Method for simulating electrical characteristics of semiconductor devices based on machine learning and models

Abstract

The application discloses a semiconductor device electrical characteristic simulation method based on machine learning and a model, comprising the following steps: step 1, determining a simulation model; step 2, determining structure parameters or process parameters; step 3, obtaining a data set; step 4, establishing a machine learning regression model; step 5, predicting input model parameters; and step 6, electrical characteristic simulation. The application realizes the simulation of a semiconductor device from structure parameters and process parameters to electrical characteristics by using machine learning and the determined simulation model, and has the advantages of fast simulation speed, good convergence, saving of computing resources and the like. Meanwhile, the regression model constructed in the application can realize the prediction from process parameters to model parameters, from process parameters to structure parameters and from structure parameters to model parameters, and can improve the design efficiency of designers and save design time.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device electrical characteristic simulation, and in particular to a semiconductor device electrical characteristic simulation method based on machine learning and models. Background Technology

[0002] In the 21st century, the field of semiconductor devices has developed rapidly. Among these advancements, the electrical characteristics of semiconductor devices are crucial for designers in determining device structural and process design parameters. Designers aim to obtain the electrical characteristics of semiconductor devices early in the design process, thereby adjusting structural and process parameters to achieve the design goals.

[0003] Simulation of the electrical characteristics of semiconductor devices is of great significance. While computer-aided design (TCAD) tools are used for this simulation, they suffer from problems such as long simulation times, poor convergence, and difficulties in circuit simulation. Semiconductor device models can more simply describe the device's operating characteristics and can be called upon in circuit simulation software, offering faster simulation speeds, better convergence, and the ability to perform large-scale circuit simulations. However, extracting model parameters is often extremely complex. For example, the static parameters of the Hefner model for IGBTs include: the lifetime τ of the large-injection excess carriers. hl Saturated electron current I in the emission region sne MOSFET saturation region transconductance K p Threshold voltage V t The empirical scaling factor K between the transconductance in the linear region and the transconductance in the saturation region f And the transverse field transconductance factor θ. For example, the ASM and statically related model parameters of GaN HEMTs: cutoff voltage V off Subthreshold swing NFACTOR, low-field mobility μ0, mobility degradation factor μ a Channel length modulation factor λ, saturation velocity V sat Source region contact resistance R sc Drain contact resistance R dc Traditional methods require obtaining output characteristic curves, turn-off characteristic curves, and capacitor voltage characteristic curves under different gate voltages through device structure simulation or experiments, and then extracting parameters from them. Although the parameter extraction process can be achieved by programs and tools such as IC-CAP, TCAD simulation of the relevant characteristic curves of each semiconductor device still takes several minutes and there are cases of non-convergence.

[0004] In addition, there are methods that use machine learning techniques to build machine learning models to directly predict the electrical characteristics of devices. While parameter extraction is relatively simple, these methods still have the following shortcomings and require improvement:

[0005] 1. For devices with complex structures, when using the ANN method, the neural network is huge, resulting in a large amount of simulation computation and slow speed. Furthermore, it can lead to non-convergence problems when the bias voltage is continuously iterated, which also limits and affects the accuracy of the simulation.

[0006] 2. Using ANN for direct modeling is a black-box modeling approach. The relationship between input and output is not clearly represented, and there is a lack of semiconductor physical interpretation, which makes it difficult for designers to adjust the device structure based on simulation results.

[0007] Therefore, there is an urgent need for a fast, simple and accurate method for simulating the electrical characteristics of semiconductor devices. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a simulation method for the electrical characteristics of semiconductor devices based on machine learning and models. This simulation method for the electrical characteristics of semiconductor devices based on machine learning and models has advantages such as high speed and good convergence, and can establish the relationship between device process parameters, structural parameters and model parameters, which is beneficial for designers to analyze device characteristics and reduce the design cycle.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] A method for simulating the electrical characteristics of semiconductor devices based on machine learning and models includes the following steps.

[0011] Step 1: Determine the simulation model: Based on the type of semiconductor device, determine the simulation model used for its circuit simulation, and determine the input model parameters of the simulation model.

[0012] Step 2: Determine structural or process parameters: Determine the structural or process parameters of the semiconductor device that affect the input model parameters in Step 1.

[0013] Step 3: Obtain the dataset: Obtain the dataset from structural parameters to input model parameters or from process parameters to input model parameters through simulation or experimentation.

