A method and system for rapid lifetime prediction of a gan device

Abstract

The application relates to the technical field of semiconductor device testing, in particular to a GaN device life rapid prediction method and system. The method comprises the following steps: acquiring physical structure attributes and use scene information of a to-be-tested GaN device; determining key working condition parameters and corresponding target stress conditions based on the physical structure attributes and the use scene information; performing adaptive compound accelerated stress testing on the to-be-tested GaN device based on the target stress conditions to obtain a degradation fingerprint vector; constructing a joint probability distribution of the key working condition parameters based on the use scene information; constructing a life prediction model; and obtaining a life distribution of the to-be-tested GaN device by using the life prediction model based on the degradation fingerprint vector and the joint probability distribution. The method solves the problem that the life rapid prediction of the GaN device in the prior art is not accurate enough. By combining the device structure and the scene information, adaptive accelerated testing and modeling, the testing time is reduced, and the accuracy of the life prediction of the GaN device is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor device testing, and in particular to a GaN device life prediction method and system. BACKGROUND

[0002] Gallium nitride (GaN) as the third generation of wide bandgap semiconductor material, with high breakdown field, fast switching speed and excellent high temperature resistance, is widely used in new energy vehicles, charging piles, aerospace and other high frequency and high voltage scenes. The reliability of device life directly determines the operation safety and service cycle of terminal products, so life prediction has become one of the core key technologies for GaN device industrialization.

[0003] The current GaN device life evaluation faces multiple technical bottlenecks: traditional high temperature operating life (HTOL) test needs thousands of hours or even ten thousand hours, which seriously lags behind the rapid iteration pace of products; the existing acceleration test mostly uses a single extreme stress (such as too high temperature or voltage), which easily changes the dominant failure mechanism of the device, resulting in significant deviation of the prediction result from the actual use scene; the extrapolation method based on Arrhenius and other traditional models is difficult to adapt to the multi-physical field coupling failure mechanism of GaN device gate trap accumulation and barrier layer degradation.

[0004] Therefore, there is an urgent need for a GaN device life prediction method that can reduce detection time and be accurate. SUMMARY

[0005] In view of the defects in the prior art, the present application provides a GaN device life prediction method and system, which solves the problem of inaccurate GaN device life prediction in the prior art.

[0006] In order to achieve the above purpose, one aspect of the present application provides a GaN device life prediction method, which comprises: obtaining the physical structure attribute and use scene information of the GaN device to be tested; determining the key working condition parameters and the corresponding target stress conditions based on the physical structure attribute and the use scene information; performing adaptive compound acceleration stress test on the GaN device to be tested based on the target stress conditions to obtain a degradation fingerprint vector; constructing a joint probability distribution of the key working condition parameters based on the use scene information; constructing a life prediction model, and obtaining the life distribution of the GaN device to be tested by using the life prediction model based on the degradation fingerprint vector and the joint probability distribution.

[0007] The application solves the problems of long time consumption and large deviation of single stress test in traditional test by obtaining physical structure attributes and use scenario information of a GaN device to be tested, determining key working condition parameters and target stress conditions, extracting a degradation fingerprint vector through adaptive composite accelerated stress test, and obtaining a life distribution by combining the key working condition parameters with a joint probability distribution and a life prediction model, thereby greatly shortening the test period and improving the accuracy of GaN device life prediction.

[0008] Optionally, the determining the key working condition parameters and the corresponding target stress conditions based on the physical structure attributes and the use scenario information comprises: determining the key working condition parameters according to the physical structure attributes and the use scenario information; determining a value range of the key working condition parameters based on the use scenario information; and converting the value range into the target stress conditions based on a preset acceleration model.

[0009] The application accurately determines the key working condition parameters by combining the physical structure attributes and the use scenario information of the GaN device to be tested, determines the parameter value range according to the use scenario, and converts the value range into the target stress conditions through the preset acceleration model, so that the target stress conditions are consistent with the device failure mechanism and the actual application scenario, and effective acceleration is achieved, the prediction deviation caused by single extreme stress is avoided, and the accuracy of the target stress conditions is improved.

[0010] Optionally, the adaptive composite accelerated stress test based on the target stress conditions comprises: obtaining device degradation data by performing adaptive composite accelerated stress test on the GaN device to be tested based on the target stress conditions; and extracting a degradation fingerprint vector from the device degradation data.

[0011] The application obtains device degradation data by performing adaptive composite accelerated stress test on the GaN device to be tested based on the target stress conditions, and extracts a degradation fingerprint vector from the degradation data, thereby efficiently capturing the real degradation process of the device, accurately refining the core degradation characteristics, avoiding the one-sidedness of single stress test, and improving the accuracy of the degradation fingerprint vector.

[0012] Optionally, the obtaining device degradation data by performing adaptive composite accelerated stress test on the GaN device to be tested based on the target stress conditions comprises: setting an initial composite accelerated stress test cycle based on the key working condition parameters and the target stress conditions; optimizing the initial composite accelerated stress test cycle to obtain an optimized composite accelerated stress test cycle by using a Bayesian optimization algorithm within a preset safety boundary; and obtaining device degradation data by applying the optimized composite accelerated stress test cycle to the GaN device to be tested.

[0013] The application sets an initial compound accelerated stress test cycle according to a key working condition parameter and a target stress condition, optimizes the initial compound accelerated stress test cycle in a preset safety boundary by a Bayesian optimization algorithm, applies the optimized cycle to the GaN device to obtain degradation data, avoids changing the dominant failure mechanism of the device by a single extreme stress, guarantees test safety, maximizes the device degradation rate, and improves the authenticity and accuracy of the degradation data.

[0014] Optionally, the optimization of the initial compound accelerated stress test cycle in the preset safety boundary by the Bayesian optimization algorithm to obtain the optimized compound accelerated stress test cycle comprises: setting an initial optimizable value range for stress parameters of the initial compound accelerated stress test cycle according to the preset safety boundary and the target stress condition; selecting a plurality of initial stress parameter combination samples in the initial optimizable value range; obtaining comprehensive degradation indexes by performing short-time stress cycles on the GaN device based on the initial stress parameter combination samples; constructing a Gaussian process regression proxy model based on the comprehensive degradation indexes and the corresponding initial stress parameter combination samples; determining optimal stress parameter combinations by iteratively optimizing the Gaussian process regression proxy model to maximize an expected improvement function; and obtaining the optimized compound accelerated stress test cycle based on the optimal stress parameter combinations.

[0015] The application sets an initial optimizable value range according to a preset safety boundary and a target stress condition, selects a plurality of initial stress parameter combination samples in the range, obtains comprehensive degradation indexes by performing short-time stress cycles, constructs a Gaussian process regression proxy model based on the comprehensive degradation indexes, iteratively optimizes the model to maximize an expected improvement function to determine optimal stress parameter combinations, and then obtains the optimized compound accelerated stress test cycle, which strictly follows the safety boundary to avoid device damage, locks the optimal stress configuration by precise modeling and iterative optimization, maximizes the device degradation rate without changing the dominant failure mechanism, significantly shortens the test cycle, and improves the scientificity and pertinence of the accelerated stress test.

[0016] Optionally, the obtaining of the comprehensive degradation indexes by performing short-time stress cycles on the GaN device based on the initial stress parameter combination samples comprises: obtaining a plurality of degradation characteristic slopes by performing short-time stress cycles on the GaN device based on the initial stress parameter combination samples; and calculating the comprehensive degradation indexes according to the degradation characteristic slopes.

[0017] The application performs short-time stress cycles on the GaN device based on the initial stress parameter combination samples, obtains a plurality of degradation characteristic slopes, and then calculates the comprehensive degradation indexes, which efficiently captures the degradation law of the device under different stress configurations, quantifies the overall degradation degree comprehensively, avoids the one-sidedness of a single feature, shortens the data acquisition time, and improves the accuracy of the comprehensive degradation indexes.

[0018] Optionally, the extracting the degradation fingerprint vector from the device degradation data comprises: extracting voltage dynamic waveforms and current dynamic waveforms of switch transients from the device degradation data; extracting time domain features from the voltage dynamic waveforms and the current dynamic waveforms respectively; performing synchronous compressed wavelet transform on the voltage dynamic waveforms and the current dynamic waveforms respectively to obtain time-frequency features; determining physical feature weights based on the physical structure attributes; weighting and fusing the time domain features and the time-frequency features by using the physical feature weights to obtain a high-dimensional feature vector; and performing dimension reduction on the high-dimensional feature vector by using a pre-trained sparse autoencoder to obtain the degradation fingerprint vector.

