Semiconductor device analysis method, system, and program product therefor

Abstract

The present application provides a semiconductor device analysis method, system and program product thereof, comprising: a training data collection step, obtaining at least one training data, the training data including: at least one characteristic parameter, time series data of at least one characteristic parameter measured at a fixed time interval in a first period and a second period longer than the first period; a training model step, training an artificial intelligence model based on the training data to learn the pattern of the change of the characteristic parameter with time; an input data obtaining step, obtaining the time series data of the characteristic parameter measured in the first period of the new plan as input data when the semiconductor device is used for the new plan; a prediction generation step, inputting the input data into the trained artificial intelligence model, and generating a predicted time series data of the characteristic parameter of the second period of the new plan by the artificial intelligence model.

Description

Technical Field

[0001] This invention relates to a semiconductor device analysis method, system, and program product thereof, particularly to a method, system, and computer-readable storage medium or program product thereof for analysis and prediction using reliability test data.

[0002] More specifically, it relates to a technique for predicting the long-term reliability characteristics of semiconductor devices (such as analog integrated circuits (AICs) manufactured using a specific process) by analyzing short-term reliability test data. Background Technology

[0003] The reliability of semiconductor devices, especially integrated circuits (ICs), is one of the key factors affecting the quality and lifespan of end electronic products. To ensure that semiconductor devices can operate stably for extended periods in their real-world application environments, the industry has developed various accelerated life test methods to assess the long-term reliability of products in a shorter timeframe.

[0004] High Temperature Operating Life (HTOL) testing is a standard test method widely used to evaluate the reliability and lifespan of semiconductor devices during long-term continuous operation under specific high-temperature environments and operating voltages.

[0005] According to relevant industry standards, such as JESD22-A108 specification developed by JEDEC (Joint Electron Device Engineering Council), HTOL testing typically involves specific high temperatures (commonly ranging from 125 to 175 degrees Celsius) and the component's rated operating voltage (e.g., 3.3V). During testing, a certain number of samples (e.g., 10 to several dozen to ensure statistical representativeness) are selected and subjected to prolonged continuous operation testing under the aforementioned stress conditions. The total testing duration often depends on product specifications or customer requirements, such as 1000 hours (hereinafter referred to as HTOL1000). During this period, key electrical parameters of the component are measured periodically or at the end of the test to determine whether failure or significant characteristic degradation has occurred.

[0006] In semiconductor industry practice, a proven silicon intellectual property (IP) design (such as a specific analog IP functional block) is often reused in multiple different design projects or products developed based on the same process technology. Theoretically, since these different projects use the same process technology and IP design (and may even have identical layouts), the analog IPs produced should have similar characteristics and reliability performance.

[0007] However, the actual semiconductor manufacturing process is extremely complex. Even with the same nominal process technology, minor process variations cannot be completely avoided due to differences in time, lot, or slight variations in equipment conditions. These minor variations may cause slight differences in the electrical characteristics or aging behavior of the final analog IP produced.

[0008] Therefore, to ensure that the reliability of each project or product incorporating the aforementioned analog IP meets the expected specifications and requirements, in the prior art, semiconductor manufacturers or design companies often still need to sample and perform complete and time-consuming HTOL tests (e.g., HTOL1000 tests lasting up to 1000 hours) for each new project or a representative production batch. While this ensures the reliability of individual projects, the lengthy testing time is clearly detrimental to rapid product development and time-to-market. Summary of the Invention

[0009] The main technical problem that this invention aims to solve is: how to significantly shorten the time required for reliability verification while ensuring sufficient confidence in the long-term reliability of analog IP (or other semiconductor devices) (i.e., effectively assessing the risk of process variation), so as to overcome the time burden caused by performing a complete long-term HTOL test in the prior art.

[0010] In other words, the present invention aims to provide a more efficient method for analyzing semiconductor devices, which is expected to accurately predict the results of long-term tests using test data from a shorter period of time, thereby reducing or avoiding the need to perform complete long-term tests.

[0011] This invention provides a semiconductor device analysis method, comprising: a training data collection step, acquiring at least one set of training data, the training data including: time series data of at least one characteristic parameter of the semiconductor device measured at fixed time intervals during a first time period and a second time period longer than the first time period; a model training step, training an artificial intelligence model based on the training data to learn the patterns of change of the multiple characteristic parameters over time; an input data acquisition step, acquiring the time series data of the characteristic parameter measured during the first time period of the new plan when the semiconductor device is used in a new plan as input data; and a prediction generation step, inputting the input data into the trained artificial intelligence model, and generating a predicted time series data of the characteristic parameter of the semiconductor device during the second time period of the new plan by the artificial intelligence model.

