Mini / micro led wafer expansion coordinate dynamic compensation method and aoi equipment

Abstract

The present application relates to the technical field of LED packaging test and die bonding process, in particular to a Mini / Micro LED expansion die coordinate dynamic compensation method and AOI equipment, the present application establishes a time-space prediction model of expansion die film deformation and dynamically compensates die picking coordinates in real time, avoids frequent re-scanning and filing, reduces the number of re-scanning from 2-4 times to 0-1 times within 8 hours of production shift, and saves a large amount of effective production time; and through dynamic compensation, the die picking accuracy is kept stable throughout the production process, overcoming the problem that the die picking accuracy gradually degrades over time in the prior art, and the average deviation can be maintained at a low level; and through abnormal detection and local recalibration mechanism, it is automatically judged when local update is needed, finally reducing the number of repeated handling and unloading of wafer rings between the AOI equipment and the die bonder, and reducing the risk of die damage.

Description

Technical Field

[0001] This invention relates to the field of LED packaging testing and die bonding technology, specifically to a method for dynamic compensation of Mini / Micro LED die expansion coordinates and AOI equipment. Background Technology

[0002] With the development of Mini / Micro LED displays, the requirements for display quality and the uniformity of LED light emission are becoming increasingly stringent. Currently, Mini / Micro LED direct-view products contain tens of thousands to hundreds of thousands of LED chips. To ensure the uniformity of brightness, chromaticity, and color coordinates of the LEDs in the display, LEDs with the same brightness, chromaticity, and color coordinates are first distinguished—a process known in the industry as LED wafer binning. During the testing phase, optical testing equipment records the position coordinates and binning parameters of each chip, forming a bin map. The die bonder then uses the position coordinates in the bin map to perform chip picking and mounting operations.

[0003] AOI (Automated Optical Inspection) is a device that uses optical principles to detect common defects encountered in welding production. During automatic inspection, the system automatically scans and acquires images using a camera. The tested weld points are compared with qualified parameters in a database. After image processing, defects are detected and displayed / marked on a monitor or with automatic markers. Currently, the industry typically uses AOI matching algorithms to convert the coordinates of the test points into the physical coordinates required for die bonding. The specific process is as follows: After die expansion, the wafer ring is clamped onto the optical stage. Mark points are printed on the carrier film using a laser marking machine. The AOI equipment scans all LED chip positions, uses a visual algorithm to fit row and column values, and establishes a physical coordinate system based on the reference mark points. This maps the Bin-level information to the actual physical positions, creating a die-attaching file for the die bonder.

[0004] However, expanded crystal film is a viscoelastic polymer material. After expansion, the film is not in static equilibrium but continues to creep and relax. However, after expansion, the spacing between the grains on the film gradually changes over time, with faster changes in the first 30 minutes, followed by gradual convergence but not complete cessation. Within an 8-hour production shift, the cumulative drift of the peripheral grains can reach tens to hundreds of micrometers. Simultaneously, the deformation characteristics of the central and peripheral regions of the film differ. The central region is relatively stable, while the region near the edge of the wafer ring exhibits significant differences in drift direction and amplitude due to different constraint conditions. Finally, changes in temperature and humidity can accelerate or slow down the creep process of the film.

[0005] When the die bonder detects that the die pick-up misalignment rate has increased beyond the set threshold, the operator needs to put the wafer ring back into the AOI equipment for a complete rescan and re-file. Each rescan is time-consuming, causing the die bonder to stop and wait; multiple rescans may be required within a shift, accumulating a significant waste of effective production time; the multiple loading and unloading of the wafer ring during the rescanning process increases the risk of die damage; and it is impossible to accurately determine when a rescan is needed, usually relying on operator experience. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for dynamic compensation of Mini / Micro LED expansion coordinates and an AOI device. This avoids the misuse of NG (no good) chips with other Bin-level chips, prevents large-area blackouts or impacts on optical visual effects, thereby improving process yield, reducing product repair time, and enhancing overall efficiency.

[0007] This invention is achieved through the following technical solution:

[0008] A method for dynamic compensation of Mini / Micro LED die expansion coordinates includes the following steps:

[0009] Perform an initial scan and record-keeping on the expanded wafer to obtain the initial time. The set of physical coordinates of each grain ,in This represents the total number of effective grains on the wafer, and also records the ambient temperature at the initial scan time. and humidity , For the first The initial coordinates of each grain;

[0010] A parameterized spatiotemporal prediction model for the deformation of the expanded crystal film is established. The spatiotemporal prediction model is used to predict the positional drift of the grain relative to the initial coordinates based on the spatial position of the grain and the current time.

[0011] During the die bonding process, the die-taking offset data generated by the die bonding machine's visual positioning system is acquired, and the die-taking offset data is used as a feedback signal to calibrate the parameters of the spatiotemporal prediction model online.

