Dnn-based large signal model modeling method, system, device and storage medium
By using a segmented modeling method based on DNN, the input and output layer data of GaN microwave devices are extracted, and deep neural network models for high and low power regions are trained. This solves the problem of low accuracy in large signal modeling models and achieves accurate fitting of high and low power regions, making it suitable for precise modeling of transistors and integrated circuit modules.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 长三角集成电路工业应用技术创新中心
- Filing Date
- 2025-03-11
- Publication Date
- 2026-05-12
AI Technical Summary
Existing large-signal modeling models for GaN microwave devices have low accuracy, especially in the low-power and high-power regions where performance curves exhibit abrupt changes, affecting the accuracy of the models.
A large-signal modeling method based on DNN is adopted. By performing large-signal tests on GaN transistors, the real and imaginary data of the input and output layers are extracted. A high-power region deep neural network model is trained using DNN, and a low-power region deep neural network model is trained by combining amplitude and phase data. The resulting partitioned deep neural network model is transformed into voltage and current characteristics to establish a large-signal behavior model.
It improves the accuracy of large-signal modeling, can simultaneously fit the characteristics of low-power and high-power ranges, expands the dynamic range of the model, and has the ability to fit wideband and multiple bias conditions, making it suitable for accurate modeling at the transistor level and integrated circuit module level.
Smart Images
Figure CN120217983B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of integrated circuit technology, and in particular to a method, system, device and storage medium for large signal modeling based on DNN. Background Technology
[0002] DNN is short for Deep Neural Network. It is a multi-layer neural network structure designed to enhance the expressive power of a model and its ability to learn complex features by stacking multiple hidden layers.
[0003] With the development of modern infrastructure, communications industry, and aerospace technology, increasingly higher demands are being placed on microwave systems for equipment: miniaturization, high temperature resistance, radiation resistance, high power, ultra-high frequency, and suitability for operation in harsh environments. Wide-bandgap semiconductor materials and devices, represented by GaN and SiC, have become a research hotspot. Researching and developing high-performance semiconductor materials and devices that can operate at higher frequencies and with greater power is of great significance.
[0004] With the continuous improvement of epitaxial material quality and the refinement of device fabrication processes, the development of AlGaN / GaN HEMT devices has been rapid. Due to their spontaneous polarization and piezoelectric polarization effects, AlGaN / GaN heterojunctions can generate highly mobile two-dimensional electron gases, which can be controlled by gate voltage. In recent years, the device's performance characteristics have improved rapidly, especially the microwave power characteristics of AlGaN / GaN HEMT devices, with significant increases in output power and power density. Besides manufacturing processes and device characteristics, modeling AlGaN / GaN HEMTs has also been a key research focus. Due to their prominent applications in the radio frequency and microwave fields, AlGaN / GaN HEMT device modeling has always been an important part of device research.
[0005] Device models include small-signal models and large-signal models. For small-signal models, both domestically and internationally, significant progress has been made in small-signal modeling, with the S-parameter method being the most widely used.
[0006] Large-signal modeling has always been a challenge in microwave power device analysis. Commonly used GaN HEMT device models can generally be categorized into: empirical analytical models, physical models, tabular-based models, and behavioral models. Existing models suffer from abrupt changes in transistor performance curves between low and high power regions, which severely impacts the accuracy of large-signal modeling, resulting in low model accuracy. Summary of the Invention
[0007] To improve the accuracy of large-signal modeling for GaN microwave devices, this application provides a large-signal modeling method, system, device, and storage medium based on DNN.
[0008] Firstly, this application provides a large signal modeling method based on DNN, employing the following technical solution:
[0009] Large-signal testing is performed on the GaN transistor to obtain its line wave characteristics, which include input layer characteristics and output layer characteristics.
[0010] The input layer characteristics are processed to obtain input layer data, and the output layer characteristics are processed to obtain output layer data;
[0011] Extract the real and imaginary parts of the input layer data and the output layer data, and use a DNN to train a high-power regional deep neural network model based on the real and imaginary parts.
[0012] The amplitude and phase data of the input layer data and the output layer data are extracted, and a low-power region deep neural network model is obtained by training a DNN based on the amplitude and phase data.
[0013] The line wave characteristics are trained based on the high-power region deep neural network model and the low-power region deep neural network model to obtain the trained line wave characteristics, and the trained line wave characteristics are converted into voltage and current characteristics.
[0014] A large-signal behavior model is constructed based on the voltage and current characteristics.
[0015] Through the above technical solutions, segmented modeling can simultaneously fit the characteristics of large signals in both low-power and high-power ranges. The use of deep neural networks to construct behavioral models highlights the weights of traveling wave parameters in high-power and low-power regions, enhances the learning sensitivity of neural networks, and improves the accuracy of large-signal modeling.
[0016] In one specific implementation, the GaN transistor undergoes large-signal testing to obtain line wave characteristics, including:
[0017] Obtain the fundamental and harmonic conditions for testing;
[0018] Under the fundamental wave condition and the harmonic wave condition, the GaN transistor is scanned by an input power analyzer to obtain the scan results;
[0019] The line wave characteristics are obtained by reading the scan results through a coupler.
[0020] By employing the above technical solution, large-signal testing can be performed under fixed fundamental and harmonic scanning range conditions using vector load traction testing, resulting in more accurate traveling wave testing of transistors.
[0021] In one specific implementation, processing the input layer characteristics to obtain input layer data and processing the output layer characteristics to obtain output layer data includes:
[0022] Calculate the target phase of the fundamental incident wave at the input port;
[0023] The input layer data is obtained by normalizing the remaining incident wave according to the target phase.
[0024] The reflected wave is processed according to the target phase normalization to obtain the output layer data.
[0025] The above technical solution transforms the dynamically changing input and output layer characteristics into time-invariant data, eliminating the periodic changes in phase data and making it easier to compare and analyze.
