[paper] Automatic Parameter Extraction of MOSFET Compact Models

Gazmend Alia^1,2, Andi Buzo¹, Hannes Maier-Flaig¹, Klaus-Willi Pieper¹, 

Linus Maurer²  and Georg Pelz¹

Automatic Parameter Extraction of MOSFET Compact Models using Differential Evolution with Population Prediction (DEpred)

6th EDTM; March 6 to 9, 2022 

   
1 Infineon Technologies AG, Munich (D)
2 Bundeswehr University Munich (D)

Abstract: Parameter extraction of MOSFET compact models with hundreds of parameters is not a trivial task. Differential evolution (DE) has proven to be very effective in such highly dimensional parameter spaces. However, DE needs a large number of iterations to converge. This paper proposes a novel method to accelerate the convergence of DE by predicting tens of iterations ahead where the population will be, based on the knowledge from the already finished iterations. The method is validated with BSIM4 and HiSIM-HV compact models, where up to 50% of the iterations are saved.

Fig: DE vs DEpred cost function for BSIM4 and HiSIM-HV models.
DEpred reaches the target 50% faster.

[paper] Automated Compact Model Parameter Extraction

Marc Huppmann∗, Klaus-Willi Pieper†, Andi Buzo†, Linus Maurer∗ and Georg Pelz†

Utilizing Differential Evolution for an Automated Compact Model Parameter Extraction

In 2021 International Semiconductor Conference (CAS), pp. 231-234. IEEE, 2021.

   
∗ Universitat der Bundeswehr Munchen, Neubiberg, Germany
† Infineon Technologies AG, Neubiberg, Germany

Abstract: Parameter extraction is a challenging task, as it searches for a solution inside a high dimensional plus non- convex space. To be able to apply well known gradient based optimizers, the problem is dissected into multiple simpler yet intertwined tasks, which yields a complex and manual labour intensive procedure. On the contrary to gradient based methods, genetic algorithms perform excellent on global search problems, which eliminates the need for such a sophisticated workflow. In this paper, a highly automated methodology is presented that is capable of replacing the standard manual extraction sequence for the BSIM MOSFET compact model. Due to its superior extreme finding behaviour, the Differential Evolution algorithm is applied in combination with a special error metric to ensure a high fitting quality, in all regions of the output and transfer curves. Repeatably good results for 20k measurement points are obtained, with a reduction of factor 10 in total fitting duration, while coincidentally consuming mostly computation instead of manual labour time.

Fig: With every iteration, the errors approach each other till

they meet in roughly one point and σi terminates the fitting.