[0019] The application extracts voltage and current dynamic waveforms of switch transients from device degradation data, extracts time domain features and time-frequency features respectively, determines feature weights based on physical structure attributes, and weights and fuses to obtain a high-dimensional feature vector, and then performs dimension reduction by using a pre-trained sparse autoencoder to accurately capture device core degradation information, avoid irrelevant feature interference, and improve the pertinence and accuracy of degradation representation.

[0020] Optionally, the constructing the life prediction model comprises: obtaining a physical simulation training set and a measured failure training set.

[0021] constructing a first loss function and a second loss function; training an initial prediction model by using the physical simulation training set based on the neural network and the first loss function; and calibrating the initial prediction model by using the measured failure training set based on the second loss function to obtain a life prediction model.

[0022] The application obtains a physical simulation training set and a measured failure training set, constructs a double-loss function, trains an initial prediction model by using a first loss function and a physical simulation training set, and then fine-tunes the model by using a second loss function and a measured failure training set to obtain a life prediction model, thereby fusing large-scale simulation data and real failure data, taking into account physical law constraints and actual scene adaptation, and improving the accuracy of model prediction.

[0023] Optionally, the obtaining the life distribution of the GaN device under test based on the degradation fingerprint vector and the joint probability distribution using the life prediction model comprises: generating a plurality of sets of working condition parameter samples based on a Monte Carlo sampling method using the joint probability distribution; inputting the degradation fingerprint vector and each set of the working condition parameters into the life prediction model to obtain a plurality of single-point predicted life values; determining probability weights of the working condition parameter samples based on the joint probability distribution; generating a reliability life distribution function of the GaN device under test by fitting based on the single-point predicted life values and the probability weights; and determining the life distribution of the GaN device under test according to the reliability life distribution function.

[0024] The present application generates a plurality of sets of working condition parameter samples from the joint probability distribution by a Monte Carlo sampling method, inputs the degradation fingerprint vector into the life prediction model to obtain a plurality of single-point predicted life values, determines the probability weights of the samples according to the joint probability distribution, and generates a reliability life distribution function by fitting to determine the life distribution, which fully considers the statistical characteristics and correlation of the working condition parameters, avoids the limitations of single working condition prediction, and improves the accuracy and reliability of the life distribution prediction of the GaN device.

[0025] Another aspect of the present application also provides a GaN device life fast prediction system, comprising: a processor, an input device, an output device and a memory, which are connected to each other, wherein the memory is used to store a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute the GaN device life fast prediction method of any one of the previous aspect of the present application.

[0026] The GaN device life fast prediction system of the present application has compact structure, stable performance, high integration and simple constitution, can stably execute the GaN device life fast prediction method provided in the previous aspect of the present application, and further improves the overall applicability and practical application ability of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A GaN device life fast prediction method flow chart of an embodiment of the present application;

[0028] Figure 2 A GaN device life fast prediction system structure schematic diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0029] Specific embodiments of the present application will now be described in detail with reference to the drawings, which are provided by way of example and not limitation. Like reference numbers refer to like elements throughout. In the following description, numerous specific details are recited to provide a thorough understanding of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced without the specific details. In other instances, well-known circuits, software or methods have not been described in detail in order to avoid obscuring the present application. As used in this application, the term "exemplary" means serving as an example, instance or illustration. Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations.

[0030] Reference throughout this specification to "one embodiment", "an embodiment", "one example" or "an example" means that a particular feature, structure or characteristic described in connection with the embodiment or example is included in at least one embodiment of the application. Thus, appearances of the phrases "in one embodiment" or "in an embodiment" or "one example" or "an example" in various places throughout this specification are not necessarily referring to the same embodiment or example. Furthermore, the particular features, structures or characteristics can be combined in any suitable

[0031] See Figure 1 , to solve the problems in the prior art, in an optional embodiment, as shown in a method for rapid life prediction of a GaN device includes the following steps: Figure 1

[0032] Step S1, obtaining the physical structure attribute and use scenario information of the GaN device to be tested.

[0033] ​In this embodiment, the physical structure properties are obtained through a combination of datasheets provided by device suppliers, detailed design documents from the chip design stage, and laboratory microscopic analysis methods (such as scanning electron microscopy (SEM) and transmission electron microscopy (TEM)). Specifically, these properties cover gate structure types (including p-GaN gate, MOSFET gate, Schottky gate, etc.), epitaxial layer structure parameters (including epitaxial layer material composition, thickness of each layer, and Al composition ratio), packaging forms (such as TO-247, DFN, power module, and other commonly used industrial packages), and core critical dimensions (gate length range 0.1μm-2μm, gate width range 100μm-1000μm). The usage scenario information is obtained through system specifications of the target application scenario, actual working condition task profile analysis report, and statistical data collection of terminal device work logs. Specifically, it includes electrical working conditions (voltage range 200-600V, current range 5-50A, switching frequency 10kHz-1MHz, duty cycle 10%-90%), thermal environment conditions (ambient temperature range -40°C-125°C, cooling method is natural cooling, forced air cooling or liquid cooling), load operation type (continuous load operation or pulse load operation, pulse load duty cycle 5%-50%), and the expected life target proposed by the end user.

[0034] Step S2: Determine key operating parameters and corresponding target stress conditions based on the physical structural properties and the usage scenario information.

[0035] The determination of key operating parameters and corresponding target stress conditions based on the physical structural properties and the usage scenario information specifically includes the following sub-steps:

[0036] Step S201: Determine key operating condition parameters based on the physical structure properties and the usage scenario information.

[0037] In the embodiment, firstly, a mapping rule of "device structure-stress sensitivity-failure mechanism" is established based on failure physical analysis, the resistance characteristics and sensitivity of the to-be-tested GaN device to different stresses are judged by analyzing the physical structure attributes of the to-be-tested GaN device, for example, the p-GaN gate structure is most sensitive to the gate-source voltage stress because the gate insulating layer and the semiconductor interface are prone to trap charges, the risk of barrier breakdown of the thin barrier layer device significantly increases with the increase of the drain-source voltage, and therefore the thin barrier layer device is sensitive to the drain-source voltage stress, the power module packaging structure is sensitive to the temperature cycle stress because the thermal expansion coefficients of different materials (chip, bonding wire and substrate) do not match due to temperature changes, and the device in a high-frequency application scenario is sensitive to the switching frequency stress because of the accumulation of switching losses. Subsequently, combined with the core elements such as electrical conditions, thermal conditions and load types extracted from the use scenario information, the stress types that continuously act or frequently occur in the use process are screened out, the stress variables that have little effect or occur occasionally are excluded, and finally the key operating condition parameters that play a leading role in the device life are determined, the key operating condition parameters include one or more of the junction temperature, the drain-source voltage, the gate-source voltage and the switching frequency, so as to ensure that the selected parameters are directly related to the device structure failure mechanism and are consistent with the stress distribution characteristics of the actual application scenario.

[0038] The determination of the leading failure mechanism is based on the PoF (Physical of Failure) theory, and the material characteristics and structure design rules of the GaN device are combined to determine that for the p-GaN gate structure, the defect states existing between the gate insulating layer and the GaN semiconductor interface are prone to capture carriers, and trap charge accumulation occurs, which is essentially an interface charge migration and capture process dominated by the reaction-diffusion theory, corresponding to the gate-source voltage sensitive failure, for the thin barrier layer structure, the carriers obtain enough energy to bombard the barrier layer under the action of high electric field, which leads to the increase of the defect density of the barrier layer until breakdown, which conforms to the energy accumulation failure mechanism described by the hot electron injection theory, corresponding to the drain-source voltage sensitive failure, for the power module packaging structure, the thermal expansion coefficients of different materials (chip, bonding wire and substrate) do not match due to temperature cycles, which produces periodic mechanical stress, and follows the fatigue accumulation law of the Coffin-Manson theory, corresponding to the temperature cycle stress sensitive failure, for the device in a high-frequency application scenario, the increase of the switching frequency leads to the accumulation of switching losses, which causes the fluctuation of the junction temperature and the thermal carrier damage, which is consistent with the energy dissipation-thermal damage coupling theory, corresponding to the switching frequency sensitive failure. The above failure mechanisms are verified by the authoritative literature in the GaN device field, the failure analysis report of the manufacturer and the TCAD simulation, so as to ensure consistency with the actual failure law.

