Pure single-photon mode recognition and design via hybrid learning framework

Single-photon sources generated by spontaneous parametric downconversion (SPDC) have been widely applied in quantum information science. Although analytical optimization methods have provided general and effective approaches for achieving high-purity single-photon sources, the design of such sources remains a complex task due to the intricate physical principles involved. We propose a hybrid learning framework to enable mode recognition and inverse design of pure single photons based on the joint spectral amplitude (JSA) in SPDC. This framework first enhances inter-mode differences and employs a probabilistic clustering approach to accurately recognize phase-matching modes, achieving a test accuracy of 91.92%. Subsequently, the multimodal information from the JSA and SPDC parameters is encoded into vectors and mapped into a shared high-dimensional embedding space to align the two modalities. Based on this alignment, a linear regression model is constructed to predict SPDC parameters, achieving accuracies of 90.66%, 85.03%, and 91.94% for the pump optical bandwidth, crystal length, and signal optical wavelength, respectively. Our framework confirms the feasibility of inferring phase-matching modes and key configurable information in reverse from the target JSA with the assistance of artificial intelligence, thereby offering new perspectives for photon source design. We hope it will be a useful and promising method for constructing high-purity single-photon sources.