CN115688611A - Small space model real-time training method based on semiconductor device structure - Google Patents
Small space model real-time training method based on semiconductor device structure Download PDFInfo
- Publication number
- CN115688611A CN115688611A CN202211703517.8A CN202211703517A CN115688611A CN 115688611 A CN115688611 A CN 115688611A CN 202211703517 A CN202211703517 A CN 202211703517A CN 115688611 A CN115688611 A CN 115688611A
- Authority
- CN
- China
- Prior art keywords
- semiconductor device
- device structure
- space
- model
- small space
- 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
Links
Images
Landscapes
- Testing Or Measuring Of Semiconductors Or The Like (AREA)
Abstract
The invention discloses a small space model real-time training method based on a semiconductor device structure, which comprises the following steps of S1: determining a target semiconductor device structure on a user interface, inputting the target semiconductor device structure into a cloud database, and acquiring the numerical value of each parameter of the target semiconductor device structure asX(ii) a S2: setting a self-adaptive threshold based on the similarity, and automatically screening a small space training sample data set in a threshold space in a cloud database; s3: transmitting the small-space sample data set to edge computing equipment, and constructing a prediction model from the structure to the electrical performance in real time on the edge computing equipment by using a machine learning algorithm; s4: structure for storing edge terminalXInputting the model in the step 3, and outputting the electrical property of the device. The invention selects a small space based on the structure to be detected, constructs in real time through edge calculation and further constructs a model by utilizing a machine learning algorithm, and can realize the prediction of the device structure to the electrical performance. The method not only can build the model quicklyThe speed is fast, and the precision is higher.
Description
Technical Field
The invention belongs to the field of semiconductor device electrical characteristic prediction, and particularly relates to a small space model real-time training method based on a semiconductor device structure.
Background
Semiconductor devices are widely used in the fields of household devices, power supplies, lighting, communication, automotive electronics and the like, and the prediction of electrical properties of the semiconductor devices is an important link in device design. At present, the acquisition of the electrical performance of the semiconductor device mainly depends on a Computer Aided Design tool (Technology Computer Aided Design, abbreviated as TCAD), which mainly solves a semiconductor equation set by a finite element/finite difference method to acquire the electrical performance of the device, but has the problems of long simulation time, poor convergence, and the like. In recent years, an intelligent prediction model is established by a machine learning algorithm through learning the relation between input structure parameters and output electrical properties, and the model established by the method is high in precision and speed and completely free of convergence problems. For machine learning models, the quality of the data is the basis for high precision model construction. The current method is usually based on a large amount of data for establishing the model, and the following defects still need to be improved:
1. for the construction process of the machine learning model, all data are generally adopted for model training, the training period of the model is long, and the complexity of the model is high.
2. Due to the fact that the data coverage range is too large, the problem that the local test accuracy is not high possibly exists for a specific test structure.
Disclosure of Invention
The invention aims to: the invention aims to provide a small space model real-time training method based on a semiconductor device structure.
The technical scheme is as follows: the invention discloses a small space model real-time training method based on a semiconductor device structure, which comprises the following steps of:
step 1: determining a target semiconductor device structure, inputting the target semiconductor device structure into an edge database and a cloud database, and acquiring the numerical value of each parameter of the target semiconductor device structure as
；
And 2, step: based on the similarity, the value of each parameter according to the target semiconductor device structureXPresetting a threshold, and screening a small space training sample data set within the threshold from a cloud database;
and step 3: transmitting the small-space sample data set to edge computing equipment, and establishing a machine learning prediction model based on the small-space sample data set by using a machine learning algorithm;
and 4, step 4: the value of each parameter of the target semiconductor device structure in the step 1 is measuredXThe machine learning predictive model in step 3 is input,
further, in step 1, the type of the target semiconductor device is an NMOS, PMOS, field effect transistor, high mobility transistor, IGBT, or FinFET nano device, and the parameters of the structure of the target semiconductor device include device structure parameters and process parameters.
Further, the step 2 specifically includes:
step 21: calculating the similarity gamma between the stored data points in the cloud database and the known structure through a similarity function;
step 22: setting a similarity threshold value delta and a small space training data volume threshold value N;
step 23: when gamma is larger than delta, acquiring the total amount M of data in the threshold space;
step 24: when M is larger than N, N data are selected from large to small, and the data are added into a small space training data set; and otherwise, taking the M data as small space training samples.
