MOSFET Model Parameter Extraction Using Reinforcement Learning

As the field of modeling and simulation continues to evolve, the demand for more accurate and complex models has grown. The increased complexity of these models leads to higher dimensionality of the model parameters space. The compact BSIM (Berkeley Short-Channel IGFET Model) for MOSFETs, widely utilized in SPICE and SPECTRE, relies on hundreds of parameters to describe device behavior. Parameter extraction for such models is traditionally a time-consuming and labor-intensive process. In this paper we propose a new method for BSIM model parameter extraction using Reinforcement Learning (RL). Our approach employs an RL agent with continuous action space to extract parameters that align with multiple measurements, offering the promise of a more automated and expedited parameter extraction process. This approach introduces a training process for multiple reference datasets, generalizing the solution to the problem.