[0012] For the current grain to be extracted The coordinate drift of the grain at the current moment is calculated based on the calibrated spatiotemporal prediction model. The coordinate drift is then superimposed onto the initial physical coordinates to obtain the compensated coordinates. The compensation coordinates are output to the die bonder control system.

[0013] Optionally, the first Initial coordinates of each grain It is obtained in the following way:

[0014] The wafer is scanned using the camera of an AOI device, and the positions of N of the dies are identified;

[0015] Superimpose an identical rectangle on each of the N said grain positions, and obtain the four midpoints of the four sides of each rectangle;

[0016] Connect the midpoints of adjacent rectangles according to the following requirements: the connecting line segment is the shortest, the difference between the angle between the connecting line segment and the side of the rectangle and 90° is the smallest, and there is exactly one connecting line segment for each midpoint.

[0017] Based on the connecting line segments, fit N row and column values ​​of the grains to determine the point coordinates of each grain;

[0018] The initial coordinates of each die are determined based on the position of the point coordinates of each die and the position of the wafer reference mark point. .

[0019] Optionally, the spatiotemporal prediction model is an analytical model, which decomposes the deformation of the expanded crystal film into radial drift components and angular drift components, and introduces an environmental correction factor to compensate for the influence of environmental temperature and humidity changes on the creep rate of the film material.

[0020] The coordinate drift is obtained by transforming the radial drift component and the angular drift component into coordinates and multiplying them by the environmental correction factor.

[0021] Optionally, the spatiotemporal prediction model is constructed based on the Burgers viscoelastic mechanical model, specifically including:

[0022] The radial drift component is ,in, For grains radial distance to the center of the wafer The time difference from the initial scan. The delay elasticity coefficient, It is the reciprocal of the delay time constant. The viscosity coefficient;

[0023] The angular drift component is ,in, For grains Angular position, Here is the anisotropy coefficient. The angular creep attenuation constant;

[0024] The environmental correction factor is ,in, This refers to the temperature sensitivity coefficient. Humidity sensitivity coefficient and These are the current temperature and humidity, respectively. and These are the temperature and humidity at the initial scan time, respectively;

[0025] The coordinate drift is: , .

[0026] Optionally, the spatiotemporal prediction model is constructed using a neural network model;

[0027] The neural network model takes the radial distance, angular position, time difference from the initial scan, current temperature and current humidity of the grain as input features, and outputs the coordinate drift of the corresponding grain.

[0028] The neural network model is pre-trained offline using historical production data and continuously fine-tuned through incremental learning during online production.

[0029] Optionally, the method for online calibration of the parameters of the spatiotemporal prediction model specifically includes:

[0030] Maintain a capacity of A time-sliding window is used to collect the measured displacement data of each grain within the time-sliding window;

[0031] The parameters of the spatiotemporal prediction model are adaptively updated using a recursive parameter estimation algorithm.

[0032] Anomaly detection is performed on the model prediction residuals. When the prediction residuals of a preset number of consecutive grains exceed the residual threshold, it is determined to be a model mismatch and local recalibration is triggered. The calibration points near the mismatch area are quickly scanned and updated.

[0033] Optionally, the recursive parameter estimation algorithm is a recursive least squares algorithm with a forgetting factor, and the parameter update formula is: ,in, Here is the gain matrix. The observation matrix is ​​composed of the Jacobian matrices of the model with respect to the parameters. Forgetting factor, For the first The updated model parameter set. For the first The updated model parameter set. The residual is the difference between the measured crystal offset and the predicted offset.

[0034] Optionally, the parameters of the spatiotemporal prediction model are initialized in one of the following ways:

[0035] Historical data-driven initialization: Retrieve historical records of the same film material type and crystal expansion parameters as the current batch from the historical production database, and extract the statistical average parameters as the initial values;

[0036] Rapid calibration initialization: Within a preset time window after the initial coordinate filing is completed, multiple preset calibration dies on the wafer are rapidly positioned to obtain the measured offset, and the initial parameters of the spatiotemporal prediction model are solved by least squares fitting.

[0037] Optionally, in the rapid calibration initialization, the selection strategy for calibration grains is as follows: the wafer is divided into multiple rings with equal radial spacing, and multiple calibration points are distributed in each ring at equal angles, with the total number of calibration grains being no less than 12.

[0038] Furthermore, the coordinate compensation method also includes a compensation effect evaluation step: continuously monitoring the crystal extraction success rate, average offset, standard deviation of model prediction residuals, and parameter drift rate, and recording the compensation data and model parameters into the production database.

[0039] An AOI device includes a computer program / instructions that, when executed by a processor, implement the Mini / Micro LED crystal expansion coordinate dynamic compensation method as described above.

