Electrostatic protection DTSCR device performance prediction method based on neural network

Abstract

The invention discloses an electrostatic protection DTSCR device performance prediction method based on a neural network. The problems that in the prior art, a traditional device simulation and experiment testing method is low in efficiency, time-consuming and labor-consuming are solved. The method comprises the steps that an electrostatic protection DTSCR device model is established, and an initial data set is obtained by simulating the electrostatic protection DTSCR device model; performing TLP simulation on the DTSCR device model and the initial data set to obtain an electrical characteristic curve and characteristic parameters, and establishing a data set according to the characteristic parameters; constructing a plurality of electrostatic protection DTSCR device performance prediction neural networks with the same framework and different parameters, and training to obtain a trained network; utilizing the trained network to predict characteristic parameters of a design window of the electrostatic protection DTSCR device; according to the method, high efficiency and short period are realized, and the association between the electrostatic protection DTSCR device and the performance index can be quickly and conveniently established.

Description

Technical Field

[0001] The present invention relates to the field of microelectronic devices and artificial intelligence technology, and in particular to a method for predicting the performance of an electrostatic protection DTSCR device based on a neural network. Background Art

[0002] The generation and discharge of static charge is a very common electrical phenomenon. However, when an electrostatic discharge event occurs on an integrated circuit chip, the chip is likely to be damaged by the ESD pulse due to the extremely small physical size of the semiconductor device. Therefore, ESD behavior will seriously deteriorate the reliability of IC products. Therefore, it is very necessary to carry out effective electrostatic protection for integrated circuit products.

[0003] SCR devices have been widely used and studied for electrostatic protection because of their large discharge current capacity per unit area and small area. As most of the working environment of chips has changed to a low-voltage environment, the ESD electrostatic protection design window has also been reduced to a low-voltage environment. The excessively high trigger voltage in traditional SCR devices limits their application in advanced low-voltage ESD protection projects. Among the many variants of optimizing the trigger voltage of SCR devices, the diode-triggered silicon-controlled rectifier (DTSCR) can achieve a very low and adjustable trigger voltage, and is therefore widely used in ESD protection of various advanced low-voltage integrated circuits. There are also many devices that have been improved according to the DTSCR device structure, so the method of predicting the performance of DTSCR devices has certain research representativeness and research significance.

[0004] At present, the design of DTSCR devices is mostly carried out through TCAD tools. First, the device model is established by numerical simulation tools, and then the electrical characteristics of the device are simulated. During the simulation, physical models such as drift diffusion transmission model, mobility model, and non-complete ionization model need to be considered. In addition, the grid establishment of the critical path also needs to consider the design to be denser, while the grids of other parts should be sparser. Only after a series of parameter and structure modifications can a more satisfactory device structure corresponding to the electrical characteristics be obtained. After completing the above work, the structural parameters and process parameters corresponding to the optimal device are extracted to reduce the high cost risk brought by subsequent tape-out.

[0005] When building a DTSCR device model based on TCAD, in order to obtain a more accurate and ideal structure, the computational grid needs to be set densely, but this will bring the disadvantage of long simulation time and high computational cost. When the required sample size is very large, the time will increase exponentially, greatly increasing the time cost of scientific research. In addition, the simulation of the electrical characteristics of DTSCR devices based on TCAD simulation software requires technicians to master certain semiconductor physics knowledge, and the learning cost will also increase.

[0006] In summary, the existing DTSCR device simulation methods are time-consuming, inefficient, and greatly affected by human factors. Neural networks provide another option for quickly predicting device performance. They can automatically extract features based on data, eliminating the high entry cost and interference from human factors, while ensuring accuracy, reducing computing costs and shortening development cycles. Summary of the invention

[0007] The present invention provides a method for predicting the performance of an electrostatic protection DTSCR device based on a neural network, thereby solving the problems of low efficiency, time-consuming and labor-intensive traditional device simulation and experimental testing methods in the prior art, achieving high efficiency and short cycle, and being able to quickly and conveniently establish a correlation between an electrostatic protection DTSCR device and performance indicators.

