Long short-term memory neural network combined with a hybrid-modular clockwork structure for transient modeling of nonlinear circuits

This paper presents a novel macromodeling method and neural network structure called Clockwork Long Short-Term Memory (CWLSTM), tailored for high-speed nonlinear circuits. The proposed CWLSTM method is considered a more powerful yet simpler model than conventional LSTM due to its reduced parameter count and more efficient structure and training strategy. This structure promotes improved model generalization, resulting in better model accuracy and training time due to its unique modular gating connections. Additionally, the required training data is considerably reduced for generating a model with similar accuracy compared to the conventional LSTM. To further improve the proposed method, a hybrid version of CWLSTM, known as Hybrid-Modular CWLSTM, is introduced, utilizing various module types to enhance the model’s accuracy further. The reported experimental results reveal the superior performance of the proposed methods compared to the conventional LSTM in modeling high-speed nonlinear circuits. On top of the above advantages, the proposed methods can produce models that execute much faster than those based on existing simulation tools (LTspice and NGspice). The performance of the proposed methods is validated by modeling two high-speed nonlinear circuits.

Mots clés

• Computer-aided design (CAD)
• clockwork LSTM (CWLSTM)
• hybrid structure
• macromodeling
• nonlinear component
• recurrent neural network (RNN)