[0012] Additionally, the present invention provides a semiconductor device analysis system, comprising: a memory storing at least one instruction; and a processor coupled to the memory, the processor being configured to execute the plurality of instructions, causing the system to perform the following operations: a training data collection step, acquiring at least one training data, the training data being time series data of at least one characteristic parameter of a semiconductor device measured at fixed time intervals during a first period of a historical plan using the semiconductor device, the first period including a second period and the first period following the second period; a model training step, training an artificial intelligence model stored in the memory and accessed by the processor based on the plurality of training data to learn the pattern of the plurality of characteristic parameters changing over time; an input data acquisition step, acquiring time series data obtained by measuring at least one characteristic parameter at fixed time intervals during the second period of the first period of the new plan when the semiconductor device is used in a new plan, as input data; and a model generation data step, inputting the input data into the trained artificial intelligence model, the artificial intelligence model generating a predicted value of the at least one characteristic parameter of the semiconductor device during the first period of the first period of the new plan.

[0013] Additionally, this invention provides a semiconductor device analysis program product for enabling a computer to perform the following operations: a training data collection step, acquiring at least one set of training data, the training data being time-series data of at least one characteristic parameter of a semiconductor device measured at fixed time intervals during a first period of a historical plan using the semiconductor device, the first period including a second period and the first period following the second period; a model training step, training an artificial intelligence model based on the plurality of training data to learn the pattern of the plurality of characteristic parameters changing over time; an input data acquisition step, acquiring time-series data obtained by measuring at least one characteristic parameter at fixed time intervals during the second period of the first period of the new plan when the semiconductor device is used in a new plan, as input data; and a model generation step, inputting the input data into the trained artificial intelligence model, and having the artificial intelligence model generate a predicted value for the at least one characteristic parameter of the semiconductor device during the first period of the first period of the new plan. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the training architecture of the training phase of a semiconductor device analysis method according to an embodiment of the present invention.

[0015] Figure 2 The illustration shows a circuit module being applied in multiple planning scenarios according to an embodiment of the present invention.

[0016] Figure 3 The diagram illustrates a method for collecting information on the changes in characteristic parameters over time from historical planning data during reliability testing, according to an embodiment of the present invention.

[0017] Figure 4 An example graph illustrating the change over time of characteristic parameters of circuit module 20 collected under different historical planning scenarios according to an embodiment of the present invention is shown.

[0018] Figure 5 A schematic flowchart illustrating the "use phase" in one embodiment of the present invention is shown.

[0019] Figure 6 Displayed graphically Figure 5 The forecasting process during the usage phase is illustrated, particularly how to bridge short-term data from actual measurements with long-term forecast data generated by AI models.

[0020] Figure 7A The illustration depicts the "data collection" and model training phase according to an embodiment of the present invention.

[0021] Figure 7BThe prediction and decision-making process of the "use phase" according to an embodiment of the present invention is illustrated.

[0022] Symbol Explanation

[0023] 100: Training Architecture

[0024] 12, 52: Artificial Intelligence Model

[0025] A, B, C: Plan

[0026] 11A, 11B, 11C: Historical planning data

[0027] 11A1, 11B1, 11C1, 511: Characteristic parameter data for the first time period

[0028] 11A2, 11B2, 11C2: Characteristic parameter data for the second time period

[0029] 11A3, 11B3, 11C3, 513: Planning parameters

[0030] 512G: Predicted characteristic parameter data for the second time period

[0031] T: Fixed time interval

[0032] S701-S707: Steps

[0033] 51: New Plan Detailed Implementation

[0034] To make the objectives, technical content, features, and effects of this invention clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be noted that the following embodiments and drawings are merely illustrative and are not intended to limit the scope of this invention. Various modifications, alterations, and refinements made without departing from the spirit and principles of this invention should still fall within the protection scope of this invention. Furthermore, the same or similar element symbols in the drawings represent the same or similar functions or structures.

[0035] This invention provides a semiconductor device analysis method, system, and program product, which aims to use an artificial intelligence model to predict the long-term reliability characteristics of a target semiconductor device based on short-term reliability test data, and to determine whether time-consuming long-term reliability testing can be omitted, thereby saving verification time and accelerating the product development process.

[0036] In one embodiment, the semiconductor device may be a circuit module, such as an analog integrated circuit (Analog IC) or a resistive-capacitive oscillator (RC oscillator) circuit, and its characteristic parameter may be the output frequency of the oscillator. Embodiments of the present invention will now be described with reference to the accompanying drawings.

[0037] Please see Figure 1 The diagram illustrates a training architecture 100 for the training phase of a semiconductor device analysis method according to an embodiment of the present invention. The main objective of this phase is to establish and train an artificial intelligence model 12, enabling it to predict the long-term aging behavior of a target semiconductor device.

