🔬 optics

Closing the Loop in Epitaxy with Machine Learning: Joint Optimization of Growth and Geometry in On-Chip Lasers

This paper presents a machine learning workflow combining multi-objective Bayesian optimization and variational autoencoders to jointly optimize growth and geometry parameters for III-V microring lasers, achieving 100% lasing yield with a 73% reduction in threshold variance while successfully decoupling geometric from material sources of device variability.

Original authors: Mihir R. Athavale, Stephen A. Church, Wei Wen Wong, Andre KY Low, Hark Hoe Tan, Kedar Hippalgaonkar, Patrick Parkinson

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mihir R. Athavale, Stephen A. Church, Wei Wen Wong, Andre KY Low, Hark Hoe Tan, Kedar Hippalgaonkar, Patrick Parkinson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ 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 a master baker trying to create thousands of identical, perfect chocolate chip cookies. You have a precise recipe (the ingredients) and a specific mold (the shape). In an ideal world, every cookie would come out exactly the same: the same size, the same crispiness, and the same perfect chocolate distribution.

But in the real world of high-tech manufacturing, things are messier. When you try to bake these "cookies" (which, in this case, are tiny lasers on a computer chip), subtle differences in the oven temperature, the humidity, or how the dough spreads can make some cookies slightly burnt, some slightly undercooked, and some the wrong size.

For a single cookie, this might not matter. But for a computer chip that needs thousands of these lasers working together perfectly (like a choir singing in harmony), even tiny differences cause the whole system to fail. Some lasers might be too weak, or they might sing a slightly different note (wavelength) than their neighbors. This is the "reproducibility bottleneck" the paper tackles.

Here is how the researchers solved this problem, broken down into simple steps:

1. The "Smart Chef" (Machine Learning Optimization)

Instead of guessing the perfect recipe, the researchers used a Machine Learning "Smart Chef" (specifically, a technique called Bayesian Optimization).

The Old Way: Bakers would tweak the recipe, bake a batch, taste a few cookies, and guess what to change next. This is slow and often misses the perfect combination.
The New Way: The Smart Chef looks at thousands of past baking attempts. It doesn't just try to make the best cookie; it tries to make the most consistent batch. It asks: "If I change the oven temperature by 1 degree and the baking time by 2 seconds, will I get a batch where every single cookie is perfect, not just the lucky ones?"

The result? They found a "Goldilocks" recipe that produced lasers with the lowest possible energy needed to turn on (the "threshold") and, crucially, made sure 100% of the lasers in a batch worked perfectly.

2. The "Magic Mirror" (Variational Autoencoders)

Even with the perfect recipe, some batches still had tiny differences. Why? Because the "dough" (the material) didn't spread exactly the same way every time, creating microscopic bumps or weird shapes that the recipe didn't account for.

To find these invisible flaws, the researchers used a Variational Autoencoder (VAE). Think of this as a Magic Mirror or a super-smart art critic.

How it works: You show the mirror a picture of a laser. The mirror doesn't just measure the diameter (like a ruler); it looks at the entire personality of the shape. It compresses the complex image into a secret code (a "latent vector") that captures every tiny bump, curve, and imperfection.
The Discovery: The researchers used this mirror to compare batches. They found that when the "secret codes" of the shapes were very different from each other, the performance of the lasers was also very different.
The Breakthrough: They proved that these tiny, invisible shape differences were the "smoking gun" causing the lasers to behave differently. By using this mirror, they could predict how a laser would perform just by looking at its shape, even before turning it on.

3. The "Two Different Goals"

The paper also made a fascinating discovery about what makes these lasers tick:

The "Note" (Wavelength): This is like the pitch of a musical instrument. It is mostly determined by the size of the laser (the mold). If the mold is big, the note is low. If the mold is small, the note is high. The tiny bumps on the surface don't change the note much.
The "Volume" (Threshold): This is how much energy it takes to make the laser sing. This is extremely sensitive to the tiny bumps and imperfections. A tiny scratch can make the laser work much harder to turn on.

The Big Picture: Closing the Loop

Before this study, scientists would optimize the recipe, make the lasers, and hope for the best. If the batch was inconsistent, they would just throw away the bad ones.

This paper creates a closed loop:

Optimize: Use AI to find the best recipe for consistency.
Diagnose: Use the "Magic Mirror" (VAE) to see why the remaining differences exist.
Improve: Use that knowledge to refine the recipe further.

Why does this matter?
Imagine building a massive city of fiber-optic internet cables. If every laser in that city sings a slightly different note or is slightly dimmer, the internet slows down or breaks. This new method ensures that when you build a million lasers, they all act like a single, perfect unit. It turns the chaotic art of "growing" tiny lasers into a reliable, predictable science, paving the way for faster computers and better internet.

