Designing strongly coupled polaritonic structures via statistical machine learning

Strong coupling photonics, in which coupling strengths significantly exceed decay factors, holds great promise for applications in energy conversion and information processing. However, the effective and scalable design of strongly coupled polaritonic structures requires precise determination of the transition boundaries between strong and weak coupling regions. Traditional trial-and-error methods and classical machine learning (ML) algorithms struggle to achieve this due to the inherent difficulty in measuring decay factors during coupling. To address these challenges, we propose a hybrid ML framework that integrates physics-informed modeling with uncertainty quantification. This approach enables accurate determination of strong-weak coupling transition boundaries-even with inaccessible decay factors. By leveraging this capability, our method facilitates efficient and large-scale design of strongly coupled polaritonic structures with sparse data, achieving a computational speedup of ~10⁴ times compared to conventional simulations. Guided by this framework, we experimentally constructed a hexagonal boron nitride (hBN) polariton coupling structure and observed strong coupling via near-field spectroscopy. This work establishes a generalizable optimization methodology for strongly coupled photonic devices, opening a broad avenue for polariton-enhanced energy conversion and optical information modulation.