Method for predicting service life of GaN transistor

Abstract

The invention provides a GaN transistor service life prediction method, which comprises the following steps of: establishing an electric-thermal-force simulation model of a GaN transistor, performing power and thermal resistance simulation to obtain an associated data set among simulation electrical parameters of the GaN transistor, and analyzing the associated data set to obtain original characteristic data; the junction temperature in the simulation electrical parameters is substituted into a Coffin-Manson thermal fatigue model, the electromobility is substituted into a Black electromigration model, decline curves of device performance output by the two physical models are subjected to linear fusion and sampling, and the obtained change data of the junction temperature and the electromobility along with the working time of the device serve as new feature data; the original feature data and the new feature data serve as training data and are input into a machine learning model to complete GaN transistor life prediction training, and the machine learning model outputs a predicted performance degradation curve; and finally, carrying out final fusion on the predicted performance degradation curve and the degradation curve after linear fusion according to the fusion weight of the hybrid model to obtain a final performance degradation curve, thereby obtaining more accurate predicted transistor life.

Description

Technical Field

[0001] This invention belongs to the field of transistor electrical prediction technology, and specifically relates to transistor lifespan prediction technology. Background Technology

[0002] With the widespread application of gallium nitride (GaN) transistors in radar, electronic warfare, communications, and power, there is a growing demand for high performance, high reliability, long lifespan, and even "zero defects," posing new requirements for GaN transistor lifetime assessment. Currently, predicting transistor lifetime primarily employs physical model-based methods and data-driven methods. Model-based methods rely on the device's structure, operating mechanism, and evolution process to accurately construct physical models of the device's failure time range. However, due to the complexity and randomness of the actual operating environment, obtaining quantitatively accurate and reliable physical models is difficult. Therefore, physical model-based methods often result in significant prediction errors. Furthermore, physical models themselves are susceptible to external interference. Consequently, research using physical models is gradually decreasing. Data-driven methods, on the other hand, do not require precise physical models or knowledge of the potential aging mechanisms of electronic devices. They predict and analyze device lifespan based on historically accumulated service data. Data-driven methods employ machine learning and deep learning models, requiring a large amount of data. However, actual data related to transistor lifetime is often scarce, resulting in insufficient data sample size and inadequate training depth for machine learning and deep learning models, leading to significant prediction bias. Summary of the Invention:

[0003] The technical problem to be solved by the present invention is to provide a method for more accurately predicting the lifetime of GaN transistors by combining simulation and machine learning.

[0004] The technical solution adopted by the present invention to solve the above-mentioned technical problems is a GaN transistor lifetime prediction method, comprising:

[0005] An electro-thermal-mechanical simulation model of GaN transistors was established, and power and thermal resistance simulations were performed to obtain a correlation dataset between the simulated electrical parameters of GaN transistors. The correlation dataset was analyzed to obtain the variation data of the simulated electrical parameters with the device operating time, and then the original feature data was obtained after data preprocessing. The simulated electrical parameters in the correlation dataset include power, electromobility, junction temperature, and stress at the edge interface.

[0006] Substituting the junction temperature from the simulated electrical parameters into the Coffin-Manson thermal fatigue model, we can output the device performance degradation curve. Similarly, substituting the electromobility into the Black electromobility model, we can output the device performance degradation curve.

[0007] The degradation curves of device performance output by the Coffin-Manson thermal fatigue model and the Black electromigration model are linearly fused using preset physical model fusion weights. The junction temperature and electromigration data obtained by sampling the linearly fused degradation curves are used as new feature data.

[0008] The original feature data and the new feature data are used as training data to input into the machine learning model to complete the GaN transistor lifetime prediction training. The machine learning model outputs the predicted performance degradation curve. Then, the predicted performance degradation curve and the linearly fused degradation curve are finally fused according to the fusion weight of the hybrid model to obtain the final performance degradation curve. The time when the performance degrades to the preset exhaustion ratio in the final performance degradation curve is the GaN transistor lifetime.

[0009] This invention obtains some lifetime-related parameter data by performing thermo-mechanical-electric simulation on transistors, and then combines the actual physical characteristic data (prediction results of the Black electromigration model and the Coffin-Manson thermal fatigue model) to train a machine learning model. This enables the machine learning model to learn the residuals of the physical model and capture complex nonlinear relationships that are not fully described by the physical equations. As a result, it can accurately predict the remaining lifetime of GaN transistors under complex operating conditions of high power and high temperature. This solves the problems of large prediction errors and susceptibility to external interference in methods based solely on physical models, and the high data sample capacity requirements of methods based solely on data-driven approaches.

