CN116151174A - General device model optimization method and system - Google Patents

General device model optimization method and system Download PDF

Info

Publication number
CN116151174A
CN116151174A CN202310397733.2A CN202310397733A CN116151174A CN 116151174 A CN116151174 A CN 116151174A CN 202310397733 A CN202310397733 A CN 202310397733A CN 116151174 A CN116151174 A CN 116151174A
Authority
CN
China
Prior art keywords
model
training
data
current
source voltage
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310397733.2A
Other languages
Chinese (zh)
Inventor
黄洪云
蒲朝斌
李林保
刁龙平
李林
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sichuan Huadun Defense Technology Co ltd
Original Assignee
Sichuan Huadun Defense Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sichuan Huadun Defense Technology Co ltd filed Critical Sichuan Huadun Defense Technology Co ltd
Priority to CN202310397733.2A priority Critical patent/CN116151174A/en
Publication of CN116151174A publication Critical patent/CN116151174A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/30—Circuit design
    • G06F30/36—Circuit design at the analogue level
    • G06F30/373—Design optimisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/30—Circuit design
    • G06F30/36—Circuit design at the analogue level
    • G06F30/367—Design verification, e.g. using simulation, simulation program with integrated circuit emphasis [SPICE], direct methods or relaxation methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Hardware Design (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Geometry (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Microelectronics & Electronic Packaging (AREA)
  • Liquid Crystal Display Device Control (AREA)

Abstract

The invention discloses a general device model optimization method and a general device model optimization system, which relate to the technical field of semiconductor device modeling, and are characterized in that firstly, device parameters and voltage bias are subjected to data preprocessing; secondly, constructing a neural network model by adopting a mean square relative error and a multi-objective loss function; and finally, randomly selecting data under a plurality of groups of different parameter combinations from the data, performing preliminary training on the model, and importing all the data into the neural network model on the basis of the preliminary training model to perform training on the neural network. The invention integrates the methods of data normalization processing, multi-objective loss function, model pre-training and the like, and the model obtained by the method can ensure the current precision of the device under various parameter combinations and bias voltage and can effectively shorten the training time.