[0026] In one specific implementation, the extraction of the real and imaginary parts of the input layer data and the output layer data includes:
[0027] Extract the target amplitude of the fundamental incident wave at the input port, extract the real and imaginary data of the input layer data, and extract the real and imaginary data of the output layer data;
[0028] The step of training a high-power region deep neural network model using a DNN based on the real part data and the imaginary part data includes:
[0029] The target amplitude, the real and imaginary parts of the input layer data, and the real and imaginary parts of the output layer data are normalized to obtain high-power region training data;
[0030] The high-power region training data is imported into a preset first neural network for training and a first loss value is calculated.
[0031] Determine whether the first loss value is less than a preset value;
[0032] If the first loss value is less than the preset value, then training is stopped and a high-power region deep neural network model is obtained based on the trained first neural network.
[0033] The above technical solution involves extracting real and imaginary data to train a high-power region deep neural network model, which is then processed in a targeted manner to improve the accuracy of large signal modeling.
[0034] In one specific implementation, the extraction of amplitude and phase data from the input layer data and the output layer data includes:
[0035] Extract the dB amplitude of the fundamental incident wave at the input port, wherein the unit of the dB amplitude is dB; extract the amplitude data and phase data of the input layer data; extract the amplitude data and phase data of the output layer data.
[0036] The step of training a low-power region deep neural network model using a DNN based on the amplitude data and the phase data includes:
[0037] The dB amplitude, the amplitude data and phase data of the input layer data, and the amplitude data and phase data of the output layer data are normalized to obtain low-power region training data;
[0038] The training data of the low-power region is imported into a preset second neural network for training, and a second loss value is calculated.
[0039] Determine whether the second loss value is less than the second preset value;
[0040] If the second loss value is less than the second preset value, then training is stopped and a low-power region deep neural network model is obtained based on the trained second neural network.
[0041] The above technical solution extracts amplitude and phase data and trains a deep neural network model for low-power regions, which is then processed in a targeted manner to improve the accuracy of large signal modeling.
[0042] In one specific implementation, the step of training the line wave characteristics based on the high-power region deep neural network model and the low-power region deep neural network model to obtain the trained line wave characteristics includes:
[0043] Set the determination points for high and low power regions;
[0044] Construct a partitioned deep neural network model based on the decision points, the high-power region deep neural network model, and the low-power region deep neural network model;
[0045] The row wave characteristics are trained based on the partitioned deep neural network model and then inversely normalized to obtain the trained row wave characteristics.
[0046] By using the above technical solution, the trained high-power region deep neural network model and low-power region deep neural network model can be used to generate an overall partitioned deep neural network model by setting high and low power region judgment points and a single-pole double-throw switch. This allows the line wave characteristics of high-power and low-power regions to be trained at a large order of magnitude.
[0047] In one specific implementation, after constructing the large-signal behavior model based on the voltage and current characteristics, the method further includes:
[0048] Receive the verification row baud characteristic, wherein the verification row baud characteristic includes the actual data of the verification row baud characteristic;
[0049] Simulation data was obtained by simulating the verification row wave characteristics using the large signal behavior model:
[0050] Determine whether the error between the actual data and the simulation data is within the allowable range;
[0051] If the error is within the allowable range, the large signal behavior model is deemed ready for use.
[0052] If the error is outside the allowable range, the large signal behavior model is deemed unusable.
[0053] The above technical solution is used to simulate and compare the obtained large signal behavior model, verify the function of the large signal behavior model, and ensure the accuracy of the large signal behavior model.
[0054] Secondly, this application provides a large-signal modeling system based on DNN, employing the following technical solution: the system is based on a GaN microwave device including GaN transistors, and the system includes:
[0055] A line wave characteristic acquisition module is used to perform large signal testing on the GaN transistor to obtain the line wave characteristics, which include input layer characteristics and output layer characteristics;
[0056] A line wave characteristic processing module is used to process the input layer characteristics to obtain input layer data and process the output layer characteristics to obtain output layer data;
[0057] A high-power region model training module is used to extract the real and imaginary parts of the input layer data and the output layer data, and to train a high-power region deep neural network model using a DNN based on the real and imaginary parts of the data.
[0058] The low-power region model training module is used to extract the amplitude and phase data of the input layer data and the output layer data, and to train a low-power region deep neural network model using a DNN based on the amplitude and phase data.
[0059] The voltage-current characteristic conversion module is used to train the row wave characteristics based on the high-power region deep neural network model and the low-power region deep neural network model to obtain trained row wave characteristics, and convert the trained row wave characteristics into voltage-current characteristics.
[0060] The behavior model construction module is used to construct a large-signal behavior model based on the voltage and current characteristics.
[0061] Thirdly, this application provides a computer device that adopts the following technical solution: it includes a memory and a processor, wherein the memory stores a computer program that can be loaded and executed by the processor, such as the above-described method for modeling large signal models based on DNN.
[0062] Fourthly, this application provides a computer-readable storage medium, which employs the following technical solution: storing a computer program that can be loaded by a processor and executed as described above regarding a DNN-based large signal model modeling method.
[0063] In summary, this application has the following beneficial technical effects:
[0064] (1) By using segmented modeling, the characteristics of large signals in both low-power and high-power ranges can be fitted simultaneously. The use of deep neural networks to construct behavioral models highlights the weights of traveling wave parameters in high-power and low-power regions, enhances the learning sensitivity of neural networks, and improves the accuracy of large-signal modeling models.
[0065] (2) A behavior model is constructed using a deep neural network. The input and output layers are processed by complex real / imaginary parts and amplitude / phase algorithms, which overcomes the limitation of neural networks in fitting data with abrupt changes. The line wave characteristics of high-power and low-power regions can be trained at a large order of magnitude, thereby expanding the dynamic range of the model.
[0066] (3) By adding frequency and voltage variables to the input layer of the neural network and combining them with the partitioning algorithm modeling method, the behavioral model can fit wide-band and multi-bias conditions.
