Method for predicting el wavelength hit rate of LED chip based on epitaxial PL data

Abstract

This invention discloses a method for predicting the EL wavelength hit rate of LED chips based on epitaxial photoluminescence (PL) data. By collecting historical photoluminescence data from epitaxial wafers and corresponding electroluminescence data from the chips, and combining this data with the target parameters of the target LED chip product, a cascaded prediction model is constructed to achieve direct prediction from photoluminescence characteristics to electroluminescence hit rate. This invention can predict the electroluminescence wavelength hit rate in advance during the epitaxial stage, enabling early classification and rational utilization of epitaxial wafers, significantly improving the utilization rate of a single batch, and reducing production costs.

Description

Technical Field

[0001] This invention relates to the field of semiconductor technology, and in particular to a method for predicting the EL wavelength hit rate of LED chips based on epitaxial PL data. Background Technology

[0002] MOCVD equipment is currently the mainstream growth equipment for LED epitaxial wafers. After epitaxial growth is completed, photoluminescence (PL) testing is typically used to perform rapid, non-destructive online detection of the epitaxial wafer, obtaining PL spectral parameters such as peak wavelength, full width at half maximum (FWHM), and luminous intensity. After the chip fabrication process is completed, electroluminescence (EL) testing is required to obtain the true wavelength of the device under powered conditions.

[0003] EL wavelength is the core parameter that determines the color and color coordinates of LED light emission, while EL wavelength hit rate (i.e., the proportion of actual EL wavelengths falling within the product specification range) is a key indicator for measuring the yield and production stability of epitaxial batches.

[0004] In large-scale production, to maintain stability and repeatability, although the mass production process parameters of each MOCVD equipment are usually kept constant, differences still exist between different MOCVD equipment and different carriers within the same MOCVD equipment. Therefore, the PL wavelength and the final EL wavelength are not a simple one-to-one correspondence. The main reasons are: firstly, PL testing is based on band-edge radiative recombination under photoexcitation, while EL testing involves the transport and recombination of electrically injected charge carriers in the active region; their physical mechanisms are different. Secondly, process steps such as etching, electrode fabrication, and annealing in chip manufacturing also modulate the final EL wavelength to some extent. Therefore, directly determining the hit rate of the final EL wavelength based on the PL wavelength has a significant deviation.

[0005] Currently, the industry practice is to perform a rough screening at the epitaxial stage by setting an empirical threshold based on the mean and distribution of PL wavelengths, and then make a final judgment through EL testing after the chip manufacturing process is completed. This approach has the following obvious shortcomings: (1) Prediction lag leads to cost waste: EL testing can only be carried out after the chip manufacturing process is completed. If the final hit rate is not up to standard, the large amount of chip manufacturing costs (materials and labor hours) already invested cannot be recovered.

[0006] (2) Low prediction accuracy of empirical thresholds: The thresholds of PL single parameters or a few parameters set based on human experience cannot fully reflect the complex nonlinear relationship between PL spectral characteristics (such as peak shape, symmetry, full width at half maximum, multi-peak distribution, etc.) and the final EL hit rate.

[0007] (3) Lack of adaptability to multiple product specifications: Different product specifications correspond to different test currents and wavelength ranges (such as center value and width). Changes in these parameters will directly affect the calculation of EL wavelength distribution and hit rate. Existing methods cannot automatically adapt and accurately predict based on specific product specification parameters.

[0008] Therefore, there is an urgent need in this field for a method that can predict the EL wavelength hit rate in advance and accurately based on epitaxial PL data and in combination with product specification parameters. Summary of the Invention