[0010] The beneficial effect of this invention is that, compared with traditional lifetime prediction methods driven by failure physics and data, this invention can predict transistor lifetime more accurately without needing to understand the microscopic failure physics of transistors and a large amount of lifetime-related data. The transistor reliable lifetime prediction considers the initial distribution of sensitive parameters, which is more consistent with the actual situation of transistor characteristic parameters, and based on this, it provides lifetime predictions under a specified reliability level, which is more instructive for the engineering application of transistors. Attached image description:

[0011] Figure 1 This is a flowchart illustrating the steps of a GaN transistor lifetime prediction method according to an embodiment of the present invention.

[0012] Figure 2 This is a flowchart illustrating the implementation of a hybrid model prediction algorithm in an embodiment of the present invention.

[0013] Figure 3 This is the prediction result and error of a GaN transistor lifetime prediction method according to an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. To enable those skilled in the art to fully understand how the invention is specifically implemented, this section provides an explanatory description of the embodiments that expand upon the technical solutions of the claims.

[0015] like Figure 1 As shown, a GaN transistor lifetime prediction scheme based on a hybrid model includes the following steps:

[0016] 101 Data Preparation: An electro-thermal-mechanical simulation model of GaN transistors was established, and power and thermal resistance simulations were performed. The correlation dataset between simulation electrical parameters such as GaN transistor structure, stress at the edge interface, stress at the material interface, power, thermal resistance, junction temperature and electromobility was obtained.

[0017] The GaN transistor electro-thermal-mechanical simulation model is established by using the multiphysics simulation software Comsol and the simulation electrical parameters in the associated dataset to create an electrothermal coupling simulation model and a thermo-mechanical coupling simulation model.

[0018] The electrothermal coupling simulation model calculates the temperature field by combining the current continuity equation and the solid heat transfer equation, and obtains a dataset of electromobility as a function of power and junction temperature. This yields the correlation between electromobility and power and junction temperature. The junction temperature can be calculated from the ambient temperature, thermal resistance, and power, thus establishing the correlation between electromobility and power, junction temperature, and thermal resistance.

[0019] The thermo-mechanical coupling simulation model combines the heat transfer equation with the linear elasticity equation, drives thermal expansion strain through the temperature field, and calculates stress distribution by combining boundary conditions, thus obtaining a dataset showing that the stress at the edge of the GaN transistor or the material interface changes periodically with the junction temperature fluctuation.

[0020] In actual simulations, the associated dataset can also include parameters such as size, material, and packaging form.

[0021] The associated datasets construct features with clear physical meaning. Power and thermal resistance reflect the coupling effect between power dissipation and heat dissipation capacity, while current density and temperature describe the accelerating effect of current density on electromobility under high-temperature conditions. These interactive features not only enrich the feature space but also preserve the physical essence of device failure.

[0022] 102 Simulation Parameter Extraction: The simulation electrical parameters of the GaN transistor electro-thermal-mechanical simulation model established by Comsol software are extracted, and the variation data of each parameter with the device operating time are analyzed.

[0023] 103 Data Cleaning: Handling missing, discrete, and outlier simulation data values, and normalizing the data values.

[0024] Data cleaning removes outliers from simulation parameters (such as unreasonable temperature rise data in the simulation), and the extracted simulation parameters are normalized to the [0,1] distribution range using the z-score normalization algorithm.

[0025] 104 Feature Parameter Screening: Based on the degradation trend of electrical parameters, select the key factors that affect transistor lifetime.

[0026] Feature parameter selection is based on the failure physics and degradation mechanism of GaN transistors. From the extracted simulation parameters, power, electromobility, junction temperature, and stress at the edge interface are selected as feature parameters of the data-driven machine learning model in the prediction model.

[0027] Optionally, the number of selected feature parameters can be adjusted according to specific needs. Experiments show that if only three parameters with a significant impact on lifetime are selected, they are power, electromobility, and junction temperature. The stress at the GaN transistor edge interface and material interface in the thermo-coupling simulation model has a relatively insignificant impact on the lifetime of the GaN transistor compared to the parameters in the electro-thermal coupling simulation model.