[0014] Step 4: Establish a machine learning regression model: Based on the dataset obtained in Step 3, establish a machine learning regression model; the machine learning regression model can predict the input model parameters in Step 1 based on the structural or process parameters of semiconductor devices.

[0015] Step 5: Predict the input model parameters: Input the structural or process parameters of the semiconductor device to be simulated into the machine learning regression model established in Step 4 to predict the corresponding input model parameters of the semiconductor device to be simulated.

[0016] Step 6, electrical characteristic simulation: inputting the input model parameters predicted in step 5 into the simulation model determined in step 1, and using the simulation model to simulate the electrical characteristics of the semiconductor device to be simulated.

[0017] In step 1, the simulation model is a basic electrical performance model or a multi-physical field coupling model; the semiconductor device is a field effect transistor, a thyristor, an IGBT or a diode.

[0018] When the semiconductor device is an IGBT, the simulation model is a Hefner model; when the semiconductor device is a field effect transistor and is a GaN HEMT, the simulation model is an ASM model.

[0019] For the Hefner model of the IGBT, the input model parameters include the large injection excess carrier lifetime, the saturation electron current of the emitter region, the MOSFET saturation region transconductance K p , the threshold voltage V t , the empirical proportion factor K f of linear region transconductance and saturation region transconductance, and the lateral field transconductance factor θ; for the ASM model of the GaN HEMT, the input model parameters include the cutoff voltage V off related to static characteristics, sub-threshold swing NFACTOR, low-field mobility μ0, mobility degradation factor μ a , channel length modulation factor λ, saturation velocity V sat , source region contact resistance R sc , and drain region contact resistance R dc .

[0020] In step 3, the data set is a data set from the structure parameters to the input model parameters; for the IGBT, the structure parameters include the metallurgical base region width, the metallurgical base region doping concentration, the channel length, the channel doping concentration and the anode region doping concentration; for the GaN HEMT, the structure parameters include the gate-drain length, the gate-source length, the gate length and the AlGaN barrier layer thickness.

[0021] In step 3, the data set is a data set from the process parameters to the structure parameters or a data set from the process parameters to the input model parameters;

[0022] wherein, for the IGBT, the process parameters are one or a combination of ion implantation process parameters, oxidation and diffusion process parameters, photolithography process parameters, etching process parameters and metallization process parameters; for the GaN HEMT, the process parameters are one or a combination of PECVD parameters, etching parameters, photolithography parameters, electron beam evaporation parameters and sputtering parameters.

[0023] In step 3, the method of obtaining the data set by simulation or experiment includes the following steps:

[0024] Step 31, determine the variation range of the semiconductor device structure parameters or process parameters.

[0025] Step 32, determine the structure parameters: realize the semiconductor device under different process parameters through process simulation or experiment, and extract the structure parameters determined in step 2.

[0026] Step 33, obtain the electrical property curve: obtain the electrical property curve of the semiconductor device with different structure parameters or process parameters through simulation or experiment.

[0027] Step 34, extract the input model parameters: according to the model parameter extraction method or using the parameter extraction tool, extract the input model parameters corresponding to the semiconductor device from the electrical property curve obtained in step 33.

[0028] Step 35, form a data set, specifically:

[0029] A, one-to-one correspondence between the process parameters and the structure parameters of each semiconductor device in step 32, forming a data set of process parameters to structure parameters.

[0030] B, one-to-one correspondence between the process parameters of each semiconductor device in step 32 and the input model parameters extracted in step 34, forming a data set of process parameters to input model parameters.

[0031] C, one-to-one correspondence between the structure parameters of each semiconductor device in step 32 and the input model parameters extracted in step 34, forming a data set of structure parameters to input model parameters.

[0032] In step 4, the method for establishing the machine learning regression model is one or a combination of deep neural network, Gaussian process regression, support vector machine, linear regression, logistic regression, Lasso regression and CART regression tree.

[0033] In step 4, the specific method for establishing the machine learning regression model includes the following steps:

[0034] Step 41, data set classification: the data set obtained in step 3 is divided into training set, validation set and test set according to the set proportion; wherein, the training set is used to train the machine learning model; the validation set is used to adjust the weight value in the training process; the test set is used to test the machine learning model.

[0035] Step 42, data processing: standardize or normalize the data in the training set, validation set and test set.

