Closing the Loop in Epitaxy with Machine Learning: Joint Optimization of Growth and Geometry in On-Chip Lasers
Mihir R. Athavale, Stephen A. Church, Wei Wen Wong, Andre KY Low, Hark Hoe Tan, Kedar Hippalgaonkar, Patrick Parkinson · 2026
Read Closing the Loop in Epitaxy with Machine Learning: Joint Optimization of Growth and Geometry in On-Chip Lasers by Mihir R. Athavale, Stephen A. Church, Wei Wen Wong, Andre KY Low, Hark Hoe Tan, Kedar Hippalgaonkar, Patrick Parkinson (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-08390.txt.
Introduction
Athavale and six colleagues close the loop between growing and shaping tiny on chip lasers, where small variations ruin repeatability. Their machine learning workflow first tunes growth and shape together, then traces leftover variation to its sources with autoencoder models. Every design lased, the median emission wavelength was 1333 nanometres in the telecom O band, and threshold spread dropped 73 percent against the best earlier values.
Catalog record
| CREATOR | Mihir R. Athavale, Stephen A. Church, Wei Wen Wong, Andre KY Low, Hark Hoe Tan, Kedar Hippalgaonkar, Patrick Parkinson |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35213 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.08390
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Closing the Loop in Epitaxy with Machine Learning: Joint Optimization of Growth and Geometry in On-Chip Lasers Mihir R. Athavale, Stephen A. Church, Wei Wen Wong, Andre KY Low, Hark Hoe Tan, Kedar Hippalgaonkar, Patrick Parkinson Abstract Achieving device-to-device reproducibility is a critical bottleneck for scalable photonic integrated circuits, as subtle variations in bottom-up epitaxial growth and fabrication severely limit yield. We present a machine learning workflow for III-V multi-quantum well microring lasers that first optimizes growth and geometry parameters via multi-objective Bayesian optimization, then leverages variational autoencoders (VAEs) to attribute residual device-to-device variability to its underlying sources. By explicitly targeting threshold variance alongside absolute performance, we demonstrate 100 % lasing yield across all designs. The optimized multi-quantum well microring laser fields achieved a median lasing threshold of 16 μ J cm − 2 pulse − 1 16~\mu\mathrm{J}\,\mathrm{cm}^{-2}\,\mathrm{pulse}^{-1} , a 73 % 73\% reduction in threshold variance relative to the previously reported best values, and a median emission wavelength of 1333 nm 1333~\mathrm{nm} , in the telecommunications O-band. Furthermore, to diagnose residual performance dispersion under nominally identical conditions, VAEs were used to isolate the key components of device morphology that impact performance. This analysis successfully decoupled geometric from material disorder, quantitatively linking previously unmeasured morphological variations to population-level threshold fluctuations. This data-driven workflow bridges the gap between fundamental epitaxy and reliable manufacturing, establishing a generalizable blueprint for designing and yield-optimizing complex, non-linear optoelectronic devices. Keywords: Integrated Photonics , Microring lasers , Variance Reduction , Multiobjective Bayesian Optimization , Variational Autoencoders , Maximum Mean Discrepancy † † email: [email protected] † † affiliation: Department of Physics and Astronomy and The Photon Science Institute, University of Manchester, Oxford Road, Manchester, M13 9PL, United Kingdom † † affiliation: Institute of Materials Research and Engineering (IMRE), Agency for Science, Technology and Research (A ∗ STAR), 2 Fusionopolis Way, Innovis #08-03, Singapore, 138634, Republic of Singapore † † affiliation: Department of Physics and Astronomy and The Photon Science Institute, University of Manchester, Oxford Road, Manchester, M13 9PL, United Kingdom † † affiliation: Centre for Engineering the Future, School of Science, Engineering and Environment, University of Salford, Salford, M5 4QJ, United Kingdom † † affiliation: ARC Centre of Excellence for Transformative Meta-Optical Systems, Department of Electronic Materials Engineering, Australian National University, Mills Road, Canberra, ACT 2601, Australia † † affiliation: School of Materials Science and Engineering, Nanyang Technological University, Singapore, 639798, Singapore † † affiliation: Institute of Materials Research and Engineering (IMRE), Agency for Science, Technology and Research (A ∗ STAR), 2 Fusionopolis Way, Innovis #08-03, Singapore, 138634, Republic of Singapore † † affiliation: ARC Centre of Excellence for Transformative Meta-Optical Systems, Department of Electronic Materials Engineering, Australian National University, Mills Road, Canberra, ACT 2601, Australia † † affiliation: School of Materials Science and Engineering, Nanyang Technological University, Singapore, 639798, Singapore † † affiliation: Department of Physics and Astronomy and The Photon Science Institute, University of Manchester, Oxford Road, Manchester, M13 9PL, United Kingdom Introduction Achieving functional uniformity across large populations of nominally identical