Microwave And RF Device Modeling Based On Neuro-Space Mapping Technique

With the rapid development in the semiconductor device manufacture technology, new devices constantly emerge for commercial use, and the models which match previous devices may not match new device behaviors well. Conventional approach to solve such modeling problem is to manually modify the existing models to fit new device characteristic. However, this is often a trial and error process and time consuming. Recently, Neuro-Space Mapping(Neuro-SM) technique was widely used for microwave and RF device modeling. By adding neural network mappings between the coarse model and the fine model, the overall Neuro-SM model can match the device data more closely than the coarse model alone. However, due to the increased modeling complexity and the industry’s increasing need for tighter accuracy in order to reduce design time, in such case, further enhancement is necessary for the traditional Neuro-SM method. In this thesis, three enhanced Neuro-SM methods are presented based on the traditional Neuro-SM.First, an enhanced analytical Neuro-SM method is presented aimed at using neural networks to automatically enhance the accuracy of the nonlinear device model. In our proposed models, a new mapping for current at gate and drain, in additional to the original mapping for voltage is used to produces better modeling accuracy compared to traditional Neuro-SM method. Analytical mapping representation of our proposed Neuro-SM model and training methods for mapping neural networks are also proposed. Application examples on modeling MESFET devices and the use of new models in DC,small signal S-parameter and large signal HB simulation demonstrate the correctness and effectiveness of our proposed enhanced Neuro-SM model.Secondly, a novel dynamic analytical Neuro-SM model is also proposed in this thesis. Our dynamic analytical Neuro-SM model is a significant advance over the traditional analytical Neuro-SM, due to the application of Time Delay Neural Network(TDNN) as mapping neural network which have the ability of learning then representing dynamic behavior of the nonlinear device. In this thesis, necessity of using dynamic mapping is demonstrated by a transistor modeling example. Analytical mapping representation and a novel two-phase training algorithm for fast training of our proposed dynamic Neuro-SM based on gradient optimization are also introduced. Application examples on modeling metal-semiconductor-field-effect transistor(MESFET), A real 2?50 GaAs Pseudomorphic high-electron mobility transistor(pHEMT) and a different HEMT generated from a physics-based device simulator Medici and the use of the dynamic Neuro-SM models in DC, small signal S-parameter and large-signal HB simulation demonstrate that our dynamic analytical Neuro-SM is an effect approach for modeling various types of advanced microwave and RF devices and allow us to exceed the present capabilities of the existing device model.Finally, advanced static and dynamic Neuro-SM eletro-thermal models considering self-heating effect for microwave and RF device modeling under different ambient temperature is proposed. The affect to the charactersitc of the nonlinear device due to the changes in working temperature is considered and electro-thermal dynamic and temperature variables are incorporated into Neuro-SM model, for the first time. Analytical mapping representation for our proposed static and dynamic Neuro-SM eletro-thermal models is derived to train the mapping neural networks to learn DC, small and large-signal data. Application examples on modeling high power transistor and GaAs MESFET devices demonstrate that our static and dynamic Neuro-SM eletro-thermal models allows us to exceed the present capabilities of device models built by existing equivalent circuit methods.