Advanced MOSFET Model Based on Artificial Neural Network

JH. Wei1, W. Mao 1, H. Fang 2, Z. Zhang 2, JX. Zhang 2, BJ. Lan 2 and J. Wan 1*
1 State key lab of ASIC and System, School of Information Science and Engineering, Fudan

University, Shanghai, China

2 Suzhou Foohu Technology Co., Ltd.

* Corresponding Author’s Email: jingwan@fudan.edu.cn

ABSTRACT advanced MOSFET is difficult and time consuming.

In this work, we develop a novel MOSFET model for Besides, novel devices with totally different device
circuit simulation purpose. Instead of using traditional physics need enormous endeavor and good understanding
physics-driven model, such as BSIM model, our work on its physics. Thus, models based purely on data and
uses ANN to model the electrical behavior of the involves no device physics have been attracting lots of
transistor. With unique pre and post-processing interests[6], mainly due to their precision and ease of
procedures, the ANN is trained to model the drain current development[7]. Artificial neural network, which is used
precisely under various applied voltage, device size and in many different fields[8][9], has also been used in
temperature. The model is further successfully MOSFET modeling, which shows good precision
implemented in SPICE through Verilog-A language. Both [10][11]. In this work, we develop a data-driven
n-type and p-type MOSFETs show good fitting between MOSFET model based on ANN. With pre and post
the BSIM and ANN models. Eventually, an inverter and processing on the training data, the model precisely
ring oscillator based on ANN model are demonstrated models the relation between the output drain current and
with static and transient simulations, showing good six variants, including applied voltage on four terminals,
agreement with the results from BSIM model. gate length and width and the operating temperature. The
model is further plugged in Cadence through the
INTRODUCTION Verilog-A code. The simulation results using conventional
The integrated circuit (IC) has been developing BSIM model and ANN model are compared on MOSFET,
rapidly in last decades following the Moore’s law [1]. It inverter and ring oscillator, all shows very good
has been playing important role in modern information agreement.
technology fields. As the Moore’s law proceeds, the
transistor in the IC continues scaling down. In order to DEVELOPMENT OF ANN MODEL
boost the transistor performance in small size, many new Figure 1 shows the procedure of developing the ANN
technologies have been used, such as high-k/metal gate, model and implementation of it in Cadence for circuit
FinFET, FD-SOI, nanowire and nanosheet. As the simulation. A complete six parameters are used as input,
structure of the MOSFET evolves, modeling of the including gate voltage (VG ), drain voltage (VD ), bulk
MOSFET with new elements are constantly needed. voltage (VB), gate length (LG ), width (WG ) and
Initial level 1 model of the MOSFET based on an ideal temperature. The input parameters are pre-processed
device physics has long been outdated. More precise before they are fed into ANN whose output is
BSIM models, such as BISM 3 and BSIM 4, have been post-processed in order to obtain the drain current (I D ).
developed for advanced technology nodes [2]. There are The training data, which includes I D values under various
also special models developed for FinFET and FD-SOI six parameters, is generated from SPICE simulation. The
technologies [3]. Besides, many novel semiconductor ANN has three layers, each of which has 32 neurons. The
devices emerge recent years as complement or training of the model takes place in python.
replacement to conventional MOSFET, such as tunneling
field effect transistor (TFET) and negative capacitance
FET (NC-FET) [4]. In order to be used in SPICE
simulation, compact models are needed for these devices.
(such as NC-FET, TFET) without studying their complex
physical characteristics [5].
Conventional model is mainly based on device
physics. However, as the MOSFET structure becomes
more and more complicated, developing models for

Authorized licensed use limited to: UNIVERSITY OF BATH. Downloaded on May 14,2021 at 04:48:43 UTC from IEEE Xplore. Restrictions apply.
Figure1. A schematic view showing the development of model (0.17mA). Figure 2 (c) and (d) compare the
ANN model and its implementation in Cadence for circuit simulation results on the pFET between BSIM and ANN
simulation models, which also show good fitting. This fully
After trained, the ANN model is translated into demonstrates the precisely fitting capability of our
Verilog-A language which can be executed in Cadence, as developed ANN model in the transistor level.
shown schematically in Fig. 1. For simplicity, the BSIM
charge model is preserved in AC simulation, and only I D is CIRCUIT SIMULATION BASED ON ANN
replaced by the ANN model. Both n-type and p-type MODEL
transistor models are obtained in such similar way.
The ANN models developed for n-type and p-type
Figure 2 shows the simulation results of a n-type
MOSFETs are further used for circuit simulation in
MOSFET with conventional BSIM model and ANN model
SPICE. Both inverter and ring oscillator are simulated to
under a temperature of 27o C. The simulated MOSFET has
demonstrate the application of ANN model in circuit
LG = 1μm and WG = 1μm. Fig. 2(a) compares the transfer
simulation. Figure 3(a) shows the simulated inverter
(ID -VG ) characteristics of the MOSFET with three
circuit with a supply voltage of 5V. The pull-up transistor
different models under a fixed VD =1V, BSIM 3, ANN
(pFET) has LG = 1μm and WG = 2.5μm, and the
without pre/post-process and ANN with pre/post process.
pull-down transistor (nFET) has LG = 1μm and WG = 1μm.
It is clearly observed that the ANN model with the
Figure 3(b) shows the simulated voltage transfer
pre/post process fits the BSIM 3 model very well in both
characteristics of this inverter. The simulation results
logarithm and linear scales. However, without pre/post
with BSIM model and ANN model show exactly the same
process, the ANN model fails to fit the current under low
voltage transfer characteristics, as the input voltage (Vin)
VG when the transistor is in subthreshold region. Thus,
sweeps from 0V to 5V. The switching point is also
the pre/post process is mandatory here to have good
precisely predicted by the ANN model.
fitting in large dynamic range of I D .
The same nFET and pFET are used to form a ring
oscillator which consists of three inverters. Loading
capacitors with capacitance of 1pF are placed at the
output port of each inverter. The transient simulation is
performed to obtain the output waveform of the ring

Figure 2. (a) Comparison of I D -V G characteristics on

nFET. (b) ID -V D characteristics between ANN model and
BSIM model on a nFET. (c) Comparison of I D -V G
characteristics on pFET. (d) ID -V D characteristics
between ANN model and BSIM model on a pFET.