[0009] The purpose of this invention is to provide a method for predicting the EL wavelength hit rate of LED chips based on epitaxial PL data, in order to address the existing technical situation. This method uses the PL data of the epitaxial wafer to predict the EL wavelength hit rate of the LED chip in advance.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting the EL wavelength hit rate of an LED chip based on epitaxial PL data, comprising: Step S1: Obtain the PL test data of the epitaxial wafer under test and the target parameters of the target LED chip product; wherein, the target parameters include the target wavelength λspec, the target wavelength range width Wspec, and the target current density Jspec. Step S2: Input the PL test data, target current density Jspec, and target wavelength range width Wspec into the trained first-level prediction model to obtain the predicted value of the EL wavelength statistical feature value. Step S3: Determine the predicted wavelength offset S0 based on the difference between the predicted value of the EL wavelength statistical characteristic value and the target wavelength λspec. Step S4: Input the PL test data, target current density Jspec, target wavelength range width Wspec, and predicted wavelength offset S0 into the trained second-level prediction model to obtain the predicted value of the EL wavelength hit rate corresponding to the epitaxial wafer under test. The construction steps of the first-level prediction model and the second-level prediction model include: Step A1: Obtain historical datasets, which include historical epitaxial wafer PL test data and corresponding historical LED chip EL test data; wherein, the EL test data includes at least current density J; Step A2: Set N different wavelength range widths W. For each wavelength range width W, perform distribution statistics on the EL test data to determine the EL wavelength statistical characteristic value; the wavelength range width W includes at least the target wavelength range width Wspec. Step A3: Using the PL test data, the corresponding current density J and wavelength range width W as training inputs, and the EL wavelength statistical feature value as training output, train the initial first-level prediction model to obtain the first-level prediction model. Step A4: Based on the EL wavelength statistical characteristic value, the wavelength range width W, and the preset wavelength offset S, construct the wavelength hit interval; calculate the proportion of EL test data that falls within the wavelength hit interval, and record it as the EL wavelength hit rate. Step A5: Using the PL test data and the corresponding current density J, wavelength range width W, and preset wavelength offset S as training inputs, and the EL wavelength hit rate as training output, train the initial second-level prediction model to obtain the second-level prediction model.

[0011] The EL wavelength hit rate prediction method provided by this invention starts from the PL test data of the epitaxial wafer and achieves early prediction of the EL wavelength hit rate through a cascaded prediction model (a first-level prediction model and a second-level prediction model). The core idea of ​​constructing the cascaded prediction model is as follows: First, divide multiple wavelength range widths, and statistically calculate the EL wavelength statistical feature values, PL test data, preset wavelength offset, and EL wavelength hit rate corresponding to the wavelength range widths to obtain the training dataset for the prediction model. First, the first-level prediction model establishes the relationship between the EL wavelength statistical feature values, PL test data, and corresponding current density. Then, the second-level prediction model establishes the relationship between the EL wavelength hit rate, PL test data, corresponding current density, wavelength range width W, and preset wavelength offset. Based on the data collected from multiple widths and offsets, the cascaded prediction model is trained, and the prediction accuracy is significantly better than the traditional empirical threshold method. When using the first-level / second-level prediction model for prediction, the predicted wavelength offset is fully considered, and automatic adaptation is achieved by introducing the target parameter (current density) of the target LED chip product, effectively solving the problems of existing technologies that cannot predict EL hit rate from PL, prediction lag, and inability to adapt to multi-specification production.

[0012] As an optional scheme of the EL wavelength hit rate prediction method of the present invention, in step A2, the statistical characteristic value of the EL wavelength is the EL mode wavelength λ. EL .

[0013] As an optional scheme of the EL wavelength hit rate prediction method of the present invention, the step of constructing the wavelength hit interval in step A4 is as follows: for each wavelength range width W, set T different preset wavelength offsets S, based on the obtained EL mode wavelength λ EL Construct the wavelength hit interval [λ] EL -SW / 2,λ EL -S+W / 2].

[0014] As an optional solution of the EL wavelength hit rate prediction method of the present invention, the value range of the wavelength range width W is 0<W≤50nm; the value range of the wavelength offset S is from -50nm to 50nm.

[0015] As an optional solution of the EL wavelength hit rate prediction method of the present invention, the PL test data includes: peak wavelength, dominant wavelength, full width at half maximum, luminous intensity, and total number of data points per epitaxial wafer.