[0067] (4) The model constructed in this application can be modeled not only at the transistor level, but also at the integrated circuit module level to complete the system design; it is also compatible with various integrated circuit design software, making it convenient for circuit and system engineers to call and simulate without additional hardware and software drivers. Attached Figure Description
[0068] Figure 1 This is a flowchart of an embodiment of this application.
[0069] Figure 2 This is a cross-sectional view of the GaN device under test.
[0070] Figure 3 This is an example diagram of the fundamental and second harmonic sweep at the output port.
[0071] Figure 4 This is a dynamic scanning range diagram of input power.
[0072] Figure 5 It is the row wave characteristic of the real / imaginary part pattern of complex numbers in linear coordinates.
[0073] Figure 6 It is the row wave characteristic of the real / imaginary part pattern of complex numbers in logarithmic coordinates.
[0074] Figure 7 It is the line wave characteristic of the amplitude / phase mode in linear coordinates.
[0075] Figure 8 It is the line wave characteristic of the amplitude / phase mode in logarithmic coordinates.
[0076] Figure 9 These are simulation curves of the model in the complex real / imaginary part scheme.
[0077] Figure 10 These are the simulation curves of the model in the amplitude / phase format.
[0078] Figure 11 These are simulation results of the partitioned deep neural network behavior model.
[0079] Figure 12 It is a deep neural network model framework.
[0080] Figure 13 Simulation curves of the line wave characteristics are used to verify the behavior model of large signals.
[0081] Figure 14 This is a simulation of a GaN power amplifier module.
[0082] Figure 15 This is a structural block diagram of an embodiment of this application.
[0083] Figure reference numerals: 1501, Linear Wave Characteristic Acquisition Module; 1502, Linear Wave Characteristic Processing Module; 1503, High-Power Region Model Training Module; 1504, Low-Power Region Model Training Module; 1505, Voltage-Current Characteristic Conversion Module; 1506, Behavioral Model Construction Module. Detailed Implementation
[0084] The following is in conjunction with the appendix Figures 1-15 This application will be described in further detail.
[0085] This application discloses a large-signal modeling method based on DNN, which is used to improve the accuracy of large-signal modeling models for GaN microwave devices.
[0086] DNN is short for Deep Neural Network. It is a multi-layer neural network structure designed to enhance the expressive power of a model and its ability to learn complex features by stacking multiple hidden layers.
[0087] With the development of modern infrastructure, communications industry, and aerospace technology, increasingly higher demands are being placed on microwave systems for equipment: miniaturization, high temperature resistance, radiation resistance, high power, ultra-high frequency, and suitability for operation in harsh environments. Wide-bandgap semiconductor materials and devices, represented by GaN and SiC, have become a research hotspot. Researching and developing high-performance semiconductor materials and devices that can operate at higher frequencies and with greater power is of great significance.
[0088] With the continuous improvement of epitaxial material quality and the refinement of device fabrication processes, the development of AlGaN / GaN HEMT devices has been rapid. Due to their spontaneous polarization and piezoelectric polarization effects, AlGaN / GaN heterojunctions can generate highly mobile two-dimensional electron gases, which can be controlled by gate voltage. In recent years, the device's performance characteristics have improved rapidly, especially the microwave power characteristics of AlGaN / GaN HEMT devices, with significant increases in output power and power density. Besides manufacturing processes and device characteristics, modeling AlGaN / GaN HEMTs has also been a key research focus. Due to their prominent applications in the radio frequency and microwave fields, AlGaN / GaN HEMT device modeling has always been an important part of device research.
[0089] Device models include small-signal models and large-signal models. For small-signal models, both domestically and internationally, significant progress has been made in small-signal modeling, with the S-parameter method being the most widely used.
[0090] Large-signal modeling has always been a challenge in the analysis of microwave power devices. Commonly used GaN HEMT device models can generally be divided into the following types:
[0091] I. Empirical analytical models, also known as equivalent circuit models, are derived by employing appropriate testing and parameter extraction techniques, along with parameter fitting and optimization methods, to create various device models required for fabricating monolithic circuits, for different devices manufactured using actual processes. Examples include the EEHEMT model and the Angelov model. These models consist of parasitic parameters and intrinsic parameters. Intrinsic capacitance and current models are fitted to test data using analytical formulas. Their advantages include simple functional relationships and algorithms, making them highly suitable for large-signal RF amplifiers. However, these models require the extraction of a certain number of fitting parameters, and the higher the model's accuracy, the more parameters need to be fitted, thus increasing the complexity of the expressions. A trade-off between model complexity and accuracy is often necessary. Furthermore, because the formulas lack physical meaning, they cannot accurately fit non-ideal effects discovered during the characterization process.
[0092] II. Physical Models. Commonly used physical models include the ASM model based on surface potential and the MVSG model based on charge. Compared with equivalent circuit models, the current and charge expressions of physical models are modeled based on the structural parameters of the device and semiconductor physics formulas, thus accurately fitting the DC and AC characteristics of the device. However, for integrated circuit design companies without process lines, it is difficult to obtain the physical parameters of the device structure, and the material parameters of the epitaxial layer are even more unknown, thus greatly limiting the use of physical models. In addition, the coupling between its DC and AC parameters poses challenges for parameter extraction and optimization in modeling.
[0093] Third, the tabular-based model, also known as the data-based model, is a model built on a large number of tests. Because it is a true reflection of the device's electrical parameters, it is more closely integrated with the manufacturing process and more closely resembles the actual device characteristics. However, the tabular-based model is less flexible and practical than the empirical analytical model. Accurate modeling requires measuring more data, which reduces its scalability and increases the difficulty of statistical averaging models.
[0094] IV. Behavioral Model. By black-boxing the transistor and measuring the large-signal characteristics in the frequency or time domain, the characteristics of the device ports are obtained, thereby establishing a behavioral model, which has the highest accuracy, similar to the tabular basis model.