[0036] Step 43, establishing a machine learning regression model: selecting and establishing a machine learning regression model; the output parameter of the machine learning regression model is an input model parameter of a semiconductor device simulation model; and the input parameter of the machine learning regression model is a process parameter or a structure parameter of the semiconductor device.

[0037] Step 44, model training: training the machine learning regression model established in step 43 by using the data in the training set in step 42, to obtain the weight value of the machine learning regression model.

[0038] Step 45, model verification and testing: verifying or testing the machine learning regression model trained in step 44 by using the data in the verification set or the test set in step 42.

[0039] Step 5, an input model parameter prediction method of a semiconductor device.

[0040] Step 51, data processing: standardizing or normalizing the structure parameter or the process parameter of the semiconductor device to be simulated.

[0041] Step 52, prediction: inputting the structure parameter or the process parameter of the semiconductor device to be simulated after data processing in step 51 into the machine learning regression model established in step 4, to predict the input model parameter corresponding to the semiconductor device to be simulated.

[0042] Step 53, reverse data processing: reverse standardization or reverse normalization of the input model parameter predicted in step 52, to obtain the input model parameter of the simulation model.

[0043] The present application has the following beneficial effects:

[0044] 1. The present application can realize the simulation of the electrical characteristics of a semiconductor device according to the structure parameter or the process parameter of the semiconductor device by using a machine learning regression model and a reliable simulation model, and has better convergence, faster speed and can perform large-scale circuit simulation compared with a TCAD simulation tool.

[0045] 2. Compared with common machine learning modeling, the method of combining a machine learning regression model and a simulation model has higher reliability based on a model widely recognized by the industry, the predicted input model parameter has physical meaning, the simulation speed is not related to the complexity of the machine learning model, and the phenomenon of non-convergence and too long simulation time caused by the complexity of the machine learning model is reduced.

[0046] 3. The method of the present application can be applied to various semiconductor devices, and can directly simulate the electrical characteristics of the semiconductor device from the structure parameter and the process parameter of the semiconductor device.

[0047] 4、The method can also be used for characterizing the nonlinear relationship among semiconductor device process parameters, structure parameters, process parameters and model parameters, and structure parameters and model parameters, helping designers to quickly analyze and optimize device process and structure parameters. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A flowchart of the semiconductor device electrical characteristic simulation method based on machine learning and model of the application.

[0049] Figure 2 A structure diagram of an Insulated Gate Bipolar Transistor (IGBT) in an embodiment of the application.

[0050] Figure 3 A structure diagram of a GaN High electron mobility transistor (GaN HEMT) in an embodiment of the application.

[0051] Figure 4 A machine learning model structure diagram of three Hefner model parameter prediction schemes in IGBT on-state current characteristic simulation of an embodiment of the application, using a neural network model in machine learning to predict model parameters, and the three schemes are a multiple-input multiple-output scheme, a multiple-input single-output scheme, and a phased multiple-output single-input scheme.

[0052] Figure 5 A machine learning model structure diagram of an ASM model parameter prediction scheme in GaN HEMT on-state current characteristic simulation of an embodiment of the application, using a neural network model in machine learning to predict model parameters.

[0053] Figure 6 A comparison diagram of IGBT on-state current characteristic simulation results and TCAD simulation results in an embodiment of the application.

[0054] Figure 7 A comparison diagram of GaN HEMT on-state current characteristic simulation results and TCAD simulation results in an embodiment of the application. DETAILED DESCRIPTION

[0055] The application will be further described in detail below in combination with the drawings and specific preferred embodiments.

[0056] In the description of the present application, it should be understood that the terms "left side", "right side", "upper part", "lower part" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and "first", "second" and the like do not represent the importance of the parts, and therefore cannot be understood as a limitation on the present application. The specific dimensions used in the embodiments are only for the purpose of illustrating the technical solutions and do not limit the protection scope of the present application.

[0057] As shown in Figure 1 , a semiconductor device electrical property simulation method based on machine learning and model comprises the following steps.

[0058] Step 1, establishing a simulation model: according to the type of semiconductor device, a simulation model for its circuit simulation is established, and the input model parameters of the simulation model are determined.

[0059] The simulation model is preferably a basic electrical property model or a multi-physical field coupling model.

[0060] The semiconductor device is preferably a field effect transistor, a thyristor, an insulated gate bipolar transistor, a diode, etc.

[0061] The present application takes IGBT and GaN HEMT as examples for detailed description.