[0016] As an optional solution of the EL wavelength hit rate prediction method of the present invention, the first prediction model and the second prediction model are configured to perform prediction by extracting PL statistical features from the PL test data based on each wavelength range width W; the PL statistical features include: PL mode wavelength λ PL , proportion of PL mode wavelength R PL , the mean, standard deviation, skewness and kurtosis of PL wavelengths, and the mean and standard deviation of full widths at half maximum; wherein the wavelength is a peak wavelength or a dominant wavelength.

[0017] As an optional solution of the EL wavelength hit rate prediction method of the present invention, both the first prediction model and the second prediction model are machine learning regression models established based on a regression modeling method; wherein the first prediction model is configured to replace the input current density J with a target current density Jspec to perform prediction; the second prediction model is configured to replace the input current density J with the target current density Jspec, and replace the wavelength range width W and a preset wavelength offset S with a target wavelength range width Wspec and a predicted wavelength offset S0 respectively to perform prediction.

[0018] As an optional solution of the EL wavelength hit rate prediction method of the present invention, the model construction step comprises: combining all historical data to train a unified prediction model for different target LED chip products and different epitaxial growth equipments; or, collecting historical data respectively for different target LED chip products and different epitaxial growth equipments, training corresponding sub-models to form a model library; when performing prediction, a corresponding sub-model is automatically selected from the model library for prediction according to the information of the current LED chip and the epitaxial growth equipment.

[0019] As an optional solution of the EL wavelength hit rate prediction method of the present invention, the method further comprises grading and screening epitaxial wafers according to the predicted value of the EL wavelength hit rate corresponding to the epitaxial wafer, and executing the original chip manufacturing plan only when the predicted value of the EL wavelength hit rate is higher than a preset threshold.

[0020] In a second aspect, the present invention further provides an EL wavelength hit rate prediction system for LED chips based on epitaxial PL data, comprising: The data acquisition module is used to acquire the PL test data of the epitaxial wafer under test and the target parameters of the target LED chip product; The first-level prediction module is used to input the PL test data and target parameters into the trained first-level prediction model to obtain the predicted value of the EL wavelength statistical feature value. The offset calculation module is used to determine the predicted wavelength offset based on the difference between the predicted value of the EL wavelength statistical characteristic value and the target wavelength; The secondary prediction module is used to input the PL test data, target parameters and the predicted wavelength offset into the trained secondary prediction model to obtain the EL wavelength prediction hit rate corresponding to the epitaxial wafer under test. The model training module is used to build and train the first-level prediction model and the second-level prediction model.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, some of which will become clear as the description proceeds, and others will be learned by practicing the invention. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the construction process of the first-level prediction model and the second-level prediction model in an embodiment of the present invention.

[0023] Figure 2 This is a flowchart illustrating the EL wavelength hit rate prediction step in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the technical solutions of this invention will be described in more detail below with reference to the embodiments and accompanying drawings. It should be noted that the accompanying drawings of this invention are all simplified and not precisely to scale, and are only used to facilitate the explanation of this invention. Obviously, the following embodiments are only some embodiments of this invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0025] This application provides a method for predicting the EL wavelength hit rate of an LED chip based on epitaxial PL data, including: Step S1: Obtain the PL test data of the epitaxial wafer under test and the target parameters of the target LED chip product; wherein, the target parameters include the target wavelength λspec, the target wavelength range width Wspec, and the target current density Jspec. Step S2: Input the PL test data, target current density Jspec, and target wavelength range width Wspec into the trained first-level prediction model to obtain the predicted value of the EL wavelength statistical feature value. Step S3: Determine the predicted wavelength offset S0 based on the difference between the predicted value of the EL wavelength statistical characteristic value and the target wavelength λspec. Step S4: Input the PL test data, target current density Jspec, target wavelength range width Wspec, and predicted wavelength offset S0 into the trained second-level prediction model to obtain the predicted value of the EL wavelength hit rate corresponding to the epitaxial wafer under test. The construction steps for the first-level prediction model and the second-level prediction model include: Step A1: Obtain historical datasets, which include historical epitaxial wafer PL test data and corresponding historical LED chip EL test data; wherein, the EL test data includes at least the current density J; Step A2: Set N different wavelength range widths W. For each wavelength range width W, perform distribution statistics on the EL test data to determine the EL wavelength statistical characteristic value. The wavelength range width W should at least include the target wavelength range width Wspec. Step A3: Using PL test data, the corresponding current density J and wavelength range width W as training inputs, and EL wavelength statistical characteristic values ​​as training outputs, train the initial first-level prediction model to obtain the first-level prediction model. Step A4: Based on the statistical characteristic value of EL wavelength, the wavelength range width W, and the preset wavelength offset S, construct the wavelength hit interval; calculate the proportion of EL test data that falls within the wavelength hit interval, and record it as the EL wavelength hit rate. Step A5: Using PL test data and the corresponding current density J, wavelength range width W, and preset wavelength offset S as training inputs, and EL wavelength hit rate as training output, train the initial second-level prediction model to obtain the second-level prediction model.