[0038] In this training phase, the first step is to collect training data. For example... Figure 1 As shown, the training data includes data from a plurality of historical plans or historical production batches (e.g., the first historical plan data 11A, the second historical plan data 11B, and the third historical plan data 11C illustrated). These historical plans all used the target semiconductor device (e.g., the same circuit module 20).

[0039] In one embodiment, each piece of historical plan data (taking the first historical plan data 11A as an example) includes information collected during reliability testing (e.g., high-temperature operating life (HTOL) testing) of the circuit module 20 in the historical plan. In one embodiment, the circuit module 20 is an RC oscillator circuit, but is not limited thereto.

[0040] In one embodiment, the first historical planning data 11A includes: characteristic parameter data 11A1 for a first time period (e.g., HTOL-168); characteristic parameter data 11A2 for a second time period (e.g., HTOL-1000); and planning parameters 11A3.

[0041] Specifically, the first time period (e.g., HTOL-168) characteristic parameter data 11A1 is a measurement record of at least one characteristic parameter (e.g., the frequency of the RC oscillator) of the circuit module 20 of the first historical plan changing over time within a relatively short test time interval (i.e., the first time period, e.g., the first 168 hours of the HTOL test). In one embodiment, this data may be time series data measured at fixed time intervals.

[0042] The characteristic parameter data 11A2 for the second time period (e.g., HTOL-1000) is the characteristic parameter data obtained by measuring the first historical plan in the second time period (e.g., HTOL-1000). This is a measurement record of the change of at least one characteristic parameter over time in the circuit module 20 within a complete and relatively long test time interval (i.e., the second time period, e.g., 1000 hours of HTOL testing).

[0043] In one embodiment, the second time-period characteristic parameter data 11A2 is time-series data, and its time span covers the first time period and the period from the end of the first time period to the end of the second time period. This second time-period characteristic parameter data 11A2 represents the actual long-term aging behavior of the circuit module 20 and is the target of the artificial intelligence model 12 for learning, but is not limited thereto.

[0044] In one embodiment, the planning parameters 11A3 of the first historical plan may include other information related to the historical plan, such as, but not limited to, specific process parameters during the manufacturing of the batch of devices, subtle differences in operating voltage or temperature conditions during testing, and layout characteristics of the devices. These parameters help the artificial intelligence model 12 learn the impact of process variations or environmental factors on aging behavior.

[0045] In one embodiment, the second historical plan data 11B and the third historical plan data 11C also have a similar structure and content to the first historical plan data 11B, so repeated descriptions are omitted here.

[0046] In one embodiment, the second historical planning data 11B includes first time period characteristic parameter data 11B1, second time period characteristic parameter data 11B2, and planning parameter 11B3; the third historical planning data 11C includes first time period characteristic parameter data 11C1, second time period characteristic parameter data 11C2, and planning parameter 11C3, but is not limited thereto.

[0047] In one embodiment, the collected first historical plan data 11A to third historical plan data 11C, etc., together constitute the complete training data for training the artificial intelligence model 12. This step is an example of the "training data collection step" described in this invention.

[0048] Next, in the "model training step," the aforementioned collected training data is input into the artificial intelligence model 12 for training. In one embodiment, the artificial intelligence model 12 has a model architecture capable of processing and generating sequence data. For example, it may employ a generative pre-trained transformer (GPT) model, a recurrent neural network (RNN), a long short-term memory network (LSTM), or other machine learning or deep learning models suitable for time series prediction and generation, but is not limited to these.

[0049] The effectiveness of the artificial intelligence model 12 lies in learning the inherent patterns and trends of the characteristic parameters of the semiconductor device as they age over time, as well as the variability that may be affected by the planned parameters 11A3, 11B3, and 11C3, by processing a large amount of training data.

[0050] In one embodiment, the artificial intelligence model 12, during the training process, can learn how to generate subsequent characteristic parameter sequences starting from an initial sequence of characteristic parameters (e.g., characteristic parameter data 11A1 of the first time period), and how to make the generated sequences approximate the actual measured long-term data (e.g., characteristic parameter data 11A2 of the second time period) as closely as possible. It can be understood that the artificial intelligence model 12 receives data from the first historical plan data 11A to the third historical plan data 11C as training input, but is not limited to this; although there are three sets of historical plan data in this example, there could be fewer or more sets. Although... Figure 1 The training phase is only illustrated, but it should be understood that the AI ​​model 12 trained in this phase will be used in the subsequent prediction and decision-making phases (described later).

[0051] Please see Figure 2 The figure illustrates a schematic diagram of a circuit module 20 being applied in multiple planning scenarios according to an embodiment of the present invention. This figure illustrates the collection... Figure 1 The background context of the training data is shown. In one specific embodiment of the present invention, the semiconductor device to be analyzed is this reusable circuit module 20. For example, this circuit module 20 may be a standard cell in a semiconductor process, an analog IP block, such as the aforementioned RC oscillator design. The present invention is particularly applicable to such circuit modules 20 that are repeatedly used in different chip designs or different production batches under the same process technology.