Description

General device model optimization method and system
Technical Field
The invention relates to the technical field of semiconductor device modeling, in particular to a general device model optimization method and system.
Background
Currently, TCAD simulation plays a very important role in semiconductor process development. However, TCAD simulation is slow, and when the process parameter combination is optimized, the required simulation times are large, so that the process development efficiency is reduced. According to the historical data of TCAD simulation, a general model from the process parameters to the electrical characteristics of the device is established, and the TCAD simulator is replaced for calculation, so that the optimization efficiency of the process parameters can be improved. In order to overcome this problem of TCAD simulation, different machine learning algorithms have been used for device simulation, and the algorithms can accurately simulate the results of device simulation using the provided training data. Most importantly, the algorithm is provided with predictability, namely, the result of TCAD simulation of the device outside the provided parameter range can be predicted, so that the algorithm can be combined with the device simulation, and a predictable model is built by using a certain amount of TCAD simulation data through the algorithm, so that the time required by a large amount of TCAD simulation can be reduced.
The machine learning used by Kashyap Mehta, hiu-Y ung Wong et al in 2021 to predict FinFET current-voltage and capacitance-voltage curves demonstrates the possibility of predicting generic device IV and CV curves by machine training using limited training data (groups 25-50).
In 2021 Chandni Akbar, yiming Li et al proposed a Machine Learning (ML) assisted simulation of three-dimensional multi-channel gate all-silicon nanoflake MOSFET work function fluctuation devices. The proposed ML-RFR algorithm for predicting ID-VG curves shows the same accuracy as the device simulation.
2022 r. Butola, y.li and s.r. Kola propose a machine learning based approach to modeling the internal parameters of all silicon nanoplatelet MOSFETs, which results indicate that the output of the proposed model predictions has an R2 score of 99% and an error rate of less than 1%.
It can be seen that machine learning has been used to build device models instead of TCAD simulation, saving significant TCAD simulation time. However, the model precision, training time and the universality of the modeling algorithm have great optimization and promotion space.
Disclosure of Invention
In view of the above, the invention provides a general device model optimization method and system, so as to solve the problems of model precision, model training time and model universality existing in the device model establishment process.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
in one aspect, a method for optimizing a generic device model is provided, comprising the steps of:
step 1, preprocessing data of device parameters and voltage bias;
step 2, constructing a neural network model by adopting a mean square relative error and a multi-objective loss function;
and 3, randomly selecting data under a plurality of groups of different parameter combinations from the data, performing preliminary training on the model, and importing all the data into the neural network model on the basis of the preliminary training model to perform training on the neural network.
Optionally, the preprocessing specifically includes:
firstly, normalizing the structural parameters of a device;
secondly, performing conversion function processing on the current;
and finally, performing smoothing function processing on the gate-source voltage and the drain-source voltage.
Optionally, the formula of the normalization process is:
Figure SMS_1
wherein ,
Figure SMS_2
for normalized structural parameters, ++>
Figure SMS_3
Is a structural parameter.
Optionally, the conversion function is:
Figure SMS_4
；
wherein ,
Figure SMS_5
for outputting current, +.>
Figure SMS_6
For the drain-source voltage to be applied,yis->
Figure SMS_7
And converting corresponding data.
Optionally, the smoothing functions of the gate-source voltage and the drain-source voltage are respectively:
Figure SMS_8
；
Figure SMS_9
；
wherein ,
Figure SMS_10
is constant (I)>
Figure SMS_11
Drain-source voltage>
Figure SMS_12
Is gate-source voltage>
Figure SMS_13
、/>
Figure SMS_14
The drain-source voltage and the gate-source voltage after smoothing are respectively adopted.
Optionally, the mean square relative error is:
Figure SMS_15
；
wherein ,
Figure SMS_16
for training data volume, < >>
Figure SMS_17
For training data values, < >>
Figure SMS_18
For predicting data values.
Optionally, the multi-objective loss function is:
Figure SMS_19
；
wherein ,
Figure SMS_20
for the off-state current, +.>
Figure SMS_21
In the event of a saturated current flow,Mrepresenting current,/->
Figure SMS_22
Is the mean square relative error of the current, +.>
Figure SMS_23
Is the mean square relative error of the saturation current, +.>
Figure SMS_24
Is the mean square relative error of the off-state current,Ais a current weight coefficient,BIs a saturated current weight coefficient,CIs an off-state current weight coefficient,NsFor training data volume、iRepresent the firstiTraining samples.
In another aspect, a generic device model optimization system is provided, comprising the following modules:
the preprocessing module is used for preprocessing data of device parameters and voltage bias;
the neural network model construction module is used for constructing a neural network model by adopting a mean square relative error and a multi-objective loss function;