[0026] In some embodiments, in step A2, the statistical characteristic value of the EL wavelength is the mode wavelength λ of the EL. EL It should be noted that using the mode wavelength, rather than the mean wavelength, better reflects the location where the wavelengths are most concentrated, representing the mainstream wavelengths. This effectively avoids interference from long-tailed distributions or individual outliers on the overall wavelength representation, making it more robust and physically meaningful than the mean.

[0027] In some embodiments, the step of constructing the wavelength hit interval in step A4 is as follows: for each wavelength range width W, set T different preset wavelength offsets S, based on the obtained EL mode wavelength λ EL Construct the wavelength hit interval [λ] EL -SW / 2,λ EL-S+W / 2]. It should be noted that the interval is constructed with the EL mode wavelength shift S as the center and the specification width W as the interval. When a wavelength falls within the interval, it has clear physical significance, which directly corresponds to the product specification definition and wavelength shift amount; it can adapt to different specification widths without manually fixing the interval; all shift scenarios are fully covered by multiple S values, enabling the model to fully learn the relationship between shift and hit rate; meanwhile, it significantly expands training samples, improves data efficiency, and has simple and efficient calculation during online inference, which facilitates real-time prediction.

[0028] In some embodiments, the value range of the wavelength range width W is 0<W≤50nm; the value range of the wavelength shift S is -50nm to 50nm. It should be noted that the value range of the wavelength range width W covers the specification widths of mainstream products in the LED industry, and has good versatility. Setting the value range of the preset wavelength shift S within the above range allows collection of EL wavelength hit rates corresponding to multiple preset wavelength shift S values, enabling the second-level prediction model to fully learn the functional relationship between "shift amount—hit rate", avoiding extrapolation prediction and improving model robustness.

[0029] In some embodiments, the PL test data includes: peak wavelength, dominant wavelength, full width at half maximum, luminous intensity, and the total number of data points per epitaxial wafer.

[0030] In some embodiments, the first prediction model and the second prediction model are configured to perform prediction by extracting PL statistical features from the PL test data based on each wavelength range width W; the PL statistical features include: PL mode wavelength λ PL , proportion of PL mode wavelength R PL , mean value, standard deviation, skewness and kurtosis of PL wavelengths, as well as mean value and standard deviation of full widths at half maximum; the wavelength is peak wavelength or dominant wavelength. It should be noted that PL statistical features comprehensively characterize the photoluminescence characteristics of epitaxial wafers from multiple dimensions including wavelength central tendency, dispersion degree, distribution shape, spectral width, uniformity, etc. Using PL statistical features for prediction can effectively capture the potential law of EL wavelength shift from PL, thereby significantly improving the prediction accuracy of EL hit rate and model robustness.

