一种具备良好泛化能力的神经网络MOSFET模型  

A Neural Network MOSFET Model with Good Generalization Ability

作　　者：张济[1] 赵柏衡[1] 常胜[1] 王豪[1] 何进[1] 黄启俊[1] 

机构地区：[1]武汉大学物理科学与技术学院,武汉430072

出　　处：《固体电子学研究与进展》2014年第4期309-314,共6页Research & Progress of SSE

基　　金：国家自然科学基金资助项目(61204096;J1210061);教育部博士点基金资助项目(20100141120040;20110141120074)

摘　　要：设计了一种具有良好泛化能力的神经网络MOSFET模型。通过对超阈值区域和亚阈值区域的分段建模,得到了对器件直流特性和温度特性的精确描述。采用自适应遗传算法对网络的泛化能力进行优化,使模型对训练样本集边界外的数据也有很好的计算精度,具备了良好的预测能力。在Verilog-A平台上实现了分段模型的整合,能直接应用于基于SPICE的电路设计与分析。A neural network MOSFET model with good generalization ability is presented. Through the subsection modeling of supra-threshold region and sub-threshold region, the DC characteristics and temperature characteristics of the device are described accurately. Adopting adaptive genetic algorithm, the networkrs generalization ability is optimized. It also has high precision for the data outside the training sample set boundaries, which brings good predictive ability for the model. Upon the Verilog-A platform, the two segmented networks are integrated into a whole device model, which can be used on SPICE-based circuit design and analysis directly.

关 键 词：金属氧化物半导体场效应晶体管 神经网络 泛化能力 遗传算法 VERILOG-A 

分 类 号：TN386[电子电信—物理电子学]