[paper] Benchmarks for SPICE AI/ML Modeling

Colin C. McAndrew, Andries J. Scholten, Kiran K. Gullapalli, Yogesh Chauhan, and Kejun Xia

Benchmarks for SPICE Modeling and Parameter Extraction Based on AI/ML

IEEE Transactions on Electron Devices (2025)

DOI: 10.1109/TED.2025.3537952

1 Computer Engineering, Iowa State University, (USA)
2 NXP Semiconductors N.V., Eindhoven (NL)
3 NXP Semiconductors N.V., Austin (USA)
4 IIT Kanpur, (IN)
5 TSMC, Hsinchu (TW)

Abstract: Over the past decades, the number of submitted articles that use numerical approaches for SPICE models or for characterization (extraction) of parameters of existing SPICE models has grown significantly. Many of those articles rely on synthetic data, generated either from technology computer-aided design (TCAD) or from physical SPICE model simulations; most do not model/fit measured data. Furthermore, those articles do not evaluate the physical correctness, smoothness/monotonicity, or asymptotic correctness of the approach they propose. That is sufficient for initial evaluation of proposed techniques. However, it does not prove that they are “industrial strength.” This article presents benchmarks/guidelines for the proposed artificial intelligence (AI)/machine learning (ML) SPICE modeling and characterization techniques to try to help them become practical and useful.

TAB: CHECKLIST FOR MODELS

Capability	Existing State-of-the-Art	Proposed AI/ML Approach	Best Prior AI/ML Approach
obeys the laws of thermodynamics	✓	?	?
accurate DC modeling for all terminal currents, on relevant log/linear scale	✓	?	?
accurate capacitance/charge modeling	✓	?	?
models DC and capacitance interaction where relevant	✓	?	?
accurate modeling of high-frequency/non-quasi-static effects where relevant	✓	?	?
works for large-signal transient simulation, including delay effects	✓	?	?
accurate noise modeling	✓	?	?
has full geometry dependence	✓	?	?
has complete temperature dependence	✓	?	?
models all necessary LDEs	✓	?	?
behaves “well” for unreasonable geometry or temperature or bias	✓	?	?
exhibits physical monotonicity over bias, geometry, and temperature	✓	?	?
is smooth (ideally C∞-continuous)	✓	?	?
exhibits relevant physical symmetries (currents, charges, their derivatives)	✓	?	?
exhibits asymptotic correctness over geometry, temperature, and bias	✓	?	?
includes modeling of electrothermal effects (with frequency dependence)	✓	?	?
includes, or enables, modeling of global and local statistical variation	✓	?	?
includes, or enables, modeling of aging	✓	?	?
enables modeling of parasitics for different layouts	✓	?	?
is verified to converge reliably in at least one circuit simulator	✓	?	?

[paper] Fast and Expandable ANN-Based Extraction

Jeong, HyunJoon, SangMin Woo, JinYoung Choi, HyungMin Cho, Yohan Kim,

Jeong-Taek Kong, and SoYoung Kim

Fast and Expandable ANN-Based Compact Model and Parameter Extraction for Emerging Transistors IEEE Journal of the Electron Devices Society (2023)

DOI 10.1109/JEDS.2023.3246477

Abstract: In this paper, we present a fast and expandable artificial neural network (ANN)-based compact model and parameter extraction flow to replace the existing complicated compact model implementation and model parameter extraction (MPE) method. In addition to nanosheet FETs (NSFETs), our published ANN based compact modeling framework is easily extended to negative capacitance NSFETs (NC-NSFETs), which are attracting attention as next-generation devices. Each device is designed using a technology computer-aided design (TCAD) simulator. Using device structure parameters, temperature, and channel doping depth as input variables, we construct a dataset of electrical properties used for machine learning (ML)-based modeling. The accuracy of predicting device electrical characteristics with the proposed ANN-based compact model is less than a 1% error compared to TCAD, and simulation results of digital and analog circuits using the proposed compact model show less than a 3% error. This allows the ANN-based modeling framework to achieve accurate DC, AC, and transient simulations without restrictions on device technology. In particular, temperature and process variables such as channel doping depth, which are not defined in the compact model parameters, are easily added to the previously presented five key parameters. Instead of conventional complex compact modeling and MPE work, we propose a method to create fast, accurate, flexible, and expandable ML-based Verilog-A SPICE models with design technology co-optimization (DTCO) capabilities.

Fig A: Conventional model parameter extraction flow

Fig B: The proposed ANN-based model parameter extraction flow

Acknowledgments: We thank the reviewers for improving the contents of the paper. This work was supported by an Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No.2021-0- 00754, Software Systems for AI Semiconductor Design) and by a National Research Foundation of Korea grant funded by the Korean government (MISP) (NRF-2020R1A2C1011831). The EDA tool was supported by the IC Design Education Center (IDEC), Korea

[paper] Compact Models for Sizing Based on ANN

Husni Habal, Dobroslav Tsonev, Matthias Schweikardt 

Compact Models for Initial MOSFET Sizing Based on Higher-order Artificial Neural Networks

ACM/IEEE Workshop on Machine Learning for CAD (MLCAD ’20)

Nov. 16–20, 2020, Virtual Event, Iceland. ACM, pp. 111-116

DOI: 10.1145/3380446.3430632

¹Infineon Technologies AG Munich, Germany
²LogiqWorks Ltd. Sofia, Bulgaria
³Reutlingen University Reutlingen, Germany

Abstract: Simple MOSFET models intended for hand analysis are inaccurate in deep sub-micrometer process technologies and in the moderate inversion region of device operation. Accurate models, such as BSIM6 model, are too complex for use in hand analysis and are intended for circuit simulators. Artificial neural networks (ANNs) are efficient at capturing both linear and non-linear multivariate relationships. In this work, a straightforward modeling technique is presented using ANNs to replace the BSIM model equations. Existing open-source libraries are used to quickly build models with error rates generally below 3%. When combined with a novel approach, such as the gm/Id systematic design method, the presented models are sufficiently accurate for use in the initial sizing of analog circuit components without simulation.

Figure: ANN Model Architecture.