One-step variation included compact modeling with conditional variational autoencoder

•A machine learning assisted compact modeling approach is proposed to include variation in one step by conditional variational autoencoder.•Simplicity and high accuracy are achieved and statistical results generated by the model can reflect the actual situation.•The trained model is compatible with...

Full description

Saved in:
Bibliographic Details
Published in:Solid-state electronics Vol. 227; p. 109119
Main Authors: Wang, Shuhan, Zhou, Zheng, Tang, Zili, Xu, Jinghan, Liu, Xiaoyan, Zhang, Xing
Format: Journal Article
Language:English
Published: Elsevier Ltd 01.08.2025
Subjects:
Compact model
Machine learning
Variation modeling
Variational autoencoder
Variational autoencoder
Variation modeling
Compact model
Machine learning
ISSN:0038-1101
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•A machine learning assisted compact modeling approach is proposed to include variation in one step by conditional variational autoencoder.•Simplicity and high accuracy are achieved and statistical results generated by the model can reflect the actual situation.•The trained model is compatible with SPICE, thus can simulate variation in circuits. Efficient and accurate variation modeling serves as a critical part in circuit evaluation, which can reproduce actual electrical behavior of semiconductor devices. Conventional variation modeling usually consists of two steps: compact modeling the basic electrical properties and sub-modeling the variation sources introduced in MOSFET manufacturing process, mostly structural and doping parameters. This lengthy process results in a gap between device production and rapid circuit analysis. In order to improve modeling efficiency, in this work, we propose a one-step variation-included compact modeling approach leveraging machine learning. Utilizing conditional variational autoencoder (cVAE), currents with variation are directly constructed without the sub-modeling step, as variation sources of the cVAE model are automatically extracted. Benchmark against prior distribution of dataset generated by Monte Carlo simulation of BSIM-CMG, normalized variation in figure of merits (FoMs) of cVAE generated I-V curves are all above 0.9. After implementing the model in SPICE, high accuracy in circuit-level variation modeling also indicates the potential of the proposed model.
ISSN:0038-1101
DOI:10.1016/j.sse.2025.109119