the training module of the neural network randomly selects data under a plurality of groups of different parameter combinations from the data, performs preliminary training on the model, and guides all the data into the neural network model on the basis of the preliminary training model to perform training on the neural network.
Optionally, the preprocessing module includes: and a conversion function unit converting the current data using the following formula:
Figure SMS_25
；
wherein ,
Figure SMS_26
for outputting current, +.>
Figure SMS_27
For the drain-source voltage to be applied,yis->
Figure SMS_28
The converted corresponding data;
the device also comprises a smoothing function unit which respectively processes the gate-source voltage and the drain-source voltage:
Figure SMS_29
；
Figure SMS_30
；
wherein ,
Figure SMS_31
is constant (I)>
Figure SMS_32
Drain-source voltage>
Figure SMS_33
Is gate-source voltage>
Figure SMS_34
、/>
Figure SMS_35
The drain-source voltage and the gate-source voltage after smoothing are respectively adopted.
Optionally, the neural network model building module comprises a mean square relative error calculation unit and a multi-objective loss function calculation unit;
the mean square relative error calculating unit is as follows:
Figure SMS_36
；
wherein ,
Figure SMS_37
for training data volume, < >>
Figure SMS_38
For training data values, < >>
Figure SMS_39
For predicting data values.
The multi-objective loss function calculation unit is as follows:
Figure SMS_40
；
wherein ,
Figure SMS_41
for the off-state current, +.>
Figure SMS_42
In the event of a saturated current flow,Mrepresenting current,/->
Figure SMS_43
Is the mean square relative error of the current, +.>
Figure SMS_44
Is the mean square relative error of the saturation current, +.>
Figure SMS_45
Is the mean square relative error of the off-state current,Ais a current weight coefficient,BIs a saturated current weight coefficient,CIs closed toA state current weight coefficient,NsFor training data volume、iRepresent the firstiTraining samples.
Compared with the prior art, the invention discloses a general device model optimization method and system, which integrate the methods of data normalization processing, multi-objective loss function, model pre-training and the like, and the model obtained by the algorithm can ensure the current precision of the device under various parameter combinations and bias voltage and can effectively shorten the training time.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only embodiments of the present invention, and that other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic diagram of a neural network according to embodiment 2 of the present invention;
fig. 2a is a diagram showing comparison between TCAD simulation and model prediction of the output characteristic curve of the MOSFET device with planar structure according to embodiment 3 of the present invention;
fig. 2b is a diagram showing a comparison between TCAD simulation and model prediction of a transfer characteristic curve of a MOSFET device with a planar structure according to embodiment 3 of the present invention;
fig. 3a is a diagram showing comparison between TCAD simulation and model prediction of the output characteristic curve of the MOSFET device with the ring gate structure according to embodiment 3 of the present invention;
fig. 3b is a diagram showing a comparison between TCAD simulation and model prediction of the transfer characteristic curve of the MOSFET device with the gate-all-around structure according to the embodiment 3 of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
On the one hand, embodiment 1 of the invention discloses a general device model optimization method, which comprises the following steps:
step 1, preprocessing data of device parameters and voltage bias;
step 2, constructing a neural network model by adopting a mean square relative error and a multi-objective loss function;
and 3, randomly selecting data under a plurality of groups of different parameter combinations from the data, performing preliminary training on the model, and importing all the data into the neural network model on the basis of the preliminary training model to perform training on the neural network.
In a specific embodiment, the pretreatment specifically comprises:
firstly, normalizing the structural parameters of a device;
secondly, performing conversion function processing on the current;
and finally, performing smoothing function processing on the gate-source voltage and the drain-source voltage.
In a specific embodiment, the normalization process is formulated as:
Figure SMS_46
wherein ,
Figure SMS_47
for normalized structural parameters, ++>
Figure SMS_48
Is a structural parameter.
In a specific embodiment, the transfer function is:
Figure SMS_49
；
wherein ,
Figure SMS_50
for outputting current, +.>
Figure SMS_51
For the drain-source voltage to be applied,yis->
Figure SMS_52
And converting corresponding data.
In a specific embodiment, the smoothing functions of the gate-source voltage and the drain-source voltage are respectively:
Figure SMS_53
；
Figure SMS_54
；
wherein ,
Figure SMS_55
is constant (I)>
Figure SMS_56
Drain-source voltage>
Figure SMS_57
Is gate-source voltage>
Figure SMS_58
、/>
Figure SMS_59
The drain-source voltage and the gate-source voltage after smoothing are respectively adopted.
In one specific embodiment, the mean square relative error is:
Figure SMS_60
；
wherein ,
Figure SMS_61
for training data volume, < >>
Figure SMS_62
For training data values, < >>
Figure SMS_63
For predicting data values.
In a specific embodiment, the multiple objective loss function is:
Figure SMS_64
；
wherein ,
Figure SMS_65
for the off-state current, +.>
Figure SMS_66
