WorksPhysics and Astronomy
An improved reliability factor for quantitative low-energy electron diffraction
Abstract
Quantitative low-energy electron diffraction [LEED or LEED , the evaluation of diffraction intensities as a function of the electron energy] is a versatile technique for the study of surface structures. The technique is based on optimizing the agreement between experimental and calculated intensities. Today, the most commonly used measure of agreement is Pendry's factor . While has many advantages, it also has severe shortcomings, as it is a noisy target function for optimization and very sensitive to small offsets of the intensity. Furthermore, , which is meant to imply perfect agreement between two curves can also be achieved by qualitatively very different curves. We present a modified factor , which can be used as a direct replacement for , but avoids these shortcomings. We also demonstrate that is as good as or better in steering the optimization to the correct result in the case of imperfections of the experimental data, while another common factor, (suggested by Zanazzi and Jona) is worse in this respect.
Introduction
Quantitative low-energy electron diffraction (LEED) analyzes the diffraction intensities as a function of the electron energy or equivalently, electron acceleration voltage . Thus, the technique is also known known as LEED or LEED , and the data are commonly referred to as or curves. Quantitative LEED is the oldest method for obtaining high-accuracy data in surface crystallography and has the advantage that it uses rather simple instrumentation (LEED optics), which is available in many ultrahigh-vacuum surface science systems [1, 2, 3, 4, 5, 6, 7, 8]. The ViPErLEED project [9, 10, 11, 12] aims at simplifying the use of this important method, and also includes further method development, as in the current work.
LEED analysis is based on the comparison of the calculated curves with the experimental ones. The agreement between calculated and experimental curves is described by a reliability factor ( factor). Assuming that the structure model is qualitatively correct, variation of the structural parameters (such as atom coordinates and vibration amplitudes) to minimize the factor leads to the best-fit structure. Many different types of factors have been proposed and used for LEED [2]. The simplest one is essentially an norm
where the subscripts “exp” and “th” refer to the experimental and calculated (theory) data, respectively. The factor takes into account that experiments determine only relative intensities; the ratio between the outgoing electron current of a given diffraction maximum (named diffraction “beam” in LEED studies) and the primary current of the incident electrons is usually not known quantitatively. In most cases, is not an ideal metric for LEED analysis, however. If two curves are qualitatively very similar, but the relative peak heights differ, values can be rather poor. Different relative peak heights between experiment and calculation are very common, however, e.g., because of energy-dependent gain factors in experiment or an inaccurate description of energy-dependent inelastic scattering in the calculations. Furthermore, is not very sensitive to the positions of the minima of the curves. This is undesirable since the minima carry important information and are less susceptible to various imperfections of experimental curves that strongly influence the maxima (e.g., uneven transmission of the grids, shot noise).
A more elaborate factor was proposed by Zanazzi and Jona [13]. It is based on the first and second derivatives of the curves:
The denominator is meant to put increased weight on regions of minima and maxima, where the first derivative vanishes.1 Like , uses an energy-independent scale factor between experimental and calculated intensities. This makes it sensitive to slow variations with the energy of the experimental or calculated intensity scale factor. A further disadvantage of lies in the fact that it heavily depends on the second derivative . Experimental curves inevitably contain some noise, but also the calculated curves often exhibit numerical noise (e.g., due to matrix inversions, if not all matrices are sufficiently well-conditioned to avoid numerical inaccuracies). Since the derivatives are calculated numerically, the second derivative is more sensitive to noise and small numerical errors than .
Pendry’s factor avoids these issues since it is based on the logarithmic derivative [14]. A direct comparison of the functions of the experimental and theoretical curves is not feasible because approaches infinity if approaches zero.2 Therefore, does not use directly but rather compares the experimental and calculated functions (named and , respectively):
Note that the function approaches zero for both and . This means that is insensitive to the high (or even infinite) values that the logarithmic derivative reaches in minima with the intensity approaching 0.
is the imaginary part of the inner potential3 and describes the inelastic scattering of the electrons. It provides a natural energy scale. Assuming single scattering (kinematic diffraction theory), the peaks of are Lorentzians with a full width at half maximum of [14]. LEED must be described by multiple scattering (dynamic scattering theory), however. For typical curves it turns out that the minimum distance between two maxima (or two minima) is about ; pronounced maxima with a deep minimum in between are usually separated by or more. Typical values of are between and eV.
Pendry’s factor takes values between 0 and 2: A value of 0 is achieved when the experimental and calculated functions are identical (ignoring a constant scale factor), and indicates uncorrelated curves. Values above 1 occur for anti-correlation between the functions (usually related to anti-correlation of the curves). Depending on the type of system, is usually considered good enough to lend high credibility to a structural model; values below 0.1 are considered excellent [5, 7].