[0039] In step S202, the value range of the key operating condition parameter is determined based on the use scenario information.

[0040] In the present embodiment, firstly, the actual operation records and constraint conditions of each key working condition parameter are extracted from the system specification description corresponding to the use scene, the task profile report, the long-term operation log of the equipment and the field measurement data, the data cleaning and feature analysis are performed with the aid of MATLAB statistical toolbox, for the deterministic working condition scene of industrial power supply, charging pile, etc., the actual operation maximum and minimum values of the parameters are taken as the boundary after eliminating abnormal data, at the same time, the peak stress of the transition state such as equipment start-stop and load switching is considered, the value range is moderately expanded by 5%-10% to cover extreme working conditions, for example, the leakage source voltage is expanded from the measured 280V-420V to 280V-450V, the switching frequency is determined as 80kHz-550kHz from 100kHz-500kHz, for the dynamic working condition scene of new energy vehicle driving, unmanned aerial vehicle power supply, etc., the parameter distribution characteristics of different running stages (such as idling, cruising, accelerating) are analyzed and split through the task profile, the probability density curve of the parameter is fitted by using the kernel density estimation method, and the value range is determined as “mean ± 3 times standard deviation” or 99.7% confidence interval, for example, the junction temperature is subject to normal distribution with mean 110°C and standard deviation 10°C, and the value range is determined as 80°C-140°C, the gate-source voltage is uniformly distributed under different loads, and the value range is determined as -2V-6V, in addition, combined with the working mode (continuous operation or pulse operation) of the device, the parameter (such as pulse current, short-time peak voltage) under pulse working condition is additionally referenced the pulse duty cycle and duration statistical results, to ensure that the value range not only fully covers all stress levels of the actual use scene, but also does not contain extreme abnormal values without practical significance, to provide accurate and practical basic data for the conversion of subsequent target stress conditions.

[0041] Step S203, converting the value range into target stress conditions based on the pre-set acceleration model.

[0042] In this embodiment, the value range of the key operating condition parameters is converted into the target stress condition by using the acceleration model. According to the dominant failure mechanism corresponding to each key operating condition parameter, an industry-recognized dedicated acceleration model is matched: for temperature-related degradation parameters such as junction temperature, the Arrhenius model is used; for voltage-related degradation parameters such as drain-source voltage and gate-source voltage, the inverse power law model is used; and for fatigue-related degradation parameters such as temperature cycling, the Coffin-Manson model is used. Then, the target acceleration multiple (usually 50-100 times, i.e., traditional thousands of hours of testing is equivalent to hundreds of hours of testing) is determined in combination with the product rapid iteration requirement. The actual maximum value of the value range of the key operating condition parameter is taken as the benchmark for acceleration calculation. For example, the value range of the junction temperature is 80°C-140°C, and the maximum value 140°C (413.15K) is taken as the benchmark. The acceleration multiple is set to 100 times, and the target test temperature 175°C (448.15K) is calculated by substituting into the Arrhenius model. The value range of the drain-source voltage is 280V-450V, and the maximum value 450V is taken as the benchmark. The voltage acceleration index is 5, the acceleration multiple is set to 80 times, and the target test voltage 540V (1.2 times the maximum value) is calculated by substituting into the inverse power law model. The temperature cycling temperature variation range is 40°C-80°C, and the maximum value 80°C is taken as the benchmark. The fatigue index is 3, the acceleration multiple is set to 60 times, and the target temperature variation range 120°C is calculated by substituting into the Coffin-Manson model. Finally, the calculated target stress parameters are checked against the preset safety boundary (based on the device specification limit and material physical limit, such as the junction temperature not exceeding 200°C, the gate voltage not exceeding 80% of the breakdown voltage, and the drain-source voltage not exceeding 1.3 times the device rated voltage). If it exceeds the boundary, it is adjusted to be within the boundary to ensure test safety. Finally, the target stress condition combination including temperature, voltage, and temperature variation range is formed to ensure that the acceleration process does not change the dominant failure mechanism of the device and meets the dual requirements of rapid testing and device safety.

[0043] Step S3, performing adaptive composite accelerated stress testing on the GaN device to be tested based on the target stress condition to obtain a degradation fingerprint vector.

[0044] Wherein, performing adaptive composite accelerated stress testing on the GaN device to be tested based on the target stress condition to obtain a degradation fingerprint vector specifically includes the following sub-steps:

[0045] Step S301, performing adaptive composite accelerated stress testing on the GaN device to be tested based on the target stress condition to obtain device degradation data.

[0046] Wherein, performing adaptive composite accelerated stress testing on the GaN device to be tested based on the target stress condition to obtain device degradation data specifically includes the following sub-steps:

[0047] Step S30101, setting an initial composite accelerated stress test cycle based on the key operating parameter and the target stress condition.

[0048] In this embodiment, the cycle stages are divided according to the correspondence between stress types and failure mechanisms, ensuring that each key operating parameter corresponds to a dedicated acceleration stage: a high-temperature gate bias (HTGB) stage is set for gate-source voltage stress, a high-temperature reverse bias (HTRB) stage is set for drain-source voltage stress, a power cycle (PC) stage is set for junction temperature stress, and a high-frequency switching (HSW) stage is set for switching frequency stress. The order of applying each stage is then determined (arranged according to the logic of "low stress adaptation-main stress acceleration-stress release", i.e., HTGB stage first, then HSW stage, followed by PC stage, and finally HTRB stage), avoiding interference of different stresses leading to failure mechanisms. The duration of each stage is optimized based on the time constant of the failure mechanism corresponding to the target stress condition. The high-temperature gate bias stage lasts for 30-45 minutes (to ensure sufficient accumulation of gate trap charges), the high-frequency switching stage lasts for 15-30 minutes (to simulate the accumulation of losses under actual high-frequency operation), the power cycle stage includes 5-10 temperature variation cycles (each cycle lasting 1-2 minutes, with the temperature variation amplitude set according to the target stress condition), and the high-temperature reverse bias stage lasts for 45-60 minutes (to accelerate the degradation of the potential barrier layer). The overall cycle period is controlled to be 1-2 hours, and the target stress parameters are accurately assigned to the corresponding stages, such as setting the test temperature to 175°C and the gate-source voltage to 6V (target stress value) for the HTGB stage, setting the test temperature to 175°C and the drain-source voltage to 540V (target stress value) for the HTRB stage, setting the temperature variation range to -40°C-175°C and the temperature rise rate to 10°C / min for the PC stage, and setting the switching frequency to 1MHz (target stress value) and the duty cycle to 50% for the HSW stage. Finally, an initial composite accelerated stress test cycle is formed, which includes multiple stress types, is sequentially connected, and has clear parameters, ensuring that the cycle can comprehensively cover the acceleration requirements of the key operating parameters without changing the dominant failure mechanism of the device.

[0049] Step S30102, optimizing the initial composite accelerated stress test cycle within a preset safety boundary by a Bayesian optimization algorithm to obtain an optimized composite accelerated stress test cycle.

