Abstract
Accurate beam steering in optical phased arrays (OPAs) requires calibrated control voltages that compensate for channel-to-channel phase nonuniformity and thermal coupling while suppressing sidelobes and background. Conventional calibration typically requires iterative optimization at each steering angle. This work proposes a physics-informed graph transformer that predicts a vector of 128 control voltages from the target steering angle in a thermo-optic silicon OPA. Each channel is treated as a graph node, and pairwise antenna-element spacing is added to the attention score as a distance-based geometric bias. Training combines measured voltage labels, a differentiable array-factor reconstruction loss, and a voltage range penalty. With 7,000 calibration samples, the controller achieves a measured sidelobe suppression ratio (SLSR) at least 11.44 dB over the tested −60
∘
to +60
∘
steering range and 13.6 dB at 0
∘
. Under the same-chip experimental comparison at the tested steering angle, the proposed method achieved SLSR improvements of 2.8 dB over the deep neural network (DNN) baseline and 3.59 dB over stochastic parallel gradient descent (SPGD). The proposed algorithm generates the 128-channel voltage vector in 2 ms, and the measured end-to-end response from target-angle command input to stabilization of the detected target-angle optical power is 52.6 ms under an all-zero initial voltage state.