devices is a persistent challenge in nanoscale device engineering, impacting fields spanning neuromorphic hardware 16 ; 15 to quantum photonics 42 ; 36 . In these cases, population-level statistics rather than single-device figures of merit ultimately govern system-level performance. Integrated photonic systems exemplify this constraint: large populations of on-chip lasers whose outputs must remain within strict power and wavelength specifications are required to support dense wavelength-division multiplexing (DWDM), parallel optical interconnects, and scalable multi-laser photonic integrated circuit (PIC) architectures 11 ; 33 ; 24 . For microcavity lasers and resonators, the stochastic nature of bottom-up epitaxial growth and fabrication tolerances introduce unavoidable nanometer-scale variations in geometry and material composition across a device population 35 . Consequently, process-induced mismatch in optical resonance increases the tuning, calibration, and trimming burden in complex circuits 44 ; 6 ; 12 ; 20 . Several approaches have been pursued to reduce performance variability in on-chip lasers. High-precision top-down fabrication can tighten wavelength distributions across nominally identical cavities 23 . At the circuit level, residual scatter can be mitigated using variation-tolerant resonator designs and active stabilization, including photoconductive resonator heaters 31 ; 13 . Bottom-up growth engineering offers a complementary route by avoiding etch-induced sidewall roughness and leveraging epitaxial selectivity, including selective area epitaxy (SAE) and aspect ratio trapping concepts that have enabled wafer-scale III–V nanoridge laser diodes on 300 mm silicon wafers 10 ; 34 . Statistical screening and population-level analysis can characterise performance distributions, link morphology to threshold and spectral variation, and support selection and deterministic microassembly workflows such as binning and transfer printing 7 ; 1 ; 14 ; 30 ; 8 . Despite these strategies, addressing variability as an explicit optimization objective within the design process remains under-explored. Data-driven optimization has emerged as a means of navigating high-dimensional and experimentally expensive photonic and metamaterial design spaces 28 ; 32 , offering a natural route to incorporating population-level objectives – such as device-to-device variability – into the optimization target. The majority of photonics optimization and inverse-design studies focus on improving figures of merit for individual devices 25 ; 29 ; in contrast, scaling to large device populations requires explicit control of variability arising from process variation across a wafer or within local regions 6 ; 40 . We have recently introduced a multi-objective optimization framework for multi-quantum-well microrings (MQW MRs) to reduce lasing thresholds, improve yield (defined as the fraction of nominally identical devices that exhibit lasing under the applied measurement conditions), and push the emission wavelength toward the telecommunications O-band 2 . Here, we extend this framework by incorporating threshold variability as an explicit objective, enabling the joint optimization of (i) the median lasing threshold, (ii) the median lasing wavelength, and (iii) a design-level variance quantifying the spread of the threshold distribution. Wavelength variance was not included as a dispersion objective because nominally identical MQW MR devices can lase on different cavity modes, each with a distinct emission wavelength, such that the device-to-device wavelength spread reflects both genuine structural variation and differences in mode selection, rather than structural variation alone. Using this objective set, we achieved median thresholds as low as 16~\mu\mathrm{J}\,\mathrm{cm}^{-2}\,\mathrm{pulse}^{-1} , reduced threshold variance by 73\% , and steered median emission wavelength to 1333~\mathrm{nm} , while maintaining 100\% yield across all designs. Despite these optimizations, residual variability can persist because both spatially systematic and stochastic process fluctuations can modulate local geometry and material properties in ways not deterministically captured by growth and design parameters 40 ; 35 . We adapt a variational autoencoder (VAE) approach to diagnose the origin of residual variability in MQW MR lasers, going beyond previous applications of this approach used to detect drift in semiconductor processing 17 ; 43 . Since the relevant morphological differences are subtle, a VAE was trained to learn a compressed latent representation of MQW MR morphology from optical images, enabling comparisons in latent space that are not accessible through direct geometric measurements alone. The morphological contribution to threshold variation is isolated independently of known growth and geometry parameters, and the added predictive value of morphology was quantified. Together, these results establish an end-to-end data-driven workflow that improves consistency through variance-aware optimization and provides morphology-aware diagnostics that isolate the independent