Apart from its insensitivity to slow variations of the scale factor between experiment and calculation, and a high sensitivity to the positions of the minima and maxima, a further advantage of is the fact that it is based only on a local (point-by point) comparison of the curves. Since the function oscillates rapidly between its two extreme values (see § 4.1 for an example), the denominator of equation (5) is roughly proportional to the energy span of the integral.4 While and are dominated by the highest peaks, for the agreement of two curves in a particular region will always have the same effect on , largely independent of the remaining regions of the curve(s). These advantages of are not merely theoretical; based on real-world experimental data it has been shown that minimizing in a structure search yields more accurate results than using [15].
Finally, does not only provide a measure for the level of agreement between two curves (or two sets of such curves), it also forms the basis for an estimate of the errors of a LEED analysis. The confidence region of the fit parameters obtained by minimizing roughly reaches to the point where the factor calculated with the modified parameter (or parameter combination in the case of strong coupling between parameters [16]) reaches the value ,5 with
where is the combined energy span of all independent curves used in the analysis (also known as the size of the experimental data base used for the theory-experiment comparison) [14].6
Given all the advantages of listed above, one could consider it a perfect factor for LEED . The following section shows its weaknesses. Section 3 presents an improved factor (named ) that does not suffer from the deficiencies discussed in section 2, without sacrificing the advantages of .
Deficiencies of Pendry’s factor
Dissimilar curves can have
As mentioned above, the function, equation (4), reaches a value of 0 for both, and ; its extrema of are reached at (Figure 1a). This means that a given value can be reached for two different values of the logarithmic derivative (except for the two extrema of ). In other words, is not invertible. Therefore, one can obtain the same curve with two different curves, and Pendry’s factor between these two curves will be . This is exemplified in Figure 1b, where two qualitatively different synthetic curves (black and red) share the same function (blue). The plot was constructed by assuming the black curve (a raised cosine function), calculating the corresponding and functions, and then the inverse functions, taking the wrong branch of the inverse function in the regions marked by the red bars. This results in the red curve. It is obvious that the black and red functions are not proportional to each other and the factor between them should clearly be different from
Figure 1. (a) The function of Pendry’s factor (blue). The gray curve shows an invertible function, , which forms the basis for the factor presented in this work. The horizontal axis is the scaled logarithmic derivative of the intensity equation (3). (b) The plot shows two curves (black, red) that share the same function, shown in the bottom. This implies that Pendry’s factor between the two curves is , a value that should occur only if two curves are identical except for a constant scale factor. The red bars indicate where the two curves map onto different branches of the function; in these regions the two curves are dissimilar (not proportional to each other). The horizontal axis is the scaled energy, centered at the position of a minimum of the curves. For , the plots in (b) would span on the energy axis.
Figure 3. (a) factors and (b) their derivatives near the minimum. The horizontal axis is the displacement of the two symmetry-equivalent uppermost Fe atoms of the -FeO(1̄102)-(1 × 1) structure [9]. The vertical axis for the Zanazzi–Jona factor has been scaled by a factor of 7 since the curvature of vs. is lower than that of the two other factors by roughly this factor. The dashed lines indicate the error bar of the parameter derived from , equation (6). The plot does not include ; it is smooth, but its minimum would lie close to the right border of the plot. This is consistent with a previous study finding less accurate results when minimizing [15].
Nevertheless, since both functions share the same function, for these two dissimilar functions. One may argue that it is unlikely that such a case will occur as a pure coincidence. However, it is easy to envisage a case where a region of the experimental curve looks similar to one of the two curves in Figure 1b, and the calculated one is similar to the other curve. Then, when minimizing the factor, the curve in this energy region will “pull” the minimization process in a wrong direction, towards the correspondence given by the wrong branch of the function.
Pendry’s R factor is too sensitive to small intensity offsets
Experimental data are typically obtained by integrating the intensity of a diffraction maximum in a camera image of the LEED screen, after subtracting the background intensity caused by inelastically scattered electrons and stray light (e.g., from other bright spots) [17, 10]. The background intensity is not homogeneous and its exact distribution is not known, thus the background subtraction is inaccurate, especially for weak spots. In some cases, the intensity obtained this way can even become negative. Due to the uneven background, this problem also exist when fitting the peaks with a Gaussian (or similar function) [18] since the curvature of the background intensity will lead to a nonzero fit amplitude even in the absence of any true peak. On the computational side, the intensity reaches zero when the complex amplitude of the outgoing wave (as a function of ) passes exactly through the origin of the complex plane. Even tiny changes of the complex amplitude that would not be noticeable in other places can make a qualitative difference, whether the complex amplitude passes through the origin or misses that point. Since the minimum intensity cannot be determined with sufficient accuracy, an factor should not be overly sensitive to a small offset of the intensity at a deep minimum.