[0052] like Figure 2 As shown, the same circuit module 20 design is applied to different plans, such as "Plan A", "Plan B", and "Plan C" illustrated. These different planning scenarios correspond to... Figure 1 The training data sources provided are first historical plan data 11A, second historical plan data 11B, and third historical plan data 11C. Specifically, the first historical plan data 11A comes from measurement results obtained when the circuit module 20 is applied to scenario "Plan A"; the second historical plan data 11B comes from scenario "Plan B"; and the third historical plan data (11C) comes from scenario "Plan C". Figure 2For ease of explanation, the differences between the plans are expressed by the different relative positions of the circuit module 20 in plans A, B, and C. However, the differences between the plans are not limited to this. In one embodiment, the differences between plans A, B, and C can be expressed as the differences between plan parameters 11A3, 11B3, and 11C3.

[0053] The technical problem this invention aims to solve stems from the fact that although "Plan A," "Plan B," and "Plan C" use the same circuit module 20, slight process variations are unavoidable in their respective semiconductor manufacturing processes. Furthermore, the chip environment in which they are ultimately integrated or the operating / testing conditions they endure may differ. These factors lead to variations in the actual characteristic parameters (e.g., RC oscillation frequency) and their aging behavior over time, even for the same circuit module 20 design, across different plans. These actually measured aging data, which include individual differences, constitute the... Figure 1 The first time period characteristic parameter data 11A1, 11B1, 11C1 (representing short-term aging results such as HTOL-168) and the second time period characteristic parameter data 11A2, 11B2, 11C2 (representing long-term aging results such as HTOL-1000) are included, as well as the planned parameters 11A3, 11B3, 11C3 (such as specific process batch number, test temperature, etc.) that may affect these results.

[0054] Please see Figure 3 This figure illustrates a schematic diagram of collecting information on the changes in characteristic parameters over time from historical planning data during reliability testing, according to an embodiment of the present invention. This figure further clarifies... Figure 1 The method of generating time series data in training data (e.g., first historical plan data 11A, second historical plan data 11B, and third historical plan data 11C).

[0055] In one embodiment, in order to obtain detailed aging data for training the artificial intelligence model 12, data from different historical programs (e.g.) are processed. Figure 3 The schematic representations of "Plan A", "Plan B" and "Plan C" all use the same circuit module 20 and have undergone reliability testing for a certain period of time.

[0056] In one embodiment, this test may be a High Temperature Operating Life (HTOL) test, such as HTOL-1000, with a total test duration of 1000 hours, which is also an example of the "second time period" referred to in this invention. In one embodiment, the test conditions (such as temperature and voltage) are set according to standard specifications (such as JEDEC JESD22-A108) or specific application requirements, but are not limited thereto.

[0057] Throughout the long-term reliability test, a key step involves repeatedly measuring and recording at least one critical characteristic parameter of the circuit module 20 at fixed time intervals T. The fixed time interval T should be chosen frequently enough to capture the dynamic changes in the characteristic parameter over time; for example, it can be set to every minute, every few minutes, every hour, or other suitable intervals determined by the actual aging rate. If a ring oscillator is used as an embodiment of the circuit module 20, the measured characteristic parameter can be the frequency of its output signal. Figure 3 The waveform on the right side illustrates the output signal as it changes over time, while the fixed time interval T indicates the time interval for frequency measurement.

[0058] By performing continuous or periodic measurements at fixed time intervals T throughout the entire long-term test period (the second phase), detailed time-series data describing the aging behavior of circuit module 20 under a specific planned scenario can be obtained. These time-series data, based on their covered time range, constitute... Figure 1 The training data shown is an important component. The data sequence corresponding to a relatively short time range in the early stage of the test (e.g., from 0 hours to 168 hours, i.e., the previously mentioned "first period") constitutes the first period characteristic parameter data 11A1, 11B1, and 11C1; while the complete data sequence covering the entire test period (e.g., from 0 hours to 1000 hours, i.e., the previously mentioned "second period") constitutes the second period characteristic parameter data 11A2, 11B2, and 11C2.

[0059] This type of time-series data, obtained through measurements at fixed time intervals T, is crucial for training artificial intelligence models12 capable of understanding and generating sequential data (such as GPT, LSTM, etc.). Its advantage lies in providing rich information about dynamic changes, enabling the AI ​​model12 to learn complex patterns of drift or degradation of characteristic parameters such as frequency over time, rather than simply static values ​​at a few discrete time points. The training data obtained in this way (such as…) Figure 1 The first historical plan data (11A) to the third historical plan data (11C) shown contain time-related information, enabling the trained artificial intelligence model 12 to "continuously generate" subsequent long-term sequence data (e.g., predicted HTOL-1000 data) starting from a short initial sequence of data (e.g., HTOL-168 data from circuit module 20). This data collection and model training method provides a solid foundation for the core steps of rapid reliability prediction and decision-making in this invention.