[0008] The present invention provides a method for predicting the performance of an electrostatic protection DTSCR device based on a neural network, the method comprising:

[0009] Establish an electrostatic protection DTSCR device model based on Sentaurus TCAD, and obtain an initial data set by simulating the electrostatic protection DTSCR device model;

[0010] By performing TLP simulation on the electrostatic protection DTSCR device model and the initial data set, an electrical characteristic curve is obtained, and parameter extraction is performed on the electrical characteristic curve to obtain characteristic parameters, and a data set is established according to the characteristic parameters;

[0011] Constructing multiple electrostatic protection DTSCR device performance prediction neural networks with the same framework and different parameters, training each of the electrostatic protection DTSCR device performance prediction neural networks according to the data set and the loss function, and obtaining multiple electrostatic protection DTSCR device performance prediction neural networks that have completed initial training;

[0012] Screening a plurality of electrostatic protection DTSCR device performance prediction neural networks that have been initially trained to obtain a trained electrostatic protection DTSCR device performance prediction neural network;

[0013] The trained electrostatic protection DTSCR device performance prediction neural network is used to predict the characteristic parameters of the design window of the electrostatic protection DTSCR device.

[0014] In a possible implementation, the step of obtaining an electrical characteristic curve by performing TLP simulation on the electrostatic protection DTSCR device model, extracting parameters from the electrical characteristic curve to obtain characteristic parameters, and establishing a data set according to the characteristic parameters includes:

[0015] Changing the structural parameters and process parameters of the electrostatic protection DTSCR device model to obtain an electrical characteristic curve of the DTSCR device;

[0016] Extracting characteristic parameters of the electrical characteristic curve and integrating the characteristic parameters into a data set;

[0017] The data set is preprocessed, and the preprocessed data set is divided into a training set, a test set, and a cross-validation set.

[0018] In a possible implementation, preprocessing the data set and dividing the preprocessed data set into a training set, a test set, and a cross-validation set includes:

[0019] Traverse each piece of data in the data set and perform the following operations on each piece of data:

[0020] Determine whether the data set meets the I / V characteristic curve standard. If not, remove the data; if yes, perform normalization on the data to obtain preprocessed data;

[0021] Each processed data is integrated into a processed data set, and the processed data set is divided according to the proportion to obtain a training set, a test set, and a cross-validation set.

[0022] In a possible implementation, each piece of data in the data set includes: 4 structural parameters and 4 characteristic parameters;

[0023] The four structural parameters include: the length of the shallow trench isolation between the P well / N+ active area and the N well / P+ active area of ​​the SCR part of the electrostatic protection DTSCR device model, the length of the shallow trench isolation between the N+ active area and the P+ active area in the same well of the SCR part, the length between the N well of the SCR part and the N well of the diode DIODE1 part, and the well spacing between the two diodes;

[0024] The four characteristic parameters include: a trigger voltage Vt, a trigger current It, a holding voltage Vh, and a holding current Ih.

[0025] In a possible implementation, the framework of the electrostatic protection DTSCR device performance prediction neural network includes an input layer, an input expansion layer, a convolution layer, a pooling layer, a fully connected layer and an output layer connected in sequence;

[0026] The input layer is used to receive input data;

[0027] The input expansion layer is used to expand the dimension of the input data to obtain the input data after the dimension is expanded;

[0028] The convolution layer is used to extract data features from the input data after the dimension expansion to obtain initial feature information;

[0029] The pooling layer is used to remove redundant information from the initial feature information to obtain feature information;

[0030] The fully connected layer is used to classify the feature information to obtain a classification result;

[0031] The output layer is used to output the classification result.

[0032] In a possible implementation, the convolution layer is a one-dimensional convolution layer, and the convolution kernel size is 3×1;

[0033] The pooling layer is a one-dimensional pooling layer, and the convolution kernel size is 2×1.

[0034] In a possible implementation, the loss function is expressed as:

[0035]

[0036] Where n represents the number of samples; y i represents the actual value of the i-th sample; Represents the predicted value of the i-th sample.