Figure 2a shows synthetic curves (again, raised cosine functions) with different intensity offsets (thus different intensities at the minimum); the corresponding functions of Pendry’s factor are shown in Figure 2b. If the intensity at the minimum is exactly zero, is an almost perfectly straight line with a moderate slope in the region around the intensity minimum (black curve). When adding a very small intensity offset, the slope of becomes very high, and extrema occur next to the energy of the intensity minimum. Thus, a small intensity offset (within the uncertainty of experimental curves) leads to a large change of the function (dark blue vs. black in Figure 2b). This will strongly affect the factor. As explained above, such an overly strong response to small intensity offsets is an undesirable property of an factor.
Figure 2. Intensities and various functions at a minimum with different intensity offsets. Note that all functions are sensitive to relative changes only, but invariant to a scale factor of the intensity. Thus, the red intensity curve in (a) would be equivalent to a curve with a peak-valley height of 1.0 (like the other curves), but an intensity of 2.6 at the minimum. (b) For small, but non-zero intensity offset, of Pendry’s factor reaches sharp minima and maxima next to the minimum (“cusps”). The dots indicate the sampling interval in the case of 0.5 eV energy steps and eV. (c) The modified function of equation (8) shows a hard step at an intensity minimum with a low or negligible offset (black curve), which is also undesirable. The function in (d) avoids these issues and depends smoothly on the intensity offset at the minimum.
One should also note that an intensity minimum that reaches exactly zero has a well-defined position on the energy axis, while the position of a shallow minimum is usually less certain. Nevertheless, the black function in Figure 2b has a very gentle slope, which means that is much less sensitive to the position of this minimum than to the positions of shallower minima, where the function is very steep (blue to orange). Near the minimum, a further problem becomes obvious in Figure 2b: The function of the minimum reaching zero intensity (black) is similar to that of the very shallow minimum shown in red. This is related to the fact that the function shown in equation (4) and Figure 1a is non-invertible; this is a similar case as discussed in section 2.1.
Pendry's R factor is noisy
For structure optimization, the factor is the objective function that should be minimized by varying the parameters of the structural model of the surface. Many minimization algorithms are sensitive to noise; this is especially true for gradient-based methods, which offer the best performance as soon as a rough (exploratory) search has identified the region of the global minimum [12]. It has been noted early that minimizing is plagued by noise [19]. The noisiness of as a function of structural parameters was initially attribute to numerical noise in the calculation of the curves [19], but it is easy to show that noise in is present also when numerical noise of the underlying calculations can be excluded. Figure 3a shows Pendry’s factor of a theory–experiment comparison for -FeO(1̄102)-(1 × 1) [9]. is plotted as a function of a single atomic coordinate, the position of the topmost two (symmetry-equivalent) Fe atoms. The calculated curves were not obtained by full-dynamic or tensor LEED [19] calculations (which might lead to numerical noise) but by a linear interpolation of the complex amplitudes between pm and pm.7 We took this approach because a simple linear interpolation cannot lead to numerical noise of the curves as a function of the parameter. Nevertheless, figure 3 shows that is noisy; it does not exhibit a smooth, parabola-like minimum. The roughness of Pendry’s factor becomes even more obvious when looking at the gradient; it is clear that optimization algorithms using this gradient will struggle finding the minimum (figure 3b; the gradient was evaluated numerically with 0.01 pm spacing of the points).
One might argue that the parameter range of pm in figure 3 is very small. This range is only slightly larger than the uncertainty of the parameter determined via [cf. equation (6)]. The gradient near the boundaries of this range points in the right direction, so one should find the minimum in spite of the noise. This may be true for a one-dimensional optimization problem, but LEED studies require high-dimensional optimization, typically with dozens,
sometimes even more than 100 parameters [20]. To illustrate the impact of the noise of on multidimensional optimization, Figure 4a shows a contour plot of as a function of two optimization parameters (the and coordinates of an atom). The plot shows that the gradients of can be grossly misleading also far from the minimum (red circles). High-dimensional minimization is much more demanding than finding a minimum in one or two dimensions [21]. Many minimization algorithms use the difference of the gradients between successive iteration points to construct an approximation for the Hessian of the minimization problem [22]. If the gradients are strongly affected by noise, this approximation for the Hessian will be grossly wrong. Then, in addition to getting trapped in local minima (like those marked by red crosses in Figure 4a), minimization algorithms may get stuck at off-minimum parameter values if the algorithm cannot cope with a noisy objective function with seemingly inconsistent gradients.
Figure 4. Contour plots of (a) and (b) as a function of the and displacement of the upper two symmetry-equivalent Fe atoms of -FeO(102)-(1 1), based on tensor-LEED calculations. The red circles mark a few points where the gradients of are grossly misleading, even far from the minimum. Minima are marked by small, red crosses.
