High-Precision Large-Scale Optical Phased Array Calibration Using Physics-Informed Neural Network With Transfer Learning

As the number of elements in integrated optical phased arrays (OPAs) continues to scale up, the complexity of calibration and controlling these independent channels becomes prohibitive. To address this challenge, this study presents a physics-informed neural network (PINN)-based phase calibration algorithm that integrates transfer learning and physical constraints. By embedding the inherent periodicity and translational invariance of OPAs into the neural architecture, we design a sinusoidal activation function and a translation-invariant loss function, effectively mitigating phase ambiguity errors inherent in traditional calibration approaches. Furthermore, a sinusoidal fitting (Sin-fit) method is introduced to characterize phase shifters, enabling more accurate characterization of each phase shifter within the OPA, compared to characterizing a separate phase shifter using an external Mach-Zehnder interferometer (MZI). Experimental validation on a 128-channel silicon OPA demonstrates a sidelobe suppression ratio (SLSR) of 10.8 dB at 0° and an average SLSR of 9.7 dB across a ±12° steering range. The hybrid training strategy, combining 200,000 simulated and 80,000 measured far-field patterns, reduces real-data requirements by ∼70% compared to existing works. This study advances scalable OPA calibration with high efficiency and precision, offering critical insights for applications in LiDAR, free-space communications, and dynamic beamforming systems.