Accuracy assessment of scalar wave propagation methods for diffractive optics design: from thin elements to thick binary grating

This paper systematically evaluates the accuracy of thin-element approximation, beam propagation, and wave propagation methods against a rigorous reference for binary diffractive gratings, generating accuracy maps to guide the selection of appropriate forward models in inverse design pipelines based on spatial frequency and grating thickness.

The Gist

Imagine you are trying to design a complex, bumpy surface (like a microscopic maze) that bends light in very specific ways to create a hologram or a special lens. To do this, you need a computer program to predict exactly how light will travel through that maze.

This paper is like a "driver's test" for three different computer programs that try to predict that light's journey. The authors wanted to know: Which program gives the most accurate map, and under what conditions does each one fail?

Here is the breakdown of their findings using simple analogies:

The Three "Navigators" (The Methods)

The researchers tested three different ways to simulate light moving through these microscopic structures:

1. The "Instant Teleport" (TEA - Thin-Element Approximation):

   • How it works: This method pretends the bumpy structure is so thin it doesn't exist. It just calculates the light's path as if the light instantly "teleported" through the surface, changing direction based on the shape, but ignoring the time it takes to travel through the material.
   • The Analogy: It's like trying to predict how a car drives through a tunnel by only looking at the entrance and exit signs, ignoring the tunnel itself.
   • The Result: It's super fast and easy, but it only works if the tunnel is very short. If the tunnel gets longer (thicker), the prediction becomes wildly wrong because it forgets about the twists and turns inside.

2. The "Straight-Line Walker" (BPM - Beam Propagation Method):

   • How it works: This method breaks the tunnel into many thin slices and calculates the light step-by-step. However, it assumes the light mostly travels straight forward, only making small, gentle turns.
   • The Analogy: Imagine walking through a forest. This method assumes you are walking in a straight line and only occasionally stepping slightly left or right. If the path requires you to make a sharp 90-degree turn, this walker gets lost because they aren't programmed to handle big angles.
   • The Result: It's better than the "Teleport" method for thicker tunnels, but if the light needs to make sharp turns (large angles) or the tunnel is very long, the small errors in its "straight-line" assumption add up, and the map gets blurry.

3. The "True Navigator" (WPM - Wave Propagation Method):

   • How it works: This is the most sophisticated of the three. Like the second method, it steps through the tunnel slice by slice, but it uses a more complex math formula that allows for any angle of turn, not just small ones.
   • The Analogy: This walker knows the exact rules of physics. They can walk straight, turn sharply, or even zigzag perfectly. They don't assume the path is simple; they calculate the exact curve of every step.
   • The Result: It is the most accurate, especially for long tunnels or paths with sharp turns. It stays true to the "real" path much longer than the other two.

The "Gold Standard" (The Reference)

To know who won the race, the researchers used a super-accurate, heavy-duty method called FMM (Fourier Modal Method).

• The Analogy: Think of FMM as a high-speed drone flying over the forest, taking millions of photos to create a perfect, 3D map of exactly where every leaf and branch is. It takes a lot of computing power and time, so you wouldn't use it for every single guess, but it is the "truth" against which the other three are measured.

The Experiment: Random Mazes

The researchers didn't just test one maze. They generated 1,210 random microscopic mazes with two changing features:

1. Thickness: How deep the tunnel is (from 1 layer to 11 layers thick).
2. Complexity: How bumpy and sharp the turns are (from gentle hills to jagged, sharp peaks).

They ran all three "navigators" on these mazes and compared their maps to the "Gold Standard" drone map.

The Verdict: When to Use Which?

The paper produced "Accuracy Maps" (like weather maps showing where it's safe to drive) that tell you which method to pick:

• Use the "Instant Teleport" (TEA) only if: The structure is extremely thin (less than the width of a single light wave). If it gets any thicker, stop using it; it will give you a bad design.
• Use the "Straight-Line Walker" (BPM) if: The structure is thin, OR if the structure is thick but the light only needs to make very gentle turns. It's a good middle-ground tool.
• Use the "True Navigator" (WPM) if: You are designing thick structures that require moderate sharp turns. This is the sweet spot where the other two methods start to fail, but WPM still gets it right.

The Catch

The researchers tested these methods on "Binary Gratings," which are like mazes with very sharp, jagged walls (like a staircase). They noted that this is a "hard mode" test. If you are designing smoother, gentler structures (like a rolling hill), all three methods would likely perform even better than the results shown here.

In short: If you want to design complex, thick optical devices, don't rely on the simple "teleport" method. If the structure is thick and the light needs to turn, the "True Navigator" (WPM) is the only one that won't get you lost.

