🔢 mathematics

Adjoint-Based Gradient Evaluation for Metasurface Inverse Design via Affine Geometric Transformations

This paper presents a systematic methodology for the inverse design of large-scale metasurfaces that leverages the adjoint variable method combined with affine geometric transformations to achieve accurate electromagnetic field control with minimal computational cost.

Original authors: Vincenzo Mottola, Luisa Faella, Carlo Forestiere, Antonello Tamburrino

Published 2026-03-17

📖 4 min read🧠 Deep dive

CC0 1.0

Original authors: Vincenzo Mottola, Luisa Faella, Carlo Forestiere, Antonello Tamburrino

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an architect trying to design a giant, invisible wall made of millions of tiny, magical tiles. This wall, called a metasurface, has a superpower: it can bend light in ways that normal glass lenses cannot. You want to arrange these tiles so that a beam of light hitting the wall gets focused into a perfect, tiny dot on the other side, or perhaps split into three different dots at once.

The problem? There are so many tiles (millions of them), and each one can be rotated, stretched, squished, or moved slightly. Trying to figure out the perfect arrangement by guessing and checking is like trying to find a specific grain of sand on a beach by picking up one grain at a time. It would take forever, and your computer would crash.

This paper introduces a super-smart shortcut to solve this problem. Here is how it works, broken down into simple concepts:

1. The Old Way: The "Blind Taste Test"

In the past, to design these surfaces, engineers would use a method called "freeform optimization." Imagine trying to sculpt a statue out of clay by pinching and pulling every single molecule. It gives you total freedom, but it's incredibly slow and computationally expensive. You have to simulate the physics of the whole wall for every tiny change you make.

2. The New Way: The "Affine Dance"

The authors of this paper say, "Let's stop trying to reshape every molecule." Instead, they treat each tile (or "meta-atom") as a rigid object that can only do three specific things:

Rotate (spin like a coin).
Stretch/Squish (get wider or taller, like a rubber band).
Slide (move left, right, up, or down).

They call this an Affine Transformation. Think of it like a dance routine where every dancer (tile) is allowed to spin, stretch their arms, and step to the side, but they don't melt or change their fundamental shape. This drastically reduces the number of choices the computer has to make, making the problem much smaller and faster to solve.

3. The Secret Weapon: The "Ghost Mirror" (Adjoint Method)

Even with fewer choices, calculating the perfect arrangement is still hard. Usually, to know if you are moving in the right direction, you have to test every single tile individually. If you have 1,000 tiles, you have to run 1,000 simulations just to know which way to turn the knobs.

This paper uses a mathematical trick called the Adjoint Variable Method.

The Analogy: Imagine you are in a dark room trying to find the exit. The normal way is to walk forward, hit a wall, turn around, walk back, try a different door, hit a wall, and repeat.
The Adjoint Trick: Instead, imagine you shine a flashlight backwards from the exit (the target) into the room. This "ghost light" (the adjoint field) travels through the room and hits the walls. By looking at how this backward light interacts with the walls, you instantly know exactly how to move every single tile to get closer to the exit.

With this method, you don't need 1,000 simulations. You only need two:

One to see how the light hits the wall normally.
One "ghost" simulation to see how the light would come from the target.

From these two, the computer instantly calculates the perfect move for every single tile simultaneously.

4. Why This Matters

The authors proved that this method is not just a theory; it works in real life (or at least in very accurate computer simulations).

Speed: They showed that their "ghost mirror" math gives the exact same answer as the slow, brute-force method, but in a fraction of the time.
Complexity: They used this to design a lens that focuses light into one spot, and then even more impressively, a lens that focuses light into two or three spots at the same time.

The Big Picture

Think of this paper as giving engineers a GPS for light.
Before, designing these advanced lenses was like driving cross-country with no map, guessing every turn. Now, thanks to this new method, they have a GPS that tells them exactly which way to turn every single "tile" on the road to get to their destination instantly.

This means we can soon build better cameras, faster internet (using light instead of wires), and advanced medical imaging devices that are smaller, cheaper, and more powerful than anything we have today.

1. Problem Statement

The design of large-scale metasurfaces (arrays of sub-wavelength scatterers) for complex electromagnetic functionalities (e.g., beam steering, focusing, imaging) presents a significant computational challenge.

The Inverse Design Challenge: Optimizing the geometry of thousands of meta-atoms to achieve a target field distribution is an inverse problem with a massive number of unknowns.
Limitations of Existing Methods:
- Unit-Cell Approach: Assumes slowly varying structures and ignores near-field coupling, limiting performance for high numerical apertures.
- Freeform Optimization: Updates the dielectric permittivity at every point. While powerful, it requires complex binarization steps to make designs manufacturable and scales poorly due to the sheer number of variables.
- Finite Difference Methods: Calculating gradients via finite differences requires solving a separate "direct" electromagnetic problem for every design parameter. For large metasurfaces, this is computationally prohibitive.
Goal: Develop a methodology that enables fast, accurate gradient computation for inverse design, allowing for complex geometric transformations (rotation, scaling, translation) of meta-atoms without the prohibitive cost of freeform optimization.