In the event of a saturated current flow,Mrepresenting current,/->
Figure SMS_67
Is the mean square relative error of the current, +.>
Figure SMS_68
Is the mean square relative error of the saturation current, +.>
Figure SMS_69
Is the mean square relative error of the off-state current,Ais a current weight coefficient,BIs a saturated current weight coefficient,CIs an off-state current weight coefficient,NsFor training data volume、iRepresent the firstiTraining samples.
In another aspect, a generic device model optimization system is provided, comprising the following modules:
the preprocessing module is used for preprocessing data of device parameters and voltage bias;
the neural network model construction module is used for constructing a neural network model by adopting a mean square relative error and a multi-objective loss function;
the training module of the neural network randomly selects data under a plurality of groups of different parameter combinations from the data, performs preliminary training on the model, and guides all the data into the neural network model on the basis of the preliminary training model to perform training on the neural network.
In a specific embodiment, the preprocessing module comprises: and a conversion function unit converting the current data using the following formula:
Figure SMS_70
；
wherein ,
Figure SMS_71
for outputting current, +.>
Figure SMS_72
For the drain-source voltage to be applied,yis->
Figure SMS_73
And converting corresponding data.
The device also comprises a smoothing function unit which respectively processes the gate-source voltage and the drain-source voltage:
Figure SMS_74
；
Figure SMS_75
；
wherein ,
Figure SMS_76
is constant (I)>
Figure SMS_77
Drain-source voltage>
Figure SMS_78
Is gate-source voltage>
Figure SMS_79
、/>
Figure SMS_80
The drain-source voltage and the gate-source voltage after smoothing are respectively adopted.
In a specific embodiment, the neural network model building module comprises a mean square relative error calculation unit and a multi-objective loss function calculation unit;
the mean square relative error calculating unit is as follows:
Figure SMS_81
；
wherein ,
Figure SMS_82
for training data volume, < >>
Figure SMS_83
For training data values, < >>
Figure SMS_84
For predicting data values.
The multi-objective loss function calculation unit is as follows:
Figure SMS_85
；
wherein ,
Figure SMS_86
for the off-state current, +.>
Figure SMS_87
In the event of a saturated current flow,Mrepresenting current,/->
Figure SMS_88
Is the mean square relative error of the current, +.>
Figure SMS_89
Is the mean square relative error of the saturation current, +.>
Figure SMS_90
Is the mean square relative error of off-state current, A is the current weight coefficient,BIs a saturated current weight coefficient,CIs an off-state current weight coefficient,NsFor training data volume、iRepresent the firstiTraining samples.
Example 2 is introduced for further explanation in order to further understand the technical scheme of the present invention.
On the one hand, embodiment 2 of the invention discloses a general device model optimization method, which comprises the following steps:
1. data processing
The data contains input parameters (device parameters and voltage bias), and for the device parameters, there may be a problem that the magnitude of the different device parameters is too different, and directly taking the data as the input of the model may cause the magnitude-less parameter influence to be ignored, and the magnitude-greater parameter influence to be amplified. Based on this, the present embodiment performs normalization processing on the device parameter data.
Figure SMS_91
wherein ,
Figure SMS_92
for normalized structural parameters, ++>
Figure SMS_93
Is a structural parameter.
Meanwhile, because the magnitude change of current data from a pinch-off region to a saturation region is large, the direct adoption of a data construction model can lead to low precision of the pinch-off region, so that data are required to be converted, the selection of a conversion function is important to the precision of the model, and the current data are processed as follows:
Figure SMS_94
；
wherein ,
Figure SMS_95
for outputting current, +.>
Figure SMS_96
For the drain-source voltage to be applied,yis->
Figure SMS_97
And converting corresponding data.
Taking into account that
Figure SMS_98
When this conversion is not possible, the invention will therefore +>
Figure SMS_99
And deleting the data. The conversion ensures->
Figure SMS_100
Time->
Figure SMS_101
So the pruning of data is reasonable. This way is achieved at->
Figure SMS_102
Limited by large variations in magnitudeyIs described.
The model is used for radio frequency distortion simulation to ensure IV curve thereof
Figure SMS_103
This is particularly important when passing Gummel test, which introduces +.>
Figure SMS_104
and />
Figure SMS_105
Is a voltage smoothing function of (a).
Figure SMS_106
；
Figure SMS_107
；
wherein ,
Figure SMS_108
is constant (I)>
Figure SMS_109
Drain-source voltage>
Figure SMS_110
Is gate-source voltage>
Figure SMS_111
、/>
Figure SMS_112
The drain-source voltage and the gate-source voltage after smoothing are respectively adopted.
2. Model construction
In order to realize high accuracy of the neural network model, a loss function and a network architecture adopted in the training of the neural network are important. The neural network structure adopted in this embodiment is shown in fig. 1. The MSE (mean square error) error function commonly used by the neural network has higher precision of a near pinch-off region but poor precision of a saturation region, and the current curve has unreasonable trend when the grid voltage is larger.