[0095] Commonly used behavioral models include the X-parameter model and the Cardiff model, with the Cardiff model being more extensively researched and applied. The basic principle of the Cardiff model is the harmonic superposition and harmonic mixing principle. It extracts model coefficients by setting a large-signal operating point and then calculates the model's performance at that point using a lookup table. However, this method struggles to achieve wide bandwidth and multiple bias conditions, and the lookup table method lacks accurate formulas for interpolation between test points, thus compromising accuracy. Another approach is to model behavioral behavior using artificial neural networks in the frequency-traveling wave domain. However, due to the abrupt changes in transistor performance curves between low and high power regions, artificial neural networks cannot fit this operating mode, thus reducing the dynamic range of input power and limiting its application under multiple bias conditions.
[0096] By creating a behavioral model to black-box transistors, problems such as low accuracy, unknown physical parameters, and difficulty in parameter extraction associated with empirical and physical models can be solved. However, some issues still exist. For example, the abrupt changes in the performance curves of transistors in the low-power and high-power regions can severely affect the accuracy of large-signal modeling, resulting in low model accuracy.
[0097] Therefore, this application proposes a large-signal modeling method based on DNN, which improves the accuracy of large-signal modeling for GaN microwave devices.
[0098] like Figure 1 As shown, the method includes:
[0099] S10, large-signal testing is performed on the GaN transistor to obtain the line wave characteristics, which include input layer characteristics and output layer characteristics.
[0100] Specifically, in the traveling wave domain of the frequency domain, large-signal tests are performed on GaN transistors to obtain traveling wave characteristics, frequency, and voltage. The traveling wave characteristics include input layer characteristics and output layer characteristics. The input layer characteristics include the incident wave, and the output layer characteristics include the reflected wave.
[0101] S20: Process the input layer characteristics to obtain input layer data, and process the output layer characteristics to obtain output layer data.
[0102] Specifically, the input layer characteristics and output layer characteristics are normalized. The line wave characteristics are converted into time-invariant data by phase normalization of the incident wave of the fundamental wave at the input port, thus obtaining the input layer data and the output layer data.
[0103] S30: Extract the real and imaginary parts of the input and output layer data, and use a DNN to train a high-power region deep neural network model based on the real and imaginary parts.
[0104] Specifically, the performance curves of transistors exhibit abrupt changes in the low-power and high-power regions. Therefore, partitioned training is performed. In the high-power region of the line wave characteristics, the real and imaginary parts of complex numbers are used for partitioned deep neural network training. That is, the real and imaginary parts of the input and output layer data of the high-power region are extracted, and the deep neural network model of the high-power region is obtained by training the deep neural network based on the real and imaginary parts.
[0105] S40: Extract amplitude and phase data from the input and output layers, and use DNN to train a low-power region deep neural network model based on the amplitude and phase data.
[0106] Specifically, the performance curves of transistors exhibit abrupt changes in the low-power and high-power regions. Therefore, partitioned training is performed. In the low-power region of the line wave characteristics, amplitude / phase is used for partitioned deep neural network training. That is, amplitude and phase data of the input layer data and output layer data of the low-power region are extracted, and the deep neural network model of the low-power region is obtained by training the deep neural network based on the amplitude and phase data.
[0107] S50 obtains the training line wave characteristics based on the training line wave characteristics of the high-power region deep neural network model and the low-power region deep neural network model, and converts the training line wave characteristics into voltage and current characteristics.
[0108] Specifically, the training row wave characteristics of the trained partitioned deep neural network model are converted into voltage and current characteristics, which facilitates the establishment of behavioral models.
[0109] S60, based on voltage and current characteristics, constructs a large-signal behavior model.
[0110] Specifically, in the ADS circuit design software, the FDD control is used to define the port voltage and current, and a behavioral model is built based on the voltage and current.
[0111] By using segmented modeling, the characteristics of large signals in both low-power and high-power ranges can be fitted simultaneously. The behavioral model constructed using deep neural networks highlights the weights of traveling wave parameters in the high-power and low-power regions, enhances the learning sensitivity of the neural network, and improves the accuracy of the large-signal model.
[0112] In one embodiment, to improve the accuracy of the large-signal modeling model of GaN microwave devices, the step of obtaining the line wave characteristics by performing large-signal testing on GaN transistors can be specifically performed as follows:
[0113] First, obtain the fundamental and harmonic conditions for testing. Specifically, for GaN devices such as... Figure 2 The diagram shows a fixed frequency, voltage, and input impedance. To obtain sample data for training, the output impedance tuner is changed to scan within a range near the fundamental impedance point where power and efficiency are optimal. The fundamental impedance point and the second harmonic impedance point are then found, and the fundamental and harmonic conditions used for testing are obtained.
[0114] Next, under fundamental and harmonic conditions, the GaN transistor is scanned using a vector network analyzer to obtain the scan results. Specifically, for example... Figure 3 As shown, for each fundamental impedance point, the second harmonic impedance point is scanned from 0° to 360° with a fixed phase step within a range where the reflection coefficient amplitude is greater than 0.8, to obtain the scanning result.
[0115] Then, the line wave characteristics are obtained by reading the scan results through a coupler. Specifically, under fixed fundamental and harmonic conditions, the input power is scanned from the linear region to the nonlinear region with gain compression. The large-signal characteristics of the transistor show significant abrupt changes in the low-power and high-power regions, such as... Figure 4 As shown, the horizontal axis represents power magnitude, and the vertical axis represents line wave characteristics. It can be seen that there is a significant abrupt change in line wave characteristics between the low-power and high-power regions shown in the figure. The line wave characteristics are obtained by reading data from a vector network analyzer using a coupler. Line wave characteristics include input layer characteristics and output layer characteristics, where the input layer characteristics include the incident wave A... ph Frequency and voltage, output layer characteristics including reflected wave B php represents the port index, p=1 is the input port, p=2 is the output port, h represents the harmonic index, when h=0 is the DC component, h=1 is the fundamental f0 component, and h=2 is the second harmonic 2f0 component.
