Decoding shape-induced variability in bottom-up grown microring lasers through generative modelling and Bayesian optimisation

Bottom-up InP/InAsP microring lasers fabricated by selective-area MOCVD offer compact, integrable light sources but, like many bottom-up grown semiconductor microstructures, remain limited by growth-induced variability. Two sequential multi-objective Bayesian optimisation runs were conducted to efficiently improve lasing performance while minimising costly trial growths. The first optimisation achieved a 1.6× reduction in lasing threshold and tuned emission to 1268 nm in the O-band with 100% yield. The second optimisation targeted a lower threshold, longer wavelength, and reduced variance, with two independent growth sets fabricated under identical parameters to verify reproducibility. This run achieved a further 2× reduction in median lasing threshold, lowered variance by 73%, and shifted the median emission to 1333 nm near the C-band. Both sets showed consistent improvements, yet measurable differences in performance revealed persistent morphology-driven variability. Variational autoencoder analysis confirmed that subtle shape variations correlate with performance, emphasising morphology-aware, generative design for reproducible, high-yield photonic integration.