An improved reliability factor for quantitative low-energy electron diffraction

This paper introduces a modified reliability factor, $R_\mathrm{S}$, to replace Pendry's $R_\mathrm{P}$ in quantitative low-energy electron diffraction, addressing its sensitivity to noise and intensity offsets while demonstrating superior or comparable performance in optimizing surface structure determination.

The Gist

Imagine you are trying to solve a 3D puzzle of a crystal surface, but instead of physical pieces, you are using invisible beams of electrons. This technique is called Low-Energy Electron Diffraction (LEED).

To solve the puzzle, scientists compare two things:

1. The Real Data: The pattern of electrons bouncing off the actual surface (the "Experimental" curve).
2. The Guess: The pattern calculated by a computer based on a model of where the atoms are (the "Theoretical" curve).

The goal is to wiggle the atoms in the computer model until the "Guess" curve matches the "Real" curve as perfectly as possible. To know how good the match is, scientists use a score called an R factor. The lower the score, the better the match.

For decades, the gold standard for this score was a method called Pendry's R factor ($R_P$). It was great, but the authors of this paper (Imre et al.) found that it had some serious "glitches" that made it hard to find the perfect solution. They have created a new, improved score called $R_S$ (the "Smooth" R factor) to fix these problems.

Here is a simple breakdown of the problems they found and how they fixed them, using everyday analogies.

The Problem: Why the Old Score ($R_P$) Was Flawed

The authors identified three main ways the old scoring system could trick scientists:

1. The "Fake Twin" Problem (Dissimilar curves can have a perfect score)

• The Analogy: Imagine you are judging two singers. The old score only listened to the changes in their pitch (going up or down), not the actual notes they hit.
• The Glitch: It was possible for two singers to hit completely different notes (qualitatively different curves) but change their pitch in the exact same way. The old score would say, "Perfect match!" (Score = 0), even though the singers were singing different songs.
• The Risk: This could trick the computer into thinking a wrong atomic structure was the right one, leading to a "false positive."

2. The "Hairline Fracture" Problem (Too sensitive to tiny errors)

• The Analogy: Imagine trying to measure the depth of a pothole in a road. If the pothole is exactly 0 inches deep (perfectly flat), the measurement is easy. But if there is a tiny speck of dust in the bottom (a tiny offset), the old score goes crazy.
• The Glitch: In real experiments, data is never perfect; there is always a tiny bit of "noise" or background static. If the electron intensity hits zero (a deep minimum), the old score becomes extremely sensitive to even the tiniest bit of noise. A tiny speck of dust makes the score jump wildly, making the graph look jagged and "noisy."
• The Risk: This makes it very hard for computers to find the true bottom of the valley (the best answer) because the path is covered in fake bumps.

3. The "Jagged Mountain" Problem (Noisy optimization)

• The Analogy: Imagine you are hiking down a mountain to find a campsite (the best structure). The old score made the mountain look like a jagged, rocky cliff face full of tiny, sharp spikes.
• The Glitch: Because of the sensitivity to noise mentioned above, the "score landscape" was full of tiny, fake valleys and spikes.
• The Risk: When a computer tries to "hike" down to the best answer, it gets stuck in these tiny fake valleys or gets confused by the jagged terrain. It takes much longer to find the real campsite, and often, it gets lost.

The Solution: The New Score ($R_S$)

The authors invented a new way to calculate the score, called $R_S$. Think of it as upgrading the hiking map.

• How it works: Instead of getting confused by the "fake twins" or the "speck of dust," the new formula smooths out the terrain. It looks at the data in a way that ignores the mathematical tricks that caused the old score to fail.
• The Result:
  • No Fake Twins: If two curves are different, the new score correctly says they are different.
  • No Jagged Spikes: The "mountain" is now a smooth slope. The computer can easily slide down to the true bottom without getting stuck on tiny bumps.
  • Better Navigation: Even when the experimental data is a bit messy (noisy), the new score guides the computer to the correct answer much more reliably than the old score.

The Verdict

The paper tested this new score against the old one ($R_P$) and another common score ($R_{ZJ}$) using real data from iron oxide crystals.

• $R_{ZJ}$ (The old alternative): Was very sensitive to noise and gave the worst results when the data wasn't perfect.
• $R_P$ (The old gold standard): Worked okay, but often got stuck in "fake" solutions because of the jagged, noisy landscape.
• $R_S$ (The new champion): Performed just as well as the old gold standard when the data was perfect, but significantly better when the data had imperfections. It found the correct structure faster and more reliably.

In short: The authors didn't throw away the old system; they just polished it. They took the best parts of the famous Pendry score and fixed the parts that made it "jumpy" and unreliable, creating a smoother, more trustworthy tool for mapping the atomic world.