[0116] By employing vector load traction testing to perform large-signal testing under fixed fundamental and harmonic sweep ranges, more accurate transistor line wave characteristics can be obtained. At the same time, by adding frequency and voltage variables to the input layer and combining it with partitioning algorithm modeling, the behavioral model has the ability to fit wideband and multiple bias conditions.
[0117] In one embodiment, to improve the accuracy of the large-signal modeling model for GaN microwave devices, the step of processing input layer characteristics to obtain input layer data and processing output layer characteristics to obtain output layer data can be specifically performed as follows:
[0118] First, calculate the target phase of the fundamental incident wave at the input port. Specifically, calculate the fundamental incident wave A at the input port. 11 Phase Angle (A) 11 ).
[0119] Then, the input layer data is obtained by normalizing the remaining incident wave according to the target phase, and the output layer data is obtained by normalizing the reflected wave according to the target phase. Specifically, according to Angle(A) 11 Normalization of input layer data, excluding A 11 A other than ph and the output layer data B ph The formula is shown below:
[0120] A ph_norm =A ph / (cos(h*Angle(A 11 ))+i*sin(h*Angle(A 11 )))
[0121] B ph_m =B ph / (cos(h*Angle(A 11 ))+i*sin(h*Angle(A 11 )))
[0122] Among them, A ph_norm and B ph_m For the normalized input layer data (incident wave component) and output layer data (reflected wave component), h>0.
[0123] It should be noted that normalization is performed to convert the line wave characteristics into time-invariant data. Therefore, this step targets the non-DC components and adds the constraint that h > 0. When h = 0, for example, B 20The measured 2-port current values are in A and B. 20 Fixed values do not require normalization.
[0124] The dynamically changing characteristics of the input and output layers are transformed into time-invariant data, eliminating the periodic changes in phase data and making it easier to compare and analyze.
[0125] In one embodiment, to improve the accuracy of the large-signal modeling model for GaN microwave devices, the step of extracting the real and imaginary parts of the input and output layer data can be specifically performed as follows:
[0126] Extract the target amplitude of the fundamental incident wave at the input port, extract the real and imaginary parts of the input layer data, and extract the real and imaginary parts of the output layer data. Specifically, A 11 The phase is used for time-invariant normalization, therefore A 11 With only one variable, the first step is to extract A. 11 The amplitude is calculated using the formula shown below:
[0127] A 11_ag =abs(A 11 )
[0128] Step 2: Extract A ph_m and B ph_rm The formulas for the real and imaginary parts are shown below:
[0129] A ph_ =real(A ph_orm )
[0130] B ph_ =real(B ph_m )
[0131] A ph_ag =imag(A ph_orm )
[0132] B ph_ =imag(B ph_m )
[0133] Where A ph_ B is the real component of the incident wave. ph_ Let A be the real component of the reflected wave. ph_ Let B be the imaginary component of the incident wave. ph_ This is the imaginary component of the reflected wave. For example... Figure 5 and Figure 6The real and imaginary parts of each component of the traveling wave are displayed, where abs(Re(A21)) is the absolute value of the real part of A21, abs(Im(A21)) is the absolute value of the imaginary part of A21, abs(Re(B21)) is the absolute value of the real part of B21, and abs(Im(B21)) is the absolute value of the imaginary part of B21. The horizontal axis represents the power magnitude.
[0134] Figure 5 The ordinate represents the line wave characteristics in a linear coordinate system. Figure 6 The ordinate represents the row wave characteristics in logarithmic coordinates, from Figure 5 As can be seen, there is no significant difference between the low-power and high-power regions under linear coordinates, but... Figure 6 As can be seen, the difference between the low-power and high-power regions in logarithmic coordinates exceeds four orders of magnitude, and this difference is more pronounced in the high-power region. Therefore, complex real / imaginary parts are used to partition the deep neural network for training in the high-power region of the line wave characteristics.
[0135] The step of training a high-power region deep neural network model using a DNN based on real and imaginary data can be specifically performed as follows:
[0136] First, the target amplitude, the real and imaginary parts of the input layer data, and the real and imaginary parts of the output layer data are normalized to obtain high-power region training data. Specifically, the extracted data from the input layer and the extracted data from the output layer are normalized to obtain high-power region training data. The normalization formula is as follows:
[0137]
[0138] Among them, X i It is the i-th value of the input that varies with the input power, X min It is the minimum value of this parameter, X max It is the maximum value of this parameter, X. i_ It is the normalized value.
[0139] Then, the high-power region training data is imported into a pre-defined first neural network for training, and a first loss value is calculated. Specifically, the first neural network is a deep neural network with three hidden layers, each with 16 neurons. The high-power region training data is imported into the pre-defined first neural network for training to obtain simulated values, while the actual sample values are the test values. The loss value is calculated based on the simulated values and the test values. The formula for calculating the loss value is as follows:
[0140]
[0141] Among them, y i For simulation values, yi_ This is the test value.
[0142] Next, it is determined whether the first loss value is less than a preset value. If the first loss value is less than the preset value, training is stopped and a high-power region deep neural network model is obtained based on the trained first neural network. Specifically, the value of the loss function is calculated during training. If the value of the loss function is less than the preset value, training is stopped, and it is considered that the first neural network can be used to train the line wave characteristics of the high-power region. A high-power region deep neural network model that can be put into use is obtained based on the trained first neural network. If it is greater than the preset value, the first neural network is optimized using the stochastic gradient descent optimization method and training continues until a certain number of training cycles are reached.
[0143] By extracting real and imaginary data and training them into partitions, a high-power region deep neural network model is generated. This partitioning approach improves the accuracy of large signal modeling.