[0040] Compared with the prior art, the present invention has the following features and beneficial effects:

[0041] This invention establishes a spatiotemporal prediction model for the deformation of the expanded die film and dynamically compensates for the die-taking coordinates in real time, avoiding frequent rescanning and filing. Within an 8-hour production shift, the number of rescans is reduced from 2-4 times to 0-1 times, saving a significant amount of effective production time. Furthermore, dynamic compensation ensures that the die-taking accuracy remains stable throughout the entire production process, overcoming the problem of gradual degradation of die-taking accuracy over time in existing technologies, and the average deviation can be maintained at a low level. Moreover, through anomaly detection and local recalibration mechanisms, it automatically determines when a local update is needed, ultimately reducing the number of times the wafer ring is repeatedly handled and unloaded between the AOI equipment and the die bonder, thus reducing the risk of die damage. Attached Figure Description

[0042] The accompanying drawings illustrate exemplary embodiments of the present invention and, together with the description thereof, serve to explain the principles of the invention. These drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, but do not constitute a limitation on the embodiments of the present invention.

[0043] Figure 1 This is a flowchart illustrating a method for dynamic compensation of Mini / Micro LED crystal expansion coordinates according to the present invention.

[0044] Figure 2 This is a schematic diagram of grain position drift caused by the creep of the expanded film according to the present invention.

[0045] Figure 3 This is a schematic diagram of superimposing a standard rectangular outline onto the LED chip position using the visual algorithm of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0047] It should also be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings.

[0048] Where there is no conflict, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] Example 1

[0050] In existing Mini / Micro LED manufacturing processes, testing equipment records the location coordinates and bin-level parameters of each die before wafer cleaving, forming a bin map. However, after wafer cleaving, due to uneven film texture and varying macroscopic stretching ratios, the die spacing changes, causing the original location coordinates to lose their correspondence with the physical coordinates. If die-by-die picking is forcibly performed using a counting algorithm during die bonding, errors are easily caused by issues such as row / column tilt and voids.

[0051] This embodiment adds a dynamic coordinate compensation layer to the existing AOI coordinate merging process. This compensation layer is embedded in the data link between the AOI merging device and the die bonder, and calculates and adds the compensation amount in real time each time the die bonder requests the die coordinates. Figure 1 As shown, the method includes the following steps.

[0052] (1) Perform initial scanning and filing on the expanded wafer to obtain the initial time. The set of physical coordinates of each grain ,in This represents the total number of effective grains on the wafer, and also records the ambient temperature at the initial scan time. and humidity , For the first The initial coordinates of each grain.

[0053] During the initial scanning and documentation process, the AOI device scans the positions of all LED chips, uses visual algorithms to fit row and column values ​​and reference Mark points to establish a physical coordinate system, and maps the Bin-level information to the actual physical positions one by one to generate the initial coordinates. The methods include:

[0054] S1. Fix the expanded wafer ring onto the optical stage of the AOI equipment; then rotate the rotating support of the stage to keep the positive electrode direction of the wafer and the row and column directions at the center of the wafer parallel to the gantry of the AOI equipment.

[0055] S2. Using a laser marking machine, print reference mark points and QR codes at blank fixed positions on the wafer carrier film through an optical stage.

[0056] S3, the AOI equipment's camera scans the wafer, identifying each LED chip through a visual template. For example... Figure 3 As shown, the visual algorithm overlays a standard rectangular frame onto each identified LED chip location and extracts the four midpoints of the frame's top, bottom, left, and right sides. Specifically, the AOI device's camera scans the wafer and identifies the positions of N chips; each of the N chip locations is then overlaid with an identical rectangular frame, and the four midpoints of each rectangle's four sides are obtained.

[0057] Next, the algorithm connects adjacent midpoints in the outer frame according to the following logic: first, the connecting line segment must be the shortest; second, the angle between the connecting line segment and the outer frame must be closest to 90°; and third, each midpoint must have exactly one connecting line segment. Through the above connection logic, even if the arrangement of the chip rows and columns is messy or tilted after chip expansion, accurate row and column values ​​can be fitted, thereby determining the point coordinates of each LED chip.

[0058] S4. Assign values ​​to the Bin-level information of the wafer one by one according to the precise row and column values ​​fitted above. At the same time, calculate the positions of the LEDs scanned by the camera and the positions of the reference Mark points printed by the laser, and use the algorithm to establish a physical coordinate system, thereby clarifying the physical coordinates of each LED.

[0059] S5. Encapsulate the data containing the chip ID, precise row and column values, physical coordinates, and Bin-level information of each LED into a chip file (i.e., the initial physical coordinate set) and send it to the die bonder control system.

[0060] (2) Establish a parameterized spatiotemporal prediction model for the deformation of the expanded film. The spatiotemporal prediction model is used to predict the positional drift of the grain relative to the initial coordinates based on the spatial position of the grain and the current time.

