HEMT device IV curve fitting method and system based on hybrid modeling

Abstract

The invention discloses an HEMT device IV curve fitting method and system based on hybrid modeling, and the method comprises the steps: building an initial IV curve analysis model based on the physical mechanism of an HEMT device, and generating an initial IV curve through the initial IV curve analysis model; s2, inputting the initial IV curve and an error term of the experimental data into a full-connection neural network, and performing nonlinear correction on the error term by using the full-connection neural network to generate a correction coefficient; s3, combining the correction coefficient with physical parameters of the initial IV curve to generate an analytic correction IV formula; s4, the corrected analytic correction IV formula is packaged into an SPICE sub-circuit model, so that the SPICE sub-circuit model can support simulation of a high-frequency or power circuit, according to the HEMT device IV curve fitting method and system based on hybrid modeling, the physical significance of a traditional model can be reserved, the complex effect can be dynamically corrected through a neural network, and the method is a hybrid modeling method. And finally, an analytic IV equation which can be directly integrated to SPICE is generated.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor devices, and specifically to a method and system for fitting the IV curve of HEMT devices based on hybrid modeling. Background Art

[0002] High Electron Mobility Transistors (HEMTs) have been widely used in the fields of radio frequency and power electronics due to their high-frequency and high-power characteristics. The modeling of the current-voltage (IV) characteristics of traditional HEMT devices mostly relies on physical analytical models (such as surface potential models, carrier transport equations). Its advantage lies in strong physical interpretability, and it can intuitively reflect the internal charge transport mechanism of the device. However, with the increase in device complexity (such as trap effects, self-heating effects, etc. in gallium nitride HEMTs) and the characteristic differences brought by devices from different manufacturers and different batches, it has become difficult for traditional models to accurately describe the non-linear IV characteristics of HEMT devices. Therefore, there are significant deviations between traditional models and experimental data.

[0003] To solve this problem, researchers have tried to introduce pure data-driven neural network models (such as patent CN112084657A). Although such methods can fit the IV curve through data, they have two major defects: First, the model lacks physical interpretability, which limits its application and promotion in engineering practice. Second, the "black box" characteristic of the neural network makes it difficult to reuse the corrected formula and it cannot be directly adapted to circuit simulation tools such as the integrated circuit simulation program (SPICE).

[0004] In recent years, hybrid modeling methods (combining traditional models with data-driven corrections) have gradually received attention. For example, patent CN104915522A performs hybrid modeling by combining process priors and data-driven methods, but its correction strategy depends on fixed rules (constraint equations) and it is difficult to dynamically adapt to the non-linear responses in different working intervals. In addition, the correction results of existing hybrid models are mostly numerical outputs and cannot generate analytical physical expressions, resulting in insufficient compatibility with SPICE tools. Summary of the Invention

[0005] The purpose of this part is to outline some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Simplifications or omissions may be made in this part, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this part, the abstract, and the title, and such simplifications or omissions shall not be used to limit the scope of the present invention.

[0006] Therefore, the purpose of the present invention is to provide a method and system for fitting the IV curve of HEMT devices based on hybrid modeling, which can not only retain the physical meaning of traditional models, but also dynamically correct complex effects through neural networks in a hybrid modeling method, and finally generate an analytical IV equation that can be directly integrated into SPICE.

[0007] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0008] A method for fitting the IV curve of a HEMT device based on hybrid modeling, the steps are as follows:

[0009] S1. Based on the physical mechanism of the HEMT device, establish an initial analytical model of the IV curve, and generate an initial IV curve through the initial analytical model of the IV curve;

[0010] S2. Input the error term between the initial IV curve and the experimental data into a fully connected neural network, and use the fully connected neural network to perform non-linear correction on the error term to generate a correction coefficient;

[0011] S3. Combine the correction coefficient with the physical parameters of the initial IV curve to generate an analytical corrected IV formula;

[0012] S4. Package the corrected analytical corrected IV formula into a SPICE sub-circuit model so that it can support the simulation of high-frequency or power circuits.

[0013] As a preferred solution of a method for fitting the IV curve of a HEMT device based on hybrid modeling according to the present invention, wherein the initial analytical model of the IV curve includes an output characteristic model and a transfer characteristic model;

[0014] The output characteristic model adopts a multi-parameter non-linear empirical formula:

[0015] Among them, the parameters a, V t0 , n, b0 to b6 are determined by fitting experimental data and are used to describe the forward and reverse output characteristic curves;

[0016] The transfer characteristic model is based on an electrothermal empirical model:

[0017] Among them, the parameter K1 is the saturation current coefficient, and the parameter c1 determines the transition shape from the linear region to the saturation region.