[0062] Embodiment 1

[0063] The semiconductor device is IGBT, and the corresponding simulation model is Hefner model.

[0064] As shown in Figure 2 , the IGBT includes anode, anode doping region P + , N-type doping drift region N-drift, P-type doping drift region P-well, heavily doped P-type region P + , heavily doped N-type drift region N + , and gate and cathode all arranged on the top layer.

[0065] Embodiment 2

[0066] The semiconductor device is GaN HEMT in field effect transistor, and the corresponding simulation model is ASM model.

[0067] As shown in Figure 3 , the GaN HEMT includes substrate, GaN buffer layer, GaN channel layer, AlGaN barrier layer arranged from bottom to top, and gate G, source S and drain D all arranged on the top layer.

[0068] Step 2, determining structure parameters or process parameters: determining the structure parameters or process parameters of the semiconductor device which have influence on the input model parameters in step 1.

[0069] Example 1 IGBT-Hefner model

[0070] The input model parameters of the Hefner model preferably include six items, which are: large injection excess carrier lifetime, saturation electron current of the emitter region, MOSFET saturation region transconductance K p , threshold voltage V t , empirical proportionality factor K f of linear region transconductance and saturation region transconductance, and lateral field transconductance factor θ.

[0071] The structure parameters of the semiconductor device which have influence on the input model parameters of the Hefner model described above preferably include five items, which are: metallurgical base region width, metallurgical base region doping concentration, channel length, channel doping concentration and anode region doping concentration.

[0072] The process parameters of the IGBT described above can be one or combination of ion implantation process parameters, oxidation diffusion process parameters, lithography process parameters, etching process parameters and metallization process parameters.

[0073] When the process parameter is ion implantation process parameter, the ion implantation process parameters of the semiconductor device which have influence on the input model parameters of the Hefner model described above preferably include impurity type, dose, energy and range, etc.

[0074] When the process parameter is oxidation diffusion process parameter, the oxidation process parameters of the semiconductor device which have influence on the input model parameters of the Hefner model described above preferably include diffusion temperature, diffusion time, well implantation dose, oxide layer thickness and junction depth, etc.

[0075] When the process parameter is lithography process parameter, the lithography process parameters of the semiconductor device which have influence on the input model parameters of the Hefner model described above preferably include photoresist type, thickness, exposure time, developer type, development time, pre-bake and post-bake temperature, pre-bake and post-bake time, etc.

[0076] When the process parameter is etching process parameter, the etching process parameters of the semiconductor device which have influence on the input model parameters of the Hefner model described above preferably include etching liquid type and etching time, etc.

[0077] When the process parameter is metallization process parameter, the metallization process parameters of the semiconductor device which have influence on the input model parameters of the Hefner model described above preferably include metal type and deposition thickness, etc.

[0078] Example 2 GaN HEMT-ASM model

[0079] The input model parameters of the ASM model related to static characteristics preferably include eight items, respectively: the cut-off voltage V off , the sub-threshold swing NFACTOR, the low-field mobility μ0, the mobility degradation factor μ a , the channel length modulation factor λ, the saturation velocity V sat , the source region contact resistance R sc , and the drain region contact resistance R dc .

[0080] The structural parameters of the semiconductor device that have an impact on the input model parameters of the ASM model described above preferably include four items, respectively: the gate-drain length, the gate-source length, the gate length, and the AlGaN barrier layer thickness.

[0081] The GaN HEMT process parameters described above can be one or a combination of the following: a plasma enhanced chemical vapor deposition (PECVD) process parameter, an etching process parameter, a lithography process parameter, an electron beam evaporation process parameter, and a sputtering process parameter.

[0082] When the process parameter is a PECVD process parameter, the PECVD process parameters of the semiconductor device that have an impact on the input model parameters of the ASM model described above are preferably: the process temperature, the cavity pressure, and the atmosphere ratio.

[0083] When the process parameter is an etching process parameter, the etching process parameters of the semiconductor device that have an impact on the input model parameters of the ASM model described above are preferably: the gas type, the cavity pressure, the cavity temperature, the gas flow rate, and the radio frequency power.

[0084] When the process parameter is a lithography process parameter, the lithography process parameters of the semiconductor device that have an impact on the input model parameters of the ASM model described above are preferably: the photoresist type, the thickness, the exposure time, the developer type, the development time, the pre-baking and post-baking temperatures, and the pre-baking and post-baking times.