[0031] In some embodiments, both the first and second prediction models are machine learning regression models built using regression modeling methods. The first prediction model is configured to replace the input current density J with the target current density Jspec to perform prediction. The second prediction model is configured to replace the input current density J with the target current density Jspec, and replace the wavelength range width W and the preset wavelength offset S with the target wavelength range width Wspec and the predicted wavelength offset S0, respectively, to perform prediction. It should be noted that by using a regression model (such as random forest, gradient boosting tree, support vector regression, or neural network) instead of a classification model, the regression model can directly output the hit rate value, facilitating subsequent graded selection. Compared with traditional empirical thresholding methods, machine learning regression models can automatically learn the complex nonlinear relationship between high-dimensional features and the hit rate, significantly improving prediction accuracy.

[0032] In some embodiments, the model building steps include: training a unified prediction model by combining all historical data for different target LED chip products and different epitaxial growth equipment; or, collecting historical data separately for different target LED chip products and different epitaxial growth equipment, training corresponding sub-models to form a model library; and automatically selecting the corresponding sub-model from the model library for prediction based on the current LED chip product and epitaxial growth equipment information. It should be noted that combining historical data from different LED chip products and different epitaxial growth equipment to train a unified model can fully utilize data diversity and improve the model's generalization ability and robustness. However, when there are significant process differences between different LED chip products or epitaxial growth equipment, or when the unified model cannot achieve ideal accuracy, training a separate sub-model for each combination (LED chip product, epitaxial growth equipment) can better capture the specific offset patterns under that combination. The sub-model library strategy allows each sub-model to use fewer hidden layer nodes or a simpler structure, resulting in faster training speed and higher prediction accuracy. During online prediction, the corresponding sub-model is automatically loaded based on the current order product and machine identifier, achieving a balance between accuracy and flexibility.

[0033] In some embodiments, the method further includes classifying and screening epitaxial wafers based on the predicted EL wavelength hit rate, and only executing the original chip manufacturing plan if the predicted EL wavelength hit rate is higher than a preset threshold. It should be noted that classifying and screening epitaxial wafers based on the predicted EL wavelength hit rate allows high-hit-rate epitaxial wafers to be used in the original order, while low-hit-rate epitaxial wafers can be used as experimental wafers or for other products. This achieves early classification and rational utilization of epitaxial wafers, significantly improving the utilization rate per furnace and reducing production costs.

[0034] Secondly, embodiments of this application also provide an LED chip EL wavelength hit rate prediction system based on epitaxial PL data, comprising: The data acquisition module is used to acquire the PL test data of the epitaxial wafer under test and the target parameters of the target LED chip product; The first-level prediction module is used to input PL test data and target parameters into the trained first-level prediction model to obtain the predicted value of EL wavelength statistical characteristic value; The offset calculation module is used to determine the predicted wavelength offset based on the difference between the predicted value of the EL wavelength statistical characteristic value and the target wavelength. The secondary prediction module is used to input PL test data, target parameters and predicted wavelength offset into the trained secondary prediction model to obtain the EL wavelength prediction hit rate corresponding to the epitaxial wafer under test. The model training module is used to build and train the first-level prediction model and the second-level prediction model.

[0035] The following are some embodiments of this application, and the embodiments of the present invention will further describe the technical solution of the present invention in detail. Example 1

[0036] This embodiment provides a method for predicting the EL wavelength hit rate of an LED chip based on epitaxial PL data. This embodiment uses a 4-inch GaN-based blue LED epitaxial wafer.

[0037] First, construct and train the first-level prediction model and the second-level prediction model, including the following steps: Step A1: Obtain historical datasets: Collect PL test data of historical epitaxial wafers and EL test data of LED chips after the corresponding epitaxial wafers have undergone a complete chip manufacturing process.

[0038] This embodiment collects data from 1000 4-inch GaN-based blue LED epitaxial wafers and corresponding LED chips. Specifically, the LED chip size is 300×300μm², the test current is 20mA, and the current density J=22.2A / cm². 2 .