[0018] As a preferred solution of a method for fitting the IV curve of a HEMT device based on hybrid modeling according to the present invention, wherein in step S2, a fully connected neural network with 14 neurons is used to perform non-linear correction on the error term of the initial analytical model of the IV curve.

[0019] As a preferred solution of a method for fitting the IV curve of a HEMT device based on hybrid modeling according to the present invention, wherein the structure of the fully connected neural network includes:

[0020] Input layer: used to receive the V output by the initial IV curve analysis model ds and V gs ;

[0021] Hidden layer: including 2 fully connected layers, 14 neurons, and the ReLU activation function, used to correct the error term of the initial IV curve analysis model;

[0022] Output layer: used to output the corrected I ds predicted value.

[0023] As a preferred solution of the HEMT device IV curve fitting method based on hybrid modeling according to the present invention, in step S2, the training configuration of the fully connected neural network is as follows: the expression of the fully connected layer of the neural network is y = W2·σ(W1x + b1) + b2, where W1 ∈ R N×2 , W2 ∈ R 1×N , using the mean squared error MSE as the loss function, and combining the curve smoothness constraint where is the standard deviation of the first derivative, used to constrain the curve smoothness, is the sum of the absolute values of the second derivatives, used to constrain the curvature change;

[0024] The optimizer is the Adam algorithm, the learning rate is 0.001, iterate 50,000 times, and the input normalization is Z-score normalization where μ is the mean and σ is the standard deviation.

[0025] As a preferred solution of the HEMT device IV curve fitting method based on hybrid modeling according to the present invention, when using the fully connected neural network to perform non-linear correction on the error term, the weights of the fully connected neural network are dynamically adjusted according to the intervals of the gate voltage V gs and the drain voltage V ds input by the initial IV curve analysis model.

[0026] A HEMT device IV curve fitting system based on hybrid modeling, which includes:

[0027] Physical model module: Based on the physical mechanism of the HEMT device, an initial IV curve analysis model is established, and an initial IV curve is generated through the initial IV curve analysis model;

[0028] Neural network correction module: uses a fully connected neural network to perform non-linear correction on the error term and generate correction coefficients;

[0029] Analysis formula generation module: used to combine the correction coefficients with the physical parameters of the initial IV curve to generate an analytical corrected IV formula;

[0030] The SPICE encapsulation module is used to encapsulate the corrected analytical IV formula into a SPICE sub - circuit model, enabling it to support the simulation of high - frequency or power circuits.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: Through the hybrid modeling framework, the present invention significantly improves the fitting accuracy of the IV curve (for example, the error in the saturation region is reduced to less than 3%), while retaining the interpretability of the physical model, providing a highly reliable tool for circuit design and simulation of GaN HEMT in the fields of high - frequency communication, power electronics, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the drawings and detailed embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings. Among them:

[0033] Figure 1 It is a flowchart of the hybrid model combining the initial IV - curve analytical model and the fully - connected neural network provided by the present invention;

[0034] Figure 2 It is a fitting comparison diagram between the traditional IV - curve analytical model and the hybrid model combining the initial IV - curve analytical model and the fully - connected neural network provided by the present invention. Among them, (a) is the traditional IV - curve analytical model, and (b) is the hybrid model combining the initial IV - curve analytical model and the fully - connected neural network;

[0035] Figure 3 It is a curve diagram showing the relationship between the number of neurons and the smoothness loss provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] To make the above - mentioned objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be made with reference to the drawings.

[0037] Secondly, the present invention will be described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross - sectional views showing the device structure will be enlarged locally out of the general proportion, and the schematic diagrams are only examples and should not limit the scope of protection of the present invention herein. In addition, in actual production, three - dimensional spatial dimensions including length, width, and depth should be included.

[0038] To make the purpose, technical solutions, and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the drawings.

[0039] The present invention provides a method and system for fitting the IV curve of a HEMT device based on hybrid modeling, which can not only retain the physical meaning of the traditional model, but also dynamically correct complex effects through a neural network, and finally generate an analytical IV equation that can be directly integrated into SPICE.