[0060] Please see Figure 4It illustrates, according to an embodiment of the present invention, different historical planning scenarios (e.g., Plan A, Plan B, and Plan C shown in the figure, respectively corresponding to...) Figure 1 and Figure 2 The historical project data (11A, 11B, 11C) are used to collect example curves showing the change of characteristic parameters of circuit module 20 (taking the frequency of the RC oscillator as an example) over time. The horizontal axis of each sub-plot represents the time of reliability testing (e.g., HTOL), and the vertical axis represents the measured frequency. Data points (e.g., solid circles) schematically represent repeated measurements performed at fixed time intervals T during the testing period (e.g., 1000 hours). A baseline frequency (e.g., the design target frequency or initial measurement frequency of circuit module 20) may also be indicated in the plot to compare the frequency drift caused by aging.

[0061] from Figure 4 In the examples, it can be observed that the frequency aging curves of circuit modules 20 under different historical plans (plans A, B, and C) exhibit different patterns and degrees of drift, which intuitively reflects the individual differences caused by process variations or other factors. It is precisely this diverse time-series data, which reflects real physical behavior, that constitutes valuable training data for training the artificial intelligence model 12 (i.e., Figure 1 The first historical plan data 11A to the third historical plan data 11C include characteristic parameter data 11A1 to 11C1 for the first time period and characteristic parameter data 11A2 to 11C2 for the second time period, etc.

[0062] In one embodiment, when the Figure 4 Before the collected numerical time-series data (e.g., frequency series) is fed into the AI ​​model 12 for training, a data preprocessing step may be included. The purpose of this preprocessing step is to transform the raw data into a specific format more suitable for the AI ​​model 12 to learn. For example, the time-series numerical values ​​representing frequencies may be converted into fixed-length vector representations, or normalization may be performed to eliminate the influence of dimensions or adjust the numerical range, or techniques such as Fourier transform may be applied to extract frequency domain features. This step helps improve the stability and convergence speed of model training. This data preprocessing can be considered as part of the preparatory work for a broader "training data collection step" or "model training step," but is not limited to it.

[0063] In the "Model Training Step," the artificial intelligence model 12 learns using pre-processed training data. In one embodiment, the model learns to predict or generate a corresponding long-term sequence (targeting a second-time-period characteristic parameter data 11A2 that approximates the actual data) from the input short-term sequence (e.g., first-time-period characteristic parameter data 11A1). This process typically involves iteratively adjusting the model's internal parameters to minimize the prediction error between the model output and the actual target. In one embodiment, the prediction error may be defined by a loss function, but is not limited to this.

[0064] One or more termination conditions can be set regarding when model training should stop. In addition to monitoring whether the loss function converges, in one embodiment, a practical termination condition related to product reliability specifications and expected prediction accuracy can be set. For example, suppose the product specifications of circuit module 20 require that its frequency offset relative to the reference frequency must be controlled within ±4% after a second aging period (e.g., HTOL-1000). However, based on historical data (e.g., second-period characteristic parameter data 11A2), it is known that in actual testing, the final frequency offset of some samples may be +2%. In this case, the training termination condition for artificial intelligence model 12 can be set as follows: training stops when the model's prediction error for the long-term frequency offset of all (or most) samples in the training data, relative to the actual observed value, has narrowed to a preset tolerance range (e.g., 2%). The selection of this tolerance range needs to consider the balance between product specifications and prediction risk; for example, the prediction error can be set to be less than 1%. In this way, the model's predictions (e.g., prediction bias falling between +1% and +3%) ensure a high probability of conforming to product specifications (±4%), while not deviating too far from the actual situation (+2%). Setting such termination conditions based on prediction accuracy and specifications ensures that the final trained AI model 12 has sufficient predictive power, and its predictions have practical reference value for subsequent decisions on whether to omit long-term testing, avoiding erroneous decisions due to inaccurate predictions (overly optimistic or pessimistic).

[0065] The preceding paragraphs and corresponding diagrams detail how to collect training data containing time-series information, and how to use this data to train an artificial intelligence model 12 to enable it to simulate and predict the aging behavior of semiconductor devices, including setting appropriate training termination conditions.

[0066] In one embodiment, once the artificial intelligence model 12 has been trained in the manner described above, the trained artificial intelligence model 52 can be deployed in practical applications to perform a rapid reliability assessment of the new target plan. This is the "use" phase of the method of the present invention, which is described in detail below.