[0039] Step A2: Extract statistical feature values ​​of EL wavelengths: For each epitaxial wafer, process its EL wavelength data as follows: Divide the entire wavelength distribution range into N consecutive adjacent intervals with a width of W, and count the number of wavelength data points within the width W of each wavelength range; denot the center value of the interval with the largest number of data points in the EL wavelength data as the EL mode wavelength λ. EL The ratio of the number of data points in the width range of this wavelength range to the total number is denoted as the mode wavelength ratio R of EL. EL Using the method described above, calculate the statistical characteristic values ​​of EL wavelengths for N different wavelength range widths W.

[0040] Specifically, in this embodiment, N=3 wavelength ranges with widths W: W=6nm, 8nm, 10nm. The mode wavelength of EL is calculated using the peak wavelength.

[0041] Step A3: Construct the first-level prediction model: using PL test data, current density J, and wavelength range width W as inputs, and the mode wavelength λ of EL as input. EL As output, a first-level machine learning regression model is trained using regression modeling methods to obtain an EL mode wavelength prediction model.

[0042] Step A4: Set wavelength offset to construct the hit interval and calculate EL wavelength hit rate: Under all wavelength range widths W, take T different preset wavelength offsets S to construct the hit interval [λ]. EL -SW / 2,λ EL -S+W / 2], and calculate the proportion of EL test data falling within the hit interval, denoted as EL wavelength hit rate H(S,W), thereby obtaining the EL wavelength hit rate corresponding to T different preset wavelength offsets S.

[0043] Specifically, in this embodiment, T=51 preset wavelength offsets S (-50 nm to 50 nm) are taken under each wavelength range width W, and the EL wavelength hit rate H(S,W) corresponding to each preset wavelength offset S is calculated.

[0044] Step A5: Construct the second-level prediction model: Use the PL test data, current density J, wavelength range width W and T different preset wavelength offsets S to form a joint feature vector as input, and use the corresponding EL wavelength hit rate H(S,W) as the output label. Train the second-level machine learning regression model through regression modeling method to obtain the EL hit rate prediction model.

[0045] The first-level and second-level prediction models are configured to extract PL statistical features from the PL test data based on the width W of each wavelength range to perform prediction. Specifically, for each epitaxial wafer, its PL wavelength data is processed as follows: Within the entire wavelength distribution range, N consecutive adjacent intervals with a width of W are divided, and the number of wavelength data points within each wavelength range width W is counted; the center value of the wavelength range width interval with the largest number of data points in the PL wavelength data is denoted as the PL mode wavelength λ. PL The ratio of the number of data points in the width range of this wavelength range to the total number is denoted as PL (mode wavelength ratio). PL Following the method described above, calculate the PL multi-wavelength data for N different wavelength range widths W. In addition, it is necessary to extract other PL statistical features corresponding to each wavelength range width W: mean / standard deviation / skewness / kurtosis of the peak wavelength, and mean / standard deviation of the half-width at half-maximum.

[0046] Online prediction using the first-level and second-level prediction models described above includes the following steps: Step S1: Obtain the PL test data of the epitaxial wafer under test and the target parameters of the target LED chip product, including the target wavelength λspec, the target wavelength range width Wspec, and the target current density Jspec.

[0047] Step S2: Input the PL test data, target current density Jspec, and target wavelength range width Wspec into the first-level prediction model trained in step A3 to obtain the predicted value λ of the EL mode wavelength. EL .

[0048] Step S3: Based on the difference between the predicted value of the EL mode wavelength and the target wavelength λspec, determine the predicted wavelength offset S0, where S0 = λspec. EL -λspec.

[0049] Step S4: Input the PL test data, target current density Jspec, target wavelength range width Wspec, and predicted wavelength offset S0 into the second-level prediction model trained in step A5 to obtain the predicted value of the EL wavelength hit rate corresponding to the epitaxial wafer under test.