[0144] In one embodiment, to improve the accuracy of the large-signal modeling model for GaN microwave devices, the step of extracting the amplitude and phase data of the input layer and output layer data can be specifically performed as follows:
[0145] Extract the dB amplitude of the fundamental incident wave at the input port. The unit of dB amplitude is dB. Extract the amplitude and phase data of the input layer data. Extract the amplitude and phase data of the output layer data. Specifically, A 11 The phase is used for time-invariant normalization, therefore A 11 With only one variable, the first step is to extract A. 11 The amplitude, expressed in dB, is calculated using the following formula:
[0146] A 11_mag_ =20log(abs(A) 11 ))
[0147] Where A 11_ag_ A in dB 11 Amplitude.
[0148] Step 2: Extract A ph_m and B ph_rm The amplitude and phase are given by the following formula:
[0149]
[0150] A ph_phase_ad =atan2(A ph_ag A ph_ B ph_ase_ad =atan2(B ph_ B ph_ )
[0151] Where A ph__ B represents the amplitude of the incident wave in dB. ph__ A represents the amplitude of the reflected wave in dB. ph__ B represents the phase of the incident wave in rad. ph__ Let B be the phase of the reflected wave in rad. When h = 0, B 20_ =20log(B) 20 ), where B 20_ The current component at port 2 is expressed in dB.
[0152] Figure 7 and Figure 8 The amplitude and phase of each component of the traveling wave are displayed. abs(Mag_A21_dB) refers to the absolute value of the amplitude data of A21, abs(Phase_A21) refers to the absolute value of the phase data of A21, abs(Mag_B21_dB) refers to the absolute value of the amplitude data of B21, and abs(Phase_B21) refers to the absolute value of the phase data of B21. The horizontal axis represents the power.
[0153] Figure 7 The ordinate represents the line wave characteristics in a linear coordinate system. Figure 8 The ordinate represents the row wave characteristics in logarithmic coordinates, from Figure 7 As can be seen, the difference between the low-power region and the high-power region is not significant under linear coordinates, but... Figure 8 As can be seen, the low-power region and the high-power region exhibit a weight difference of about 4 orders of magnitude under logarithmic coordinates, and this difference is more pronounced in the low-power region. Therefore, amplitude / phase is used to train the deep neural network in the low-power region of the line wave characteristics.
[0154] The step of training a deep neural network model for the low-power region using DNN based on amplitude and phase data can be specifically performed as follows:
[0155] First, the dB amplitude, amplitude and phase data of the input layer data, and amplitude and phase data of the output layer data are normalized to obtain low-power region training data. Specifically, the extracted data from the input layer and the extracted data from the output layer are normalized to obtain low-power region training data. The normalization formula is shown below:
[0156]
[0157] Among them, X i It is the i-th value of the input that varies with the input power, X min It is the minimum value of this parameter, X max It is the maximum value of this parameter, X. i_ It is the normalized value.
[0158] Next, the low-power region training data is imported into a pre-defined second neural network for training, and a second loss value is calculated. The second neural network is used to train a deep neural network model for the low-power region. It is a deep neural network with three hidden layers, each with 16 neurons. The low-power region training data is imported into the pre-defined second neural network to obtain simulated values, while the actual sample values are the test values. The loss value is calculated based on the simulated values and the test values. The formula for calculating the loss value is as follows:
[0159]
[0160] Among them, y i For simulation values, y i_ This is the test value.
[0161] Then, it is determined whether the second loss value is less than the second preset value. If the second loss value is less than the second preset value, training is stopped and a low-power region deep neural network model is obtained based on the trained second neural network. Specifically, the value of the loss function is calculated during training. If the value of the loss function is less than the preset value, training is stopped, and it is considered that the second neural network can be used to train the line wave characteristics of the low-power region. A low-power region deep neural network model that can be put into use is obtained based on the trained second neural network. If it is greater than the preset value, the second neural network is optimized using the stochastic gradient descent optimization method and training continues until a certain number of training cycles are reached.
[0162] By extracting amplitude and phase data and training a deep neural network model for low-power regions, and by processing each region accordingly, the accuracy of large-signal modeling can be improved.
[0163] It should be noted that in existing technologies, there are deep neural network models that are trained using only real and imaginary data, and their simulation curves are as follows. Figure 9 As shown, the simulation curves only show good fitting results in the high-power region, while the simulation curves fluctuate significantly in the low-power region, resulting in inaccurate simulations. If a deep neural network model is obtained by training a DNN based solely on amplitude and phase data, its simulation curves are as follows: Figure 10 As shown, the simulation curves only show good fitting results in the low-power region, while the simulation curves fluctuate significantly in the high-power region, resulting in inaccurate simulations. The technical solution of this application obtains a deep neural network model through partitioned training, and its simulation curves are shown below. Figure 11 As shown, the simulation curve fitting effect is not affected by the power level, and the fitting effect is good.
[0164] In one embodiment, to improve the accuracy of the large-signal modeling model for GaN microwave devices, the step of training the row wave characteristics based on the high-power region deep neural network model and the low-power region deep neural network model can be specifically performed as follows:
[0165] First, set the determination points for high and low power regions; specifically, set P... in_ It is the point for determining the high and low power regions.
[0166] Next, based on the decision points, the high-power region deep neural network model, and the low-power region deep neural network model, a partitioned deep neural network model is constructed. Specifically, the first and second neural networks mentioned earlier are deep neural networks with three hidden layers, each with 16 neurons. The specific structure of the constructed partitioned deep neural network model is as follows: Figure 12 As shown, the activation function of the partitioned deep neural network model is set to the Sigmoid function, and the formula is as follows:
[0167]
[0168] Where z is the weighting function, as shown in the formula below:
[0169] z=∑w i x i +b
[0170] Where x u w is the independent variable of the neural network. u is the weight, and b is the bias.