[0067] Specifically, when a new project also uses the same semiconductor device, only a short-term (first time period, e.g., HTOL-168) reliability test needs to be performed on the semiconductor device to obtain its first-time characteristic parameter data as "input data". This "input data" is then fed into the trained artificial intelligence model 52. Leveraging its learned aging patterns and generation capabilities, the artificial intelligence model 52 can generate a "prediction result" for the target device over a long time period (second time period, e.g., HTOL-1000), i.e., the second-time characteristic parameter data, based on this short-term input data. Since the model outputs continuous data, the values ​​corresponding to the end time of the first time period (HTOL-168) to the end time of the second time period (HTOL-1000) can be extracted from the sequence of continuous data as the prediction result, but this is not limited to this.

[0068] Please see Figure 5 The diagram illustrates a schematic flowchart of the "use phase" in one embodiment of the invention, demonstrating how a trained AI model is used to predict a new plan 51. This new plan 51 also incorporates the same circuit modules 20 as the historical plan data 11A to 11C. Figure 5 The central "AI Model" square represents the aforementioned artificial intelligence model 52, which has been trained based on historical data.

[0069] During this usage phase, for new plan 51, an operation equivalent to the "input data acquisition step" is performed. For example... Figure 5 As shown on the left, only a short-term reliability test (e.g., HTOL-168) needs to be performed on the circuit module 20 in the new plan 51 to obtain its characteristic parameter time series data within a first time period, which is the first time period (HTOL-168) characteristic parameter data marked in the figure. This first time period characteristic parameter data 511 is then provided as the actual input to the artificial intelligence model 52. In one embodiment, if there are specific parameters related to the new plan 51, they can also be provided to the model as plan parameters 513.

[0070] Then, an operation equivalent to the "prediction generation step" is performed. The first time-period characteristic parameter data 511 (and possible plan parameters 513) from the new plan 51 are input into the trained artificial intelligence model 52, and their connection relationship is as follows: Figure 5 As indicated by the arrow.

[0071] In one embodiment, the artificial intelligence model 52 functions by using its learned aging pattern knowledge to process the input sequence and generate (predict) a sequence of characteristic parameter evolutions of the circuit module 20 over a longer subsequent time period (until the end of the second time period, e.g., HTOL-1000). The output generated by this model is the "prediction result," such as... Figure 5The 512G of characteristic parameter data for the second time period (HTOL-1000) shown on the right is an example of the "predicted time series data" of this invention.

[0072] Figure 6 It is presented in a more intuitive graphical way. Figure 5 The diagram illustrates the prediction process of a fully trained model during its usage phase, particularly how it connects short-term, actual measured data to long-term prediction data generated by the AI ​​model. This figure depicts the prediction process for a "new plan" (corresponding to...). Figure 5 The curves showing the change of characteristic parameters (again, RC frequency) of the new plan 51) over time are also shown in the figure, with a reference frequency marked for comparison.

[0073] like Figure 6 As shown in the shaded area on the left, the solid data points represent the sequence of characteristic parameter data obtained through actual measurements during short-term reliability testing (e.g., HTOL-168, corresponding to the "first period" as defined in this invention) of the new plan. This sequence is... Figure 5 The first time period characteristic parameter data 511 is shown and labeled as "Actual Measurement Result (First 168 Hours)". This first time period characteristic parameter data 511 is the basic input provided to the artificial intelligence model 52 for subsequent prediction.

[0074] like Figure 6 As shown in the shaded area on the right, the hollow data points represent the subsequent characteristic parameter data sequence generated (predicted) by the artificial intelligence model 52 based on the input first-period characteristic parameter data 511 (i.e., the actual measurement results on the left). This generated sequence (labeled "AI model generated result") is... Figure 5 The predicted second-period characteristic parameter data 512G shown represents the model's prediction of the aging behavior of the new plan over a longer period (corresponding to the "second period" as defined in this invention, such as extending from 168 hours to 500 hours (HTOL 500), 1000 hours (HTOL 1000) or longer).

[0075] Figure 6 The core effectiveness and data connection of the method in the application phase of this invention are clearly demonstrated: using the characteristic parameter data 511 of the first time period, which is actually executed and has a relatively short duration, as input "hints," the artificial intelligence model 12, which has previously learned the aging pattern, is driven to generate predictive characteristic parameter data 512G representing long-term aging behavior in the second time period. The figure also schematically shows that the predicted values ​​corresponding to specific target time points (e.g., 500 hours (HTOL 500), 1000 hours (HTOL 1000)) can be extracted from the generated predictive characteristic parameter data 512G in the second time period.

[0076] In one embodiment, the subsequent "decision step" is based on the predicted value of the second-period characteristic parameter data 512G. By comparing the predicted value with the product specifications, it can be determined whether the actual long-term testing should be omitted.