[0037] In a possible implementation, the frameworks of the multiple electrostatic protection DTSCR device performance prediction neural networks with the same framework and different parameters all include an input layer, an input expansion layer, a convolution layer, a pooling layer, a fully connected layer and an output layer connected in sequence, including:

[0038] Among them, the input layer and the output layer of the electrostatic protection DTSCR device performance prediction neural network are fixed parameters; the parameters of the input extension layer, the convolution layer, the pooling layer and the fully connected layer of the electrostatic protection DTSCR device performance prediction neural network are non-fixed parameters; the electrostatic protection DTSCR device performance prediction neural network with different parameters is obtained by randomly setting the parameters of the input extension layer, the convolution layer, the pooling layer and the fully connected layer.

[0039] In a possible implementation, the method of screening a plurality of initially trained electrostatic protection DTSCR device performance prediction neural networks to obtain a trained electrostatic protection DTSCR device performance prediction neural network includes:

[0040] Calculate the loss value of each electrostatic protection DTSCR device performance prediction neural network in each round of training, and calculate the mean square error value of each electrostatic protection DTSCR device performance prediction neural network based on multiple loss values;

[0041] The electrostatic protection DTSCR device performance prediction neural network corresponding to the minimum mean square error value is determined to be the trained electrostatic protection DTSCR device performance prediction neural network.

[0042] In a possible implementation, the training of each electrostatic protection DTSCR device performance prediction neural network according to the data set and the loss function includes:

[0043] According to the data set, the electrostatic protection DTSCR device performance prediction neural network is trained using the SGD optimizer method and the loss function.

[0044] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0045] The present invention utilizes the characteristics of deep learning of neural networks to automatically extract features according to data, and applies it to the prediction of the performance of electrostatic protection DTSCR devices, so as to quickly predict the performance of the devices and adjust the input structure according to the prediction results to achieve the effect of optimizing the performance of the devices; on the other hand, this prediction method can provide design references for engineers and shorten the design cycle; in the experiment, the network structure used has extremely high prediction accuracy, the mean square error of the electrical parameters is as low as 0.004, and the average prediction accuracy reaches 96.9%;

[0046] The first two steps of the present invention require professional background knowledge of microelectronics to complete the work, and once the prediction network is designed, it can be provided to non-professional technicians, so that they can use the neural network to quickly understand the size of the DTSCR device design window and accurately and quickly predict the performance indicators of the electrostatic protection DTSCR device;

[0047] The method proposed by the present invention is to predict the design window of the electrostatic protection device, and this design window is what all ESD devices need to examine. The present invention is applicable to other electrostatic protection devices, provides a different research direction for the development of electrostatic protection ESD devices, and improves research efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flowchart of the steps of a method for predicting the performance of an electrostatic protection DTSCR device based on a neural network provided in an embodiment of the present invention;

[0049] Figure 2 A structural diagram of an electrostatic protection DTSCR device provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of SDE structure modeling of an electrostatic protection DTSCR device provided in an embodiment of the present invention;

[0051] Figure 4 The I / V characteristic curve of the electrostatic protection DTSCR device provided by the embodiment of the present invention under the TLP transmission line pulse test;

[0052] Figure 5 A Kneed method based on an embodiment of the present invention is provided to find the trigger point on the I / V characteristic curve of an electrostatic protection DTSCR device;

[0053] Figure 6 A Kneed method based on an embodiment of the present invention is provided to find a holding point on an I / V characteristic curve of an electrostatic protection DTSCR device;

[0054] Figure 7 A schematic diagram of the structure of a neural network for predicting the performance of an electrostatic protection DTSCR device provided in an embodiment of the present invention;

[0055] Figure 8 Comparison between the predicted value and the true value of the trigger current It after the test set provided by the embodiment of the present invention passes through the neural network;

[0056] Fig. 9 Comparison between the predicted value and the actual value of the trigger voltage Vt after the test set provided by the embodiment of the present invention passes through the neural network;

[0057] Fig.10 Comparison between the predicted value and the true value of the holding current Ih after the test set provided by the embodiment of the present invention passes through the neural network;

[0058] Fig.11 The test set provided in the embodiment of the present invention is a comparison between the predicted value and the actual value of the holding voltage Vh after passing through the neural network. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The present invention provides a method for predicting the performance of an electrostatic protection DTSCR device based on a neural network. Figure 1 The method includes the following steps S101 to S105.