[0171] The following formula illustrates the training variable settings for the input and output layers:
[0172]
[0173] Among them, P in_ This is the decision point between high and low power regions. An if-else statement is used to implement the single-pole double-throw switch function, enabling intelligent switching between high and low power. DNN_L is a trained deep neural network behavioral model for the low-power region, and DNN_H is a trained deep neural network behavioral model for the high-power region. Freq and V... gs These are the frequency and the gate voltage, respectively, where the frequency is measured in GHz and the voltage in V.
[0174] Next, the training line wave characteristics are obtained by training the deep neural network model based on the partitioned deep neural network model and then performing inverse normalization. Specifically, after constructing the deep neural network model for the partitioned region, inverse normalization is performed to restore the line wave characteristic format, which yields the training line wave characteristics of the trained partitioned neural network model.
[0175] By setting high-power and low-power region deep neural network models and using a single-pole double-throw switch to generate an overall partitioned deep neural network model, the line wave characteristics of high-power and low-power regions can be trained simultaneously by a large order of magnitude.
[0176] In one embodiment, to improve the accuracy of the large-signal modeling model for GaN microwave devices, after constructing the large-signal behavior model based on the voltage and current characteristics, the following steps can be performed:
[0177] First, receive the verification row wave characteristics. The verification row wave characteristics include the actual data of the verification row wave characteristics. Specifically, in order to verify whether the model is accurate, it is necessary to receive new row wave characteristics, i.e., the verification row wave characteristics, and the verification row wave characteristics include the actual data of the verification row wave characteristics.
[0178] Then, simulation data is obtained by simulating and verifying the row wave characteristics using a large-signal behavior model. Specifically, such as... Figure 13 As shown in the figure, the output power, gain, and efficiency of the DNN model during verification are displayed. The simulation data are obtained by simulating the large signal behavior model based on the verification row wave characteristics.
[0179] Next, it is determined whether the error between the actual data and the simulation data is within the allowable range. If the error is within the allowable range, the large signal behavior model is deemed usable; if the error is not within the allowable range, the large signal behavior model is deemed unusable. Specifically, determining whether the error between the actual data and the simulation data is within the allowable range is to verify the simulation accuracy of the large signal behavior model. If the error is within the allowable range, it indicates that the large signal behavior model has high accuracy and can be used; conversely, it indicates that the large signal behavior model has low accuracy and cannot be used.
[0180] The obtained large signal behavior model is simulated and compared to verify its functionality, thus ensuring its accuracy.
[0181] Additionally, it should be noted that when performing large-signal tests on GaN transistors, the line wave characteristics, frequency, and voltage are obtained. Through simulation training at different frequencies, the fitting accuracy over a wide dynamic range can be improved, such as... Figure 14 As shown in the figure, the output power, gain, and efficiency of the DNN model are illustrated. It can be seen that large-signal models can be modeled not only at the transistor level but also at the integrated circuit module level to complete system design.
[0182] Based on the above method, this application also discloses a large-signal modeling system based on DNN, the system being based on a GaN microwave device including GaN transistors. Figure 15 The system includes the following modules:
[0183] The line wave characteristic acquisition module 1501 is used to perform large signal testing on GaN transistors to obtain line wave characteristics, which include input layer characteristics and output layer characteristics.
[0184] The line wave characteristic processing module 1502 is used to process the input layer characteristics to obtain input layer data and process the output layer characteristics to obtain output layer data.
[0185] The high-power region model training module 1503 is used to extract the real and imaginary data of the input layer data and the output layer data, and to obtain a high-power region deep neural network model by training a DNN based on the real and imaginary data.
[0186] The low-power region model training module 1504 is used to extract amplitude and phase data from the input layer data and the output layer data, and to train a low-power region deep neural network model using DNN based on the amplitude and phase data.
[0187] The voltage-current characteristic conversion module 1505 is used to obtain the training row wave characteristics based on the training row wave characteristics of the high-power region deep neural network model and the low-power region deep neural network model, and convert the training row wave characteristics into voltage-current characteristics.
[0188] Behavioral model building module 1506 is used to build a large-signal behavioral model based on voltage and current characteristics.
[0189] In one embodiment, the line wave characteristic acquisition module 1501 is specifically used to acquire the fundamental wave conditions and harmonic conditions for testing; to obtain the scanning result by performing an input power scan on the GaN transistor under the fundamental wave conditions and harmonic conditions using a vector network analyzer; and to obtain the line wave characteristics by reading the scanning result through a coupler.
[0190] In one embodiment, the line wave characteristic processing module 1502 is specifically used to calculate the target phase of the fundamental incident wave at the input port; normalize the remaining incident wave according to the target phase to obtain input layer data; and normalize the reflected wave according to the target phase to obtain output layer data.
[0191] In one embodiment, the high-power region model training module 1503 is specifically used to extract the target amplitude of the fundamental incident wave at the input port, extract the real and imaginary data of the input layer data, and extract the real and imaginary data of the output layer data; normalize the target amplitude, the real and imaginary data of the input layer data, and the real and imaginary data of the output layer data to obtain high-power region training data; import the high-power region training data into a preset first neural network for training and calculate a first loss value; determine whether the first loss value is less than a preset value; if the first loss value is less than the preset value, stop training and obtain a high-power region deep neural network model based on the trained first neural network.
[0192] In one embodiment, the low-power region model training module 1504 is specifically used to extract the dB amplitude of the fundamental incident wave at the input port, where the unit of dB amplitude is dB; extract the amplitude and phase data of the input layer data; extract the amplitude and phase data of the output layer data; normalize the dB amplitude, the amplitude and phase data of the input layer data, and the amplitude and phase data of the output layer data to obtain low-power region training data; import the low-power region training data into a preset second neural network for training and calculate a second loss value; determine whether the second loss value is less than a second preset value; if the second loss value is less than the second preset value, stop training and obtain a low-power region deep neural network model based on the trained second neural network.