[0077] In one embodiment, after obtaining the predicted second-period characteristic parameter data 512G, the "decision step" can be performed. The predicted characteristic parameter value (e.g., the predicted HTOL-1000 frequency offset) represented by the predicted second-period characteristic parameter data 512G is compared with the predetermined product reliability specification or acceptance standard. If the comparison result shows that the predicted value 512G meets the standard, a decision can be made to determine that the circuit module 20 in the new plan 51 has sufficient long-term reliability, thus eliminating the need for the time-consuming and laborious actual long-term HTOL test (e.g., HTOL-1000).

[0078] Although the illustrations and descriptions use an RC oscillator (referred to as RC in the figure) and its frequency as an example, the method disclosed in this invention for training an AI model using historical time series data and predicting long-term aging behavior based on short-term measured data can be extended to other analog ICs or semiconductor devices. It is only necessary to select appropriate key characteristic parameters and collect the corresponding time series data.

[0079] Furthermore, to achieve better predictive results, as disclosed in the invention, the data type input to the artificial intelligence model should be consistent with the data type of the expected output. For example, if the goal is to predict frequency changes over time, the input first-period characteristic parameter data should primarily consist of time-series data of the frequency, so that the model can focus on learning and generating the aging dynamics of the specific parameter.

[0080] Please see Figure 7A and Figure 7B The flowchart shown. Figure 7A and Figure 7B The flowchart fully connects the core steps of the semiconductor device analysis method of the present invention, and integrates the concepts illustrated in the foregoing embodiments and figures. Figure 7A The diagram illustrates the "data collection" stage, or model training stage, in the method of this invention. Figure 7B It illustrates the forecasting and decision-making process during the "use phase".

[0081] exist Figure 7A In the data collection and training phase, the process may begin with step S701, which represents a semiconductor device design to be analyzed (e.g., corresponding to...). Figure 2 , Figure 3 Circuit module 20, taking an RC oscillator as an example (hereinafter referred to as RC in the figure), already exists or is completed.

[0082] Then, step S702 is performed, which is an example of the "training data collection step" described in this invention. In this step, data is collected from at least one (preferably multiple, such as...) Figure 1 The plans A, B, and C shown use training data from historical examples (e.g., historical production batches or historical plans) of the circuit module 20 design. The key component of this training data is time-series data of at least one characteristic parameter (e.g., frequency) of the circuit module 20. This time-series data is obtained by measuring at fixed time intervals T during a first time period (e.g., HTOL-168) and a second time period longer than the first time period (e.g., HTOL-1000). This collected data corresponds to… Figure 1 The historical planning data 11A, 11B, and 11C shown include characteristic parameter data 11A1, 11B1, and 11C1 for the first time period, characteristic parameter data 11A2, 11B2, and 11C2 for the second time period, and planning parameters 11A3, 11B3, and 11C3.

[0083] Subsequently, in step S703 (this step is an example of the "training model step" described in this invention), the training data collected in step S702 (historical plan data 11A, 11B, 11C, etc.) is used to train the artificial intelligence model 12 (e.g., Figure 1 The training architecture 100). The purpose of training is to enable the artificial intelligence model 12 to learn the pattern of change of the characteristic parameters (e.g., frequency) over time (covering the first and second time periods). This training process may include the aforementioned data preprocessing and termination conditions based on prediction accuracy and specifications.

[0084] Next, we will enter the usage phase. Figure 7B This corresponds to applying the trained artificial intelligence model 52 to analyze a new plan (e.g.) Figure 5 The scenario is described in the new plan 51). Step S704 marks the beginning of this scenario. In step S705 (this step is an example of the "input data acquisition step" described in this invention), for the circuit module 20 in this new plan, a reliability test corresponding to the first time period (e.g., HTOL-168) is performed, and its characteristic parameters (e.g., frequency) are measured at fixed time intervals T to obtain the time series data of the first time period. This data is used as the input data (corresponding to...). Figure 5 The first time period characteristic parameter data (511).

[0085] In step S706 (an example of the "prediction generation step" described in this invention), the input data from step S705 (first time period characteristic parameter data 511) is input to the artificial intelligence model 52, which has been trained in step S703. Based on this input short-term time series data, the artificial intelligence model 52 generates predicted characteristic parameter time series data covering the subsequent second time period (e.g., HTOL-1000). This output is the "predicted time series data" (corresponding to...). Figure 5 The predicted characteristic parameter data for the second time period is 512G.

[0086] Finally, in step S707 (an example of the "decision step" described in this invention), the prediction result generated in step S706 (predicted second-period characteristic parameter data 512G) is compared with a preset reliability acceptance standard. If the prediction result shows that it meets the standard ("acceptable"), it can be decided to omit the actual, time-consuming second-period (e.g., HTOL-1000) test for the new plan, thereby achieving the effect of saving time and cost.

