Hybrid Nonlinear Modeling Using Adaptive Sampling

Abstract

Abstract-This paper proposes a direct method for the extraction of empirical-behavioral hybrid models using adaptive sampling. The empirical base is responsible for the functionality over a wide range of variables, especially in the extrapolation range. The behavioral part corrects for the errors of the empirical part in the region of particular interest, thus, it improves the accuracy in the desired region. Employment of response surface methodology and adaptive sampling allows full automation of the hybrid model extraction and assures its compactness. We used this approach to build a hybrid model composed of a robust empirical model available in CAD tools and a Radial Basis Functions interpolation model with Gaussian basis function. We extracted the hybrid model from measurements of a 0.15 µm GaAs HEMT and compared it to the pure behavioral and pure empirical models. The hybrid model yields higher accuracy while maintaining extrapolation capabilities. Additionally, the extraction time of the hybrid model is relatively low and we show that a good accuracy level can be achieved with a small number of measurements.

Uncertainty in Large-Signal Measurements Under Variable Load Conditions

“…Either using such models or measurement results directly, the designers can guarantee that circuits comply with the design requirements even in the worst-case scenarios [11]. If we could provide load-pull uncertainty data in real-time, another application could be adaptive sampling for modeling purposes [40]. In this case, the information provided by the uncertainty data could be used for safely choosing subsequent measurement points without violating the DUT's operational constraints [41].…”

Knowledge‐based neural network with Bayesian optimization for efficient nonlinear RF device modeling

“…[15][16][17] To address this challenge, optimization of sampling data have been proposed to cut down the number of measurement data needed. [18][19][20] However, these methods may not fully cover the diversity and complexity of the entire data set. On the other side, using simulation data in place of actual test data may have disadvantages, such as model bias and missing data.…”

“…To address this challenge, optimization of sampling data have been proposed to cut down the number of measurement data needed 18–20 . However, these methods may not fully cover the diversity and complexity of the entire data set.…”