[0050] In this embodiment, the cycle stages are divided according to the correspondence between stress types and failure mechanisms, ensuring that each key operating parameter corresponds to a dedicated acceleration stage: a high-temperature gate bias (HTGB) stage is set for gate-source voltage stress, a high-temperature reverse bias (HTRB) stage is set for drain-source voltage stress, a power cycle (PC) stage is set for junction temperature stress, and a high-frequency switching (HSW) stage is set for switching frequency stress. The order of applying each stage is then determined (arranged according to the logic of "low stress adaptation-main stress acceleration-stress release", i.e., HTGB stage first, then HSW stage, followed by PC stage, and finally HTRB stage), avoiding interference of different stresses leading to failure mechanisms. The duration of each stage is optimized based on the time constant of the failure mechanism corresponding to the target stress condition. The high-temperature gate bias stage lasts for 30-45 minutes (to ensure sufficient accumulation of gate trap charges), the high-frequency switching stage lasts for 15-30 minutes (to simulate the accumulation of losses under actual high-frequency operation), the power cycle stage includes 5-10 temperature variation cycles (each cycle lasting 1-2 minutes, with the temperature variation amplitude set according to the target stress condition), and the high-temperature reverse bias stage lasts for 45-60 minutes (to accelerate the degradation of the potential barrier layer). The overall cycle period is controlled to be 1-2 hours, and the target stress parameters are accurately assigned to the corresponding stages, such as setting the test temperature to 175°C and the gate-source voltage to 6V (target stress value) for the HTGB stage, setting the test temperature to 175°C and the drain-source voltage to 540V (target stress value) for the HTRB stage, setting the temperature variation range to -40°C-175°C and the temperature rise rate to 10°C / min for the PC stage, and setting the switching frequency to 1MHz (target stress value) and the duty cycle to 50% for the HSW stage. Finally, an initial composite accelerated stress test cycle is formed, which includes multiple stress types, is sequentially connected, and has clear parameters, ensuring that the cycle can comprehensively cover the acceleration requirements of the key operating parameters without changing the dominant failure mechanism of the device.

[0051] Step S3010201, setting an initial optimizable value range for the stress parameters of the initial composite accelerated stress test cycle according to the preset safety boundary and the target stress condition.

[0052] In this embodiment, when setting the initial optimizable value range based on the preset safety boundary and the target stress condition, the safety boundary is first defined by the device datasheet derating rules (e.g., gate voltage not exceeding 80% of the breakdown voltage, drain-source voltage not exceeding 1.3 times the rated value) and material limits (e.g., junction temperature ≤ 200°C, temperature change range ≤ 150°C). Then, the adjustable core stress parameters in the initial cycle (including temperature and gate voltage in the HTGB stage, drain voltage in the HTRB stage, and temperature change range and heating current in the PC stage) are selected. Based on the target stress value corresponding to each parameter, a floating range of 10%-20% is set. Finally, it is checked whether the floating range exceeds the safety boundary. If it does, it is corrected by the upper limit of the boundary, forming an initial optimizable value range that meets the acceleration requirements and ensures device safety. At the same time, the initial optimizable value range should ensure that the optimal solution has enough space to avoid touching the safety boundary in the subsequent optimization process. It is generally required that the upper limit of the optimizable range of each parameter does not exceed 90% of the upper limit of the safety boundary.

[0053] Step S3010202: Select multiple sets of initial stress parameter combination samples within the initial optimizable value range.

[0054] In this embodiment, when selecting multiple sets of initial stress parameter combination samples within the initial optimizable value range, the Latin hypercube sampling (LHS) method is used to ensure uniform coverage of the samples across all parameter dimensions. The number of initial samples is dynamically determined based on the dimensions of the optimizable stress parameters (including temperature / gate voltage in the HTGB stage, drain voltage in the HTRB stage, temperature change amplitude / heating current in the PC stage, and switching frequency in the HSW stage, totaling N dimensions), typically 3 to 5 times N, for example, selecting 15-30 sets of initial samples. Each set of samples is a complete vector containing all adjustable stress parameters, such as temperature and gate voltage in the high-temperature gate bias stage, drain voltage in the high-temperature reverse bias stage, and temperature change amplitude and heating current in the power cycling stage. This avoids excessive correlation between parameters and ensures that the samples can comprehensively reflect the stress configuration characteristics within the initial optimizable value range, providing representative basic data for the subsequent construction of a Gaussian process regression surrogate model.

[0055] Step S3010203: Based on the initial stress parameter combination sample, the GaN device under test is subjected to short-time stress cycling to obtain a comprehensive degradation index.

[0056] The comprehensive degradation index obtained by performing short-time stress cycling on the GaN device under test based on the initial stress parameter combination sample includes:

[0057] Step S301020301: Based on the initial stress parameter combination sample, perform short-time stress cycling on the GaN device under test to obtain multiple degradation feature slopes.

[0058] In the embodiment, the initial stress parameter combination sample is used to perform three short-time stress cycles on the GaN device to be tested. The stage type and application sequence of the short-time stress cycle are completely the same as those of the initial composite accelerated stress test cycle, but the duration of each stage is set to 10% to 20% of the duration of the corresponding stage in the initial composite accelerated stress test cycle. Before and after the start of each cycle, the real-time values of the threshold voltage, on-resistance and gate leakage current of the device are immediately collected by using a precision measuring device. Thus, for each electrical parameter, the measurement values at four time points, i.e., the initial state and after the three short-time stress cycles, can be obtained. Then, for each parameter data point sequence, a linear regression method is used for fitting, and the slope of the obtained fitting straight line is defined as the degradation characteristic slope of the parameter. The degradation characteristic slope quantifies the average relative change rate of the device performance per short-time cycle under the corresponding stress combination, and a positive slope indicates that the parameter increases, and a negative slope indicates that the parameter decreases. Finally, each stress parameter combination sample can obtain three degradation characteristic slopes corresponding to different electrical parameters. All the slopes are positive, representing the degradation rate.

[0059] In step S301020302, a comprehensive degradation index is calculated according to the degradation characteristic slopes.

[0060] In the embodiment, the weight of each degradation characteristic slope is determined according to the mapping relationship between the degradation parameter and the dominant failure mechanism. For example, by analyzing historical failure data, the Pearson correlation coefficient of each degradation characteristic slope and the key parameter of the dominant failure mechanism (such as trap density and barrier layer thickness) is calculated, the absolute value of the correlation coefficient is normalized to obtain the initial weight of each degradation characteristic slope. Combined with the physical structure attributes of the GaN device to be tested and the analysis of historical failure data, a higher weight (0.3-0.5) is assigned to the degradation characteristic (such as the degradation characteristic slope of the gate leakage current) that has a strong correlation with the dominant failure mechanism (such as gate trap and barrier breakdown), a lower weight (0.1-0.2) is assigned to the characteristic (such as the degradation characteristic slope of the secondary electrical parameter) that has a weak correlation, all the weights are normalized to satisfy the sum of the weights being 1, then the absolute value of each degradation characteristic slope is taken to avoid the cancellation of the degradation amplitude due to the positive and negative signs, and finally the weighted geometric mean is used to calculate the comprehensive degradation index, which can quantitatively quantify the overall degradation rate of the device under the corresponding stress configuration, and provide an accurate output label for subsequent proxy model construction.

[0061] The comprehensive degradation index satisfies the following formula:

[0062]

[0063] wherein, is the comprehensive degradation index, is the total number of degradation characteristics, For the first The absolute value of the slope of each degradation feature is used to eliminate the influence of the sign of the degradation on the magnitude of degradation. For the first The weights of the slopes of each degenerate feature. It is a positive decimal, taking the value 1e-10, used to avoid the entire product being zero when the slope of a certain degenerate feature is zero, thus ensuring the stability of numerical calculation.

[0064] The above formula uses a weighted geometric mean. First, it takes the maximum absolute value of the slope of each degradation feature to ensure that it is not lower than a very small positive number. This effectively avoids the problem of the entire product being zero due to the slope of a certain degenerate feature being too small or zero, thus ensuring the stability of the calculation. Secondly, by applying weights to each feature... By exponentiation, the contribution of each degradation feature to the overall degradation rate can be differentiated. Moreover, the weighted geometric mean is better able to capture the nonlinear coupling relationship between multiple feature degradations than the arithmetic mean, and is not sensitive to extreme values. Thus, it provides a robust and physically meaningful optimization objective for subsequent Bayesian optimization.

[0065] Step S3010204: Based on the comprehensive degradation index and the corresponding initial stress parameter combination sample, construct a Gaussian process regression surrogate model.

[0066] In this embodiment, when constructing a Gaussian process regression surrogate model based on the comprehensive degradation index and the corresponding initial stress parameter combination samples, the input initial stress parameter combination samples (multidimensional vectors) are first Z-score standardized to eliminate dimensional differences. Simultaneously, the comprehensive degradation index is used as the model output label. A Matern kernel (smoothness parameter nu = 2.5) is selected as the core kernel function (adapting to nonlinear degradation relationships and engineering data noise characteristics). A noise term alpha of 1e-6 is set to characterize the measured data error. The maximum likelihood estimation method is used to optimize the kernel function hyperparameters. Local optima are avoided by restarting the optimizer 10 times. The preprocessed sample data is input into the model for fitting training. After training, 5-fold cross-validation is used to evaluate the model performance. A root mean square error of less than 0.05% per cycle and a coefficient of determination greater than 0.9 are considered acceptable. Finally, a Gaussian process regression surrogate model that can accurately map the nonlinear relationship between stress parameter combinations and comprehensive degradation index, and can quantify the uncertainty of prediction, is constructed.