The noise of is caused by the extrema of near a minimum (Figure 2b). As mentioned in § 2.2, the occurrence of these extrema depends sensitively on whether the intensity at the minimum reaches exactly zero or not. In practice, another important issue is the fact that the integration of equation (5) in current LEED codes is performed as a discrete sum, with an energy step of typically 0.5 or 1 eV.8 Figure 2b shows that even 0.5 eV steps are rather coarse compared to the sharpness of the sharpest extrema (the cusps) of (dark blue curve). Thus, depending on the position of the sampling points with respect to the cusps, a slight energy shift can lead to large changes of the values at the points sampled. This is another source of noise, which comes in addition to the overly sensitive response to the residual intensity at deep minima.
Modifying Pendry’s factor
The results of § 2 point at the function as the root of most deficiencies of Pendry’s factor, especially the fact that “folds back” towards 0 after reaching its extrema, which makes it non-invertible: the inverse function is ambiguous. It is easy to create modified functions that do not fold back for high values of and still suppress the infinity of at ; e.g.
This function is shown as the gray curve in Figure 1a. and the other functions discussed below have the same slope as near , and also the minimum and maximum function values are the same for all functions, .
avoids the sharp cusps near a deep intensity minimum and the overly strong dependence on small intensity offsets, see Figure 2c. This figure shows an undesirable property of , however: For minima with a small or vanishing offset (black and dark blue curves), has a step at the position of the minimum. When using the factor to compare two curves, the position of the minimum will usually be slightly different between the two curves. In the energy range between the minima of the two curves, their functions will be almost maximally different, . If the energy difference of the minima is less than the energy step used in the numerical evaluation, it comes down to coincidence whether one of the discrete energies used in the evaluation is between the minima of the two curves. If so, this energy will lead to a clear increase of the factor. Otherwise, if the energy grid happens to have no point the region between the minima of the two curves, the difference will affect the factor only marginally. As an example, when using 0.5 eV energy steps, an 0.4 eV difference between the minima of the two curves may go unnoticed in some cases, while in other cases an 0.1 eV difference gets penalized. It is clear that such an “all or nothing” measure with some degree of arbitrariness is not suitable as an objective function for minimization.
In principle, the steps of at deep minima of could be avoided by smoothing the function. This approach, however, would introduce a problem at the beginnings and ends of an curve: smoothing would require to access data outside the available range. Linear extrapolation (with weights quickly decreasing with increasing distance from the edge) followed by smoothing can sometimes help [23], but is not applicable in the current case.9 Furthermore, since the functions for curves with some offset are very similar to what one would obtain by smoothing the step-like obtained for zero offset (see Figure 2c), smoothing will make all the functions very similar, irrespective of the offset. Such an almost complete independence on the offset is undesirable, since the offset carries information (on how close the complex amplitudes come to the origin of the complex plane). This information should not be discarded (which would be the other extreme to 's overreaction on a small offset).
As an alternative to smoothing, a purely local approach would be preferable. One can avoid the division by (almost) zero at the minima by adding a term to the denominator. Since the function of a Pendry-style factor must be scale-invariant with respect to multiplying the intensity by a constant factor, the only nonzero quantity locally available at a minimum and suitable for this purpose is the second derivative of the intensity . One could therefore consider the following function:
The constant determines by how much the steps of the functions are smoothed out. In practice, it turns out that depends too weakly on the intensity offset, a problem that was already discussed above in the context of smoothing the function.
Before introducing an improved function, let us discuss the desirable behavior of a function near an intensity minimum. As mentioned above, it should avoid the step of (seen in Figure 2c). Since numerical evaluation is performed by sampling at discrete energy points, should be roughly linear between two evaluation points. Furthermore, the function should be sufficiently, but not excessively sensitive to the offset, i.e. to the intensity at the minimum. For a shallow minimum (large offset), the logarithmic derivative decreases in magnitude, and the slope of the function should also decrease. This desired behavior cannot be accomplished with alone, because becomes very large towards the minimum.
Our approach makes use of the fact that the offset of a parabolic minimum can be expressed as , and provides a dimensionless measure for the offset at a minimum (for how shallow a minimum is). We use this expression to obtain the desired behavior of the function:
Similar to in equation (8), determines how much the step of is smoothed out, or, in other words, how much the slope of the function at a minimum gets reduced. As described above, the first term in equation (9) is a scale-invariant measure of the offset; thus the smoothing (slope-limiting) action of equation (12) increases with increasing offset (shallower minima). Equation (11) limits this smoothing for very shallow minima because these yield only a weak modulation of the function without any additional measures (see the red curves in Figure 2; there, the logarithmic derivative is sufficiently small). The constant in equation (9) provides a minimum degree of smoothing (maximum slope) of the function at deep minima where the intensity approaches or reaches zero. The case distinction in equation (12) ensures that the correction is restricted to regions close to the minima; checking for the sign of (or, equivalently, ) provides a smooth transition to the unmodified function ( at the transition).