2. Methodology

The authors propose a hybrid approach combining parametric geometric modeling with the Adjoint Variable Method (AVM).

A. Parametric Model (Affine Transformations)

Instead of optimizing the shape of every point on a meta-atom (freeform), the authors define the optimization variables as a set of geometric parameters controlling an affine transformation applied to a base meta-atom shape.
For each meta-atom $i$ , the transformation is defined by a vector $p_i = [\theta_i, \lambda_{x,i}, \lambda_{y,i}, x_i, y_i]$ :

$\theta_i$ : Rotation angle.
$\lambda_{x,i}, \lambda_{y,i}$ : Anisotropic scaling factors (expansion/contraction).
$x_i, y_i$ : Rigid translation (position).

This reduces the dimensionality of the problem significantly compared to freeform optimization while preserving the ability to explore a wide design space.

B. Adjoint Variable Method (AVM)

To compute the gradient of the cost function $I$ with respect to all design parameters efficiently:

Forward Problem: Solve for the total field $E_f$ produced by the metasurface under the incident excitation.
Adjoint Problem: Solve for an auxiliary field $E_a$ (the adjoint field) generated by a specific "adjoint source" $J_a$ defined on the target plane $S$ .
Gradient Calculation: The gradient is computed via a surface integral over the boundary of the meta-atoms, decoupling the cost function from the geometric transformation.

The core formula derived is:
$\frac{\partial I}{\partial p_j^i} = 2 \text{Re} \left[ \beta \oint_{\partial \Omega_i} E_a(x) \cdot E_f(x) w(x) \, dl \right]$
Where:

$E_a$ and $E_f$ are the adjoint and forward fields evaluated on the current geometry.
$w(x)$ is a geometric weighting function derived analytically based on the specific transformation (rotation, scaling, or translation).
$\beta$ is a scalar coefficient determined by the specific cost function.

Key Insight: The adjoint source $J_a$ depends only on the cost function, while the weighting function $w$ depends only on the geometric transformation. This allows the same adjoint solution to be reused for different geometric parameters, requiring only one additional direct simulation (the adjoint problem) regardless of the number of design parameters.

C. Analytical Derivations

The paper rigorously derives the weighting functions $w(x)$ for:

Rotation: $w = \mathbf{r} \cdot \hat{t}$ (involving the cross product of position and tangent).
Anisotropic Expansion: $w = (\mathbf{r} \cdot \hat{d})(\hat{d} \cdot \hat{n})$ .
Translation: $w = \hat{d} \cdot \hat{n}$ .
It also derives the adjoint sources $J_a$ for various cost functions, including field magnitude matching, norm differences, and focal point maximization.

3. Key Contributions

General Affine Framework: Extends previous parametric methods (which were limited to isotropic scaling) to include general affine transformations (rotation, anisotropic scaling, and translation) within the adjoint framework.
Decoupled Gradient Computation: Demonstrates that the gradient computation can be separated into a geometry-dependent part (weighting function) and a cost-function-dependent part (adjoint source). This enables efficient reuse of the adjoint field for different objective functions.
Manufacturability by Design: By optimizing geometric parameters of a fixed base shape, the resulting designs are inherently manufacturable, avoiding the need for post-processing binarization required in freeform topology optimization.
Rigorous Mathematical Formulation: Provides a complete derivation using Lorentz reciprocity and boundary element methods, explicitly handling the discontinuities of fields at dielectric interfaces.

4. Results and Validation

The authors validated the method using numerical simulations in a TEz propagation framework with TiO $_2$ meta-atoms at a wavelength of 660 nm.

Gradient Accuracy:
- The gradients computed via the proposed adjoint method were compared against finite-difference approximations.
- Result: The two methods showed excellent agreement for rotation, width deformation, and translation, confirming the mathematical correctness of the derived weighting functions.
Inverse Design Examples:
- Single Point Focusing: Optimized a metasurface of 128 meta-atoms to focus light at a specific point ( $37\lambda_0$ ). The resulting design achieved a focal spot with a Full Width at Half Maximum (FWHM) of $0.84\lambda_0$ .
- Multi-Point Focusing: Successfully designed metasurfaces to focus light simultaneously at two and three distinct points.
- Observation: The optimized structures exhibited complex, non-symmetric patterns with rapidly varying rotation angles, demonstrating the method's ability to handle high-complexity designs that unit-cell approaches cannot achieve.

5. Significance

This work addresses the scalability bottleneck in metasurface inverse design.

Computational Efficiency: It reduces the computational cost of gradient evaluation from $O(N)$ (where $N$ is the number of parameters, as in finite differences) to $O(1)$ (requiring only one adjoint solve), making the optimization of large-scale metasurfaces feasible.
Design Flexibility: It bridges the gap between simple unit-cell designs and complex freeform optimization, offering a "sweet spot" where designs are complex enough for high-performance applications (like high-NA metalenses) but remain computationally tractable and manufacturable.
Broad Applicability: The derived framework is general and can be extended to 3D problems and other geometric transformations, providing a robust tool for the next generation of photonic device design.