[0067] Step S3010205: With the goal of maximizing the expected improvement function, iterative optimization is performed by updating the Gaussian process regression surrogate model to determine the optimal combination of stress parameters.

[0068] In this embodiment, when performing iterative optimization aiming at maximizing the expected improvement function (EI), first, based on the constructed Gaussian process regression surrogate model, the EI values of all potential stress parameter combinations in the initial optimizable value range are calculated. The EI function realizes the dynamic balance of the predicted mean (reflecting the "exploitation" of the high degradation rate area) and the predicted variance (reflecting the "exploration" of the high uncertainty area) by fusing them. In each iteration, a new stress parameter combination with the maximum EI value is selected, and 1-3 short-time stress cycles are performed on it. The key electrical parameters are measured and the corresponding comprehensive degradation index is calculated to form a new (stress parameter combination-comprehensive degradation index) data pair. The new data pair is added to the training set, and the kernel function hyperparameters are optimized again through maximum likelihood estimation to update the Gaussian process regression surrogate model to improve the prediction accuracy. In the iteration process, the maximum number of iterations is set to 5-10, and the change amplitude of the EI maximum value of two consecutive iterations is less than or equal to 5% as the convergence condition. If it is met, the iteration is stopped, and it is checked whether the current optimal stress parameter combination is too close to the safety boundary. If any parameter value exceeds 90% of the upper limit of the safety boundary, the iteration is terminated and the current suboptimal solution is selected. Finally, from all the stress parameter combinations and corresponding comprehensive degradation indexes generated in the iteration, the combination with the maximum comprehensive degradation index is selected as the optimal stress parameter combination, which ensures that the device degradation rate is maximized without changing the dominant failure mechanism, and the subsequent accelerated test period is shortened.

[0069] Step S3010206, obtaining an optimized composite accelerated stress test cycle based on the optimal stress parameter combination.

[0070] In this embodiment, first, the phase structure (such as the order of high-temperature gate bias HTGB, high-frequency switching HSW, power cycle PC, and high-temperature reverse bias HTRB) and the duration of each phase of the initial composite accelerated stress test cycle are kept unchanged, and only the corresponding adjustable core parameters (including the test temperature and gate-source voltage in the HTGB phase, the drain-source voltage in the HTRB phase, the temperature variation amplitude and heating current in the PC phase, and the switching frequency in the HSW phase) in the optimal stress parameter combination are replaced with the corresponding parameters in the initial cycle. Then, it is checked again whether all the stress parameters after replacement still strictly meet the preset safety boundary (such as junction temperature ≤ 200°C and gate voltage not exceeding 80% of breakdown voltage). After confirming that there is no exceeding, an optimized composite accelerated stress test cycle with fixed structure and accurate parameters is formed, which can maximize the device degradation rate and ensure that the dominant failure mechanism is consistent with the actual use scenario.

[0071] Step S30103, applying the optimized composite accelerated stress test cycle to the GaN device to be tested to obtain device degradation data.

[0072] In the embodiment, first, the GaN device to be tested is fixed to a high-temperature test seat and connected to a comprehensive test platform, and 100-200 test cycles are continuously performed according to optimized cycle parameters (including stress values, duration and application sequence of each stage), the total test time is controlled to be 100-400 hours, and every 5 cycles is set as a data acquisition node. At each node, the drain-source voltage and drain current dynamic waveforms of the switching transient are synchronously captured by a high-speed oscilloscope (sampling rate greater than 100 MSa / s), the evolution trajectory of the waveforms with the cycle number is recorded throughout the process, and a device degradation data set containing the switching transient waveform is formed. The data set will be used as an input source for subsequent extraction of the degradation fingerprint vector.

[0073] Step S302, extracting a degradation fingerprint vector from the device degradation data.

[0074] In the embodiment, the step of extracting a degradation fingerprint vector from the device degradation data specifically includes the following sub-steps:

[0075] Step S30201, extracting a voltage dynamic waveform and a current dynamic waveform of the switching transient from the device degradation data.

[0076] In the embodiment, the device degradation data contains switching transient waveform data of multiple acquisition nodes, and the voltage dynamic waveform and the current dynamic waveform can be directly extracted therefrom. The voltage dynamic waveform and the current dynamic waveform in the switching transient process directly reflect the real working state and internal degradation physical process of the GaN device under high-speed switching action, the evolution of the voltage waveform (such as the drain-source voltage) represents the dynamic change of the electric field in the on and off processes of the device channel, and the distortion of the overshoot, ring and rising / falling edge is closely related to the charging and discharging of the gate interface trap charge, the current collapse effect and the change of the parasitic parameters. The dynamic characteristics of the current waveform (such as the drain current) directly reveal the effects of the channel carrier transport capability, the degradation of the on-resistance and the hot electron injection, and the two together constitute the original information carrier representing the dynamic degradation mechanism of the device.

[0077] Step S30202, extracting time domain features from the voltage dynamic waveform and the current dynamic waveform, respectively.

[0078] In the embodiment, firstly, feature extraction is independently performed on each waveform: the rise time, the fall time and the overshoot amplitude of the voltage waveform and the current waveform are calculated respectively. Secondly, the steady-state values of each waveform are determined: after the waveform enters the stable conduction stage, the data in the middle 10% period is selected, and the steady-state average value of the voltage and the steady-state average value of the current are calculated respectively. Finally, the composite features are calculated using the double waveform relationship: the two steady-state values obtained in the previous step are combined, and the key steady-state conduction resistance is calculated by dividing the voltage steady-state value by the current steady-state value. Through the above process, a time domain feature set composed of voltage rise time, voltage fall time, voltage overshoot, current rise time, current fall time, current overshoot, voltage steady-state value, current steady-state value and steady-state conduction resistance is finally output.

[0079] In step S30203, the voltage dynamic waveform and the current dynamic waveform are subjected to synchronous compression wavelet transform to obtain time-frequency features.

[0080] In the embodiment, when the voltage dynamic waveform and the current dynamic waveform are subjected to synchronous compression wavelet transform to obtain time-frequency features, Morlet wavelet is selected as the base function, and according to the typical switching frequency range (10 kHz-1 MHz) of the GaN device to be measured, the scale range is set to cover the frequency band of the main harmonic components of the switching frequency, usually 0.1 times to 10 times of the switching frequency, for example, 1 kHz-10 MHz (adapted to the frequency characteristics of the switching transient of the device), the preprocessed two waveforms are subjected to synchronous compression wavelet transform, the one-dimensional time domain signal is mapped to a two-dimensional time-frequency diagram, and three types of core features are extracted from the time-frequency diagram: the energy value of a specific frequency band (10-50 MHz) (representing the signal intensity of the frequency band), the time-frequency entropy (quantifying the complexity of the time-frequency distribution of the signal), and the instantaneous frequency gravity center (reflecting the time scale of the dominant dynamic degradation process), these features are arranged in a predetermined order to form a time-frequency feature vector that can comprehensively capture the time-frequency evolution law of the waveform.

[0081] In step S30204, the physical feature weight is determined based on the physical structure attribute.

[0082] In this embodiment, the physical feature weight can be realized by combining analytic hierarchy process (AHP) and failure physical analysis. First, a three-layer analysis structure is established: the target layer is "accurately representing the degradation state", the criterion layer is the dominant failure mechanism identified according to the physical structure attribute (such as gate trap accumulation, barrier layer breakdown), and the scheme layer is all time-domain and time-frequency features extracted from the waveform. Subsequently, 3-5 experts in the field are organized, and according to historical failure data, simulation results and the strength of the physical correlation between features and failure mechanisms, for example, by calculating the correlation coefficient of the change amount of a specific feature and the key parameter of a specific failure mechanism (such as trap density) in simulation or experiment, it is quantified as a value in the range of 1-9 scale, which is an important reference for expert scoring. For each failure mechanism, the 1-9 scale method is used to compare the importance of each feature, a judgment matrix is constructed, and consistency test is performed (consistency ratio CR < 0.1). Then, the local weight of each feature relative to different failure mechanisms is calculated, and then the global weight of each feature is determined according to the proportion of each failure mechanism in the historical failure, and the initial global weight of each feature is obtained by weighted integration. Finally, the initial weights of all features are normalized so that their sum is 1, thereby forming a physical feature weight vector that strictly corresponds to the dimension of the high-dimensional feature vector and is strongly bound to the physical structure of the device.