Equation (9) has two parameters: , which determines the influence of the offset of the minimum, and , which determines the behavior at a minimum reaching . The choice of and is motivated by the desired behavoir of an factor discussed above. With these values, as Figure 2d shows, (i) there are clear differences between the curves with zero offset (black) and larger offsets (blue to orange), and (ii) the steepest curve (black) at the intensity minimum varies sufficiently slowly to yield a smooth variation of the factor when sampled at discrete energy with 0.5 eV steps (violet tick marks in Figure 2). Since the minimum shown in Figure 2 is roughly the sharpest that can occur in real data and a low value of eV was assumed, the figure should be seen as a worst case; for higher values, the tick spacing would be shorter and for wider minima the function gets smoothed out. Nevertheless, we recommend using an energy step of 0.5 eV or less also for larger values of .
Apart from the behavior near the minima, the function is very similar to , thus one can use it in equation (5). We call the factor obtained this way . The subscripts “S” stands for “smooth”, which is justified since is not plagued by the noise of , see Figure 3 and Figure 4. Both and it’s derivative react smoothly to small changes of structure parameters.
Comparison of factors
Similarities of and
Except for the minima of the intensity, the functions of Pendry’s factor and the newly presented are very similar, as shown in Figure 5, which presents these two functions for a typical experimental curve. Therefore, it is not astonishing that the values of these two factors are also similar. As Figure 6 shows, on average, values are slightly lower than , because is not plagued by the cusps at deep minima, which exaggerate minuscule differences of the curves.
Figure 5. Experimental curve and its functions. The black curve at the bottom is of the (1,2) beam of -FeO(1̄102)-(1 × 1). The upper curves are the functions of Pendry’s factor (blue) and of the factor introduced in this work (orange). Blue arrows mark cusps of , where the function reaches an extremum and folds back (cf. Figure 1a). The inset around eV shows the functions affected by experimental noise due to insufficiently smoothed data (smoothing parameter 1.0 eV, far below the recommended range [10]).
Figure 6. Correlation between and . Each data point gives the factors between the calculated and experimental curve of a particular beam; also the overall factors for all beams are included (marked by black circles). The data are from three systems, -FeO(102)-(1 × 1) [9] (blue squares), Ir(100)-(3 × 1)-MnO [25] (pink diamonds), and Pt(111)-(10 × 10)-49Te [20] (brown crosses).
Since the new factor is based on the same construction principle as and has similar values, one can also determine the error bars for fit parameters the same way as for , using as defined in equation (6), by simply replacing with in this equation. This is justified since the derivation of equation (6) in [14] is based on peaks only, where the and functions are essentially identical (the regions with negative slope of the functions in Figure 5). Since the numerical and values are similar (Figure 6) and the curvatures of the factor at the parameter minima are also very similar, using the procedure based on equation (6) results in very similar error estimates, see Figure 3.
Sensitivity to experimental noise
In contrast to , the function depends on the second derivative , though only in the vicinity of sufficiently deep minima. Since most sources of noise in the experimental curves increase with increasing intensity (e.g., shot noise, modulation by the grids), this dependence of on is less severe than for , which is sensitive to also at the maxima. This means that the noise of the second derivative will have a smaller effect on than on . In practice, the functions of moderately noisy (weakly smoothed) curves do not show any noticeable increase of the noise at deep minima where comes into play.
Concerning noise, the main difference between and is not the use of the second derivative. As Figure 1a shows, the function reaches saturation earlier than . ( and are identical with the exception of the regions around deep minima.) When reaches saturation, i.e. when the derivative of vanishes, the noise of the underlying logarithmic derivative is suppressed. It turns out that values near , where reaches saturation, are very common. In Figure 5, these are the regions where reaches an extremum of ; this occurs next to all intensity minima except for the very shallow ones. In these regions, does not yet reach its saturation value , which means that the slope of is nonzero and any noise of the data (and the corresponding logarithmic derivative ) will affect to some degree. This can be seen in the inset of Figure 5, where the functions from noisy (insufficiently smoothed) experimental data are plotted. Nevertheless, in practice, the impact of this higher noise sensitivity of is marginal. First, it is partly compensated by regions where is more sensitive to noise, since the cusps at deep intensity minima are sensitive to small changes of the input data, as discussed in § 2.2. Second, the noise of in the near-saturation regions is less than in regions where the logarithmic derivative is small and the slope of is much higher. Since the initial slopes of and are identical, there is no difference between the two factors in these regions, which often dominate the noise. The comparison of the factors between experimental and calculated curves in Figure 6 also indicates that is not plagued by experimental noise: Since the functions of both factors are very similar, a higher impact of noise in typical experimental data on would lead to higher values compared with . In reality, the opposite is true: on average, is lower than . Only for very noisy data, obtained by far too weak smoothing,10 the situation is reversed, with higher than .
