Artificial Neural Network Surrogate Models for Efficient Design Space Exploration of 14-nm FinFETs

Journal article

Surila Guglani, A. Dasgupta, M. Kao, Chenming Hu, Sourajeet Roy
Device Research Conference, 2022

Cite

APA Click to copy
Guglani, S., Dasgupta, A., Kao, M., Hu, C., & Roy, S. (2022). Artificial Neural Network Surrogate Models for Efficient Design Space Exploration of 14-nm FinFETs. Device Research Conference.

Chicago/Turabian Click to copy
Guglani, Surila, A. Dasgupta, M. Kao, Chenming Hu, and Sourajeet Roy. “Artificial Neural Network Surrogate Models for Efficient Design Space Exploration of 14-Nm FinFETs.” Device Research Conference (2022).

MLA Click to copy
Guglani, Surila, et al. “Artificial Neural Network Surrogate Models for Efficient Design Space Exploration of 14-Nm FinFETs.” Device Research Conference, 2022.

BibTeX Click to copy

@article{surila2022a,
  title = {Artificial Neural Network Surrogate Models for Efficient Design Space Exploration of 14-nm FinFETs},
  year = {2022},
  journal = {Device Research Conference},
  author = {Guglani, Surila and Dasgupta, A. and Kao, M. and Hu, Chenming and Roy, Sourajeet}
}

Abstract

For contemporary technology nodes, Fin Field Effect Transistors (FinFETs) as shown in Fig. 1 are considered to be the device of choice as they offer superior electrostatic control of the channel [1]. For design space explorations, device optimizations, and efficient circuit designs of FinFETs, we rely on various mathematical models ranging from Technology Computer-Aided Design tools (TCAD) which are based on accurate device physics but are computationally expensive to solve, to compact models [2], which prioritize localized accuracy and computational efficiency over high generalizability and predictive ability. For the high accuracy and predictability required for proper design optimizations, TCAD is used as the tool of choice. However, the high computational cost associated with the large number of TCAD simulations required for parametric sweeps is a major bottleneck. Here, we present a novel methodology using artificial neural network (ANN) based surrogate models that meets both the criteria of numerical efficiency and predictive accuracy simultaneously.