Gen-Fab: A Variation-Aware Generative Model for Predicting Fabrication Variations in Nanophotonic Devices

This paper introduces Gen-Fab, a variation-aware conditional generative adversarial network that accurately predicts diverse, high-resolution fabrication outcomes for nanophotonic devices by modeling process-induced uncertainties, outperforming existing deterministic and probabilistic baselines in both accuracy and distribution alignment.

The Gist

The Big Picture: The "Perfect" vs. The "Real"

Imagine you are an architect designing a beautiful, intricate sandcastle. You draw a perfect blueprint on paper (this is the GDS layout). In a perfect world, if you built it exactly as drawn, it would look exactly like your drawing.

But in the real world, sand is tricky. The wind might blow a little grain here, the tide might wash away a corner there, or your shovel might slip slightly. When you build the castle, it never looks exactly like the blueprint. It looks "close," but with tiny, random imperfections.

In the world of Silicon Photonics (chips that use light instead of electricity), these tiny imperfections are a huge problem. A chip designed to be perfect might fail if the factory "sand" (the manufacturing process) shifts the edges by just a few nanometers.

The Problem: The "One-Size-Fits-All" Crystal Ball

Scientists have tried to build "Digital Twins"—computer programs that predict what the factory will actually produce.

• The Old Way (Deterministic Models): Imagine a crystal ball that always shows you one specific outcome. If you ask, "What will my sandcastle look like?" it shows you one specific, slightly imperfect castle.
  • The Flaw: In reality, if you built that same sandcastle 35 times, you'd get 35 slightly different versions. The old crystal ball only shows you one, so it misses the range of possibilities. It's like predicting the weather by saying, "It will be 72°F," when in reality, it could be sunny, rainy, or windy.

The Solution: Gen-Fab (The "Imagination Engine")

The authors of this paper created a new tool called Gen-Fab. Think of it not as a crystal ball, but as a creative artist with a magic wand.

1. The Input: You give the artist your perfect blueprint (the GDS layout).
2. The Magic Wand (Latent Noise): This is the secret sauce. Before the artist starts drawing, you tap a magic wand that adds a tiny bit of "randomness" (like a gust of wind or a shaky hand).
3. The Output:
   • If you tap the wand once, the artist draws Version A of the sandcastle.
   • If you tap the wand again (with a different random shake), the artist draws Version B.
   • If you do this 35 times, you get 35 unique, realistic versions of the castle, all slightly different from each other, just like real life.

How It Works (The "Teacher and Student" Game)

The paper uses a type of AI called a Generative Adversarial Network (GAN). You can think of this as a game between two characters:

• The Forger (The Generator): This AI tries to draw fake SEM images (photos of the real chips) that look so real, you can't tell them apart from the truth. It uses the blueprint and the "magic wand" (random noise) to create these images.
• The Detective (The Discriminator): This AI is a strict art critic. Its job is to look at the Forger's drawings and the real photos and say, "That's a fake!" or "That's real!"

The Training Loop:

1. The Forger tries to make a fake chip image.
2. The Detective tries to catch it.
3. If the Detective catches it, the Forger learns, "Oh, I made the corners too sharp. I need to make them rounder."
4. They play this game millions of times. Eventually, the Forger becomes so good at mimicking the "randomness" of the factory that the Detective can no longer tell the difference.

Why Gen-Fab is Better Than the Others

The researchers compared Gen-Fab to three other methods:

1. The Strict U-Net: This is the old crystal ball. It gives one answer. It's accurate on average, but it's boring and doesn't show the variety.
2. The "Gambler" U-Net (MC-Dropout): This tries to guess by randomly turning parts of its brain off during the test. It's like a student guessing answers on a test. It creates variety, but the variety is random noise, not realistic factory errors.
3. The "Crowd" (Ensemble): This uses 35 different U-Nets trained separately. It's like asking 35 different architects to guess the outcome. It's better, but it's slow and expensive to train 35 separate models.

The Winner: Gen-Fab wins because it learns the rules of the factory's randomness. It doesn't just guess; it understands that "corners get rounded" and "edges get rough" in a specific way.

The Results: A Perfect Match

When they tested Gen-Fab on new chip designs it had never seen before:

• Accuracy: It matched the real factory photos better than anyone else (89.8% accuracy vs. 85% for the others).
• Realism: The "shape" of the mistakes Gen-Fab made looked exactly like the mistakes real factories make.
• Speed: It can generate 35 different realistic outcomes in seconds, helping engineers design chips that won't fail when the factory inevitably makes a tiny mistake.

The Takeaway

Gen-Fab is a "Digital Twin" that understands uncertainty.

Instead of asking, "What will this chip look like?" and getting one answer, Gen-Fab asks, "What are all the possible ways this chip could look?" and shows you a whole gallery of them. This allows engineers to design chips that are robust enough to survive the messy, imperfect reality of the real world.

Technical Summary

1. Problem Statement

Silicon photonic devices are highly sensitive to fabrication-induced variations (e.g., over-etching, under-etching, corner rounding) caused by systematic and stochastic effects in CMOS-compatible processes.

• The Challenge: These variations are spatially non-uniform and depend on feature size and shape. Traditional design methods (statistical corner analysis) lack the spatial resolution to capture complex, stochastic deviations.
• The Gap: Existing deep learning approaches (e.g., deterministic U-Nets) map a design layout (GDS) to a single predicted outcome. This fails to represent the one-to-many mapping inherent in real-world fabrication, where a single layout can result in a distribution of physically distinct outcomes.
• Goal: Develop a "fabrication-aware digital twin" capable of predicting the distribution of possible fabricated outcomes from a single design layout, capturing both deterministic structure and stochastic variability.