The “early” saturation of reached in many regions of the curves may suppress noise to a larger degree than the weak, but nonzero slope of [or equivalently ], at the same values. On the other hand, this vanishing derivative is a disadvantage of Pendry’s factor. It means that is insensitive to the slope of the curves in these regions. In other words, information present in the slopes of is discarded by Pendry’s factor. The function reaches saturation only at higher absolute values of the logarithmic derivative ; these values are reached only close to deep minima. There, the construction of the function given by equation (9)–(12) ensures that contains information on the depth of the minima, as discussed in § 3 and exemplified in Figure 2d. Figure 5 shows that for realistic data reaches its saturation value at very few points, which means that the information lost at these few points is negligible.
Which factor yields the most accurate result?
The new factor, , may be slightly more sensitive than to strong noise in the experimental data. The relevant question, however, is not the influence of noise on the factor but to which degree the outcome of the structure optimization is affected by noise and other imperfections of the experimental data. The factor should not steer the minimization away from the true value in the presence of noise and imperfections. Figure 3 and Figure 4 show that minimization of different factors leads to slightly different parameter values. Unfortunately, we cannot determine which of these results is closer to reality, since there is no method to determine the structure with better accuracy than a LEED study. We therefore take a different approach: We use data with more imperfections than our standard data set for -FeO( 102)-(1 1) and examine how well the resulting fit parameters agree with the result of our best data set. Minimizing a given factor with imperfect data should yield fit parameters that are as close as possible to the result obtained with the best data.
For this comparison, 30 structural parameters (all independent atom coordinates in the upper two layers) of -FeO( 102)-(1 1) have been optimized with different experimental data. We analyze the resulting deviations of the parameters between the optimization with the imperfect data and the best data set for a given factor. Since different parameters influence the factor to a different degree,11 the deviations have been normalized: Each deviation has been divided by the uncertainty of the respective parameter, derived from , cf. equation (6). Statistics of these 30 normalized parameter deviations, obtained by minimizing the three factors , , and have been plotted in Figure 7.12 The results obtained with different imperfect input data are discussed in the following paragraphs:
Figure 7. Deviations of the best-fit parameters for data sets with imperfect curves. The plot shows statistics of the normalized deviations of 30 fit parameters obtained using imperfect experimental data sets, with respect to the parameters derived from the best data set, when using a given factor for both. It therefore provides a measure for how sensitive structure optimization with a given factor is with respect to experimental imperfections. The horizontal axis is the absolute value of the deviation, normalized by the error bar for the respective parameter as derived from . The plot shows the median, mean, and maximum of these normalized, absolute deviations. See the text for a description of the data sets named at the left.
Experimental noise—The best data set (used as the reference) suppresses the impact of instrumental imperfections (modulation by the grids of the LEED optics, screen defects) and sensor noise by averaging over two measurements with the beams impinging on slightly different positions of the LEED screen. In addition, this reference data set was obtained using standard image processing techniques to correct for inhomogeneities of the LEED screen and sensor (flat field and dark frame corrections [10]). The imperfect experimental curves used for the top section of Figure 7 were created without these corrections and without averaging, which results in increased noise. Figure 7 shows that the parameters derived from these noisy curves show the least deviation from the reference when is used for the optimization, closely followed by . performs worst. The difference between and is not necessarily significant since the values obtained with might not exactly correspond to the global minimum, in spite of running all minimizations with more than 50 different starting values (random starting points as well as the best-fit parameters of the reference set) and selecting the parameters corresponding to the best factor obtained in these runs. (This problem is due to the noise of , see Figure 4a and § 4.4.)
Undersmoothed data—The experimental data were based on the same measurement as the best data set, but the smoothing applied was much weaker than recommended [10]. The plot summarizes the results obtained with two data sets, with smoothing parameters of 1.0 and 1.5 eV (the recommended value is eV [10]). Insufficient smoothing increases especially the high-frequency noise, which strongly affects the second derivative [26]. Since is the only factor under consideration that does not make use of , one should expect that it is the least sensitive to this type of noise. Nevertheless, the parameter values obtained with come closer to the reference value of the best data set than those from minimization. Again, due to the noisiness of , the difference between and is not necessarily significant. leads to the largest deviations from the reference.
Oversmoothed data—The best data set was strongly smoothed by an additional smoothing step (smoothing parameter 6 eV; this is more than recommended). This smoothing step eliminates fine detail, e.g., double peaks with a separation of turn into a single maximum. The sequence of the parameter deviations obtained with the different factors is the same as in the previously discussed cases, with best and worst.
Energy-dependent intensity scaling—This problem may occur in experiment due to improper (or missing) correction of the intensity data for the incident electron beam current , or in the case of an energy-dependent gain of a channelplate LEED system. To simulate it, the intensities of the best data set were multiplied or divided by . The statistics in Figure 7 summarize these two data sets. By design, neither nor should be sensitive to such a “slow” intensity variation, and the results confirm this.13 One might expect that should be very sensitive to an energy-dependent intensity factor, because it uses an energy-independent intensity scale factor between experiment and calculation, see equation (2). Nevertheless, steers the optimization to reasonable parameter values (though not as good as the other factors).