[0083] Step S30205, the time-domain features and the time-frequency features are fused by using the physical feature weight to obtain a high-dimensional feature vector.

[0084] In this embodiment, when the time-domain features and the time-frequency features are fused by using the physical feature weight to obtain a high-dimensional feature vector, first, the extracted time-domain feature vector (including rise time, dynamic on-resistance, etc.) and the time-frequency feature vector obtained by synchronous compression wavelet transform (including specific frequency band energy, time-frequency entropy, etc.) are spliced in a predetermined order to form a complete-dimensional original feature vector, ensuring that the physical feature weight vector and the original feature vector are completely matched in dimension (each element of the physical feature weight corresponds to a specific feature in the original feature vector, not a feature group, and one-to-one matching is required for accurate weighting), then through Hadamard product (element-by-element multiplication) operation, each feature component is multiplied by the corresponding physical feature weight, the contribution of the features strongly related to the physical structure attribute and the dominant failure mechanism of the device is strengthened, and the interference of the secondary features is weakened, and finally a high-dimensional feature vector of complete degradation information is obtained.

[0085] Step S30206, the high-dimensional feature vector is reduced by using a pre-trained sparse autoencoder to obtain a degradation fingerprint vector.

[0086] In the embodiment, the sparse autoencoder is an unsupervised deep neural network with sparsity constraint, which is used for feature dimension reduction. The network structure includes three parts: an input layer, a symmetric structure composed of an encoder and a decoder. The encoder includes several hidden layers and a low-dimensional bottleneck layer, which is responsible for compressing the input into a low-dimensional representation, and the decoder attempts to reconstruct the original input from the low-dimensional representation.

[0087] The training data of the autoencoder is a set of high-dimensional feature vectors extracted from historical degradation data of the same type of GaN devices with the same gate structure and application scenario as the device under test by the same feature extraction process as steps S30201 to S30205. These vectors constitute the sample set for the autoencoder learning.

[0088] The training principle is unsupervised learning, and the objective function minimizes the mean square error (MSE) between the input high-dimensional feature vector and the decoder reconstruction output, and introduces a sparse penalty term (such as KL divergence) to constrain the activation of the bottleneck layer neurons to be as sparse as possible. This mechanism forces the network to retain only the most significant features that contribute to reconstruction during information compression, suppressing the transmission of redundant or noisy information, thereby learning a low-dimensional representation of the data essence.

[0089] After training, only the encoder part is retained as a feature extractor. In the application stage, the high-dimensional feature vector of the GaN device under test is input into the encoder, and the encoder maps it to the low-dimensional bottleneck layer, and the output is the degradation fingerprint vector. The degradation fingerprint vector is a compact representation of the original high-dimensional feature after network refinement, which retains the core information representing the device degradation state, while achieving dimension reduction. This extraction method is universal and suitable for waveform data generated by simulation, accelerated testing and actual monitoring.

[0090] Step S4, constructing a joint probability distribution of the key operating condition parameters based on the use scenario information.

[0091] In the embodiment, for each key operating parameter (such as junction temperature, drain-source voltage), samples are extracted based on historical operation data in the use scenario information, the Kolmogorov-Smirnov test or Akaike information criterion is used to select an optimal marginal distribution from candidate distributions (such as normal distribution, Weibull distribution, uniform distribution, etc.), and the maximum likelihood estimation method is used to fit the marginal distribution (for example, normal distribution, Weibull distribution, or uniform distribution). Secondly, the Kendall correlation coefficient between all parameters is calculated to analyze the nonlinear dependence structure, the square Euclidean distance between the empirical Copula and different theoretical Copulas (such as Gaussian Copula, t-Copula) is calculated, the Copula function with the highest goodness of fit is selected, the Gaussian Copula is selected if the dependence relationship is approximately symmetric and linear, and the t-Copula is selected if there is tail correlation. Then, based on the fitted marginal distribution function and the selected Copula type, the maximum likelihood estimation method is used again to jointly estimate the parameters (such as the correlation coefficient matrix) of the Copula. The joint probability distribution model obtained in this way can accurately represent the statistical characteristics and internal correlation of the key operating parameters in actual application. The effectiveness of the model can be verified by graphical comparison (such as Q-Q plot).

[0092] Step S5, constructing a life prediction model, obtaining the life distribution of the GaN device to be measured by using the life prediction model based on the degradation fingerprint vector and the joint probability distribution.

[0093] The constructing a life prediction model specifically includes the following sub-steps:

[0094] Step S501, obtaining a physical simulation training set and a measured failure training set.

[0095] A physical simulation training set is constructed, based on the physical structure attributes (such as gate structure, epitaxial layer parameters, packaging details) of the GaN device under test, an accurate two-dimensional or three-dimensional physical model is established using process computer-aided design (TCAD) software (such as Sentaurus TCAD). In this model, the Poisson equation, carrier continuity equation, heat conduction equation, etc. are coupled and solved, and key degradation physical mechanism models (for example, gate interface trap generation model based on reaction-diffusion theory, bond wire damage model based on thermal fatigue accumulation) are implanted. Through batch simulation, under a wide range of stress condition combinations (such as junction temperature 125°C to 250°C, gate voltage +5V to +10V, drain voltage 400V to 700V), thousands of "virtual devices" are simulated from the initial state to the complete degradation process until failure (such as a 30% increase in dynamic on-resistance or gate leakage exceeding a threshold value). During the degradation process of each virtual device, sampling is performed at multiple different simulation time points, and a complete data sample is recorded at each sampling point, which includes three parts, namely the degradation state feature, the working condition parameter and the life label. The degradation state feature is that at this simulation time point, the simulated switching transient voltage and current dynamic waveform is output by the simulation model, then the high-dimensional feature vector is extracted according to the method of steps S30201 to S30205, and the same pre-trained sparse autoencoder as step S30206 is used for dimension reduction to obtain the degradation fingerprint vector. The working condition parameter is the parameter combination (such as junction temperature, gate voltage, drain voltage) applied at this simulation time point to simulate the actual working condition, and the life label is the remaining simulation time (i.e. remaining life) from this sampling point to the failure of the virtual device. Through the above method, a physical simulation training set containing thousands of samples (for example, 5000 to 10000 samples) is constructed. Each sample establishes a mapping relationship between a specific degradation state (characterized by the degradation fingerprint vector) under the condition of working condition parameters and the corresponding remaining life.

[0096] A measured failure training set is constructed by selecting a small number (e.g., 10 to 20) of actual GaN devices from the same batch or design as the device under test. These samples are subjected to high-accelerated stress tests of varying intensities (e.g., high temperature and high gate voltage, power cycling, etc.) until each device reaches a predetermined failure criterion. During the testing process, data is collected periodically at fixed intervals or fixed stress cycle counts. At each collection point, a sample with the same format as the simulation data is recorded: the degradation state characteristic is the measured degradation fingerprint vector obtained at that collection point using a process identical to the simulation data extraction method; the operating condition parameter is the actual usage scenario parameter corresponding to that collection point (i.e., the measured operating condition); the stress condition applied during testing is based on the transformed operating condition parameter (used to obtain degradation data); and the lifetime label is the remaining test time from that collection point to the actual device failure (i.e., the measured remaining lifetime). This yields a relatively small but accurate measured failure training set, which is used to calibrate potential biases in the simulation model. The physical simulation training set and the measured failure training set use the exact same sample data structure (degradation fingerprint vector, operating parameters, remaining life) to ensure the effectiveness and consistency of subsequent model training.

[0097] The training set constructed in the above manner enables the lifetime prediction model to learn the complex mapping relationship between the device degradation state (degradation fingerprint vector) and the remaining lifetime under any given stress conditions (operating conditions).

[0098] Step S502: Construct the first loss function and the second loss function.

