Gaussian-Excited Causal Behavioral Modeling of RFICs From Real-World Measurements Using Neural Networks

With the advancement of radio frequency integrated circuit (RFIC) technology, neural network-based behavioral modeling of RFICs has become an emerging research focus. However, most existing studies focus on optimizing neural network architectures but rely on training and testing datasets with specific and similar characteristics, which constrains the generalization capability of the resulting models. Moreover, simulation-generated data are widely used, further limiting the models’ ability to capture real-world RFIC behavior. To address this issue, this paper introduces an information-rich, band-limited Gaussian signal as the excitation signal for RFIC feature extraction. The datasets are obtained from real-world measurements rather than simulations, enhancing the generalization performance of the trained models. Considering the limited dynamic range of oscilloscopes in practical measurements, we further propose a novel “Weighted Band-Selective Gaussian” (WBS-Gauss) signal, which effectively expands the dynamic range of the collected datasets. Additionally, we develop a causal regression behavioral modeling framework that is compatible with various neural network architectures. Modeling experiments are conducted on RF amplifiers with high nonlinearity and filters with wide dynamic range. The results demonstrate that the proposed excitation signal and modeling methodology enable the construction of behavioral RFIC models with excellent generalization capabilities.