[0040] Among them, the method for fitting the IV curve of a HEMT device based on hybrid modeling is as follows:

[0041] S1. Based on the physical mechanism of the HEMT device, establish an initial analytical model of the IV curve, and generate an initial IV curve through the initial analytical model of the IV curve;

[0042] S2. Input the error term between the initial IV curve and the experimental data into a fully connected neural network, and use the fully connected neural network to perform non-linear correction on the error term to generate a correction coefficient. When using the fully connected neural network to perform non-linear correction on the error term, dynamically adjust the weights of the fully connected neural network according to the intervals of the gate voltage V gs and the drain voltage V ds input by the initial analytical model of the IV curve. In this embodiment, the structure of the fully connected neural network includes:

[0043] Input layer: used to receive V ds and V gs output by the initial analytical model of the IV curve;

[0044] Hidden layer: including 2 fully connected layers, 14 neurons, and a ReLU activation function, used to correct the error term of the initial analytical model of the IV curve;

[0045] Output layer: used to output the predicted value of the corrected I ds [[ID=2 "]]Predicted value

[0046] S3. Combine the correction coefficient with the physical parameters of the initial IV curve to generate an analytical corrected IV formula;

[0047] S4. Package the corrected analytical corrected IV formula into a SPICE sub-circuit model so that it can support the simulation of high-frequency or power circuits.

[0048] In this embodiment, the initial analytical model of the IV curve includes an output characteristic model and a transfer characteristic model;

[0049] The output characteristic model adopts a multi-parameter non-linear empirical formula:

[0050] Among them, the parameters a, V t0 , n, b0 to b6 are determined by fitting experimental data and are used to describe the forward and reverse output characteristic curves;

[0051] The transfer characteristic model is based on the electrothermal empirical model:

[0052] Among them, the parameter K1 is the saturation current coefficient, and the parameter c1 determines the transition shape from the linear region to the saturation region.

[0053] In this embodiment, the fully connected neural network training is configured as follows: the expression of the fully connected layer of the neural network is y = W2·σ(W1x + b1)+b2, where W1 ∈ R N×2 , W2 ∈ R 1×N , and the mean squared error MSE is used as the loss function, combined with the curve smoothness constraint Among them is the standard deviation of the first derivative, which is used to constrain the curve smoothness, is the sum of the absolute values of the second derivatives, which is used to constrain the curvature change;

[0054] The optimizer is the Adam algorithm, the learning rate is 0.001, the iteration is 50000 times, and the input normalization is Z-score normalization where μ is the mean and σ is the standard deviation.

[0055] The present invention also provides an HEMT device IV curve fitting system based on hybrid modeling, which is characterized by including:

[0056] A physical model module, based on the physical mechanism of the HEMT device, establishes an initial IV curve analytical model, and generates an initial IV curve through the initial IV curve analytical model;

[0057] A neural network correction module, which uses a fully connected neural network to perform non-linear correction on the error term and generates a correction coefficient;

[0058] An analytical formula generation module, which is used to combine the correction coefficient with the physical parameters of the initial IV curve to generate an analytical corrected IV formula;

[0059] A SPICE packaging module, which is used to package the corrected analytical corrected IV formula into a SPICE sub-circuit model so that it can support the simulation of high-frequency or power circuits.

[0060] In order to verify the technical effects of the above solutions of the present invention, the following experiments are provided for verification.

[0061] A GaN HEMT chip (model INN650D150A) manufactured by Inno-science is used, and the packaging form is DFN8×8, which is suitable for high-power scenarios. The measurement device uses a Keysight B1500A semiconductor analyzer to measure the I-V characteristics of the device at room temperature, including: the forward output characteristic curve (Vds Range: 0 to 20 V); Reverse output characteristic curve (V ds Range: -20 to 0 V); Transfer characteristic curve (Vgs range: 0 to 20 V). Input the above gate voltage V gs and drain voltage V ds into the hybrid model combining the initial IV curve analysis model and the fully connected neural network as shown in Figure 1 . Verify the results of the hybrid model by fitting the output characteristic curve. The experimental results are as shown in Figure 2 . Practically, it shows that the error of the hybrid model in the saturation region is reduced to less than 3% (the error of the physical model > 15%); The verification of the transfer characteristic curve shows that after the correction of the fully connected layer of the neural network, although the adjustment range of the parameter c1 of the physical model is reduced, the overall change trend still satisfies physical interpretability, and R 2 > 0.99.

