International Conference on Solid State Devices and Materials

In BiCMOS technology, the shrinking size of SiGe heterojunction bipolar transistors (HBTs) makes doping profile determination less reliable and device development more complex. This paper proposes a comprehensive methodology to help address this issue. Surrogate models based on decoder Long Short-Term Memory (LSTM) neural networks are trained on a suitable database created with TCAD. They allow fast and accurate prediction of electrical characteristics from 1D profiles. They are then used to recover profiles that match silicon measurements. Finally, we show how they can guide profile improvement to achieve more ambitious performance.