[0085] When the process parameter is an electron beam evaporation process parameter, the electron beam evaporation process parameters of the semiconductor device that have an impact on the input model parameters of the ASM model described above are preferably: the substrate temperature and the electron gun current.

[0086] When the process parameter is a sputtering process parameter, the sputtering process parameters of the semiconductor device that have an impact on the input model parameters of the ASM model described above are preferably: the power supply power, the cavity pressure, the cavity temperature, the atmosphere ratio, and the gas flow rate.

[0087] Step 3, obtaining data set: obtaining data set from structure parameter to input model parameter, from process parameter to input model parameter directly or indirectly by simulation or experiment.

[0088] The data set from process parameter to input model parameter indirectly is preferably obtained from process parameter to structure parameter and then from corresponding structure parameter to input model parameter in the embodiment.

[0089] In the embodiment, the method for obtaining data set by simulation or experiment preferably includes the following steps.

[0090] Step 31, determining the variation range of structure parameter or process parameter of semiconductor device, taking IGBT structure parameter and GaN HEMT structure parameter as examples, then:

[0091] Table 1 IGBT structure parameter and variation range

[0092] Structural parameter Range Metallurgical base region doping concentration, N b (cm -3 )]]> [[1 x 10 13 ,1 x 10 15 ]]]> Channel doping concentration, N well (cm -3 )]]> [[1 x 10 17 ,5 x 10 17 ]]]> Anode region doping concentration, N P+ (cm -3 )]]> [[1x10 17 ,1x10 19 ]]]> Metallurgical base width, W B (μm) [50,100] Channel length, L (pm) [1,5]

[0093] Table 2 GaN HEMT structure parameter and variation range

[0094] Structural parameter Range Gate leakage length, L gd (μm) [4,10] Gate-source length, L sg (μm) [1,4] Gate length, L (cm -3 )]]> [0.5,4] AlGaN barrier layer thickness, T bar (μm) [0.015,0.025]

[0095] Step 32, determining structure parameter: obtaining semiconductor device under different process parameters by process simulation or experiment, and extracting structure parameter determined in step 2.

[0096] Step 33, obtaining electrical characteristic curve: obtaining electrical characteristic curve of semiconductor device with different structure parameters or process parameters by simulation or experiment.

[0097] In the embodiment, for IGBT device, the off-state current characteristic curve and on-state current characteristic curve of different IGBT structures are simulated according to the need of Hefner model parameter extraction; for GaN HEMT device, the output characteristic curve and transfer characteristic curve are simulated according to the need of ASM model parameter extraction.

[0098] Step 34, extracting input model parameter: extracting input model parameter corresponding to semiconductor device from electrical characteristic curve obtained in step 33 according to model parameter extraction method or using parameter extraction tool.

[0099] Step 35, forming data set, specifically:

[0100] A, corresponding process parameter and structure parameter of each semiconductor device in step 32 to form data set from process parameter to structure parameter.

[0101] B. One-to-one correspondence between the process parameters of each semiconductor device in step 32 and the input model parameters extracted in step 34, forming a data set of process parameters to input model parameters.

[0102] C. One-to-one correspondence between the structure parameters of each semiconductor device in step 32 and the input model parameters extracted in step 34, forming a data set of structure parameters to input model parameters. In this embodiment, it is assumed that the structure parameters are known, so it is preferred to select this way of obtaining the data set.

[0103] Step 4, establish a machine learning regression model: according to the data set obtained in step 3, establish a machine learning regression model; the machine learning regression model can predict the input model parameters in step 1 according to the structure parameters or process parameters of the semiconductor device.

[0104] The establishment method of the above machine learning regression model is preferably one or a combination of deep neural network, Gaussian process regression, support vector machine, linear regression, logistic regression, Lasso regression and CART regression tree.

[0105] In this embodiment, taking IGBT and GaN HEMT as an example, the specific method of establishing the machine learning regression model preferably includes the following steps.

[0106] Step 41, data set classification: the data set obtained in step 3 is divided into training set, validation set and test set according to the set proportion (such as 7:2:1); wherein, the training set is used to train the machine learning model; the validation set is used to adjust the weight value in the training process; the test set is used to test the machine learning model.

[0107] Step 42, data processing: the data in the training set, validation set and test set are all standardized or normalized. In this embodiment, normalized processing is preferred. Deep neural network

[0108] Step 43, establish a machine learning regression model: select and establish a machine learning regression model, specifically:

[0109] A. The machine learning regression model is preferably selected as a deep neural network.