[0061] S101, establish an electrostatic protection DTSCR device model based on Sentaurus TCAD, and obtain an initial data set by simulating the electrostatic protection DTSCR device model;

[0062] Exemplary, reference Figure 2 , Figure 3 and Figure 4 , the SDE tool in the Sentaurus TCAD tool was used to complete the structural modeling of the DTSCR device, and the doping type and doping method were added according to the design drawing. Finally, the calculation grid was designed. The grid needs to be constantly modified, especially the SCR path grid in the DTSCR device needs to be encrypted; the length of the shallow trench isolation between the P well / N+ active area and the N well / P+ active area of ​​the SCR part of the device is defined as LPath, the length of the shallow trench isolation between the N+ active area and the P+ active area in the well connected to the positive electrode of the SCR part is defined as LA_STI, the length of the shallow trench isolation between the N+ active area and the P+ active area in the well connected to the positive electrode of the SCR part is defined as LC_STI, the length of the active area is defined as Lactive, the length of the well edge area including the shallow trench isolation is defined as Lspace, and the length of the wells on the left and right sides of the device from the edge of the device is defined as Loute r, define the length between the SCR part N well and the DIODE1 part N well as LSCR_DIODE, define the well spacing between the two diodes as LD_D, define the length of the shallow trench isolation in the diode well as LDIODE_STI, define the depth of the active area as Pdiff / Ndiff, define the depth of the well as Hwell, define the depth of the shallow trench isolation as HSTI, and define the substrate depth as Hsub. After the structure is established, the process doping area and grid division are completed through empirical design, and the doping concentration of the SCR part well is defined as Pwelldoping / Nwelldoping, and the doping concentration of the diode part well is defined as DPwelldoping / DNwelldoping. The doping concentration of the active area is defined as Pdiff_Doping / Ndiff_Doping. After the doping model is established, Figure 2 shown.

[0063] After the DTSCR device model is built, the Sdevice tool in Sentaurus TCAD is used to simulate and analyze the device. Here, quasi-static testing is used to replace TLP testing. The characteristic curve obtained in this way is similar to that of TLP testing, but faster. Then, the I / V characteristic curve is changed by adjusting the parameters and grid size to reach a reasonable value. The physical models used here include mobility model, bandgap narrowing model, composite model, etc. After the I / V characteristic curve is changed by adjusting the parameters and grid size to reach a reasonable value, as shown in the figure below: Figure 3 shown.

[0064] S102, obtaining an electrical characteristic curve by performing TLP simulation on an electrostatic protection DTSCR device model and an initial data set, extracting parameters from the electrical characteristic curve to obtain characteristic parameters, and establishing a data set according to the characteristic parameters;

[0065] Here, each data in the data set includes: 4 structural parameters and 4 characteristic parameters; the structural parameters include: the length of the shallow trench isolation between the P-well / N+ active area and the N-well / P+ active area of ​​the SCR part of the electrostatic protection DTSCR device model, the length of the shallow trench isolation between the N+ active area and the P+ active area in the same well of the SCR part, the length between the N-well of the SCR part and the N-well of the diode DIODE1 part, and the well spacing between the two diodes; the characteristic parameters include: trigger voltage Vt, trigger current It, holding voltage Vh, and holding current Ih.

[0066] Specifically, in step S102, an electrical characteristic curve is obtained by performing TLP simulation on the electrostatic protection DTSCR device model, and parameters of the electrical characteristic curve are extracted to obtain characteristic parameters, and a data set is established according to the characteristic parameters, including the following steps S1021 to S1023.

[0067] S1021, changing the structural parameters and process parameters of the electrostatic protection DTSCR device model to obtain an electrical characteristic curve of the DTSCR device;

[0068] S1022, extracting characteristic parameters of the electrical characteristic curve, and integrating the characteristic parameters into a data set;

[0069] S1023, preprocessing the data set, and dividing the preprocessed data set into a training set, a test set, and a cross-validation set.