Intensity offset—A constant of was added to all intensities, where is the overall maximum of the set of curves. Note that this intensity offset corresponds to hardly more than one least-significant bit in legacy 8-bit image acquisition systems, if the brightest beams just reach the saturation limit. 14 Nevertheless, for a few low-intensity beams in our data set [9], the offset is higher than the average intensity of the curve. The effect of such an intensity offset on the parameters is comparable for and . With , the effect of an intensity offset is significantly worse.
Incomplete data—These curves contained only half of the beams in the full data set. (The statistics in the bottom section of Figure 7 are based on two such partial data sets, one half of the beams and the other half, and the statistics summarize the results for the two.) Optimization with incomplete data sets is meant as a test for how well the factor makes use of the information in the data. Ignoring important aspects of the data would lead to overfitting and therefore a poor agreement with the fit using the full data set. At least for the data set used here, the results of the three factors are comparable; performs slightly worse than the other factors. A closer examination of the individual parameter values shows that and yield very similar values, substantially deviating from those of (not shown). This is not unexpected since and share the same concept. It should be noted that, like all the test data used here, also this data set is a rather moderate deterioration of the best data: The redundancy factor [5, 6] for these incomplete data sets is , so one should not expect severe overfitting (total energy span eV, eV, 31 parameters including the offset of the energy axis or ).
Summarizing all results in Figure 7, the fit results are significantly more sensitive to imperfections of the data when using compared with the other factors. This may be related to a comparably low sensitivity of to the fit parameters [26], which is also seen as a lower ratio between the curvature and minimum factor in Figure 3a. The newly introduced wins the comparison with , but the individual results should be seen with care since the noisiness of also adds noise to the values obtained with .
Impact on computational performance
The choice of the factor has a major impact on the computational performance of the minimization procedure. The computing time is usually dominated by the calculation of the curves; the time for the -factor evaluation is comparably low and the increased complexity of vs. has no substantial impact on the overall performance. The number of trial structures required to find the minimum is much more important. The noise of makes gradient-based minimization less efficient since a noisy gradient can steer the minimization in the wrong direction. In the minimization runs performed for Figure 7, the average computing time required for optimization with as an objective function is 30–50% higher than when optimizing or . While this is not particularly bad, the main problem comes from the fact that -based minimization with efficient gradient-based methods will often fail to find the minimum of ; thus one needs many search runs. Due to many local minima in close proximity to each other (Figure 4a), finding the global minimum is even more difficult than just finding a “reasonable” minimum well within the error bars of the parameters. For the reference structure of Figure 7, when using , the best minimum was found in only one of 150 optimization runs. Although the average deviations of the parameters at the lowest minima are only in the order of one tenth of their error bars, some minimization runs (and also the deviations of some parameters in “good” runs) end much further out. This is an unsatisfying situation. In principle, one could select minimization methods that are more robust with respect to the -induced noise, but these require substantially more computing time [12]. Clearly, a better solution is using for the optimization. With , more than 98% of all minimization runs land closer to the minimum than the best one percent of all minimization runs using .
Conclusions
For the structural search by LEED , Pendry’s factor has many attractive properties, but also suffers from severe shortcomings. We have presented a new factor, named , that retains the advantages of , but avoids its shortcomings. The differences between and are limited to a different handling of the high (or even diverging) values of the logarithmic derivative at the minima of the curves. The numerical values of and are similar, and also the -based method of estimating the error bars of the fit parameters can be directly transferred to . We show that avoids the noisiness of , thus it is much more suitable as an objective function for minimization. Based on minimization runs with imperfect experimental data, we show that is at least as good as in steering the minimization towards the right structure, whereas the optimization results obtained using are clearly more sensitive to experimental imperfections.
Funding
This research was funded in part by the Austrian Science Fund (FWF) under project 10.55776/F81, Taming Complexity in Materials Modeling (TACO). The computational results presented have been achieved in part using the Austrian Scientific Computing (ASC) infrastructure. For the purpose of open access, the authors have applied a CC BY public copyright license to any Author Accepted Manuscript version arising from this submission. The authors acknowledge TU Wien Bibliothek for financial support through its Open Access Funding Programme.
Data availability
Supplementary file ImprovedReliabilityFactorData.zip contains the computer code to create Figure 1 and Figure 2 as well as the data used for the other figures.