[0110] B. Set the model parameters: the output parameters of the machine learning regression model are the input model parameters of the semiconductor device simulation model; the input parameters of the machine learning regression model are the process parameters or structure parameters of the semiconductor device.

[0111] C. Model setting, specifically: set the number of hidden layers of the neural network, the number of neurons, use the linear rectifier function as the activation function, select the Adam algorithm as the optimization algorithm of the neural network parameter value, and set the loss function to represent the gap between the actual model parameters and the predicted model parameters.

[0112] Figure 4 The neural network structure diagram of the three neural network parameter prediction schemes proposed for the Hefner model parameter prediction of IGBT is given. Among them, (a) is a multiple-input multiple-output scheme; (b) is a multiple-input single-output scheme; and (c) is a multiple-input single-output scheme in stages.

[0113] The above multiple-input multiple-output means that only one neural network needs to be trained, however, the input parameters are multiple, in the embodiment, the input parameters are five structure parameters; and the outputs are also multiple, in the embodiment, six input model parameters.

[0114] The above multiple-input single-output means that the number of trained neural networks is the same as the number of input model parameters, a neural network is established for each input model parameter, and six input model parameters correspond to six neural networks. The input parameters of each neural network are five structure parameters, and the output is a corresponding single input model parameter.

[0115] The above multiple-input single-output in stages means that the number of trained neural networks is the same as the number of input model parameters,

[0116] A neural network is established for each input model parameter, and six input model parameters correspond to six neural networks. Different from the multiple-input single-output, according to the order, the output of the previous neural network is added to the input of the second neural network during prediction, that is, the number of input parameters of the six neural networks is 5, 6, 7, 8, 9, and 10 respectively. Then the prediction order is according to the order of parameter extraction, because the input model parameter extracted in the previous step may have an impact on the subsequent input model parameter. Therefore, the prediction is also in this order, and the output of the previous step is placed in the input of the subsequent step, so that the prediction accuracy is higher. It has been proved that the scheme of multiple-input single-output in stages is more accurate and can accurately predict the structure parameters outside the training range.

[0117] Figure 5 The neural network structure diagram of the neural network parameter prediction scheme proposed for the ASM model parameter prediction of GaN HEMT is given.

[0118] Step 44, model training: the data in the training set in step 42 is used to train the machine learning regression model established in step 43, and the weight value of the machine learning regression model is obtained.

[0119] Step 45, model verification and testing: the data in the verification set or the test set in step 42 is used to verify or test the machine learning regression model after the training in step 44.

[0120] Step 5, predicting input model parameters: inputting the structure parameters or process parameters of the semiconductor device to be simulated into the machine learning regression model established in step 4 to predict the input model parameters corresponding to the semiconductor device to be simulated.

[0121] In this embodiment, the input model parameter prediction method of the semiconductor device IGBT and GaN HEMT is preferably:

[0122] Step 51, data processing: standardizing or normalizing the structure parameters or process parameters of the semiconductor device to be simulated.

[0123] Step 52, prediction: inputting the structure parameters or process parameters of the semiconductor device to be simulated after data processing in step 51 into the machine learning regression model established in step 4 to predict the input model parameters corresponding to the semiconductor device to be simulated.

[0124] Step 53, reverse data processing: reverse standardization or reverse normalization of the input model parameters predicted in step 52 to obtain the input model parameters of the simulation model.

[0125] Step 6, electrical characteristic simulation: substituting the input model parameters predicted in step 5 into the simulation model established in step 1, and using the simulation model to simulate the electrical characteristics of the semiconductor device to be simulated.

[0126] The semiconductor device electrical characteristic simulation method of the present application is based on machine learning and reliable models, and can quickly simulate the electrical characteristics of semiconductor devices given the structure parameters or process parameters of the device. This method will improve simulation efficiency and provide better guidance for designers.

[0127] In this embodiment, for IGBT, taking two groups of different structure IGBTs as an example, the structure parameters of the two groups of IGBTs are given in Table 3. The two groups of structure parameters are input into the neural network to predict the Hefner model parameters, and then the model is called in the circuit simulation software and the parameters are substituted to simulate and obtain the electrical characteristics of the device.