[0028] 105 Prediction Model Selection and Training: Prediction model selection defines a hybrid prediction model that includes data-driven machine learning models and physical models based on materials science theory.

[0029] Lifespan prediction was performed by comparing and selecting physical models based on materials science theory, including the Coffin-Manson thermal fatigue model and the Black electromigration model, as well as data-driven machine learning models. In this embodiment, the machine learning model used is Support Vector Regression (SVR).

[0030] First, physical model parameters based on materials science theory were defined, including the Coffin-Manson thermal fatigue model and the Black electromigration model.

[0031] Substituting the simulated junction temperature and electromobility into the Coffin-Manson and Black models, the degradation curves of device performance were calculated:

[0032] Coffin-Manson thermal fatigue model:

[0033]

[0034] Where, N fFor fatigue life, A is the proportional material constant, ΔT is the junction temperature fluctuation range, m is the Coffin-Manson index, or thermal fatigue index, exp is the exponent, and E is the coefficient of performance. a1 T is the activation energy material constant, k is the Boltzmann constant, and T is the activation energy material constant. αvg This represents the average absolute temperature. Specific value: A = 1.2 * 10⁻⁶ 6 m = 3, Ea1 = 0.55 eV.

[0035] The Coffin-Manson model uses a proportional material parameter A = 1.2 * 10⁻⁶. 6 The thermal fatigue index m=3 and the activation energy material constant Ea1=0.55eV are used to describe the material fatigue failure mechanism caused by temperature cycling. The core is the power-law relationship between lifetime and junction temperature fluctuation amplitude and the exponential relationship with mean absolute temperature.

[0036] Black electromigration model:

[0037]

[0038] t 50 The mean time to failure (MTBF) is defined as the statistical time when 50% of a batch of identical samples fail, where B is a proportionality constant, J is the current density, n is the current density exponent, and E... a2 T is the activation energy material constant, k is the Boltzmann constant, and T is the activation energy material constant. max This is the absolute temperature. Specific value: B = 5.6 * 10⁻⁶ -12 n = 2, Ea² = 0.7 eV. Current density J and activation energy material constant E a2 It is obtained indirectly through electromobility.

[0039] The Black model uses B = 5.6 * 10 -12 The failure process induced by metal atom migration under high current density is described by the current density exponent n=2 and the activation energy Ea2=0.7eV. The lifetime is inversely proportional to the square of the current density and decreases exponentially with the highest temperature.

[0040] The Coffin-Manson and Black models only require junction temperature and electromobility, which cannot fully represent the degradation trends of various device parameters over time or describe the complex nonlinear relationships of parameter degradation. Introducing the SVR algorithm allows for the use of more feature parameters, enabling interactions between features. For example, power and thermal resistance can reflect the coupling effect of power dissipation and heat dissipation capacity, while current density and temperature can describe the accelerating effect of current density on electromobility at high temperatures. Using more feature parameters in the SVR algorithm leads to more accurate predictions. Furthermore, substituting the predictions from the physical model into the SVR algorithm can accelerate model convergence and avoid the influence of discrete and outlier values ​​in the multi-parameter data obtained from COMSOL simulations on the prediction results.

[0041] Comprehensive physical life prediction = w1*N f +w2*t 50 The predictions from these two physical models are linearly fused using a weighting of 0.6:0.4 to form an initial lifetime prediction based on the physical mechanism. Specifically, the weight w1 of the Coffin-Manson thermal fatigue model is 0.6, and the weight w2 of the Black electromigration model is 0.4. In other words, the degradation curves calculated by the physical models are used as prediction results and linearly fused using a weighting of 0.6:0.4. Data on the changes in junction temperature and electromigration over time are sampled from the fused degradation curves as new feature data. Optionally, sampling from the degradation curves can be performed at equal time intervals or in other forms.

[0042] The predictions from the physical model are added as new features to the feature parameters. Specifically, the data on degradation over time (new features) and the data on the changes in power, thermal resistance, and stress parameters at the material interface obtained from COMSOL simulations over device operating time (original feature parameters) are combined to form the dataset for the machine learning model. This allows the subsequent machine learning model to learn the residuals of the physical model and capture complex nonlinear relationships not fully described by the physical equations.

