Neural Network Based Microwave And RF MOSFET Modeling

With the continuous development of wireless communication technology, radio frequency integrated circuit plays an increasingly important role in it. Bulk CMOS technology has become one of the most feasible candidates for building reliable circuits for RF applications. The progress of CMOS technology has made MOSFET transistors show excellent microwave performance. In addition, CMOS technology has such advantages relative to other technologies as mature manufacturing process, low cost, low power, high integration, good integration with high performance digital circuits and high-speed analog circuits, and successful scalability. In order to ensure the circuit performance for the required frequency bands and also to shorten the ratio of time to market, device models are very critical. Artificial neural networks (ANNs) have already been applied to RF and Microwave computer-aided design (CAD) tasks as an unconventional alternative. Neuromodeling is efficient in comparison to conventional modeling methods, such as numerical modeling methods, which could be computationally expensive, or analytical methods, which could be difficult to obtain for new devices, or empirical models, whose range and accuracy could be limited. Furthermore, neural models are easier to develop for new devices than conventional models. Thus, neural network based MOSFET modeling is studied in this thesis. The main content of the thesis is divided into four parts as follows:Firstly, the characterization and principal modeling method for MOSFET is reviewed. The structure and operation of MOSFET is presented. The characteristic and characterization of MOSFET at high frequency is also studied. Various kinds of modeling methods for MOSFET are summarized, where we focus on neural network based modeling method.Secondly, neural network based small-signal modeling for MOSFET is studied. The test structure and de-embedding method of MOSFET are presented. Small-signal equivalent circuit and direct parameter extraction method for MOSFET are also studied. Parameters in equivalent circuit are extracted and then optimized, and corresponding bias-dependent empirical model is determined. Bias-dependent small-signal modeling approach based on neuro-space mapping is proposed for MOSFET. Good agreement is obtained between the simulated and measured results for a 130 nm MOSFET in the frequency range of 100MHz-40GHz confirming the validity and effectiveness of our approach.Thirdly, neural network based DC modeling for MOSFET is studied. Two approaches for modeling DC characteristics for MOSFET based on neuro-space mapping (SM) are proposed. The first approach makes use of classical neuro-SM technology, while the second combines neuro-SM with prior knowledge input and source difference method. The formulas for obtaining the transconductance and output conductance in two approaches are derived. TheⅠ-Ⅴcharacteristics as well as their conductances obtained by the formulas in two approaches are compared to the measured data. Experimental results, which confirm the validity of our approaches, are also presented.Finally, neural network based large-signal modeling for MOSFET is studied. A large-signal modeling approach based on the combination of equivalent circuit and neuro-space mapping modeling techniques is proposed for MOSFET. In order to account for the dispersion effects, two neuro-space mapping based models are employed to model the drain current at DC and RF conditions respectively. Corresponding training process in our approach is also presented. Good agreement is obtained between the model and data of the DC, S parameter, and harmonic performance for a 0.13μm gate length,5μm gate width per finger and 20 fingers MOSFET over a wide range of bias points, demonstrating the proposed model is valid for DC, small-signal and nonlinear operation. Comparison of DC, S-parameter and harmonic performance between proposed model and empirical model further reveals the better accuracy of the proposed model.