[0128] Table 3 Structure parameters of two groups of different IGBTs

[0129]

[0130] In this embodiment, for GaN HEMT, taking a group of GaN HEMT as an example, the structure parameters thereof are given in Table 4. The structure parameters are input into the neural network to predict the ASM model parameters, and then the model is called in the circuit simulation software and the parameters are substituted to simulate and obtain the electrical characteristics of the device.

[0131] Table 4 Structure parameters of a group of GaN HEMT

[0132]

[0133] Figure 6 The simulation results of the output characteristic curves of two different structure IGBTs (IGBT1 and IGBT2 respectively) in circuit simulation software are given. Figure 6 The left three figures are the output characteristic curves of the same IGBT (IGBT1); wherein, the (a) figure, the (b) figure and the (c) figure correspond to the (a), the (b) and the (c) three neural network prediction models in Figure 4

[0134] The right three figures are the output characteristic curves of IGBT2; wherein, the (d) figure, the (e) figure and the (f) figure correspond to the (a), the (b) and the (c) three neural network prediction models in Figure 6 Figure 4

[0135] The difference between the above IGBT1 and IGBT2 is that the structure parameters of IGBT1 are all within the training set range, and the structure parameters of IGBT2 are all outside the training set range. It can be seen that the multi-input single-output scheme in stages has good extrapolation ability.

[0136] It can be seen from Figure 6 that the curve is highly fitted compared with the TCAD simulation. The three schemes can effectively simulate the electrical characteristics, wherein the curve fitting of the multi-input single-output scheme in stages is better than that of the single-output scheme, and still has high accuracy for the structure outside the training set range, which shows that the multi-input single-output scheme in stages has higher expandability. The simulation time of the two groups of IGBTs in TCAD is 121.10s and 116.05s, while the simulation time of the semiconductor device electrical characteristic simulation method based on ANN and model is 16.5s and 12.91s respectively, which shows that the method saves the simulation time and is beneficial to accelerate the device design.

[0137] Figure 7 The simulation results of the output characteristic curves of GaN HEMT in circuit simulation software are given. The neural network model in Figure 5

[0138] It can be seen from Figure 7 that the curve is highly fitted compared with the TCAD simulation, and the simulation time is within 10 seconds. Due to the material characteristics, the TCAD simulation of a group of GaN HEMT output characteristic curves takes tens of minutes or even hours.

[0139] ​​​To sum up, the semiconductor device electrical characteristic simulation method can predict the model parameters according to the structure parameters of the device by using the neural network for the IGBT structure and the GaN HEMT structure, and then the model parameters are substituted into the simulation software for simulation, so that the rapid simulation of the semiconductor device electrical characteristics based on the machine learning and the model is realized.

[0140] The preferred embodiments of the application are described in detail above, but the application is not limited to the specific details in the above-described embodiments, and various equivalent transformations can be made to the technical solutions of the application within the technical concept of the application, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. A method for simulating the electrical characteristics of semiconductor devices based on machine learning and models, characterized in that: The method comprises the following steps: Step 1, determining a simulation model: according to the type of the semiconductor device, determining a simulation model for circuit simulation, and determining input model parameters of the simulation model; The semiconductor device is a field effect transistor, a thyristor, an IGBT or a diode; The simulation model is a basic electrical performance model or a multi-physics field coupling model; When the semiconductor device is an IGBT, the simulation model is a Hefner model; input model parameters of the Hefner model include large-injection excess carrier lifetime, emitter saturation electron current, MOSFET saturation region transconductance K p , threshold voltage V t , empirical scaling factor for linear region transconductance and saturation region transconductance K f , and lateral field transconductance factor θ ; When the semiconductor device is a field effect transistor and is a GaN HEMT, the simulation model is an ASM model; input model parameters of the ASM model include a threshold voltage VTH related to a static characteristic V off , a sub-threshold swing NFACTOR, a low field mobility µ 0, a mobility degradation factor µ a , a channel length modulation factor λ , a saturation velocity V sat , a source region contact resistance R sc , and a drain region contact resistance R dc ; Step 2, determining a structure parameter or a process parameter: determining a structure parameter or a process parameter of the semiconductor device which has an influence on the input model parameters in step 1; Step 3, obtaining a data set: obtaining a data set from the structure parameter to the input model parameter or from the process parameter to the input model parameter by simulation or experiment; Step 4, establishing a machine learning regression model: establishing a machine learning regression model according to the data set obtained in step 3; The machine learning regression model can predict the input model parameters in step 1 according to the structure parameter or the process parameter of the semiconductor device; wherein the machine learning regression model adopts a phased multi-input single-output, and the number of trained neural networks is the same as the number of input model parameters; Step 5, predicting input model parameters: inputting the structure parameter or the process parameter of the semiconductor device to be simulated into the machine learning regression model established in step 4, and predicting the input model parameters corresponding to the semiconductor device to be simulated; Step 6, electrical characteristic simulation: substituting the input model parameters predicted in step 5 into the simulation model determined in step 1, and simulating the electrical characteristics of the semiconductor device to be simulated by using the simulation model. 2.The method of claim 1, wherein: In step 3, the data set is a data set from the structure parameter to the input model parameter; for IGBT, the structure parameters include metallurgical base width, metallurgical base doping concentration, channel length, channel doping concentration and anode region doping concentration; for GaN HEMT, the structure parameters include gate-drain length, gate-source length, gate length and AlGaN barrier layer thickness.