1. Problem Statement

The primary bottleneck in scaling photonic integrated circuits (PICs) is device-to-device reproducibility. While individual devices can be optimized, achieving functional uniformity across large populations of nominally identical devices is difficult due to:

Stochastic Variations: Bottom-up epitaxial growth (specifically Selective Area Epitaxy, SAE) and fabrication tolerances introduce unavoidable nanometer-scale variations in geometry and material composition.
Yield Limitations: These variations lead to mismatches in optical resonance, increasing the burden of tuning and calibration.
Optimization Gaps: Traditional design processes often optimize for single-device figures of merit (e.g., lowest threshold) rather than population-level statistics (e.g., variance reduction). Furthermore, residual variability often persists even under "optimized" conditions due to subtle morphological differences not captured by standard growth parameters.

2. Methodology

The authors propose a closed-loop, data-driven workflow combining Multi-Objective Bayesian Optimization (MOBO) and Variational Autoencoders (VAEs) to optimize III-V multi-quantum well (MQW) microring lasers.

A. Multi-Objective Bayesian Optimization (MOBO)

Objective: To jointly optimize three competing metrics:
1. Median lasing threshold (minimize).
2. Median emission wavelength (targeting the telecommunications O-band, ~1330 nm).
3. Threshold variance (minimize the spread of thresholds within a field).
Parameter Space: A 7-dimensional space comprising:
- Growth Parameters: Number of quantum wells, growth temperature, As/P ratio, V/III ratio, capping time.
- Geometry Parameters: Microring diameter and pitch (center-to-center spacing).
Process:
- Used Gaussian Process surrogate models trained on 95 historical field-level data points (Batch 1 & 2).
- Employed the qN-ParEGO acquisition function to identify Pareto-optimal trade-offs.
- Generated new growth recipes (Batch 3) and fabricated paired samples (Batch 3A and 3B) to test reproducibility.

B. Variational Autoencoder (VAE) for Morphology Analysis

To diagnose residual variability that MOBO could not explain, the authors used a VAE to analyze optical images of the microrings.

Image Processing: RGB images were converted to binary rim-edge maps using OpenCV (CLAHE normalization, morphological black-hat filtering).
Latent Representation: A convolutional VAE compressed 128×128 binary images into 32-dimensional latent vectors.
Metric Comparison:
- Morphology Gap: Measured using Squared Maximum Mean Discrepancy ( $MMD^2$ ) between latent distributions of paired samples.
- Threshold Gap: Measured using Wasserstein distance ( $W_1$ ) between lasing threshold distributions.
Predictive Modeling: Ridge regression was used to determine if VAE-derived morphology features could predict lasing thresholds better than growth recipes and nominal geometry alone.

3. Key Contributions

Explicit Variance Optimization: The study is among the first to treat device-to-device variance as an explicit objective function in the design loop, rather than just a post-hoc metric.
Decoupling Geometry from Material Disorder: By using VAEs, the authors successfully isolated the contribution of subtle morphological variations (shape disorder) from nominal growth parameters, quantitatively linking them to performance fluctuations.
End-to-End Workflow: Established a generalizable framework that bridges fundamental epitaxy with reliable manufacturing, moving from "hero device" optimization to population-level yield optimization.
Morphology-Aware Diagnostics: Demonstrated that standard geometric descriptors (e.g., diameter) fail to capture complex facet variations in SAE growth, whereas latent space representations capture these nuances effectively.

4. Key Results

Performance Metrics (Batch 3 vs. Previous Best):
- Median Lasing Threshold: Reduced to 16 µJ cm⁻² pulse⁻¹ (a 52% reduction from previous bests).
- Threshold Variance: Reduced by 73% (from previous bests), achieving a variance of 24 µJ² cm⁻⁴.
- Emission Wavelength: Shifted to 1333 nm, successfully targeting the telecommunications O-band.
- Yield: Achieved 100% lasing yield across all designs.
Morphology-Performance Correlation:
- A statistically significant positive correlation was found between the morphology gap ( $MMD^2$ ) and the threshold gap ( $W_1$ ) ( $\beta = 70, p = 0.00189$ ).
- Fields with larger morphological shifts in latent space exhibited larger shifts in threshold distributions.
Predictive Power Analysis:
- Adding VAE latent features to a baseline model (Recipe + Nominal Size) increased predictive power ( $\Delta R^2$ ) by 0.054 for lasing threshold.
- This confirmed that the VAE captures independent, physically meaningful shape information not encoded in the growth recipe or nominal dimensions.
- Conversely, lasing wavelength was well-predicted by nominal parameters alone, suggesting wavelength is robust to subtle morphological noise, while threshold is highly sensitive to it.

5. Significance

This work represents a paradigm shift in nanomanufacturing and optoelectronic design:

From Single-Device to Population-Centric: It establishes that for scalable PICs, consistent ensemble behavior is more critical than isolated high-performance devices.
Closing the Loop: The integration of machine learning allows for the "closing of the loop" between growth conditions, resulting morphology, and final device performance.
Generalizability: The framework is not limited to III-V microrings; it offers a blueprint for optimizing any complex, nonlinear optoelectronic device where performance inhomogeneity stems from growth kinetics and structural variations (e.g., quantum dots, 2D materials, heterogeneous integration).
Manufacturing Impact: By reducing the need for post-fabrication tuning and trimming, this approach directly addresses the yield and cost barriers preventing the widespread adoption of dense wavelength-division multiplexing (DWDM) and large-scale photonic circuits.