2. Methodology

The authors propose Gen-Fab, a conditional Generative Adversarial Network (cGAN) based on the Pix2Pix architecture, modified to handle stochasticity.

Core Architecture

• Generator ($G$): A U-Net encoder-decoder with skip connections.
  • Key Modification: Unlike standard Pix2Pix, Gen-Fab injects a latent noise vector ($z$) at the bottleneck layer of the generator.
  • Mechanism: The $d$-dimensional noise vector is tiled to match the bottleneck feature map dimensions. This allows the model to sample different noise vectors for the same input layout, generating diverse, plausible SEM images (one-to-many mapping).
• Discriminator ($D$): A PatchGAN architecture that evaluates $70 \times 70$ pixel patches to distinguish between real fabricated SEM images and generated predictions, ensuring local realism and high-frequency consistency.

Training Objectives

The model is trained using a minimax game with a combined loss function:

1. Adversarial Loss ($L_{GAN}$): Encourages the generator to produce images indistinguishable from real SEMs.
2. Reconstruction Loss ($L_{L1}$): An $L1$ loss penalizes pixel-level errors between the generated image and the ground truth.
   • Weighting: The total generator loss is $L_G = L_{GAN} + \lambda L_{L1}$, with $\lambda = 100$. This high weight ensures structural fidelity while the adversarial term sharpens details.
3. Training Procedure: The noise vector $z$ is resampled for every training example in every iteration, forcing the generator to learn a distribution rather than a fixed mapping.

Baselines for Comparison

Gen-Fab is benchmarked against three baselines:

1. Deterministic U-Net: A standard encoder-decoder producing a single output.
2. MC-Dropout U-Net: A U-Net with dropout layers active during inference to approximate Bayesian uncertainty.
3. Varying U-Nets (Ensemble): An ensemble of 35 independently trained U-Nets to approximate variability.

3. Key Contributions

1. Stochastic cGAN Framework: The introduction of a latent noise vector into the Pix2Pix bottleneck to explicitly model the one-to-many nature of nanophotonic fabrication.
2. Digital Twin Capability: The ability to generate a distribution of high-resolution SEM-like images from a single GDS layout, enabling uncertainty quantification and robust design analysis.
3. Comprehensive Evaluation: A rigorous evaluation using both pixel-level metrics (IoU) and distribution-level metrics (KL-Divergence, Wasserstein Distance) on an out-of-distribution (OOD) dataset.
4. Uncertainty Decomposition: A novel analysis separating aleatoric uncertainty (inherent fabrication randomness) from epistemic uncertainty (model parameter uncertainty), demonstrating that Gen-Fab primarily captures the former.

4. Experimental Results

The model was trained on the ANT-NanoSOI dataset (GDS-SEM pairs of inverse-designed structures) and tested on six unseen nominal structures (Cross, Square, Target patterns with varying dimensions).

Quantitative Performance

• Intersection-over-Union (IoU): Gen-Fab achieved the highest average IoU of 89.8% (using Greedy matching), significantly outperforming:
  • Varying U-Nets (85.8%)
  • Deterministic U-Net (85.3%)
  • MC-Dropout U-Net (83.4%)
• Distributional Fidelity: Gen-Fab showed the lowest divergence from real data:
  • Kullback-Leibler Divergence (KL-D): 0.2657 (Cross-25) vs. 0.9471 for Varying U-Nets.
  • Wasserstein Distance (W-D): 0.1693 (Cross-25) vs. 0.1831 for Varying U-Nets.
  • Note: MC-Dropout performed poorly, indicating it captures model uncertainty rather than data variability.

Qualitative Analysis

• Visual Realism: Gen-Fab outputs exhibited realistic edge roughness, asymmetry, and corner rounding, whereas deterministic U-Nets produced overly smooth, symmetric results.
• Variance Mapping: Variance maps of Gen-Fab outputs closely matched the variance patterns of real SEMs (concentrated at edges), whereas MC-Dropout showed weak variance and Varying U-Nets showed exaggerated, unrealistic variance.

Generalization

• Distribution Shift: Analysis using t-SNE and Fréchet Distance confirmed a significant domain shift between training and test sets. Gen-Fab maintained high performance on these OOD structures, proving strong generalization capabilities.

5. Significance and Implications

• Robust Design: Gen-Fab enables "Design for Manufacturability" (DfM) by allowing designers to simulate a distribution of outcomes. This supports optimization strategies based on yield, variance reduction, and percentile constraints rather than idealized single-point predictions.
• Digital Twin Integration: The model serves as a fast, data-driven surrogate for physical fabrication, bridging the gap between ideal layouts and manufactured reality.
• Future Applications: The generated binary SEM-like images can be converted back to polygonal geometries (GDSII) for integration into electromagnetic simulators (e.g., Lumerical), enabling automated, fabrication-aware inverse design loops.
• Scalability: The approach is computationally efficient compared to training massive ensembles, offering a practical path toward stochastic modeling in nanophotonics.

In conclusion, Gen-Fab represents a significant advancement in photonic digital twins by moving beyond deterministic predictions to capture the stochastic reality of nanofabrication, thereby enabling more robust and reliable silicon photonic device design.