[0050] In this embodiment, the target current density Jspec of the target LED chip product is 22.2 A / cm². 2 The target peak wavelength range is 450nm-460nm. The target wavelength is calculated based on the center value, yielding λspec=455nm. The target wavelength range width Wspec=5nm. Using the first-level prediction model, the predicted value λ of the EL mode wavelength is obtained. EL =461nm. The predicted wavelength offset S0 was calculated to be -6nm. The final EL wavelength prediction hit rate H(S,W) was 35%. This is lower than the preset threshold of 60%, so production of this target LED chip product was abandoned.

[0051] The extraeptaxial wafer will be redirected to a new target LED chip product, with a new target current density Jspec of 15 A / cm². 2 The wavelength range is 480nm-490nm. The target wavelength is calculated based on the center value, resulting in λspec = 485 nm. The relevant parameters are then re-input into the first-level prediction model trained in step A3 to obtain the predicted value λ of the EL mode wavelength. EL =485.5nm. The predicted wavelength offset S0 = 0.5nm is calculated, and the relevant parameters are then input into the second-level prediction model trained in step A5. The prediction hit rate is 90%, which is higher than the preset threshold, so this target LED chip product can be put into production. Example 2

[0052] The basic steps in this embodiment are the same as those in Embodiment 1, with the main difference being: This embodiment uses a 4-inch GaN-based yellow LED epitaxial wafer as an example. The LED chip size is 27 mil, the test current is 20 mA, and the current density J = 5.71 A / cm². The target LED chip product requires an EL dominant wavelength range of 550 nm to 557.5 nm. Five wavelength ranges are selected: W = 6 nm, 7.5 nm, and 12 nm. PL and EL test data are collected from 300 epitaxial wafers. For both PL and EL, the dominant wavelength is used to calculate the mode wavelength and its proportion.

[0053] Calculate the mode wavelength λ of PL for each wavelength range width W. PL PL mode wavelength ratio R PL EL mode wavelength λ EL and the proportion of the mode wavelength of EL R EL Within each wavelength range width W, take T=31 preset wavelength offsets S (-50 nm to 50 nm), and calculate the EL wavelength hit rate H(S,W) corresponding to each preset wavelength offset S. Extract PL statistical features.

[0054] Online prediction: Inputting PL test data from the new epitaxial wafer into the first-level prediction model yields the predicted value λ of the EL mode wavelength. EL =553nm, target wavelength λspec=553.75 nm, predicted wavelength offset S0=0.75nm, the EL wavelength hit rate predicted by the second-level prediction model is 75%, which is higher than the preset threshold and belongs to high yield and can be deployed.

[0055] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Any act of predicting EL wavelength hit rate and EL wavelength based on the method of the present invention is considered an infringement of the patent rights of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for predicting the EL wavelength hit rate of an LED chip based on epitaxial PL data, characterized in that, Comprising: Step S1: acquiring PL test data of an epitaxial wafer to be tested and target parameters of a target LED chip product; wherein the target parameters comprise a target wavelength λspec, a target wavelength range width Wspec and a target current density Jspec; Step S2: inputting the PL test data, the target current density Jspec and the target wavelength range width Wspec into a trained first-level prediction model to obtain a predicted value of an EL wavelength statistical characteristic value; Step S3: determining a predicted wavelength offset S0 according to a difference between the predicted value of the EL wavelength statistical characteristic value and the target wavelength λspec; Step S4: inputting the PL test data, the target current density Jspec, the target wavelength range width Wspec and the predicted wavelength offset S0 into a trained second-level prediction model to obtain a predicted value of an EL wavelength hit rate corresponding to the epitaxial wafer to be tested; Wherein, construction steps of the first-level prediction model and the second-level prediction model comprise: Step A1: acquiring a historical data set, wherein the historical data set comprises PL test data of historical epitaxial wafers and EL test data of corresponding historical LED chips; wherein the EL test data at least comprises a current density J; Step A2: setting N different wavelength range widths W, performing distribution statistics on the EL test data for each wavelength range width W, and determining an EL wavelength statistical characteristic value; the wavelength range width W at least comprises the target wavelength range width Wspec; Step A3: taking the PL test data, the corresponding current density J and the wavelength range width W as training input, and taking the EL wavelength statistical characteristic value as training output, training an initial first-level prediction model to obtain the first-level prediction model; Step A4: constructing a wavelength hit interval based on the EL wavelength statistical characteristic value, the wavelength range width W and a preset wavelength offset S; calculating a proportion of data of the EL test data falling within the wavelength hit interval, which is recorded as an EL wavelength hit rate; Step A5: taking the PL test data, the corresponding current density J, the wavelength range width W and the preset wavelength offset S as training input, and taking the EL wavelength hit rate as training output, training an initial second-level prediction model to obtain the second-level prediction model.

