Calibration of Wideband Multiport Junction Receivers Using Memory-Polynomial-Informed Neural Network

Multiport junction receivers experience performance degradation due to frequency-dependent behaviour of passive components and power detectors for wideband modulated signals. To calibrate these effects, a technique using Memory-Polynomial-Informed Neural-Network model is being proposed. The proposed calibration operates in two steps: a memory polynomial block captures dominant nonlinearities and memory effects, while a neural network refines predictions by learning residual and higher-order interactions. This dual-stage approach compensates wideband impairments, including nonideal multiport junctions, diode nonlinearities, and memory effects. The proposed technique is validated using a fabricated wideband sixport junction receiver with passive junction and four power detectors. The receiver is tested using 16-QAM signals at 200 Mbps operating over RF frequencies between 1GHz and 3GHz , and RF input powers from -25dBm to -15dBm, and an LO power of -5dBm. Operating under these test conditions, an EVM of less than 0.07% is achieved, significantly outperforming traditional calibration approaches.