[0087] This achieves the goal of saving testing time and costs, as demonstrated by this invention. If the prediction results do not meet the standards, it indicates potential risks, and further long-term testing may be required for final confirmation. The complete process described above, including short-term testing, model prediction, and decision-making based on the prediction results, reflects the practical application value of the method of this invention.

[0088] In summary, through detailed descriptions of various embodiments and figures, this invention fully discloses a method, system, and process for analyzing the reliability of semiconductor devices using artificial intelligence models. The method, by learning aging patterns from historical data, can effectively predict the long-term reliability characteristics of a target device based on short-term test data, and thereby optimize the testing process, demonstrating significant practicality and industrial value. The content disclosed in this specification is sufficient to support the claims of this invention. The above descriptions are merely preferred embodiments of this invention and should not be construed as limiting the scope of the invention. Any simple equivalent changes and modifications made in accordance with the claims and description of this invention shall still fall within the scope of this patent.

Claims

1. A method for analyzing semiconductor devices, characterized in that, Include: The training data collection step involves obtaining at least one training data set, which includes: time series data of at least one characteristic parameter of a semiconductor device, measured at fixed time intervals during a first time period and a second time period longer than the first time period. The training model step involves training an artificial intelligence model based on at least one of the training data to learn the pattern of change of at least one of the characteristic parameters over time. The input data acquisition step involves acquiring the time series data of the characteristic parameters measured during the first time period of the new plan when the semiconductor device is used for a new plan, as input data. The prediction generation step involves inputting the input data into the trained artificial intelligence model, which then generates a predicted time series of data for the characteristic parameters of the semiconductor device during the second time period of the new plan.

2. The semiconductor device analysis method as described in claim 1, characterized in that, The artificial intelligence model includes a generative model.

3. The semiconductor device analysis method as described in claim 1, characterized in that, The first time period corresponds to a 168-hour high-temperature operating life test period; The second time period corresponds to a 1000-hour high-temperature operating life test period.

4. The semiconductor device analysis method as described in claim 3, characterized in that, It also includes: The decision-making steps include comparing the predicted time series data with a predetermined reliability criterion; and When the predicted time series data meets the predetermined reliability standard, it is determined that the operation life test corresponding to the second time period can be omitted for the semiconductor device.

5. The semiconductor device analysis method as described in claim 1, characterized in that, The training data comes from time-series data of the characteristic parameters collected during previous high-temperature operation life tests of the semiconductor device.

6. The semiconductor device analysis method as described in claim 1, characterized in that, The artificial intelligence model generates sequence data representing the continuous change of the at least one characteristic parameter over time based on the input data.

7. The semiconductor device analysis method as described in claim 1, characterized in that, The semiconductor device is an analog integrated circuit.

8. The semiconductor device analysis method as described in claim 1, characterized in that, The second time period includes the period that overlaps with the first time period.

9. A semiconductor device analysis system, characterized in that, Include: A memory that stores at least one instruction; and A processor, coupled to the memory, is configured to execute at least one of the instructions, causing the system to perform the following operations: The training data collection step involves acquiring at least one training data point, which is time-series data of at least one characteristic parameter of a semiconductor device measured at fixed time intervals within a first period of a historical plan for using the semiconductor device. The first period includes a second period and the first period following the second period. The training model step involves training an artificial intelligence model stored in the memory and accessed by the processor based on at least one of the training data, in order to learn the pattern of at least one of the characteristic parameters changing over time. The input data acquisition step involves acquiring time-series data obtained by measuring at least one characteristic parameter at fixed time intervals during the second time period of the first time period of the new plan when the semiconductor device is used for a new plan, as input data. The model generates data by inputting the input data into the trained artificial intelligence model, which then generates a predicted characteristic parameter of the semiconductor device for the first time period of the new plan.

10. A semiconductor device analysis program product, characterized in that, To enable a computer to perform the following operations: The training data collection step involves acquiring at least one training data point, which is time-series data of at least one characteristic parameter of a semiconductor device measured at fixed time intervals within a first period of a historical plan for using the semiconductor device. The first period includes a second period and the first period following the second period. The training model step involves training an artificial intelligence model based on at least one of the training data to learn the pattern of change of at least one of the characteristic parameters over time. The input data acquisition step involves acquiring time-series data obtained by measuring at least one characteristic parameter at fixed time intervals during the second time period of the first time period of the new plan when the semiconductor device is used for a new plan, as input data. The model generates data by inputting the input data into the trained artificial intelligence model, which then generates a predicted characteristic parameter of the semiconductor device for the first time period of the new plan.

Images