Transient Modeling of Nonlinear Components Using Global Gated Deep Recurrent Neural Network Masked With Dropout

This article introduces a novel approach for nonlinear electronic circuit modeling called Global Gated Deep Recurrent Neural Network (GGDRNN). GGDRNNs leverage the stacking of multiple recurrent layers in a manner that each hidden layer is fully connected to all hidden layers of the previous time step, rather than just the same hidden layer in conventional deep RNN, and then controlled by a global gating unit. This layer to layer recurrent interconnection allows time-dependent information to pass through multiple layers in a much more efficient way and they finally controlled by a gating unit. Therefore, the GGDRNN can model complex time dependencies considerably better than conventional RNNs. To train the proposed GGDRNN structure, input-output waveforms of nonlinear components are extracted from circuit simulators. Additionally, a dropout technique is applied to GGDRNN, further increasing the model accuracy and reducing overfitting. The results have shown that the proposed model outperforms conventional deep RNNs in accuracy for modeling nonlinear components by more effectively learning complex time dependencies. Furthermore, the evaluation time of the GGDRNN-based model is shorter than transistor-level models. The validity of the proposed method is verified by modeling two transmission lines with nonlinear drivers and receivers, as well as the commercial TI SN74AHCT540 device.