[0062] In addition, in order to verify the influence of the number of neurons in the fully connected neural network on the smoothness loss of the test neurons, optimize the test on the number of neurons in the fully connected neural network. As shown in Figure 3 , the results show that the loss function is the smallest when there are 14 neurons (Loss = 0.023), which is better than 8 neurons (Loss = 0.041) and 20 neurons (Loss = 0.038). 14 neurons are the best configuration, balancing the model complexity and the anti-overfitting ability.

[0063] Although the present invention has been described above with reference to the embodiments, various improvements can be made to it and components thereof can be replaced with equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention can be combined with each other in any way. The exhaustive description of these combinations is omitted in this specification only for the sake of saving space and resources. Therefore, the present invention is not limited to the specific embodiments disclosed in the text, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for fitting the IV curve of a HEMT device based on hybrid modeling, characterized in that, The steps are as follows: S1. Based on the physical mechanism of HEMT devices, establish an initial analytical IV curve model, and generate an initial IV curve through the initial analytical IV curve model; S2. Input the error term between the initial IV curve and experimental data into a fully connected neural network, and use the fully connected neural network to perform non-linear correction on the error term to generate a correction coefficient; S3. Combine the correction coefficient with the physical parameters of the initial IV curve to generate an analytical correction IV formula; S4. Package the corrected analytical correction IV formula into a SPICE sub-circuit model so that it can support the simulation of high-frequency or power circuits.

2. A method for fitting the IV curve of a HEMT device based on hybrid modeling according to claim 1, characterized in that, The initial analytical IV curve model includes an output characteristic model and a transfer characteristic model; The output characteristic model adopts a multi-parameter non-linear empirical formula: Among them, the parameters a, V t0 , n, b0 to b6 are determined by fitting experimental data and are used to describe the forward and reverse output characteristic curves; The transfer characteristic model is based on the electrothermal empirical model: Among them, the parameter K1 is the saturation current coefficient, and the parameter c1 determines the transition shape from the linear region to the saturation region.

3. A method for fitting the IV curve of a HEMT device based on hybrid modeling according to claim 1, characterized in that, In step S2, a fully connected neural network with 14 neurons is used to perform non-linear correction on the error term of the initial analytical IV curve model.

4. A method for fitting the IV curve of a HEMT device based on hybrid modeling according to claim 1, wherein The structure of the fully connected neural network includes: Input layer: used to receive V output by the initial IV curve analysis model ds and V gs ; Hidden layer: including 2 fully connected layers, 14 neurons, and a ReLU activation function, which is used to correct the error term of the initial analytical IV curve model; Output layer: used to output the corrected I ds predicted value.

5. A method for fitting the IV curve of a HEMT device based on hybrid modeling according to claim 1, characterized in that In step S2, the training configuration of the fully connected neural network is as follows: The expression of the fully connected layer of the neural network is y = W2·σ(W1x + b1) + b2, where W1 ∈ R N×2 , W2 ∈ R 1×N , the mean squared error MSE is used as the loss function, and combined with the curve smoothness constraint where is the standard deviation of the first derivative, used to constrain the curve smoothness, is the sum of the absolute values of the second derivatives, used to constrain the curvature change; The optimizer is the Adam algorithm, with a learning rate of 0.001, iterating 50,000 times, and the input is normalized by Z-score normalization where μ is the mean and σ is the standard deviation.

6. The method for fitting the IV curve of a HEMT device based on hybrid modeling according to claim 1, characterized in that, When using a fully connected neural network to perform non-linear correction on the error term, the gate voltage V gs and drain voltage V ds input according to the initial IV curve analytical model are used to dynamically adjust the weights of the fully connected neural network.

7. A HEMT device IV curve fitting system based on hybrid modeling, characterized in that, including: Physical model module: Based on the physical mechanism of HEMT devices, establish an initial analytical IV curve model, and generate an initial IV curve through the initial analytical IV curve model; Neural network correction module: Use a fully connected neural network to perform non-linear correction on the error term to generate a correction coefficient; Analytical formula generation module: used to combine the correction coefficient with the physical parameters of the initial IV curve to generate an analytical correction IV formula; SPICE packaging module: used to package the corrected analytical correction IV formula into a SPICE sub-circuit model so that it can support the simulation of high-frequency or power circuits.

Images