An Automatic Integration Network Approach for Generic Device Charge Modeling

The artificial neural network method has been explored for compact device modeling. In this work, an automatic integration network approach is presented for generic charge modeling. By utilizing the availability of partial derivatives for a given network, a loss function including only the mutual capacitance is designed for a terminal charge neural network. The network is trained against the capacitance data to obtain an accurate terminal charge model. A flow for sizing the network is proposed from considerations of both the accuracy and circuit simulation speed. It is found that a unified network with two hidden layers serves for the complete charge model with balancing the accuracy and simplicity. The approach has been used to develop a charge model for emerging tunneling transistors with hybrid transport mechanisms, and its applicability is verified.