2. The EL wavelength hit rate prediction method according to claim 1, characterized in that: In step A2, the statistical characteristic value of the EL wavelength is the mode wavelength λ of the EL. EL .

3. The EL wavelength hit rate prediction method according to claim 2, characterized in that: The step A4 for constructing the wavelength hit interval is as follows: For each wavelength range width W, set T different preset wavelength offsets S, based on the obtained EL mode wavelength λ. EL Construct the wavelength hit interval [λ] EL -SW / 2,λ EL -S+W / 2].

4. The EL wavelength hit rate prediction method according to claim 1, characterized in that: The value range of the wavelength range width W is 0 < W ≤ 50 nm; the value range of the wavelength offset S is -50 nm to 50 nm.

5. The EL wavelength hit rate prediction method according to claim 1, characterized in that: The PL test data comprises: a peak wavelength, a dominant wavelength, a full width at half maximum, a luminous intensity, and a total number of data points per epitaxial wafer.

6. The EL wavelength hit rate prediction method according to claim 1, characterized in that: The first prediction model and the second prediction model are configured to, based on each wavelength range width W, extract PL statistical characteristics from the PL test data to perform prediction; The statistical characteristics of PL include: the PL mode wavelength λ. PL PL mode wavelength ratio R PL The mean, standard deviation, skewness, kurtosis, and mean and standard deviation of the full width at half maximum (FWHM) of the PL wavelength; the wavelength is the peak wavelength or the dominant wavelength.

7. The EL wavelength hit rate prediction method according to claim 1, characterized in that: Both the first prediction model and the second prediction model are machine learning regression models established based on regression modeling methods. The first prediction model is configured to replace the current density J in the input with the target current density Jspec to perform prediction. The second prediction model is configured to replace the current density J in the input with the target current density Jspec, and replace the wavelength range width W and the preset wavelength offset S with the target wavelength range width Wspec and the predicted wavelength offset S0, respectively, to perform prediction.

8. The EL wavelength hit rate prediction method according to claim 1, characterized in that, The model building steps include: training a unified prediction model by combining all historical data for different target LED chip products and different epitaxial growth equipment; or, collecting historical data for different target LED chip products and different epitaxial growth equipment respectively, training corresponding sub-models to form a model library; and automatically selecting the corresponding sub-model from the model library for prediction based on the current LED chip and epitaxial growth equipment information.

9. The EL wavelength hit rate prediction method according to claim 1, characterized in that, It also includes classifying and screening epitaxial wafers based on the predicted EL wavelength hit rate, and only executing the original chip manufacturing plan when the predicted EL wavelength hit rate is higher than a preset threshold.

10. A system for predicting the wavelength hit rate of an LED chip EL based on epitaxial PL data, characterized in that, include: The data acquisition module is used to acquire the PL test data of the epitaxial wafer under test and the target parameters of the target LED chip product; The first-level prediction module is used to input the PL test data and target parameters into the trained first-level prediction model to obtain the predicted value of the EL wavelength statistical feature value. The offset calculation module is used to determine the predicted wavelength offset based on the difference between the predicted value of the EL wavelength statistical characteristic value and the target wavelength; The secondary prediction module is used to input the PL test data, target parameters and the predicted wavelength offset into the trained secondary prediction model to obtain the EL wavelength prediction hit rate corresponding to the epitaxial wafer under test. The model training module is used to build and train the first-level prediction model and the second-level prediction model.

Images