[0070] Here, the data set is preprocessed and divided into a training set, a test set, and a cross-validation set, including:

[0071] (1) Traverse each piece of data in the data set and perform the following operations on each piece of data:

[0072] (2) determining whether the data set meets the I / V characteristic curve standard; if not, removing the data; if yes, normalizing the data to obtain preprocessed data;

[0073] (3) Integrate each processed data into a processed data set, and divide the processed data set according to the ratio to obtain a training set, a test set, and a cross-validation set.

[0074] For example, a data set is established by simulating a TLP test of a DTSCR device. Based on the DTSCR device model, the structural parameters and process parameters of the design are changed to obtain the electrical characteristic curve of the DTSCR device, and the characteristic parameters on the curve are extracted, and the required parameters are integrated into a data set; after the data set is preprocessed, the data set is divided into a training set, a test set, and a cross-validation set.

[0075] Based on the established DTSCR device model, the structural parameters and process parameters are appropriately changed to obtain different device structures and electrical characteristics results; after comparing the impact of various parameter changes on the electrical characteristics, it is decided to use the following four parameters as the input neurons of the neural network: the length LPath of the shallow trench isolation between the P-well / N+ active area and the N-well / P+ active area of ​​the SCR part of the device, the length LA / C_STI of the shallow trench isolation between the N+ active area and the P+ active area in the same well of the SCR part, the length LSCR_DIODE between the N-well of the SCR part and the N-well of the DIODE1 part, and the well spacing LD_D of the two diodes.

[0076] After determining the input neuron, these four values ​​are set: Lpath takes values ​​of 0.6, 0.7, 0.8, 0.9, 1.0; LA / C_STI takes values ​​of 0.8, 0.9, 1.0, 1.1, 1.2; LSCR_DIODE takes values ​​of 1.2, 1.3, 1.4, 1.5, 1.6; LD_D takes values ​​of 1.2, 1.3, 1.4, 1.5, 1.6; simulate the above data, and finally use the Svisual tool in Sentaurus TCAD to collect the I / V characteristic curves corresponding to each device structure.

[0077] refer to Figure 5 and Figure 6, use the Kneed method of Python statement to extract electrical parameters. Here, you need to reverse the ordinate and abscissa values ​​of the extracted curve first, because the Kneed method is to find the inflection points on convex and concave set functions. After reversing the curve data, it is easier to extract the trigger point and the holding point. The electrical parameters are Vt trigger voltage, It trigger current, Vh holding voltage, and Ih holding current. After extracting the electrical parameters, the data set can be established. One sample in the data set contains 4 structural parameters and 4 electrical parameters. The structural parameters are the previous 4 input neurons, and the 4 electrical parameters are Vt trigger voltage, It trigger current, Vh holding voltage, and Ih holding current. 8 data are 1 sample, so many samples constitute a data set.

[0078] There are many inappropriate data in the data set, so data preprocessing is required. First, the data that does not meet the I / V characteristic curve standard of the ESD device is eliminated. This step can be done by filtering the data values ​​in the Excel table. Then the data is normalized. Using Python statements, normalization is achieved by calculating the minimum and maximum values ​​of the feature columns.

[0079] After data preprocessing, the data in the 564 datasets were divided into training set, test set and cross-validation set in a ratio of 8:1:1.

[0080] S103, constructing multiple electrostatic protection DTSCR device performance prediction neural networks with the same framework and different parameters, training each electrostatic protection DTSCR device performance prediction neural network according to the data set and the loss function, and obtaining multiple electrostatic protection DTSCR device performance prediction neural networks that have completed initial training;

[0081] Specifically, in step S103, multiple electrostatic protection DTSCR device performance prediction neural networks with the same framework and different parameters are constructed, including the following steps S1031 to S1033.

