System Identification with SINDy Neural Networks for Transistor Modeling

Accurate models of semiconductor devices are a key component in modern circuit development. Although parameters of such models are often derived from fundamental physical laws, device geometry or material properties, empirical models from extensive measurements have become increasingly popular. Sparse identification of non-linear dynamics (SINDy) is one approach that emphasizes interpretability. Although several extensions have been developed for this procedure, it still lacks the ability to accommodate function compositions that are very common amongst established semiconductor models. This work introduces an approach fusing the capabilities of neural networks with the core principles of SINDy to address this shortcoming. The results are compared to well-known bipolar transistor models such as Gummel-Poon and indicate promising approximation capabilities.