3. The method of claim 2, wherein: In step 3, the data set is a data set from the process parameter to the structure parameter or a data set from the process parameter to the input model parameter; Wherein, for IGBT, the process parameters are one or a combination of ion implantation process parameters, oxidation diffusion process parameters, photolithography process parameters, etching process parameters and metallization process parameters; for GaN HEMT, the process parameters are one or a combination of PECVD parameters, etching parameters, photolithography parameters, electron beam evaporation parameters and sputtering parameters.

4. The method of claim 1, wherein: In step 3, the method for obtaining the data set by simulation or experiment comprises the following steps: Step 31, determining the variation range of the structure parameter or the process parameter of the semiconductor device; Step 32, determining the structure parameter: realizing the semiconductor device under different process parameters by process simulation or experiment, and extracting the structure parameter determined in step 2; Step 33, obtaining an electrical characteristic curve: obtaining the electrical characteristic curve of the semiconductor device with different structure parameters or process parameters by simulation or experiment; Step 34, extracting input model parameters: extracting the input model parameters corresponding to the semiconductor device from the electrical characteristic curve obtained in step 33 according to the model parameter extraction method or using the parameter extraction tool; Step 35, forming a data set, specifically: A. Match the process parameters and structural parameters of each semiconductor device in step 32 one by one to form a dataset of process parameters to structural parameters. B. Match the process parameters of each semiconductor device in step 32 with the input model parameters extracted in step 34 to form a dataset of process parameters to input model parameters. C. Match the structural parameters of each semiconductor device in step 32 with the input model parameters extracted in step 34 to form a dataset of structural parameters to input model parameters.

5. The method of claim 1, wherein: In step 4, the machine learning regression model is established using one or a combination of deep neural networks, Gaussian process regression, support vector machines, linear regression, logistic regression, Lasso regression, and CART regression trees.

6. The method of claim 1, wherein: Step 4, the specific method for building the machine learning regression model, includes the following steps: Step 41: Dataset Classification: Divide the dataset obtained in Step 3 into training set, validation set, and test set according to a set ratio; the training set is used to train the machine learning model; the validation set is used to adjust the weight values ​​during the training process; and the test set is used to test the machine learning model. Step 42, Data Processing: Standardize or normalize the data in the training set, validation set, and test set. Step 43: Establish a machine learning regression model: Select and establish a machine learning regression model; the output parameters of the machine learning regression model are the input model parameters of the semiconductor device simulation model; the input parameters of the machine learning regression model are the process parameters or structural parameters of the semiconductor device. Step 44, Model Training: Use the data from the training set in Step 42 to train the machine learning regression model established in Step 43, and obtain the weight values ​​of the machine learning regression model. Step 45, Model Validation and Testing: Using the data from the validation set or test set in Step 42, validate or test the machine learning regression model trained in Step 44.

7. The method of claim 1, wherein: In step 5, the method for predicting the input model parameters of semiconductor devices is as follows: Step 51, Data Processing: Standardize or normalize the structural or process parameters of the semiconductor device to be simulated; Step 52, Prediction: Input the structural or process parameters of the semiconductor device to be simulated after data processing in Step 51 into the machine learning regression model established in Step 4 to predict the input model parameters corresponding to the semiconductor device to be simulated. Step 53, Reverse Data Processing: Denormalize or denormalize the input model parameters predicted in Step 52 to obtain the input model parameters of the simulation model.

Images