[0193] In one embodiment, the voltage-current characteristic conversion module 1505 is specifically used to set the determination points for high and low power regions; construct a partitioned deep neural network model based on the determination points, the high power region deep neural network model, and the low power region deep neural network model; and train the row wave characteristics based on the partitioned deep neural network model and perform inverse normalization processing to obtain the trained row wave characteristics.
[0194] This application also discloses a computer device.
[0195] Specifically, the computer device includes a memory and a processor, with the memory storing computer programs that can be loaded and executed by the processor.
[0196] This application also discloses a computer-readable storage medium.
[0197] Specifically, the computer-readable storage medium stores a computer program that can be loaded and executed by a processor, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code.
[0198] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.
Claims
1. A large-signal modeling method based on DNN, the method being based on a GaN microwave device including GaN transistors, characterized in that, The method includes: Large-signal testing is performed on the GaN transistor to obtain its line wave characteristics, which include input layer characteristics and output layer characteristics. The input layer characteristics are processed to obtain input layer data, and the output layer characteristics are processed to obtain output layer data; Extract the real and imaginary parts of the input layer data and the output layer data, and use a DNN to train a high-power regional deep neural network model based on the real and imaginary parts. The amplitude and phase data of the input layer data and the output layer data are extracted, and a low-power region deep neural network model is obtained by training a DNN based on the amplitude and phase data. The line wave characteristics are trained based on the high-power region deep neural network model and the low-power region deep neural network model to obtain the trained line wave characteristics, and the trained line wave characteristics are converted into voltage and current characteristics. A large-signal behavior model is constructed based on the voltage and current characteristics.
2. The method according to claim 1, characterized in that, The line wave characteristics obtained by large-signal testing of the GaN transistor include: Obtain the fundamental and harmonic conditions for testing; Under the fundamental wave condition and the harmonic wave condition, the GaN transistor is scanned by an input power analyzer to obtain the scan results; The line wave characteristics are obtained by reading the scan results through a coupler.
3. The method according to claim 1, characterized in that, The input layer characteristics include incident waves, the output layer characteristics include reflected waves, and the process of processing the input layer characteristics to obtain input layer data and processing the output layer characteristics to obtain output layer data includes: Calculate the target phase of the fundamental incident wave at the input port; The input layer data is obtained by normalizing the remaining incident wave according to the target phase. The reflected wave is processed according to the target phase normalization to obtain the output layer data.
4. The method according to claim 3, characterized in that, The extraction of the real and imaginary parts of the input layer data and the output layer data includes: Extract the target amplitude of the fundamental incident wave at the input port, extract the real and imaginary data of the input layer data, and extract the real and imaginary data of the output layer data; The step of training a high-power region deep neural network model using a DNN based on the real part data and the imaginary part data includes: The target amplitude, the real and imaginary parts of the input layer data, and the real and imaginary parts of the output layer data are normalized to obtain high-power region training data; The high-power region training data is imported into a preset first neural network for training and a first loss value is calculated. Determine whether the first loss value is less than a preset value; If the first loss value is less than the preset value, then training is stopped and a high-power region deep neural network model is obtained based on the trained first neural network.
5. The method according to claim 4, characterized in that, The extraction of amplitude and phase data from the input layer data and the output layer data includes: Extract the dB amplitude of the fundamental incident wave at the input port, wherein the unit of the dB amplitude is dB; extract the amplitude data and phase data of the input layer data; extract the amplitude data and phase data of the output layer data. The step of training a low-power region deep neural network model using a DNN based on the amplitude data and the phase data includes: The dB amplitude, the amplitude data and phase data of the input layer data, and the amplitude data and phase data of the output layer data are normalized to obtain low-power region training data; The training data of the low-power region is imported into a preset second neural network for training, and a second loss value is calculated. Determine whether the second loss value is less than the second preset value; If the second loss value is less than the second preset value, then training is stopped and a low-power region deep neural network model is obtained based on the trained second neural network.
6. The method according to claim 1, characterized in that, The step of training the line wave characteristics based on the high-power region deep neural network model and the low-power region deep neural network model to obtain the trained line wave characteristics includes: Set the determination points for high and low power regions; Construct a partitioned deep neural network model based on the decision points, the high-power region deep neural network model, and the low-power region deep neural network model; The row wave characteristics are trained based on the partitioned deep neural network model and then inversely normalized to obtain the trained row wave characteristics.
7. The method according to claim 1, characterized in that, After constructing the large-signal behavior model based on the voltage and current characteristics, the method further includes: Receive the verification row baud characteristic, wherein the verification row baud characteristic includes the actual data of the verification row baud characteristic; Simulation data was obtained by simulating the verification row wave characteristics using the large signal behavior model: Determine whether the error between the actual data and the simulation data is within the allowable range; If the error is within the allowable range, the large signal behavior model is deemed ready for use. If the error is outside the allowable range, the large signal behavior model is deemed unusable.
8. A large signal modeling system based on DNN, characterized in that, The system is based on a GaN microwave device including GaN transistors, and the system includes: A line wave characteristic acquisition module (1501) is used to perform large signal testing on the GaN transistor to obtain the line wave characteristics, which include input layer characteristics and output layer characteristics; The line wave characteristic processing module (1502) is used to process the input layer characteristics to obtain input layer data and process the output layer characteristics to obtain output layer data. The high-power region model training module (1503) is used to extract the real part data and imaginary part data of the input layer data and the output layer data, and to obtain a high-power region deep neural network model by training a DNN based on the real part data and the imaginary part data. The low-power region model training module (1504) is used to extract the amplitude data and phase data of the input layer data and the output layer data, and to obtain a low-power region deep neural network model by training a DNN based on the amplitude data and the phase data. The voltage-current characteristic conversion module (1505) is used to train the row wave characteristics according to the high-power region deep neural network model and the low-power region deep neural network model to obtain trained row wave characteristics, and convert the trained row wave characteristics into voltage-current characteristics; The behavior model construction module (1506) is used to construct a large signal behavior model based on the voltage and current characteristics.
9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed according to any one of claims 1 to 7.