Further, in step 21, the similarity function is one of a gaussian kernel function, an euclidean distance, an included angle cosine, a manhattan distance, a chebyshev distance, a minkowski distance, a normalized euclidean distance, a mahalanobis distance, a hamming distance, a jackard similarity coefficient, a jackard distance, a correlation coefficient, a correlation distance, a lanchwise distance, a skew spatial distance, an exponential similarity coefficient, a non-parameterized similarity, or an information entropy.
Further, the similarity threshold δ includes a lower threshold
And an upper threshold
Obtaining sample data in an autofilter space
When the size of the design threshold is smaller than the sample space, it is a local region; when it is larger than the sample space, it covers the entire sample space.
Further, in step 3, the machine learning algorithm includes a supervised algorithm, a semi-supervised algorithm and an unsupervised learning algorithm.
Further, the supervision algorithm comprises one or a combination of Gaussian process regression, deep neural network, support vector machine, linear regression, logistic regression, lasso regression and CART regression tree.
Further, the establishing of the machine learning prediction model based on the small space sample data set through the supervision algorithm comprises the following steps:
step 31, determining a data set which accords with the input specification of a target semiconductor device model;
step 32, determining characteristic items related to the target in the data set;
step 33, performing data preprocessing on the feature items through a data preprocessing method, and then constructing a model input standard data set based on the feature items;
and step 34, constructing a machine learning prediction model between the characteristic items and the target value based on the standard data set.
Has the advantages that: compared with the prior art, the invention has the following remarkable advantages:
1. the invention constructs the machine learning prediction model for the sample data in a small range through edge calculation, has low requirement on calculation resources, and has high model training speed and high safety.
2. According to the invention, based on the device structure, the small sample in the threshold value space of the device structure is utilized to construct the real-time electrical performance prediction model, and the model construction accuracy is higher.
3. The invention provides a design threshold value, the space of model training is controlled by controlling the value of the threshold value, and the flexibility of model construction is higher.
Drawings
FIG. 1 is a flow chart of a method for training a small space model in real time based on a semiconductor device structure according to the present invention;
fig. 2 is a schematic structural diagram of a Silicon-on-insulator (SOI) Laterally Diffused Metal Oxide Semiconductor (LDMOS) on an insulating substrate according to an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and specific embodiments. And (3) constructing a typical example of the power device, namely an SOI LDMOS small-space electrical performance prediction model.
Fig. 2 shows a schematic structural diagram of the SOI LDMOS, where 1 is a heavily doped region having a second conductivity type, 2 is a top silicon drift region having the second conductivity type, 3 is a channel region having a first conductivity type, 4 is a heavily doped region having the second conductivity type, 5 is a heavily doped region having the first conductivity type, 6 is a buried insulating layer, and 7 is a substrate.
As shown in FIG. 1, the method for training a small space model based on a semiconductor device structure in real time comprises the following steps:
step 1: determining a target semiconductor device structure, inputting the target semiconductor device structure into a cloud database, and acquiring the numerical value of each parameter of the target semiconductor device structure as
And will beXStored in the edge database.
Step 2: based on the similarity, the value of each parameter according to the target semiconductor device structureXPresetting a threshold value, and screening a small space training sample data set within the threshold value from a cloud database;
and 3, step 3: transmitting the small space sample data set to edge computing equipment, and establishing a machine learning prediction model based on the small space sample data set by using a machine learning algorithm;
and 4, step 4: the value of each parameter of the target semiconductor device structure in the step 1 is measuredXInputting the machine learning prediction model in the step 3, obtaining the electrical performance of the target semiconductor device structure, and embodying the quality of the machine learning regression model according to the error of the prediction result.
Example 1
Taking the predicted breakdown voltage as an example for analysis, the small space modeling process of the invention is as follows:
step 1: user interface determining device structure setXThe range of the test set structure parameters includes the concentration of the drift regionN d Length of drift regionLTop silicon thicknesst s Thickness of buried oxide layert ox . Two sets of structural parameters are given in table 1. And transmitting the structure to an edge database and a cloud database.
TABLE 1 test device configuration parameter set and ranges thereof
And 2, step: and calculating the similarity gamma between the stored structure parameters of the cloud space and the structure to be simulated in the cloud database by using the Gaussian kernel function based on the structure of the device. Meanwhile, the threshold δ is set to 0.7 and the number threshold N is set to 30. When γ is larger than the threshold δ, i.e., 0.7, the data amount M thereof is acquired as 28. The data are sorted and selected from big to small, and the number of the data does not exceed 30, so that the selected 28 data are added into the small space training data set.
And 3, step 3: and (3) processing the small space training data in the step (2) into a csv file which can be input by a machine learning algorithm, and transmitting the csv file to the edge computing equipment.
And 4, step 4: screening out structural parameters which have main influence on the breakdown voltage of the device, wherein the structural parameters comprise the concentration of a drift region, the length of the drift region, the thickness of top silicon and the thickness of an embedded oxide layer;
and 5: and carrying out normalized data processing on the data, and constructing a machine learning prediction model of the breakdown voltage in real time through edge calculation by utilizing a Gaussian process regression algorithm.
And 6: and inputting the device structure set into a machine learning model of breakdown voltage of a small space for testing to obtain the testing precision.
Table 2 shows the deviations of the electrical property predictions in the whole space and in the real-time small space under the same test structure. As can be seen from Table 2, the prediction error of the breakdown voltage can be effectively reduced by adopting small-space data to construct a model.
TABLE 2 comparison of test deviations between the global space model and the small space model under the same test structure
Table 3 shows the memory occupied situation when the machine learning model is constructed using the whole space and the real-time small space under the same test structure. As can be seen from Table 3, the use of the memory can be relatively reduced by 95% by adopting the small-space data to construct the model.
TABLE 3 comparison of test deviations between the global space model and the small space model under the same test structure
Example 2:
by taking two electrical properties of predicted breakdown voltage and on-resistance as examples for analysis, the small space modeling process of the invention is as follows:
step 1: user interface determining device structure setXThe range of the test set structure parameters includes the concentration of the drift regionN d Length of drift regionLThickness of top silicont s Thickness of buried oxide layert ox . Table 4 gives the range of each structural parameter. And transmitting the structure to an edge database and a cloud database.
Table 4 test device configuration parameter set and its range
Step 2: in the cloud database, according to the structure of the device, setting a self-adaptive spatial lower threshold value
And an upper threshold
Thereby obtaining sample data in the autofilter space
Therefore, a small space training sample space can be obtained by screening, as shown in table 5.
TABLE 5 Small sample spatial data and its Range
And 3, step 3: and (3) processing the small space training data in the step (2) into a csv file which can be input by a machine learning algorithm, and transmitting the csv file to the edge computing equipment.
And 4, step 4: screening out structural parameters which mainly affect the breakdown voltage and the on-resistance of the device, wherein the structural parameters comprise the concentration of a drift region, the length of the drift region, the thickness of top silicon and the thickness of an embedded oxide layer;
and 5: and carrying out normalized data processing on the data, and constructing a machine learning prediction model of breakdown voltage and on-resistance in real time through edge calculation by utilizing Gaussian process regression.
And 6: and inputting the device structure set into a machine learning model of breakdown voltage and on-resistance in a small space for testing to obtain the testing precision.
Table 6 shows the deviations generated by predicting the electrical properties using the entire space and the real-time small space under the same test space device structure. As can be seen from table 6, when the model is constructed using the small-space data, the error of the model prediction is reduced for both the breakdown voltage and the on-resistance.
TABLE 6 comparison of test deviations between the model in the same test space and the model in the small space
Table 7 shows the situation that the machine learning model is constructed using the entire space and the real-time small space to occupy the memory under the same test space device structure. As can be seen from table 7, the memory usage can be relatively reduced by 95% by using small-space data for model building.
TABLE 7 comparison of test deviations between the model in the same test space and the model in the small space
In summary, the method for performing edge calculation real-time training of the small space model based on the device structure, provided by the invention, includes the steps of firstly selecting the small space data set in the set threshold space in the training data set according to the device structure, and then performing real-time construction of the machine learning model in real time by utilizing edge calculation, so that the electrical performance prediction model can be simply and quickly constructed, the reference information provided by small-range data is more accurate, and the interference generated by a large amount of data is avoided. In addition, the memory requirement in the model training process can be greatly reduced due to the reduction of the data volume.
Claims (8)
1. A small space model real-time training method based on a semiconductor device structure is characterized by comprising the following steps:
step 1: determining a target semiconductor device structure on a user interface, inputting the target semiconductor device structure into a cloud database, and acquiring the numerical value of each parameter of the target semiconductor device structure as
；
Step 2: giving a value based on similarityXPresetting a threshold value, and screening a small space training sample data set within the threshold value from a cloud database;
and 3, step 3: transmitting the small space training sample data set to edge computing equipment, and establishing a machine learning prediction model based on the small space sample data set by using a machine learning algorithm;
step 4, mixing the obtained product of step 1XInputting the machine learning prediction model in the step 3, and outputting the predicted electrical property.
2. The method as claimed in claim 1, wherein in step 1, the type of the target semiconductor device is NMOS, PMOS, fet, high mobility transistor, IGBT, or FinFET nano device, and the parameters of the target semiconductor device structure include device structure parameters and process parameters.
3. The method according to claim 1, wherein the step 2 specifically comprises:
step 21: calculating the similarity gamma between the stored data points in the cloud database and the known structure through a similarity function;
step 22: setting a similarity threshold value delta and a small space training data volume threshold value N;
step 23: when gamma is larger than delta, acquiring the total amount M of data in the threshold space;
step 24: when M is larger than N, selecting N data from big to small, and adding the data into a small space training data set; and otherwise, taking the M data as small space training samples.
4. The method as claimed in claim 3, wherein in step 21, the similarity function is one of a Gaussian kernel function, an Euclidean distance, an included angle cosine, a Manhattan distance, a Chebyshev distance, a Minkowski distance, a normalized Euclidean distance, a Mahalanobis distance, a Hamming distance, a Jacard similarity coefficient, a Jacard distance, a correlation coefficient, a correlation distance, a Langmuir distance, an oblique spatial distance, an exponential similarity coefficient, a non-parameterized similarity or an information entropy.
5. The method of claim 3, wherein the similarity threshold δ comprises a lower threshold
And upper threshold
Obtaining sample data in an autofilter space
When the size of the design threshold is smaller than the sample space, it is a local region; when it is larger than the sample space, it covers the entire sample space.
6. The method as claimed in claim 1, wherein the machine learning algorithm in step 3 comprises a supervised algorithm, a semi-supervised algorithm and an unsupervised learning algorithm.
7. The method of claim 6, wherein the supervised algorithm comprises one or a combination of Gaussian process regression, deep neural networks, support vector machines, linear regression, logistic regression, lasso regression, and CART regression trees.
8. The method of claim 6, wherein the step of building a small-space sample data set-based machine learning prediction model by a supervised algorithm comprises the steps of:
step 31, determining a data set which accords with the input specification of a target semiconductor device model;
step 32, determining characteristic items related to the target in the data set;
step 33, performing data preprocessing on the characteristic items by a data preprocessing method, and then constructing a model input standard data set based on the characteristic items;
and step 34, constructing a machine learning prediction model between the characteristic items and the target value based on the standard data set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211703517.8A CN115688611A (en) | 2022-12-29 | 2022-12-29 | Small space model real-time training method based on semiconductor device structure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211703517.8A CN115688611A (en) | 2022-12-29 | 2022-12-29 | Small space model real-time training method based on semiconductor device structure |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115688611A true CN115688611A (en) | 2023-02-03 |
Family
ID=85055412
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211703517.8A Pending CN115688611A (en) | 2022-12-29 | 2022-12-29 | Small space model real-time training method based on semiconductor device structure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115688611A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108427992A (en) * | 2018-03-16 | 2018-08-21 | 济南飞象信息科技有限公司 | A kind of machine learning training system and method based on edge cloud computing |
