\Omega$, is the volume of particle (atom, molecule, cluster) coming to the surface during film deposition,
  • damp1-damp4, $a_1-a_4$, are damping coefficients, responsible for replication of underlying roughness and the growth of the new independent one,
  • maxSpatFrequency, $\nu_{max}$, is the cut-off spatial frequency of the roughness added by growth, in nm⁻¹.

This is the model described by Stearns 1993 and extended by Stearns and Gullikson 2000.

The model describes the evolution of the roughness spectrum on the top surface of a growing film. Its autocorrelation spectrum depends not only on the model parameters but also on the film thickness and the spectrum of the underlying interface. Therefore, the model cannot be applied directly to the substrate. On a layer without roughness, the film grows from a flat interface, $S_{below}=0$, and is uncorrelated with the interfaces further below. With nonzero growth noise, a finite maxSpatFrequency is required if damp3 and damp4 are both zero: otherwise the RMS diverges.

The autocorrelation spectrum is

$$ S_{above}(\nu)=S_{below}(\nu)e^{-b(\nu)t} + \Omega\dfrac{1-e^{-b(\nu)t}}{b(\nu)}, \quad for \quad \nu<\nu_{max} $$ where $t$ - film thickness, $b(\nu) = \sum{a_i\nu^i}$ - relaxation function. Above $\nu_{max}$, only the inherited term $S_{below}(\nu)e^{-b(\nu)t}$ remains.

An essential property of the model is that it describes not only autocorrelation but also cross-correlation properties. Therefore, it does not require cross-correlation to be specified separately.

The multilayer linear-growth example shows LinearGrowthModel above a self-affine substrate.

Cut-off frequency and effective rms roughness

The cut-off $\nu_{max}$ applies to the roughness that a model itself introduces at its interface, and sets it to zero above that spatial frequency: the whole spectrum of SelfAffineFractalModel and PoweredExponentialModel, the growth term of LinearGrowthModel. Inherited roughness is cut, if at all, by the model of the interface where it originates. The cut-off is a real frequency, not a circular one, in nm⁻¹. By default, SelfAffineFractalModel and PoweredExponentialModel have no cut-off, so that sigma is their rms height. LinearGrowthModel defaults to 0.5 nm⁻¹, a lateral period of 2 nm; pass math.inf to remove its cut-off when its RMS converges. No rescaling is applied to the spectrum below the cut-off. Cutting the spectrum also changes the real-space covariance.

For all three models, the effective rms height of the cut spectrum is used for the specular reflection and transmission through the transient profile, and the damping of the diffuse scattering amplitudes with the vertical momentum transfer.

For the self-affine fractal model, integrating the cut spectrum gives $$ \sigma_{\nu_{max}}^2 = \int_0^{\nu_{max}} S(\nu)\thinspace 2\pi\nu\thinspace d\nu = \sigma^2 \left[1 - \left(1 + (2\pi\xi\nu_{max})^2\right)^{-H}\right], $$ where $\sigma$ is the rms height of the full, uncut spectrum. For $\nu_{max} \gg 1/(2\pi\xi)$, $\sigma_{\nu_{max}}$ approaches $\sigma$. For $H = 0.5$ and $\xi = 12$ nm, a cut-off of 0.5 nm⁻¹ gives $\sigma_{\nu_{max}} = 0.99\thinspace\sigma$. A cut-off below $1/(2\pi\xi)$ removes most of the roughness: with the same $H$ and $\xi$, but $\nu_{max} = 0.003$ nm⁻¹, it is $\sigma_{\nu_{max}} = 0.16\thinspace\sigma$. For the powered-exponential and linear growth models, the same integral is evaluated numerically (with analytic limits where available).

With a low cut-off, the interfaces become smoother for the specular reflectivity, and the diffuse scattering changes through the reflected and transmitted amplitudes also at spatial frequencies below the cut-off. Lowering the cut-off is therefore not a way to restrict the diffuse scattering to low spatial frequencies while keeping the specular reflectivity.