Technical Summary

Technical Summary: An Improved Reliability Factor for Quantitative Low-Energy Electron Diffraction

Problem Statement
Quantitative Low-Energy Electron Diffraction (LEED) is a standard technique for determining surface structures by minimizing the discrepancy between experimental and calculated diffraction intensities ($I(E)$ curves) as a function of electron energy. The most widely used metric for this agreement is Pendry's $R$ factor ($R_P$). While $R_P$ offers significant advantages over simpler metrics (such as insensitivity to global intensity scaling and high sensitivity to peak positions), the authors identify three critical deficiencies that hinder its effectiveness as an objective function for structural optimization:

1. Non-invertibility and False Minima: The $Y_P$ function used in $R_P$ is non-invertible; distinct $I(E)$ curves can map to the same $Y_P$ curve. Consequently, two qualitatively different curves can yield $R_P = 0$, potentially misleading optimization algorithms toward incorrect structural models.
2. Sensitivity to Intensity Offsets: $R_P$ is overly sensitive to small offsets in intensity at deep minima. If an experimental minimum does not reach exactly zero (due to background noise or subtraction errors), the $Y_P$ function develops sharp "cusps." This makes the metric unstable and highly dependent on the precise sampling of these minima.
3. Noise and Optimization Failure: The combination of the non-invertible nature of $Y_P$ and its sensitivity to offsets creates a "noisy" objective function. The gradient of $R_P$ with respect to structural parameters is often erratic, causing gradient-based minimization algorithms to struggle in finding the global minimum, particularly in high-dimensional parameter spaces.

Methodology
The authors propose a modified reliability factor, $R_S$, designed to retain the advantages of $R_P$ while eliminating its specific weaknesses. The core of the methodology involves replacing the $Y_P$ function with a new function, $Y_S$.

• Derivation of $Y_S$: The authors analyze the behavior of the logarithmic derivative $L = d(\ln I)/dE$ near intensity minima. They introduce a smoothing mechanism that depends on the second derivative of the intensity ($I''$) and the depth of the minimum.
• The $Y_S$ Function: Unlike $Y_P$, which folds back to zero after reaching extrema (causing non-invertibility), $Y_S$ is constructed to be monotonic in the relevant regions. It incorporates a term derived from the offset of a parabolic minimum ($I_{min} \approx I - (I')^2/2I''$) to create a dimensionless measure of the minimum's depth.
• Smoothing Logic: The function applies a "slope-limiting" action that increases with the offset of the minimum. This ensures that for deep minima (where $I \to 0$), the function varies smoothly rather than forming sharp cusps, while still maintaining sensitivity to the offset information.
• Implementation: The new $R_S$ factor is calculated using the same integral formulation as $R_P$ (Equation 5), simply substituting $Y_{exp}$ and $Y_{th}$ with their $Y_S$ counterparts.

Key Results
The paper validates $R_S$ through theoretical analysis and comparative testing on experimental data from $\alpha\text{-Fe}_2\text{O}_3(1\bar{1}02)$, Ir(100), and Pt(111) systems.

• Smoothness and Noise Reduction: $R_S$ exhibits a smooth, parabolic-like minimum with respect to structural parameters, whereas $R_P$ shows significant roughness. The gradient of $R_S$ is stable, allowing gradient-based minimization algorithms to converge efficiently. In contrast, $R_P$ often traps algorithms in local minima or fails to find the global minimum due to noisy gradients.
• Robustness to Imperfections: When tested against imperfect experimental data (including added noise, undersmoothing, oversmoothing, energy-dependent scaling, and intensity offsets), $R_S$ consistently steered the optimization toward results closer to the "best" reference data set than $R_P$ did.
  • $R_P$ performed reasonably well but was occasionally misled by its own noise.
  • The Zanazzi-Jona factor ($R_{ZJ}$) performed significantly worse, showing high sensitivity to experimental imperfections, particularly undersmoothing and intensity offsets.
• Computational Efficiency: Due to the smoothness of $R_S$, optimization runs are more efficient. The authors report that when using $R_P$, only 1 in 150 optimization runs found the best minimum, whereas with $R_S$, over 98% of runs landed closer to the minimum than the best 1% of $R_P$ runs.
• Error Estimation: The numerical values of $R_S$ are similar to $R_P$ (typically slightly lower). Crucially, the method for estimating error bars of fit parameters (based on the variance of the $R$ factor) derived for $R_P$ remains valid for $R_S$.

Significance and Claims
The authors claim that $R_S$ serves as a direct, drop-in replacement for $R_P$ in quantitative LEED analysis. Its primary significance lies in transforming the reliability factor from a "noisy target function" into a robust objective function suitable for modern, high-dimensional gradient-based optimization.

The paper asserts that $R_S$ avoids the severe shortcomings of $R_P$ (non-invertibility and offset sensitivity) without sacrificing its advantages (insensitivity to slow intensity scaling and sensitivity to peak positions). Furthermore, $R_S$ outperforms the alternative $R_{ZJ}$ metric in guiding optimization toward the correct structure in the presence of real-world experimental imperfections. The authors conclude that the use of $R_S$ should lead to more accurate and reliable structural determinations in surface crystallography.