Therefore, the embodiment adopts the mean square relative error and considers the off-state current
Figure SMS_113
) And saturation current%
Figure SMS_114
) Precision requirements, using a multi-objective loss function:
Figure SMS_115
；
wherein ,
Figure SMS_116
for the off-state current, +.>
Figure SMS_117
In the event of a saturated current flow,Mrepresenting current,/->
Figure SMS_118
Is the mean square relative error of the current, +.>
Figure SMS_119
Is the mean square relative error of the saturation current, +.>
Figure SMS_120
Is the mean square relative error of the off-state current,Ais a current weight coefficient,BIs a saturated current weight coefficient,CIs an off-state current weight coefficient,NsFor training data volume、iRepresent the firstiTraining samples.
The mean square relative error is:
Figure SMS_121
；
wherein ,
Figure SMS_122
for training data volume, < >>
Figure SMS_123
For training data values, < >>
Figure SMS_124
For predicting data values. />
3. Model training
For training of the neural network, the initial weight and bias settings of the neural network are particularly important for the training duration, and generally the initial parameters of the neural network are set randomly, and then the loss function is continuously adjusted and reduced through the training process to find the figure of merit, but a long training time is caused in the case of a large amount of training data. Based on this, the embodiment adopts a model pre-training method, randomly extracts data under 50 groups of different parameter combinations, performs preliminary training on the model, has short training time due to small data volume, and finally performs model training on a large amount of data on the basis of the model.
Example 3 is introduced for further explanation in order to further understand the technical scheme of the present invention.
The embodiment 3 of the invention provides an optimization algorithm for an N-type MOSFET device model which can be used for a planar structure and a gate-all-around structure, and the model obtained by the algorithm is verified to show that:
comparison with TCAD simulation dataModel predictive current [ ]
Figure SMS_125
) Error less than 1%, saturation current (>
Figure SMS_126
) Error less than 3%, off-state current (+)>
Figure SMS_127
) Error less than 11%, threshold voltage (>
Figure SMS_128
) The error is less than 0.005V, and the accuracy of the algorithm and the universality of the model are verified.
The specific steps of this embodiment 3 include data preprocessing, model construction, and model prediction.
(1) And (5) data processing. According to the provided training data parameter combination file, corresponding voltage and current data under different device structure parameter combinations are searched out and are arranged into a data table, the front column corresponds to the device structure parameter, and the rear three columns respectively correspond to the grid source voltage
Figure SMS_129
Drain-source voltage->
Figure SMS_130
And output current +.>
Figure SMS_131
. Then, device structure parameter normalization processing and current logarithmization processing are carried out, and in order to make the model pass Gummel test, the method is carried out on +>
Figure SMS_132
and />
Figure SMS_133
And (5) performing smoothing function processing.
(2) And (5) constructing a model. According to the data table obtained in the first step, parameters are divided into input parameters and output parameters, firstly, pre-training of a model is carried out, data under 50 groups of different device parameter combinations are randomly extracted from the data, preliminary training of a neural network is carried out, then on the basis, all training data are imported into the model, training of the neural network is carried out, and the final model is obtained after the proposed multi-objective loss function based on relative errors is utilized for continuous iterative optimization until the loss is smaller than 0.00001.
(3) Model prediction. And according to the provided test data parameter combination file, corresponding voltage data under different device structure parameter combinations are searched out, and are used as input to finally obtain a prediction result of the model.
Example planar structure MOSFET device current output characteristics and transfer characteristics TCAD simulation and model prediction pairs such as shown in fig. 2a, 2b, and gate-all-around structure MOSFET device current output characteristics and transfer characteristics TCAD simulation and model prediction pairs such as shown in fig. 3a, 3 b.
The model training and testing time of the MOSFET device with the two structures of the plane and the ring gate in the embodiment are shown in the table 1, and the model precision index off-state current is shown in the specification
Figure SMS_134
) Saturation current (+)>
Figure SMS_135
) And threshold voltage->
Figure SMS_136
The specific parameter errors of (2) are shown in table 2.
Table 1 model training and test time for MOSFET devices of two structures, planar and circular gate
Device structure Training time Test time Total time of
Ring grid 762.62s 878.23s 1640.85s
Plane surface 2210.53s 1442.42s 3652.95s
Table 2 specific parameter error of model accuracy index of MOSFET device with two structures of plane and ring gate
Figure SMS_137
In summary, the present embodiment provides a TCAD-oriented fast and high-precision general device model algorithm, which solves the problems of model precision and long model training time by adopting a data processing, a multi-objective loss function based on relative errors, and a model pre-training method.
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. A method for optimizing a generic device model, comprising the steps of:
step 1, preprocessing data of device parameters and voltage bias;
step 2, constructing a neural network model by adopting a mean square relative error and a multi-objective loss function;
and 3, randomly selecting data under a plurality of groups of different parameter combinations from the data, performing preliminary training on the model, and importing all the data into the neural network model on the basis of the preliminary training model to perform training on the neural network.
2. The method for optimizing a generic device model according to claim 1, wherein the preprocessing specifically comprises:
firstly, normalizing the structural parameters of a device;
secondly, performing conversion function processing on the current;
and finally, performing smoothing function processing on the gate-source voltage and the drain-source voltage.
3. The method for optimizing a generic device model according to claim 2, wherein the normalization process is formulated as:
Figure QLYQS_1
wherein ,
Figure QLYQS_2
for normalized structural parameters, ++>
Figure QLYQS_3
Is a structural parameter.
4. A generic device model optimization method according to claim 2, characterized in that the transfer function is:
Figure QLYQS_4
；
wherein ,
Figure QLYQS_5
for outputting current, +.>
Figure QLYQS_6
For the drain-source voltage to be applied,yis->
Figure QLYQS_7
And converting corresponding data.
5. The method for optimizing a generic device model according to claim 2, wherein the smoothing functions of the gate-source voltage and the drain-source voltage are respectively:
Figure QLYQS_8
；
Figure QLYQS_9
；
wherein ,
Figure QLYQS_10
is constant (I)>
Figure QLYQS_11
Drain-source voltage>
Figure QLYQS_12
Is gate-source voltage>
Figure QLYQS_13
、/>
Figure QLYQS_14
The drain-source voltage and the gate-source voltage after smoothing are respectively adopted.
6. The method for optimizing a generic device model of claim 1, wherein the mean square relative error is:
Figure QLYQS_15
；
wherein ,
Figure QLYQS_16
for training data volume, < >>
Figure QLYQS_17
For training data values, < >>
Figure QLYQS_18
For predicting data values.
7. The method of claim 1, wherein the multi-objective loss function is:
Figure QLYQS_19
；
wherein ,
Figure QLYQS_20
for the off-state current, +.>
Figure QLYQS_21
In the event of a saturated current flow,Mrepresenting current,/->
Figure QLYQS_22
Is of current typeMean square relative error, +.>
Figure QLYQS_23
Is the mean square relative error of the saturation current, +.>
Figure QLYQS_24
Is the mean square relative error of the off-state current,Ais a current weight coefficient,BIs a saturated current weight coefficient,CIs an off-state current weight coefficient,NsFor training data volume、iRepresenting the ith training sample.
8. A generic device model optimization system comprising the following modules:
the preprocessing module is used for preprocessing data of device parameters and voltage bias;
the neural network model construction module is used for constructing a neural network model by adopting a mean square relative error and a multi-objective loss function;
the training module of the neural network randomly selects data under a plurality of groups of different parameter combinations from the data, performs preliminary training on the model, and guides all the data into the neural network model on the basis of the preliminary training model to perform training on the neural network.
9. The generic device model optimization system of claim 8, wherein the preprocessing module comprises: and a conversion function unit converting the current data using the following formula:
Figure QLYQS_25
；
wherein ,
Figure QLYQS_26
for outputting current, +.>
Figure QLYQS_27
For the drain-source voltage to be applied,yis->
Figure QLYQS_28
The converted corresponding data;
the device also comprises a smoothing function unit which respectively processes the gate-source voltage and the drain-source voltage:
Figure QLYQS_29
；
Figure QLYQS_30
；
wherein ,
Figure QLYQS_31
is constant (I)>
Figure QLYQS_32
Drain-source voltage>
Figure QLYQS_33
Is gate-source voltage>
Figure QLYQS_34
、/>
Figure QLYQS_35
The drain-source voltage and the gate-source voltage after smoothing are respectively adopted.
10. The general device model optimization system of claim 8, wherein the neural network model building module comprises a mean square relative error calculation unit and a multi-objective loss function calculation unit;
the mean square relative error calculating unit is as follows:
Figure QLYQS_36
；
wherein ,
Figure QLYQS_37
for training data volume, < >>
Figure QLYQS_38
For training data values, < >>
Figure QLYQS_39
Is a predicted data value;
the multi-objective loss function calculation unit is as follows:
Figure QLYQS_40
；
wherein ,
Figure QLYQS_41
for the off-state current, +.>
Figure QLYQS_42
In the event of a saturated current flow,Mrepresenting current,/->
Figure QLYQS_43
Is the mean square relative error of the current, +.>
Figure QLYQS_44
Is the mean square relative error of the saturation current, +.>
Figure QLYQS_45
Is the mean square relative error of off-state current, A is the current weight coefficient,BIs a saturated current weight coefficient,CIs an off-state current weight coefficient,NsFor training data volume、iRepresent the firstiTraining samples. />
CN202310397733.2A 2023-04-14 2023-04-14 General device model optimization method and system Pending CN116151174A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310397733.2A CN116151174A (en) 2023-04-14 2023-04-14 General device model optimization method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310397733.2A CN116151174A (en) 2023-04-14 2023-04-14 General device model optimization method and system