Technical Summary

Technical Summary: Accuracy Assessment of Scalar Wave Propagation Methods for Diffractive Optics Design

Problem Statement
The design of diffractive optical elements (DOEs) increasingly relies on gradient-based inverse design, where the accuracy of the forward propagation model fundamentally limits the performance of the resulting design. While rigorous methods like the Fourier Modal Method (FMM) and Finite-Difference Time-Domain (FDTD) provide high accuracy, they are computationally prohibitive for iterative optimization loops. Conversely, the Thin-Element Approximation (TEA) is computationally trivial but fails for thick structures or large diffraction angles due to its neglect of internal propagation effects and interface coupling. Intermediate scalar methods, specifically the Beam Propagation Method (BPM) and the Wave Propagation Method (WPM), offer a middle ground, yet a systematic, quantitative assessment of their accuracy regimes as a function of grating parameters has been lacking in the literature.

Methodology
The authors conducted a systematic accuracy assessment using the rigorous FMM (implemented via FMMAX) as a reference standard. The study utilized the following experimental design:

• Test Structures: Ensembles of random binary gratings were generated on a $7.5 \, \mu\text{m}$ periodic cell. The gratings were parameterized by spatial frequency cutoff ($f_c$) and physical thickness ($h$).
  • $f_c$ was sampled from 0 to $1.33 \, \mu\text{m}^{-1}$, corresponding to maximum diffraction angles ($\theta_{\text{max}}$) up to $\approx 45^\circ$.
  • Thickness ($h$) was sampled from $1\lambda$ to $11\lambda$ (where $\lambda = 532 \, \text{nm}$).
  • A total of 1,210 gratings (11 $\times$ 11 parameter pairs $\times$ 10 random realizations) were evaluated.
• Propagation Models: Three scalar methods were implemented in the GPU-accelerated, differentiable framework FluxOptics.jl:
  • TEA: Modeled as an infinitely thin phase mask, with propagation steps approximated in an effective medium.
  • BPM: Used a paraxial approximation with the grating volume sliced into steps of $\Delta z = 25 \, \text{nm}$.
  • WPM: Employed an exact non-paraxial propagator ($k_z = \sqrt{(nk_0)^2 - k_\perp^2}$) with a fixed number of 30 slices regardless of thickness, utilizing smooth transition layers for index discontinuities.
• Metric: Accuracy was quantified using the normalized field overlap ($\eta$) between the scalar method's transmitted field and the FMM reference field, averaged over the 10 realizations for each parameter set. Cross-polarized components were excluded to remain consistent with the scalar approximation.

Key Results
The study generated accuracy maps in the $(f_c, h)$ parameter space, revealing distinct domains of validity for each method:

• TEA: Accuracy degrades rapidly with increasing thickness. Overlap drops below 90% at $h \approx 3\lambda$ for moderate angles and falls below 50% for $h > 6\lambda$, confirming its limitation to very thin structures.
• BPM: Performs significantly better than TEA but suffers from accumulated paraxial phase errors. While it extends the valid regime to moderate thicknesses, its accuracy stabilizes at an overlap of 0.4–0.5 for large thicknesses within the tested range.
• WPM: Maintains the highest accuracy across the entire parameter space. For small angles ($\theta_{\text{max}} \lesssim 12^\circ$), WPM maintains $\eta > 0.90$ even for thicknesses up to $11\lambda$. The advantage of WPM over BPM is most pronounced for thick structures ($h > 3\lambda$) at moderate angles ($\theta_{\text{max}} \lesssim 30^\circ$), where the exact dispersion relation prevents the progressive dephasing that limits BPM.

Significance and Claims
The paper claims to provide the first systematic quantitative assessment of these scalar methods specifically for binary diffractive gratings, a conservative benchmark due to their sharp index discontinuities. The primary contribution is the generation of accuracy maps that serve as practical guidelines for selecting forward models in diffractive optics inverse design pipelines:

1. TEA is adequate only for thin structures ($h \lesssim \lambda$).
2. BPM is a competitive alternative for thin structures at large angles or moderate thicknesses at low angles.
3. WPM is identified as the method of choice for the inverse design of thick diffractive structures at moderate angles, offering a significant accuracy advantage over BPM without the prohibitive cost of rigorous methods.

The authors note that because binary gratings represent a "conservative benchmark" (maximizing spectral content at index discontinuities), the reported accuracy maps likely constitute lower bounds for the performance of these methods on smooth, continuous-relief structures. The study acknowledges that all three methods share limitations regarding unidirectional propagation and the neglect of vectorial effects, leaving extensions to split-step non-paraxial methods for future work.