[0082] S1031, the framework of the electrostatic protection DTSCR device performance prediction neural network includes an input layer, an input expansion layer, a convolution layer, a pooling layer, a fully connected layer and an output layer connected in sequence;

[0083] S1032, the input layer and output layer of the electrostatic protection DTSCR device performance prediction neural network are fixed parameters;

[0084] S1033, the parameters of the input extension layer, convolution layer, pooling layer and fully connected layer of the electrostatic protection DTSCR device performance prediction neural network are non-fixed parameters, and the parameters of the input extension layer, convolution layer, pooling layer and fully connected layer are randomly set to obtain multiple groups of parameter values ​​of the electrostatic protection DTSCR device performance prediction neural network, and then obtain multiple electrostatic protection DTSCR device performance prediction neural networks with the same framework but different parameters.

[0085] Specifically, in step S103, the framework of the electrostatic protection DTSCR device performance prediction neural network includes an input layer, an input expansion layer, a convolution layer, a pooling layer, a fully connected layer and an output layer which are connected in sequence.

[0086] Among them, the input layer and output layer of the electrostatic protection DTSCR device performance prediction neural network are fixed parameters; the parameters of the input extension layer, convolution layer, pooling layer and fully connected layer of the electrostatic protection DTSCR device performance prediction neural network are non-fixed parameters; the electrostatic protection DTSCR device performance prediction neural network with different parameters is obtained by randomly setting the parameters of the input extension layer, convolution layer, pooling layer and fully connected layer.

[0087] Here, the convolution layer is a one-dimensional convolution layer, and the convolution kernel size is 3×1; the pooling layer is a one-dimensional pooling layer, and the convolution kernel size is 2×1.

[0088] Specifically, in step S103, each electrostatic protection DTSCR device performance prediction neural network is trained according to the data set and the loss function, including: according to the data set, the electrostatic protection DTSCR device performance prediction neural network is trained using the SGD optimizer method and the loss function.

[0089] Specifically, in step S103, the loss function is expressed as:

[0090]

[0091] Where n represents the number of samples; y i represents the actual value of the i-th sample; Represents the predicted value of the i-th sample.

[0092] Exemplary, reference Figure 6The electrostatic protection DTSCR device performance prediction neural network of the present invention is constructed based on a convolutional neural network, and its architecture includes an input layer, an input expansion layer, a convolution layer, a pooling layer, a fully connected layer and an output layer; first, the input layer includes 4 neurons, namely LPath, LA / C_STI, LSCR_DIODE, and LD_D. After the input layer, the input expansion layer is connected. The expansion layer consists of n layers of fully connected layers. After each linear layer, it passes through a normalization layer. Then it enters the feature extraction process, which includes m convolutions and q poolings. The convolution kernel size is 3×1 during convolution and 2×1 during pooling. There is padding during convolution to keep the size consistent before and after convolution. In addition, each channel passes through a normalization layer after convolution. After the upper part, the neural network is flattened to 1 channel number and enters the feature classification process, which includes n layers of fully connected layers. After each linear layer, it passes through a normalization layer. All processes use the ReLU activation function. Finally, the classifier leaves 4 eigenvalues, trigger current, trigger voltage, holding current, and holding voltage. In the entire electrostatic protection DTSCR device performance prediction neural network, the values ​​of each network layer n, m, and q need to be determined by the post-training comparison loss function, that is, the mean square error. The network with the smallest mean square error is fixed as the final prediction model.

[0093] Modify the neural network structure for iterative electrostatic protection DTSCR device performance prediction, and then input the training set and cross-validation set obtained by simulating the DTSCR device into the modified neural network prediction model. Observe and compare the loss function to evaluate the training results. After the network training, stop modifying the network structure and use the network with the smallest loss function, i.e., the mean square error, as the fixed network structure.

[0094] In a specific embodiment provided by the present invention, the network structure parameters of the electrostatic protection DTSCR device prediction neural network are randomly initialized, and the first set of neural network structures are designed. The input layer and the output layer have been determined, and the middle layer needs to be designed. For example, 2 layers of input expansion layers, 3 layers of convolutional layers, 3 layers of pooling layers, and 3 layers of fully connected layers are designed. In the process of modifying the iterative network, these numbers of layers need to be modified in the same way as the control group is set.