Publications (1)

Publication Number Publication Date
CN116151174A true CN116151174A (en) 2023-05-23

Family

ID=86356477

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310397733.2A Pending CN116151174A (en) 2023-04-14 2023-04-14 General device model optimization method and system

Country Status (1)

Country Link
CN (1) CN116151174A (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109791627A (en) * 2018-06-19 2019-05-21 香港应用科技研究院有限公司 Using input pretreatment and switch target for training the semiconductor devices of deep neural network to model
US20190385047A1 (en) * 2018-06-19 2019-12-19 Hong Kong Applied Science and Technology Research Institute Company, Limited Semiconductor Device Modeling Using Input Pre-Processing and Transformed Targets for Training a Deep Neural Network
US20220114317A1 (en) * 2020-10-13 2022-04-14 Samsung Electronics Co., Ltd. Systems, methods, and computer program products for transistor compact modeling using artificial neural networks
US20230025626A1 (en) * 2021-07-20 2023-01-26 Samsung Electronics Co., Ltd. Method and apparatus for generating process simulation models

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109791627A (en) * 2018-06-19 2019-05-21 香港应用科技研究院有限公司 Using input pretreatment and switch target for training the semiconductor devices of deep neural network to model
US20190385047A1 (en) * 2018-06-19 2019-12-19 Hong Kong Applied Science and Technology Research Institute Company, Limited Semiconductor Device Modeling Using Input Pre-Processing and Transformed Targets for Training a Deep Neural Network
US20220114317A1 (en) * 2020-10-13 2022-04-14 Samsung Electronics Co., Ltd. Systems, methods, and computer program products for transistor compact modeling using artificial neural networks
US20230025626A1 (en) * 2021-07-20 2023-01-26 Samsung Electronics Co., Ltd. Method and apparatus for generating process simulation models

Similar Documents

Publication Publication Date Title
LiKamWa et al. Redeye: analog convnet image sensor architecture for continuous mobile vision
CN103678941B (en) The Forecasting Methodology of electrode air gap breakdown voltage
CN101859383B (en) Hyperspectral remote sensing image band selection method based on time sequence important point analysis
US7761275B2 (en) Synthesizing current source driver model for analysis of cell characteristics
JP5006214B2 (en) Variation simulation system, variation determination model method and apparatus, and program
Choi et al. Neural approach for modeling and optimizing Si-MOSFET manufacturing
Akbar et al. Deep learning algorithms for the work function fluctuation of random nanosized metal grains on gate-all-around silicon nanowire MOSFETs
Aminzadeh Systematic circuit design and analysis using generalised gm/ID functions of MOS devices
Thakker et al. A novel table-based approach for design of FinFET circuits
CN111079361A (en) Load modeling method of FPGA circuit
CN112580288B (en) Semiconductor device characteristic modeling method and system based on multi-gradient neural network
CN116151174A (en) General device model optimization method and system
He et al. Analytic carrier-based charge and capacitance model for long-channel undoped surrounding-gate MOSFETs
CN111859627B (en) Parameter optimization method and device for component model
US20130125080A1 (en) Circuit optimization method and apparatus for analog circuit migration
Li A simulation-based evolutionary approach to LNA circuit design optimization
CN112685958A (en) SiC MOSFET blocking voltage determination method based on neural network
CN115271154B (en) Nonlinear regression flood element prediction method based on polynomial and partial least square coupling
CN112765933B (en) Integrated circuit design method based on population optimization algorithm
Zhou et al. Prediction of SET on SRAM based on WOA-BP neural network
Wang et al. Optimization and Performance Prediction of Tunnel Field‐Effect Transistors Based on Deep Learning
Wang et al. DC-Model: A New Method for Assisting the Analog Circuit Optimization
Li et al. Capacitance characteristic optimization of germanium MOSFETs with aluminum oxide by using a semiconductor-device-simulation-based multi-objective evolutionary algorithm method
Song et al. Hardware-aware neural architecture search for stochastic computing-based neural networks on tiny devices
Malik et al. Sparse regression driven mixture importance sampling for memory design

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination