Eigenvalue-accelerated LDOS optimization of high-Q optical resonances

This paper presents an eigenvalue-accelerated optimization method that utilizes a fast shift-invert eigensolver to eliminate ill-conditioning and achieve orders-of-magnitude speedup in designing high-$Q$ optical resonant cavities for maximizing the local density of states.

The Gist

The Big Picture: Tuning a Radio in a Storm

Imagine you are trying to tune an old-fashioned radio to a specific station. You want the signal to be crystal clear and loud (this is what scientists call High-Q or high quality).

In the world of light and lasers, scientists want to build tiny "cavities" (like a room for light) that trap light so perfectly that it bounces around millions of times before escaping. This is incredibly useful for things like super-fast internet, quantum computers, or ultra-sensitive sensors.

The Problem:
When scientists try to design these perfect light rooms using computers, they run into a massive headache.

• The "Knife Edge" Problem: Imagine trying to balance a ball on the very tip of a razor blade. If you move the ball even a tiny bit to the left or right, it falls off instantly.
• In these light cavities, if the computer changes the shape of the room by a microscopic amount, the "station" (the light frequency) shifts slightly. If it shifts even a tiny bit away from the target, the signal drops to zero.
• Because the signal drops so fast, the computer gets confused. It thinks, "I'm not sure which way to go!" and it takes thousands of years (or thousands of computer hours) to find the perfect shape. It's like trying to find the top of a mountain that is so sharp it looks like a needle.

The Solution: The "Smart Navigator"

The authors of this paper (George Shaker, Steven Johnson, and colleagues) invented a new method to solve this. They call it "Eigenvalue-accelerated LDOS optimization." That's a mouthful, so let's break it down with a better analogy.

The Old Way: The Blind Hiker

Previously, the computer was like a blind hiker trying to find the top of that razor-sharp mountain. It would take a step, check the height, take another step, check again. Because the mountain is so sharp, the hiker would stumble around for a long time, making tiny, inefficient steps.

The New Way: The GPS with a Live Map

The new method is like giving that hiker a GPS that updates in real-time.

1. The Rough Draft: First, the computer does a little bit of the old, slow work. It finds a "good enough" spot where a resonance (a strong light signal) exists. It's not perfect yet, but it's close.
2. The "Shift": Once it finds that spot, the computer asks: "Hey, where exactly is the peak of this mountain right now?" It calculates the exact frequency of the light resonance.
3. The Adjustment: Instead of trying to tune the radio to the original frequency we asked for, the computer says, "Okay, the mountain moved slightly. Let's tune the radio to the new exact location of the peak."
4. The Result: By constantly adjusting the target to match the mountain's current location, the "knife edge" problem disappears. The optimization becomes smooth and fast. The computer can take giant, confident strides up the mountain instead of tiny, stumbling ones.

The Magic Analogy: The Tightrope Walker

Think of the optimization process as a tightrope walker trying to cross a canyon.

• The Old Method: The walker is blindfolded. Every time they take a step, the wind blows the rope slightly. If they step wrong, they fall. They have to move incredibly slowly to stay safe.
• The New Method: The walker puts on a pair of smart glasses. As soon as the wind moves the rope, the glasses instantly tell the walker, "The rope is now 2 inches to the left. Step 2 inches left." The walker can now sprint across the canyon because they are always perfectly balanced.

What Did They Achieve?

Using this "Smart Navigator" method, the team designed light cavities that are orders of magnitude better than before.

• Speed: They found the best designs 1,000 times faster (or more) than the old methods.
• Quality: They built 1D cavities with a "Quality Factor" ($Q$) of over 100 million ($10^8$) and 2D cavities over 1 million ($10^6$).
  • To put that in perspective: If a photon (a particle of light) entered one of these cavities, it would bounce around inside for a long time. In a normal mirror, light bounces a few times. In these new designs, the light bounces millions of times before escaping.

The "Successive Enlargement" Trick

They also used a clever trick called "Successive Enlargement."
Imagine you are trying to build a giant castle.

• The Old Way: You try to build the whole castle at once. It's overwhelming, and you might get the foundation wrong.
• The New Way: You start by building a tiny, perfect tower. Once that's done, you expand the blueprint slightly to add a second tower, using the first one as a guide. Then you add a third, and so on.
• By growing the design region step-by-step, the computer avoids getting stuck in "local traps" (bad designs that look good but aren't the best) and finds the truly perfect structure.

Why Does This Matter?

This isn't just about math; it's about building the future.

• Quantum Computing: These cavities can trap light to interact with atoms, which is essential for quantum computers.
• Lasers: Better cavities mean more efficient, powerful, and precise lasers.
• Sensors: They can detect single molecules or tiny changes in the environment.

Summary

The paper solves a problem where computer designs for light traps were too slow and unstable because the targets were too sharp. The authors fixed this by creating a method that chases the target rather than aiming at a fixed point. It's like switching from a blindfolded hiker to a GPS-guided sprinter, allowing them to build perfect light traps in a fraction of the time.

Technical Summary

1. Problem Statement

Inverse design (topology optimization) of optical resonant cavities aims to maximize the Local Density of States (LDOS) at a specific target frequency $\omega_0$ and position $x_0$. The LDOS is a critical figure of merit for processes like spontaneous emission, often approximated by the Purcell factor ($Q/V$).

The Core Challenge:
While maximizing LDOS works well for low-to-moderate quality factors ($Q$), it suffers from severe ill-conditioning when optimizing for high-$Q$ resonances ($Q \gg 100$).

• The "Knife-Edge" Effect: At high $Q$, the LDOS response becomes an extremely sharp peak (Lorentzian). Small perturbations in design parameters ($p$) that shift the resonant frequency away from the target $\omega_0$ cause a catastrophic drop in performance, while perturbations that improve the mode shape (without shifting frequency) yield slow gains.
• Hessian Ill-Conditioning: This creates a Hessian matrix (second-derivative matrix) where the eigenvalue corresponding to frequency shifts scales as $O(Q^2)$ (or $O(Q^3)$ for raw LDOS). This massive disparity between eigenvalues makes gradient-based optimization algorithms converge extremely slowly, often requiring tens of thousands of iterations to reach $Q \sim 10^4$, and making $Q \gtrsim 10^5$ practically intractable.

2. Methodology

The authors propose a new "Eigenfrequency-Shifted" optimization algorithm that decouples the optimization of the resonance frequency from the optimization of the mode shape/localization.

A. Two-Stage Optimization Strategy

1. Unshifted Initialization:
   • Run a standard LDOS optimization (maximizing $\text{LDOS}(\omega_0, x_0, p)$) for a limited number of steps ($\lesssim 10^3$).
   • Goal: Quickly identify a strong resonance near $\omega_0$ (typically achieving $Q \gtrsim 100$) without needing a perfect initial guess.
2. Shifted Optimization:
   • Once a resonance is found, the objective function is modified to track the real part of the nearest complex eigenfrequency, $\text{Re}[\omega^*(p, \omega_0)]$.
   • New Objective: Maximize $\log \text{LDOS}(\text{Re}[\omega^*], x_0, p)$.
   • Constraint: Enforce a bandwidth constraint $\text{Re}[\omega^*] \in \text{BW}(\omega_0)$ to prevent the optimizer from "jumping" to a different mode or drifting too far from the target.
   • Mechanism: By optimizing the LDOS at the resonant frequency itself rather than at a fixed $\omega_0$, the algorithm eliminates the sensitivity to frequency shifts. The optimizer is effectively "centered" on the peak of the resonance curve at every iteration.

B. Computational Implementation

• Eigensolver: The method employs a shift-invert Arnoldi eigensolver to rapidly compute the eigenfrequency $\omega^*$ and its gradient.
• Gradient Computation:
  • The gradient of the shifted objective is derived using the adjoint method.
  • Crucially, the gradient computation remains efficient. While calculating $\omega^*$ requires an additional sparse-matrix factorization (doubling the cost per iteration compared to the unshifted method), the gradient itself is computed "for free" using the existing field solutions ($e$ and $e^*$).
• Topology Optimization (TO):
  • Uses density-based TO where permittivity $\varepsilon$ is mapped from a design density $\rho$.
  • 1D: Direct optimization of $\varepsilon$ (no filtering required).
  • 2D: Employs a conic filter and Subpixel-Smoothed Projection (SSP) to enforce a minimum lengthscale and binarize the structure (0 or 1 permittivity) while maintaining differentiability.

C. Heuristic: Successive Enlargement

To escape local optima and improve convergence further, the authors introduce a "successive enlargement" heuristic:

• Start with a small design region.
• Optimize, then enlarge the region by a small step (e.g., $\lambda_0/2$).
• Use the previous optimum as the initial guess for the larger domain.
• This mimics the physical growth of a cavity (adding layers) and helps the algorithm find global optima that are topologically distinct from the initial guess.

3. Key Contributions

1. Acceleration of High-$Q$ Design: Demonstrated orders-of-magnitude acceleration in convergence for $Q > 10^5$. The method eliminates the $O(Q^2)$ ill-conditioning of the Hessian associated with frequency shifts.
2. Theoretical Analysis: Analytically and numerically proved that the dominant eigenvalue of the Hessian for the unshifted objective scales as $Q^2$ (due to frequency sensitivity), whereas the shifted objective removes this eigenvalue, leaving only the eigenvalues related to mode shape improvements.
3. Record-Breaking Results:
   • 1D: Achieved $Q > 10^8$ in a dielectric cavity.
   • 2D: Achieved $Q > 10^6$ in a tiny $1.5\lambda_0 \times 1.5\lambda_0$ cavity, even with a source located outside the design region (a challenging scenario for standard photonic bandgap designs).
4. Gap Reduction to Theoretical Bounds: Showed that the shifted algorithm can narrow the gap between achievable LDOS and theoretical upper bounds in lossy materials, outperforming previous unshifted methods which often got stuck in poor local minima due to slow convergence.

4. Results

• 1D Experiments:
  • Compared shifted vs. unshifted algorithms starting from the same initialization.
  • Result: The shifted algorithm converged to $Q \approx 1.7 \times 10^5$ in $\sim 10^5$ iterations, while the unshifted algorithm stalled at $Q \approx 5.1 \times 10^4$.
  • With successive enlargement, the shifted algorithm reached $Q \approx 4.4 \times 10^8$ (surpassing a hand-designed quarter-wave stack of $Q \approx 4.0 \times 10^7$) in fewer than 1,000 iterations.
  • Observed that LDOS increases exponentially with design region size in steps of $\approx \lambda_0/2$, corresponding to the addition of quarter-wave layers.
• 2D Experiments:
  • Designed a cavity with a source $0.05\lambda_0$ outside the design region.
  • Result: Shifted algorithm reached $Q \approx 1.6 \times 10^6$, while the unshifted algorithm only reached $Q \approx 1.2 \times 10^5$ after 500,000 iterations.
  • The shifted algorithm reached $Q=10^5$ roughly 1,000 times faster (in terms of iterations) than the unshifted method.
• Hessian Verification:
  • Explicitly computed the Hessian for a 1D case ($Q \approx 11,000$).
  • Confirmed the unshifted Hessian has a dominant eigenvalue $\sim 10^5$ times larger than the next, aligned with the frequency shift direction.
  • Confirmed the shifted Hessian removes this large eigenvalue, resulting in a well-conditioned problem.

5. Significance

• Practical Impact: This method makes the inverse design of ultra-high-$Q$ cavities ($Q > 10^6$) computationally feasible, a task that was previously prohibitively expensive due to convergence issues.
• General Applicability: While demonstrated on LDOS, the "eigenfrequency-shifted" approach is applicable to any frequency-domain resonant response optimization (e.g., nonlinear effects, Smith-Purcell radiation, lasing efficiency) where the objective depends on staying on a resonance peak.
• Physical Insight: The work highlights that the primary bottleneck in high-$Q$ optimization is not the complexity of the geometry, but the numerical ill-conditioning caused by frequency sensitivity. By dynamically adjusting the objective frequency to the resonance, the optimization landscape is flattened, allowing standard gradient-based solvers to perform efficiently.
• Future Directions: The authors suggest this approach could be combined with theoretical upper bounds to generate better initial guesses and applied to more complex nonlinear or multi-physics optimization problems.