[0095] The training set and cross-validation set obtained by simulating the DTSCR device are input into this neural network prediction model for training and mid-way verification optimization; the early stopping method is used in the training to improve the training efficiency; the cross-validation set will be appropriately arranged for verification in the training set test to obtain the loss function at each training moment; this scheme uses the SGD optimizer, and the optimizer will optimize the network according to the loss function until the neural network model converges to obtain the trained electrostatic protection DTSCR device neural network prediction model. If you return to step 4 to retrain, you need to modify the optimizer's learning rate, number of iterations and other data to allow the network to modify the internal weight bias by itself without modifying the network structure.

[0096] S104, screening a plurality of electrostatic protection DTSCR device performance prediction neural networks that have been initially trained to obtain a trained electrostatic protection DTSCR device performance prediction neural network.

[0097] Specifically, in step S104, a plurality of electrostatic protection DTSCR device performance prediction neural networks that have been initially trained are screened to obtain a trained electrostatic protection DTSCR device performance prediction neural network, including the following steps S1041 and S1042.

[0098] S1041, calculating the loss value of each electrostatic protection DTSCR device performance prediction neural network in each round of training, and calculating the mean square error value of each electrostatic protection DTSCR device performance prediction neural network based on multiple loss values;

[0099] S1042, determining that the electrostatic protection DTSCR device performance prediction neural network corresponding to the minimum mean square error value is the trained electrostatic protection DTSCR device performance prediction neural network.

[0100] Exemplary, ESD protection DTSCR device performance prediction neural network final design reference Figure 7 , the input layer has 4 neurons, the expansion layer has 3 layers, the convolution and pooling layers have 2 layers each, and the final classifier fully connected layer has 3 layers.

[0101] The test set is input into the trained prediction model. If the loss function and the prediction accuracy meet the predetermined requirements, the final training model can be used as a tool to predict the design window feature points of the DTSCR device. Otherwise, retraining is performed.

[0102] The test results of the neural network show that the average prediction accuracy of the four eigenvalues ​​reaches 96.9%, and the mean square error is 0.004, which is a relatively accurate result.

[0103] refer to Figure 8-Figure 11, is a comparison chart of the true value and the neural network predicted value during the test set test process, which are trigger current, trigger voltage, holding current, and holding voltage respectively. It can be observed that the predicted results are basically consistent with the original data, which proves the feasibility of the electrostatic protection DTSCR device performance prediction method based on neural network.

[0104] S105, predicting characteristic parameters of the design window of the electrostatic protection DTSCR device by using the trained electrostatic protection DTSCR device performance prediction neural network.

[0105] Specifically, the trained electrostatic protection DTSCR device performance prediction neural network satisfies the accuracy formula. The specific accuracy expression formula is:

[0106]

[0107] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. All or part of the present invention can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, mobile communication terminals, multi-processor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0108] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the present invention.

Claims

1. A method for predicting the performance of an electrostatic protection DTSCR device based on a neural network, characterized in that: include: Establish an electrostatic protection DTSCR device model based on Sentaurus TCAD, and obtain an initial data set by simulating the electrostatic protection DTSCR device model; By performing TLP simulation on the electrostatic protection DTSCR device model and the initial data set, an electrical characteristic curve is obtained, and parameter extraction is performed on the electrical characteristic curve to obtain characteristic parameters, and a data set is established according to the characteristic parameters; Constructing multiple electrostatic protection DTSCR device performance prediction neural networks with the same framework and different parameters, training each of the electrostatic protection DTSCR device performance prediction neural networks according to the data set and the loss function, and obtaining multiple electrostatic protection DTSCR device performance prediction neural networks that have completed initial training; Screening a plurality of electrostatic protection DTSCR device performance prediction neural networks that have been initially trained to obtain a trained electrostatic protection DTSCR device performance prediction neural network; The trained electrostatic protection DTSCR device performance prediction neural network is used to predict the characteristic parameters of the design window of the electrostatic protection DTSCR device.

2. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 1, characterized in that: The method of obtaining an electrical characteristic curve by performing TLP simulation on the electrostatic protection DTSCR device model, extracting parameters from the electrical characteristic curve, obtaining characteristic parameters, and establishing a data set according to the characteristic parameters includes: Changing the structural parameters and process parameters of the electrostatic protection DTSCR device model to obtain an electrical characteristic curve of the DTSCR device; Extracting characteristic parameters of the electrical characteristic curve and integrating the characteristic parameters into a data set; The data set is preprocessed, and the preprocessed data set is divided into a training set, a test set, and a cross-validation set.

3. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 2, characterized in that: The preprocessing of the data set, dividing the preprocessed data set into a training set, a test set and a cross-validation set, comprises: Traverse each piece of data in the data set and perform the following operations on each piece of data: Determine whether the data set meets the I / V characteristic curve standard. If not, remove the data; if yes, perform normalization on the data to obtain preprocessed data; Each processed data is integrated into a processed data set, and the processed data set is divided according to the proportion to obtain a training set, a test set, and a cross-validation set.

4. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 1, characterized in that: Each piece of data in the data set includes: 4 structural parameters and 4 characteristic parameters; The four structural parameters include: the length of the shallow trench isolation between the P well / N+ active area and the N well / P+ active area of ​​the SCR part of the electrostatic protection DTSCR device model, the length of the shallow trench isolation between the N+ active area and the P+ active area in the same well of the SCR part, the length between the N well of the SCR part and the N well of the diode DIODE1 part, and the well spacing between the two diodes; The four characteristic parameters include: trigger voltage Vt, trigger current It, holding voltage Vh and holding current Ih.

5. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 1, characterized in that: The framework of the electrostatic protection DTSCR device performance prediction neural network includes an input layer, an input expansion layer, a convolution layer, a pooling layer, a fully connected layer and an output layer connected in sequence; The input layer is used to receive input data; The input expansion layer is used to expand the dimension of the input data to obtain the input data after the dimension is expanded; The convolution layer is used to extract data features from the input data after the dimension expansion to obtain initial feature information; The pooling layer is used to remove redundant information from the initial feature information to obtain feature information; The fully connected layer is used to classify the feature information to obtain a classification result; The output layer is used to output the classification result.

6. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 5, characterized in that: The convolution layer is a one-dimensional convolution layer, and the convolution kernel size is 3×1; The pooling layer is a one-dimensional pooling layer, and the convolution kernel size is 2×1.

7. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 1, characterized in that: The loss function is expressed as: Where n represents the number of samples; y i represents the actual value of the i-th sample; Represents the predicted value of the i-th sample.

8. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 1, characterized in that: The frameworks of the electrostatic protection DTSCR device performance prediction neural networks with the same framework and different parameters all include an input layer, an input expansion layer, a convolution layer, a pooling layer, a fully connected layer and an output layer connected in sequence; Among them, the input layer and the output layer of the electrostatic protection DTSCR device performance prediction neural network are fixed parameters; the parameters of the input extension layer, the convolution layer, the pooling layer and the fully connected layer of the electrostatic protection DTSCR device performance prediction neural network are non-fixed parameters; the electrostatic protection DTSCR device performance prediction neural network with different parameters is obtained by randomly setting the parameters of the input extension layer, the convolution layer, the pooling layer and the fully connected layer.

9. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 1, characterized in that: The method of screening a plurality of electrostatic protection DTSCR device performance prediction neural networks that have been initially trained to obtain a trained electrostatic protection DTSCR device performance prediction neural network includes: Calculate the loss value of each electrostatic protection DTSCR device performance prediction neural network in each round of training, and calculate the mean square error value of each electrostatic protection DTSCR device performance prediction neural network based on multiple loss values; The electrostatic protection DTSCR device performance prediction neural network corresponding to the minimum mean square error value is determined to be the trained electrostatic protection DTSCR device performance prediction neural network.

10. The method for predicting the performance of an electrostatic protection DTSCR device based on a neural network according to claim 1, characterized in that: The step of training each of the electrostatic protection DTSCR device performance prediction neural networks according to the data set and the loss function includes: According to the data set, the electrostatic protection DTSCR device performance prediction neural network is trained using the SGD optimizer method and the loss function.

Images