[0099] The first loss function satisfies the following formula:

[0100]

[0101]

[0102]

[0103] in, For the first loss function, This is the mean square error term. For the consistency constraint terms of physical laws, The weights of the mean square error term are... The weights of the physical law consistency constraint terms. This represents the number of samples in the physical simulation training set. For the first physical simulation training set The true failure lifetime value of each sample For the first physical simulation training set The predicted failure lifetime value for each sample. The number of sample pairs constructed for the physical simulation training set. For the first Predicted lifetimes for each sample pair under moderate and operating conditions. For the first Predicted lifetime of harsh operating parameters in a sample pair. It is a positive decimal, used as a safety margin to strengthen constraints.

[0104] In the project, the weight of the mean square error term can be set. The weight of the physical law consistency constraint term is 0.7. The value is 0.3, which balances prediction accuracy with physical constraints.

[0105] The purpose of the physical law consistency constraint is to force the lifetime prediction model to adhere to the fundamental physical law that the more stringent the operating conditions, the shorter the lifetime of GaN devices. Therefore, it is necessary to dynamically construct sample pairs for constraint training within the physical simulation training set. Specifically, from the physical simulation training set, samples with the same or similar degradation fingerprint vectors are selected and paired with one set of mild operating conditions (e.g., case temperature 150°C, gate-source voltage +6V) and one set of more stringent operating conditions (e.g., case temperature 180°C, gate-source voltage +7.5V) to form training data pairs. These data pairs are then input into the network model being trained. The model will output two predicted lifetime values ​​for the same fingerprint vector under the two operating conditions. (Predicted lifespan under mild operating conditions) and (Predicted lifetime under harsh operating conditions). The physical consistency constraint penalizes violations of physical laws (e.g., the predicted lifetime under harsh operating conditions is not significantly shorter than the predicted lifetime under mild operating conditions), where... This is a small positive safety margin, whose value is based on the statistical standard deviation of the ratio of lifetimes of similar devices under mild and harsh operating conditions. It is typically taken as 0.05-0.1 to ensure that there is a reasonable order of magnitude difference between predicted lifetimes.

[0106] For example, in each training round, a sample is randomly selected from the training batch as an anchor point, and its degenerate fingerprint vector is: Operating parameters are Subsequently, another operating condition parameter is searched from the same batch or dataset. To ensure that key operating parameters (such as junction temperature and gate voltage) are met, The values ​​are all clearly greater than (or less than) The value, thus forming a ( , (mild) and ( , (Strict) data pairs.

[0107] The second loss function satisfies the following formula:

[0108]

[0109] in, For the second loss function, The number of samples in the actual failure training set. The first in the actual failure training set The true failure lifetime value of each sample The first in the actual failure training set The predicted failure lifetime value for each sample.

[0110] Step S503: Based on the neural network, the initial prediction model is obtained by training the physical simulation training set with the first loss function.

[0111] In this embodiment, a multi-layer fully connected neural network is constructed as the main body of the prediction model. Its input layer receives a feature vector composed of a degradation fingerprint vector and operating parameters. This feature vector undergoes a nonlinear transformation through three hidden layers (each containing 32 to 128 neurons, using the ReLU activation function). Finally, the predicted logarithmic lifetime value is output through a linear neuron in the output layer. The network is trained using the physical simulation training set, with a first loss function as the optimization objective, and the Adam optimizer (initial learning rate set to 0.001) is used for parameter updates. During training, batch gradient descent is used in each round, and the loss change is monitored using a validation set. When the loss function no longer decreases significantly after several consecutive rounds, an early stopping strategy is implemented to prevent overfitting. The initial prediction model is then obtained. This initial prediction model has learned the basic mapping relationship between degradation features, operating parameters, and lifetime from large-scale simulation data, and incorporates the physical constraint that higher operating parameters (corresponding to more stringent real-world scenarios) lead to shorter lifetimes.

[0112] Step S504: Based on the second loss function, the initial prediction model is calibrated using the measured failure training set to obtain a lifetime prediction model.

[0113] In the embodiment, the initial prediction model is trained in the second stage (calibration) by using a transfer learning strategy. Specifically, a progressive fine-tuning strategy is used: first, the network parameters close to the input layer (such as the input layer and the first hidden layer) in the initial prediction model are frozen, and only the subsequent network layers are trained using the measured failure training set and the second loss function, the learning rate is set to 0.0001, then all network parameters are unfrozen, and the entire network is fine-tuned with a smaller learning rate (such as 0.00001). The Adam optimizer is used during training, and the early stopping mechanism is enabled to prevent overfitting. The purpose of this stage of training is to finely calibrate the life mapping relationship of the model using valuable measured data, correct the systematic deviation between the simulation model and the real physical process, and thus obtain a final model (i.e., the life prediction model) with more accurate prediction results and stronger generalization ability.

[0114] It should be noted that the present application effectively combines the completeness of simulation data, the constraint of physical laws and the authenticity of measured data by pre-training based on large-scale physical simulation data and fine-tuning using small-sample measured data in a batch training strategy. This strategy enables the model to learn the basic life mapping relationship from simulation in a wide stress range and embed physical constraints, and to correct the simulation model deviation and prevent small-sample overfitting through measured data, thereby achieving an optimal balance between data utilization efficiency and model prediction accuracy and generalization ability, which is the key to improving the engineering practicality and reliability of the life prediction model.

[0115] In addition, the input data normalization of the model (especially the training model described above) can effectively eliminate the dimensional differences between different features, balance the contribution weights of the working condition parameters and the degradation fingerprint vectors, and thus improve the model training convergence speed and the prediction result stability, reduce the interference of extreme values on the model parameter update, which is a conventional technical means in the field. As long as the training data during training and measurement are kept the same normalization processing logic, it is not necessary to repeat it here.

[0116] The life distribution of the GaN device to be measured is obtained by using the life prediction model based on the degradation fingerprint vector and the joint probability distribution.

[0117] In step S511, a plurality of groups of working condition parameter samples are generated based on the joint probability distribution by using the Monte Carlo sampling method.

[0118] In the embodiment, the total sample amount of Monte Carlo sampling is dynamically set according to the dimension of key working condition parameters and the complexity of the joint probability distribution thereof, and is usually not less than 5,000 groups. For high-dimensional (such as more than 4 parameters) or strongly nonlinearly correlated distribution, the total sample amount can be increased to more than 20,000 groups to ensure the significance of the statistical result. According to the type of the joint probability distribution, a specific sampling strategy is adapted. If the joint probability distribution is constructed by a Copula function, uniform distribution sampling is first performed on the Copula function, and then the sampling points are converted into sample values of original variables by using the inverse cumulative distribution function of each edge distribution to retain the correlation structure between parameters. If the joint probability distribution is the product of independent edge distributions, each working condition parameter (such as junction temperature, drain-source voltage, and gate-source voltage) is independently sampled according to its edge distribution type. For normal distribution, BoxMuller transformation is used. For uniform distribution, a pseudo-random number generator is used. For Weibull distribution, an inverse distribution function method is used. After sampling, the effectiveness of the sample is verified according to the absolute safety limit of the device (such as the junction temperature not exceeding 200°C and the gate-source voltage not exceeding 80% of the breakdown voltage), and abnormal samples exceeding the boundary are removed to ensure that all samples represent actual possible working conditions. Finally, a normalized multi-dimensional working condition parameter sample data set is generated.

[0119] In step S512, the degradation fingerprint vector and each group of working condition parameters are input into the life prediction model to obtain a plurality of single-point predicted life values.

[0120] In the embodiment, first, the degradation fingerprint vector and each group of working condition parameter samples (such as a multi-dimensional vector containing junction temperature, drain-source voltage, and gate-source voltage) are spliced in a preset order to form a complete input feature vector. The feature vector is input into a trained life prediction model. The model is a multi-layer fully connected neural network, the output layer neuron of which is linearly activated and directly outputs a logarithmic life value with base 10. Through forward propagation calculation of the model, a corresponding logarithmic predicted life value is obtained for each group of input feature vectors. Subsequently, the logarithmic life value is restored to a single-point predicted life value in actual hour units through exponential operation (power of the predicted value of 10). The above process is sequentially performed on all working condition parameter samples, and finally a plurality of single-point predicted life values equal to the sample amount are obtained to constitute a discrete life prediction sample set.

[0121] In step S513, the probability weight of the working condition parameter sample is determined based on the joint probability distribution.