References
- [1]Pendry J B 1974 Low Energy Electron Diffraction: The Theory and its Application to Determination of Surface Structure (London: Academic Press)
- [2]Van Hove M A, Weinberg W H and Chan C M 1986 Low-Energy Electron Diffraction: Experiment, Theory and Surface Structure Determination softcover reprint of the hardcover 1. ed (Springer Series in Surface Sciences no 6) (Berlin: Springer)DOI
- [3]Van Hove M A, Moritz W, Over H, Rous P J, Wander A, Barbieri A, Materer N, Starke U and Somorjai G A 1993 Automated determination of complex surface structures by LEED Surf. Sci. Rep. 19 191–229DOI
- [4]Heinz K and Hammer L 1998 Surface crystallography by low energy electron diffraction Z. Kristallogr. 213 615–634
- [5]Heinz K 2012 Electron based methods: Low-energy electron diffraction (LEED) Surface and Interface Science vol 1 ed Wandelt K (Weinheim: Wiley-VCH) pp 93–150
- [6]Fauster T, Hammer L, Heinz K and Schneider M A 2020 Surface Physics: Fundamentals and Methods (De Gruyter Oldenbourg)DOI
- [7]Moritz W and Van Hove M A 2022 Surface Structure Determination by LEED and X-rays (Cambridge: Cambridge University Press)DOI
- [8]Held G 2025 Structure determination by low-energy electron diffraction–A roadmap to the future Surf. Sci. 754 122696
- [9]Kraushofer F, Imre A M, Franceschi G, Kißlinger T, Rheinfank E, Schmid M, Diebold U, Hammer L and Riva M 2025 ViPerLEED package I: Calculation of I(V) curves and structural optimization Phys. Rev. Res. 7 013005DOI
- [10]Schmid M, Kraushofer F, Imre A M, Kißlinger T, Hammer L, Diebold U and Riva M 2025 ViPerLEED package II: Spot tracking, extraction, and processing of I(V) curves Phys. Rev. Res. 7 013006DOI
- [11]Dörr F, Schmid M, Kraushofer F, Hammer L, Diebold U and Riva M ViPErLEED package III: Data acquisition for quantitative low-energy electron diffraction in preparation
- [12]Imre A M, Haidegger P, Kraushofer F, Wanzenböck R, Hable T, Tobisch S, Kienzer M, Buchner F, Carrete J, Madsen G K H, Schmid M, Diebold U and Riva M 2025 Structural optimization in tensor LEED using a parameter tree and R-factor gradients submittedarxiv.org/abs/2512.09737
- [13]Zanazzi E and Jona F 1977 A reliability factor for surface structure determinations by low-energy electron diffraction Surf. Sci. 62 61–80DOI
- [14]Pendry J B 1980 Reliability factors for LEED calculations J. Phys. C: Solid State Phys. 13 937DOI
- [15]Sporn M, Platzgummer E, Forsthuber S, Schmid M, Hofer W and Varga P 1998 The accuracy of quantitative LEED in determining chemical composition profiles of substitutionally disordered alloys: a case study Surf. Sci. 416 423–429DOI
- [16]Blum V, Hammer L, Meier W and Heinz K 2001 Quantification of substitutional disorder and atomic vibrations by LEED – the role of parameter correlations Surf. Sci. 488 219–232DOI
- [17]Roučka R, Jiruše J and Šikola T 2002 Spot intensity processing in LEED images Vacuum 65 121–126 ISSN 0042-207XDOI
- [18]Mayer A, Salopaasi H, Pussi K and Diehl R D 2012 A novel method for the extraction of intensity–energy spectra from low-energy electron diffraction patterns Computer Physics Communications 183 1443–1447
- [19]Rous P J 1992 The tensor LEED approximation and surface crystallography by low-energy electron diffraction Prog. Surf. Sci. 39 3–63DOI
- [20]Kißlinger T, Schewski A, Raabgrund A, Loh H, Hammer L and Schneider M A 2023 Surface telluride phases on Pt(111): Reconstructive formation of unusual adsorption sites and well-ordered domain walls Phys. Rev. B 108 205412
- [21]Chen S, Montgomery J and Bolufé-Röhler A 2015 Measuring the curse of dimensionality and its effects on particle swarm optimization and differential evolution Appl. Intell. 42 514–526DOI
- [22]Nocedal J and Wright S J 2006 Numerical Optimization second edition ed Springer series in operations research and financial engineering (New York, NY: Springer)
- [23]Schmid M, Rath D and Diebold U 2022 Why and how Savitzky–Golay filters should be replaced ACS Meas. Sci. Au 2 185–196DOI
- [24]Dahlquist G and Björck Å 2008 Numerical Methods in Scientific Computing, Volume I (Philadelphia: Society for Industrial and Applied Mathematics)DOI
- [25]Ferstl P, Hammer L, Sobel C, Gubo M, Heinz K, Schneider M A, Mittendorfer F and Redinger J 2016 Self-organized growth, structure, and magnetism of monatomic transition-metal oxide chains Phys. Rev. Lett. 117 046101DOI
- [26]Tear S P, Roll K and Prutton M 1981 A comparison of reliability (R) factors in a LEED structural analysis of the copper (111) surface J. Phys. C: Solid State Phys. 14 3297DOI