| WO2020249125A1 (en) * | 2019-06-14 | 2020-12-17 | 第四范式（北京）技术有限公司 | Method and system for automatically training machine learning model |
| WO2021262140A1 (en) * | 2020-06-22 | 2021-12-30 | Hewlett-Packard Development Company, L.P. | Machine learning model training |
| CN114462623A (en) * | 2022-02-10 | 2022-05-10 | 电子科技大学 | Data analysis method, system and platform based on edge calculation |
| CN114997092A (en) * | 2022-06-16 | 2022-09-02 | 南京邮电大学 | Machine learning and model-based semiconductor device electrical characteristic simulation method |
| CN115099126A (en) * | 2022-05-24 | 2022-09-23 | 南京邮电大学 | Method for automatically screening performance-related characteristic structure parameters of semiconductor device based on machine learning |
-
2022
- 2022-12-29 CN CN202211703517.8A patent/CN115688611A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108427992A (en) * | 2018-03-16 | 2018-08-21 | 济南飞象信息科技有限公司 | A kind of machine learning training system and method based on edge cloud computing |
| WO2020249125A1 (en) * | 2019-06-14 | 2020-12-17 | 第四范式（北京）技术有限公司 | Method and system for automatically training machine learning model |
| WO2021262140A1 (en) * | 2020-06-22 | 2021-12-30 | Hewlett-Packard Development Company, L.P. | Machine learning model training |
| CN114462623A (en) * | 2022-02-10 | 2022-05-10 | 电子科技大学 | Data analysis method, system and platform based on edge calculation |
| CN115099126A (en) * | 2022-05-24 | 2022-09-23 | 南京邮电大学 | Method for automatically screening performance-related characteristic structure parameters of semiconductor device based on machine learning |
| CN114997092A (en) * | 2022-06-16 | 2022-09-02 | 南京邮电大学 | Machine learning and model-based semiconductor device electrical characteristic simulation method |
Non-Patent Citations (2)
| Title |
|---|
| 刘峡壁等: "《人工智能-机器学习与神经网络》", vol. 1, 武汉大学出版社, pages: 330 - 331 * |
| 李丹丹;姚淑珍;王颖;王森章;谭火彬;: "基于半监督集成学习的多核设计空间探索", no. 04 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Encoder-X: solving unknown coefficients automatically in polynomial fitting by using an autoencoder | |
| Dhillon et al. | TCAD-augmented machine learning with and without domain expertise | |
| CN108141645A (en) | Video emphasis detection with pairs of depth ordering | |
| WO2021103761A1 (en) | Compound property analysis method and apparatus, compound property analysis model training method, and storage medium | |
| Guo | Local entropy map: A nonparametric approach to detecting spatially varying multivariate relationships | |
| AU2021245165A1 (en) | Method and device for processing quantum data | |
| CN109901058B (en) | Analysis method of semiconductor device | |
| Arif et al. | Cascade quality prediction method using multiple PCA+ ID3 for multi-stage manufacturing system | |
| Laksaci et al. | Spatial modelization: local linear estimation of the conditional distribution for functional data | |
| US11829893B2 (en) | Analysis method for semiconductor device | |
| Varghese et al. | Transmission line parameter estimation under non-Gaussian measurement noise | |
| Butola et al. | A machine learning approach to modeling intrinsic parameter fluctuation of gate-all-around si nanosheet MOSFETs | |
| CN113536661B (en) | TFET device structure optimization and performance prediction method based on neural network | |
| Pan et al. | An inferential modeling method using enumerative PLS based nonnegative garrote regression | |
| CN115688611A (en) | Small space model real-time training method based on semiconductor device structure | |
| Chavez et al. | Optimization of deep learning-based bsim-cmg iv parameter extraction in seconds | |
| Ibáñez et al. | A comprehensive characterization of the threshold voltage extraction in MOSFETs transistors based on smoothing splines | |
| Guglani et al. | Artificial neural network surrogate models for efficient design space exploration of 14-nm FinFETs | |
| US20230281363A1 (en) | Optimal materials and devices design using artificial intelligence | |
| Batsidis et al. | A necessary power divergence type family tests of multivariate normality | |
| JP5839922B2 (en) | Surface potential simulation apparatus and surface potential simulation program | |
| Shu et al. | Computation of the run-length percentiles of CUSUM control charts under changes in variances | |
| Gustafsson | On marginal particle filters with linear complexity | |
| Romero-Medrano et al. | Multinomial sampling for hierarchical change-point detection | |
| Mehmani et al. | Variable-fidelity optimization with in-situ surrogate model refinement |
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 |