Figure 2(b) shows the output characteristics (ID -VD )

of the n-type transistors under various VG values. The
ANN model captures the trend very well in both linear and
saturation regions with neglected error. For example,
when VGS = 3V, the saturated ID value of ANN model is
0.168mA, which is very close to that obtained by BSIM 3 Figure 3. (a) Schematic view of the simulated inverter

Authorized licensed use limited to: UNIVERSITY OF BATH. Downloaded on May 14,2021 at 04:48:43 UTC from IEEE Xplore. Restrictions apply.
structure. (b) Comparison of voltage transfer [3] J. Song, Y. Yuan, B. Yu, W. Xiong , Y. Taur,
characteristics of the inverter between the ANN model and "Compact modeling of experimental n- and
the BSIM model. (c)Comparison of ring oscillator output p-channel FinFets ", IEEE Transactions on Electron
characteristics between ANN model and transistor model. Devices, vol. 57, no. 6, pp. 1369-1374, Jun 2010.
[4] Yunpeng Dong, Lining Zhang, Xiangbin Li, Xinnan
oscillator. Figure 3(c) compares the output waveform of Lin, Mansun Chan, "A Compact Model for
the ring oscillator with BSIM and ANN models. The Double-Gate Heterojunction Tunnel FETs", Electron
oscillating curve with ANN model fits the BSIM model Devices IEEE Transactions on, vol. 63, no. 11, pp.
very well. The oscillating frequency of ANN model is 4506-4513, 2016.
15.810MHZ, while that of transistor model is 15.873MHZ. [5] C. Enz , Y. Cheng, "MOS transistor modeling for RF
The relative error is only 0.39%. IC design", Solid-State Circuits IEEE Journal of ,
vol. 35, no. 2, pp. 186-201, 2000.
CONCLUSION [6] Lining Zhang, Mansun Chan, "Artificial neural
We have successfully developed an ANN-based network design for compact modeling of generic
MOSFET model. With unique pre and post process transistors", Journal of Computational Electronics,
procedures , the transfer and output characteristics of 2017.
both n-type and p-type MOSFETs show good fitting [7] Yazi Cao, Xi Chen, and Gaofeng Wang, " Dynamic
between BSIM and ANN models. The ANN model is Behavioral Modeling of Nonlinear Microwave
further implemented in Verilog-A and used in circuit Devices Using real-Time Recurrent neural
simulation. The simulated inverter and ring oscillator with Netwrok," IEEE Transactions on Electron Devices,
ANN model show the same voltage transfer Vol. 56, No. 5, pp. 1020-1026, May 2009.
characteristics and output waveform as those with the [8] M. H. Weatherspoon, H. A. Martinez, D. Langoni, S.
BSIM model. Since the developed ANN model is purely Y. Foo, "Small-signal modeling of microwave
based on data without involving any device physics, it has MESFETs using RBF-ANNs", IEEE Trans. Instrum.
not only high fitting precision but also great potentiality in Meas., vol. 56, no. 5, pp. 2067-2072, Oct. 2007.
the modeling of advanced MOSFET and novel [9] J. Xu, D. Gunyan, M. Iwamoto, A. Cognata, D.
semiconductor devices. Root, "Measurement-based non-quasi-static
large-signal FET model using artificial neural
ACKNOWLEDGEMENTS networks", IEEE MTT-S Int. Microw. Symp. Dig.,
pp. 469-472, 2006.
The work at Fudan University is sponsored by
[10] Litovski, J.I. Radjenovic, Z.M. Mrcarica, S.L.
National Natural Science Foundation of China
Milenkovic, "MOS transistor modelling using neural
(61904032).
network", Electron. Lett., vol. 28, no. 18, pp.
1766-1768, 1992.
REFERENCES [11] Youngseo Ko, Patrick Roblin, Andrés Zarate-De
[1] D. Root, "Future device modeling trends, " Landa, J. Apolinar Reynoso-Hernandez, Dan Nobbe,
Microwave Magazine, IEEE, vol. 13, no. 7, pp. Chris Olson, Francisco Javier Martinez, "Artificial
45-59, nov.-dec. 2012. Neural Network Model of SOS-MOSFETs Based on
[2] Y. Cheng, M.-C. Jeng, Z. Liu, J. Huang, M. Chan, K.
Dynamic Large-Signal Measurements", IEEE T
Chen, P. K. Ko, C. Hu, " A physical and scalable I-V Transactions On Microwave Theory and Techniques,
model in BSIM3v3 for analog/ digital circuit vol. 62, no. 3, marchal 2014.
simulation ", IEEE Trans. Electron Devices, vol. 44,
pp. 277-287, Feb. 1997.

Authorized licensed use limited to: UNIVERSITY OF BATH. Downloaded on May 14,2021 at 04:48:43 UTC from IEEE Xplore. Restrictions apply.