[0043] First, the input dataset undergoes the same feature engineering and standardization processes as the training phase to ensure data consistency. Then, the SVR model outputs a prediction result—a new performance degradation curve—which is then fused with the physical model's prediction result using a weighted ratio of 0.3:0.7 to form the final performance degradation curve. The lifetime is defined as the time it takes for performance to degrade to 80%. This weighting strategy preserves the interpretability advantages of the physical model while fully leveraging the SVR model's ability to learn complex data patterns. This ensures that the final prediction result conforms to the physical laws of device failure while adapting to uncertainties arising from manufacturing process differences, material nonlinearities, and other practical factors.

[0044] Training the 106 prediction model:

[0045] Support Vector Regression (SVR) uses the Radial Basis Function (RBF) as the kernel function, with a regularization parameter C = 100 and kernel coefficient gamma = 0.01 to balance model complexity and generalization ability. Before training, the system performs Z-score normalization on all numerical features to eliminate the impact of feature scale differences on SVR performance. During training, the SVR model learns the mapping relationship between input features and actual lifetimes, paying particular attention to nonlinear regions that are difficult for physical models to accurately describe.

[0046] The results were verified by analyzing the impact of a single parameter on lifespan and the impact of multiple parameters on lifespan: the simulation data were compared with the prediction results, and the prediction error was controlled to be less than 10%.

[0047] The transistor lifetime was predicted using the method described in the examples, and the results are as follows: Figure 3 As shown in the figure, the GaN transistor lifetime prediction method of the present invention has an error of less than 1%.

Claims

1. A method for predicting the lifetime of a GaN transistor, characterized in that, Including the following steps: An electro-thermal-mechanical simulation model of GaN transistors was established, and power and thermal resistance simulations were performed to obtain a correlation dataset between the simulated electrical parameters of GaN transistors. The correlation dataset was analyzed to obtain the variation data of the simulated electrical parameters with the device operating time, and then the original feature data was obtained after data preprocessing. The simulated electrical parameters in the correlation dataset include power, electromobility, junction temperature, and stress at the edge interface. The junction temperature from the simulated electrical parameters is substituted into the degradation curve of the device performance output by the Coffin-Manson thermal fatigue model, and the electromobility is substituted into the degradation curve of the device performance output by the Black electromobility model. The degradation curves of the device performance output by the Coffin-Manson thermal fatigue model and the degradation curves of the device performance output by the Black electromobility model are linearly fused using preset physical model fusion weights. The junction temperature and electromobility variation data with device operating time obtained by sampling the linearly fused degradation curve are used as new feature data. The original feature data and the new feature data are used as training data to input into the machine learning model to complete the GaN transistor lifetime prediction training. The machine learning model outputs the predicted performance degradation curve. Then, the predicted performance degradation curve and the linearly fused degradation curve are finally fused according to the fusion weight of the hybrid model to obtain the final performance degradation curve. The time when the performance degrades to the preset exhaustion ratio in the final performance degradation curve is the GaN transistor lifetime.

2. The method as described in claim 1, characterized in that, The preset physical model fusion weight for linearly fusing the device performance degradation curves output by the Coffin-Manson thermal fatigue model and the Black electromigration model is 0.6:0.

4.

3. The method as described in claim 1, characterized in that, The hybrid model, which ultimately merges the predicted performance degradation curve with the linearly fused degradation curve, has a fusion weight of 0.3:0.

7.

4. The method as described in claim 1, characterized in that, Data preprocessing includes handling missing, discrete, and outlier simulation data values ​​and normalizing the data values; cleaning and removing outliers from simulation parameters; and normalizing the extracted simulation parameters to the [0,1] distribution range using the z-score normalization algorithm.

5. The method as described in claim 1, characterized in that, The GaN transistor electro-thermal-mechanical simulation model includes an electrothermal coupling simulation model and a thermo-mechanical coupling simulation model; The correlation between electromobility and power, junction temperature, and thermal resistance was calculated using an electrothermal coupling simulation model. A dataset showing the periodic variation of the GaN transistor edge interface with junction temperature fluctuations was obtained through a thermo-coupling simulation model.

6. The method as described in claim 1, characterized in that, An electro-thermal-mechanical simulation model of GaN transistors was established using the multiphysics simulation software Comsol.

7. The method as described in claim 1, characterized in that, The utilization rate is 80%.

8. The method as described in claim 1, characterized in that, The machine learning model is based on the Support Vector Regression (SVR) algorithm.

9. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method of claim 1.

Images