[0122] In the present embodiment, the probability weight, i.e. the probability density value of each set of operating condition parameter samples in its joint probability distribution, quantifies the relative likelihood of the set of operating conditions occurring in the actual usage scenario. For each set of operating condition parameter samples, the probability density function of the joint probability distribution is called to calculate directly: if the joint probability distribution is composed of a Copula function and marginal distributions, the probability density of each parameter in its marginal distribution is calculated first, and then multiplied by the density value of the Copula function at the sample point; if it is a multivariate analytical distribution (such as multivariate normal distribution), the probability density function of the distribution is called directly. After the calculation is completed, the probability density values of all samples are normalized so that their sum is equal to 1, thereby obtaining the standardized probability weight corresponding to each sample, which will serve as an indicator of the contribution of each life prediction value in the subsequent weighted fitting.

[0123] Step S514, based on the single-point predicted life value and the probability weight, a reliability life distribution function of the GaN device under test is generated by fitting.

[0124] In the present embodiment, each single-point predicted life value is paired with its corresponding probability weight to form a set of weighted life data. A parameter fitting method is used to fit the weighted life data to a selected life distribution model. Since the failure time of semiconductor devices usually follows a Weibull distribution, the present application selects a two-parameter Weibull distribution model to fit the weighted life data, in order to represent the statistical characteristics of the life of the GaN device under test. The fitting process uses the probability weight as an indicator of the contribution of each data point, and optimizes the characteristic life parameters and shape parameters of the Weibull distribution by the weighted maximum likelihood estimation method. The goodness of fit test is performed on the fitting result to ensure that the generated reliability life distribution function, including the cumulative distribution function and the probability density function, can accurately represent the statistical characteristics and uncertainties of the life of the device under test.

[0125] Step S515, according to the reliability life distribution function, the life distribution of the GaN device under test is determined.

[0126] In the present embodiment, based on the reliability life distribution function, key characteristic indicators representing the life of the device are calculated and extracted, including the median life, B10 life and B1 life. By calculating the cumulative distribution function, the median life, B10 life (corresponding to a life with a cumulative failure probability of 10%) and B1 life (corresponding to a life with a cumulative failure probability of 1%) and other reliability characteristic values are directly obtained. At the same time, the average expected life of the device is calculated according to the probability density function, in addition, the reliability curve of the survival probability of the device changing with time is drawn from the reliability function (i.e. 1 minus the cumulative distribution function). Finally, these characteristic life values and the reliability curve together constitute the complete life distribution report of the GaN device under test, providing quantitative basis for reliability design, life evaluation and maintenance decision-making.

[0127] As Figure 2 shown, the application also provides a GaN device life prediction system, comprising: a processor, an input device, an output device and a memory, the processor, the input device, the output device and the memory are connected with each other, wherein the memory is used to store a computer program, the computer program comprises program instructions, the processor is configured to call the program instructions, and execute the related steps of the related embodiments of the GaN device life prediction method.

[0128] The GaN device life prediction system provided by the application can integrate each functional component in a processing component, or each component can exist physically alone, or two or more components can be integrated in one component. The integrated components can be realized in the form of hardware or in the form of software function.

[0129] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit it, although the application has been described in detail with reference to the above embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the above embodiments, or make equivalent replacement for part or all of the technical features, and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application, and they should be covered in the scope of the claims and the description of the application.

Claims

1. A method for rapid lifetime prediction of GaN devices, characterized in that, The method includes: Obtain the physical structure properties and usage scenario information of the GaN device under test; Based on the physical structural properties and the usage scenario information, key operating condition parameters and corresponding target stress conditions are determined. Based on the target stress condition, adaptive composite accelerated stress testing is performed on the GaN device under test to obtain a degradation fingerprint vector, including: Based on the target stress condition, adaptive composite accelerated stress testing is performed on the GaN device under test to obtain device degradation data. Extracting degradation fingerprint vectors from the device degradation data includes: Extract the voltage and current dynamic waveforms of the switching transients from the device degradation data; Time-domain features are extracted from the voltage dynamic waveform and the current dynamic waveform, respectively. Time-frequency characteristics are obtained by performing synchronous compressed wavelet transform on the voltage dynamic waveform and the current dynamic waveform, respectively. The physical feature weights are determined based on the physical structural properties. The time-domain features and the time-frequency features are weighted and fused using the physical feature weights to obtain a high-dimensional feature vector; The high-dimensional feature vector is reduced in dimensionality using a pre-trained sparse autoencoder to obtain a degenerate fingerprint vector. Construct a joint probability distribution of the key operating condition parameters based on the aforementioned usage scenario information; Constructing a lifetime prediction model, and obtaining the lifetime distribution of the GaN device under test based on the degradation fingerprint vector and the joint probability distribution using the lifetime prediction model, including: Multiple sets of operating condition parameter samples are generated using the joint probability distribution based on the Monte Carlo sampling method. Using the degraded fingerprint vector and each set of operating condition parameters as input, the lifetime prediction model is used to make predictions and obtain multiple single-point predicted lifetime values. The probability weights of the operating condition parameter samples are determined based on the joint probability distribution; The reliability lifetime distribution function of the GaN device under test is generated by fitting based on the single-point predicted lifetime value and the probability weight. The lifetime distribution of the GaN device under test is determined based on the reliability lifetime distribution function.

2. The method for rapid lifetime prediction of GaN devices according to claim 1, characterized in that, The process of determining key operating parameters and corresponding target stress conditions based on the physical structural properties and the usage scenario information includes: Key operating parameters are determined based on the physical structure properties and the usage scenario information; The value range of the key operating condition parameters is determined based on the usage scenario information. The range of values ​​is transformed into target stress conditions based on a pre-set acceleration model.

3. The method for rapid lifetime prediction of GaN devices according to claim 1, characterized in that, The process of obtaining device degradation data by performing adaptive composite accelerated stress testing on the GaN device under test based on the target stress condition includes: An initial composite accelerated stress test cycle is set based on the key operating parameters and the target stress conditions; An optimized composite accelerated stress test cycle is obtained by optimizing the initial composite accelerated stress test cycle using a Bayesian optimization algorithm within a preset safety boundary. Optimized composite accelerated stress test cycles were applied to the GaN device under test to obtain device degradation data.

4. The method for rapid lifetime prediction of GaN devices according to claim 3, characterized in that, The process of optimizing the initial composite accelerated stress test cycle within a preset safety boundary using a Bayesian optimization algorithm to obtain an optimized composite accelerated stress test cycle includes: The initial optimizable range of stress parameters for the initial composite accelerated stress test cycle is set according to the preset safety boundary and the target stress condition. Within the initial optimizable value range, multiple sets of initial stress parameter combination samples are selected; Based on the initial stress parameter combination sample, the GaN device under test is subjected to short-time stress cycling to obtain a comprehensive degradation index; Based on the comprehensive degradation index and the corresponding initial stress parameter combination sample, a Gaussian process regression surrogate model is constructed. With the goal of maximizing the expected improvement function, the optimal combination of stress parameters is determined by iterative optimization through updating the Gaussian process regression surrogate model. An optimized composite accelerated stress test cycle is obtained based on the optimal combination of stress parameters.

5. The method for rapid lifetime prediction of GaN devices according to claim 4, characterized in that, The comprehensive degradation index obtained by performing short-time stress cycling on the GaN device under test based on the initial stress parameter combination sample includes: Based on the initial stress parameter combination sample, the GaN device under test is subjected to short-time stress cycling to obtain multiple degradation feature slopes; The comprehensive degradation index is calculated based on the slope of the degradation characteristics.

6. The method for rapid lifetime prediction of GaN devices according to claim 1, characterized in that, The constructed lifetime prediction model includes: Obtain the physical simulation training set and the measured failure training set; Construct the first loss function and the second loss function; An initial prediction model is obtained by training the physical simulation training set using the first loss function based on the neural network. Based on the second loss function, the initial prediction model is calibrated using the measured failure training set to obtain the lifetime prediction model.

7. A system for rapid lifetime prediction of GaN devices, characterized in that, include: The device includes a processor, an input device, an output device, and a memory, which are interconnected. The memory stores a computer program that includes program instructions. The processor is configured to invoke the program instructions to execute a method for rapid lifetime prediction of a GaN device as described in any one of claims 1 to 6.

Images