[0061] Because the expanded crystal film is a viscoelastic polymer material, it undergoes continuous creep and stress relaxation after expansion. The spacing between the grains on the film gradually changes over time, and this change exhibits time dependence, spatial non-uniformity, and environmental sensitivity. For example... Figure 2 As shown, grain drift has both a radial outward expansion component and an angular deflection component caused by the anisotropy of the film material. The spatiotemporal prediction model captures the spatiotemporal characteristics of the above creep behavior, taking the spatial position (radial distance and angular position) and time information of the grain as input, and outputting the predicted drift amount of the grain.

[0062] (3) During the die bonding process, the crystal picking offset data generated by the vision positioning system of the die bonding machine is acquired, and the crystal picking offset data is used as a feedback signal to calibrate the parameters of the spatiotemporal prediction model online.

[0063] The vision positioning system of the die bonder locates the target die before picking it up. This positioning process generates data representing the difference between the predicted and actual coordinates, known as the die picking offset. By acquiring this offset data from the die bonder's vision positioning system, the offset data is used as a feedback signal to calibrate the parameters of the spatiotemporal prediction model online. This forms a closed-loop control structure of "prediction → feedback → correction → prediction," enabling the model to adaptively track changes in the actual creep process. Furthermore, it utilizes the existing visual feedback data from the die bonder as the calibration signal, eliminating the need for additional hardware.

[0064] (4) For the current grain to be taken The coordinate drift of the grain at the current moment is calculated based on the calibrated spatiotemporal prediction model. The coordinate drift is superimposed onto the initial physical coordinates to obtain the compensated coordinates. The dynamic compensation module outputs compensated coordinates to the die bonder control system, replacing the original physical coordinates. The module can be deployed as software on the die bonder's host computer or as a standalone edge computing node.

[0065] Example 2

[0066] This embodiment provides a detailed explanation of steps (2) and (3) based on embodiment one.

[0067] In step (2), the spatiotemporal prediction model is an analytical model. The analytical model decomposes the deformation of the expanded film into radial drift components and angular drift components, and introduces an environmental correction factor to compensate for the influence of environmental temperature and humidity changes on the creep rate of the film material. The coordinate drift is obtained by transforming the radial drift components and angular drift components into coordinates and multiplying them by the environmental correction factor.

[0068] The film surface deformation is decomposed into two independent components, radial and angular, for modeling. This is because during the expansion process of the expanded film, the film surface mainly bears radial tensile stress, and creep is primarily manifested as radial expansion. Simultaneously, due to the anisotropy of the film material (e.g., the relationship between the stretching direction and the crystal orientation), an angular deflection component is also generated. The environmental correction factor is introduced because temperature and humidity have a significant impact on the creep rate of polymer materials.

[0069] The spatiotemporal prediction model is built upon the Burgers viscoelastic mechanics model, a four-parameter viscoelastic model that combines Maxwell and Kelvin-Voigt volumes in series. This model can simultaneously describe three deformation behaviors: instantaneous elastic deformation, delayed elastic deformation (rapid creep), and viscous flow (steady-state creep). Specifically, it includes:

[0070] The radial drift component is ,in, For grains radial distance to the center of the wafer The time difference from the initial scan. The delay elasticity coefficient, It is the reciprocal of the delay time constant. is the viscous flow coefficient.

[0071] First item The delayed elastic response corresponding to the Kelvin-Voigt body represents the rapid creep behavior in the early stage after crystal expansion, which gradually saturates over time, reflecting the actual phenomenon that the drift is relatively fast in the first 30 minutes after crystal expansion and then gradually converges.

[0072] Second item Corresponding to the viscous flow of Maxwell volumes, this represents long-term steady-state creep, a slow drift that persists even after a considerable period of time.

[0073] The typical value range is 0.001~0.01. The typical value range is 0.01 / min to 0.1 / min. The typical value range is ~ min⁻¹.

[0074] Radial drift and distance from the grain to the wafer center It is directly proportional to the physical law of film expansion—the farther the grain is from the center, the greater the drift.

[0075] Angular drift component is ,in, For grains Angular position, Here is the anisotropy coefficient. This represents the angular creep attenuation constant. Angular drift originates from the anisotropic properties of the membrane material, through... This represents directional dependence. The typical value is approximately 0.0008. The typical value is approximately 0.03 / min.

[0076] Environmental correction factor is ,in, This refers to the temperature sensitivity coefficient. Humidity sensitivity coefficient and These are the current temperature and humidity, respectively. and These are the temperature and humidity at the initial scan time, respectively;

[0077] The environmental correction factor corrects the creep rate in real time. When the ambient temperature rises, the molecular chain motion of the polymer material intensifies, and the creep rate increases accordingly. The typical value is about 0.03~0.05 / ℃, which means that for every 5℃ increase in workshop temperature, the creep rate increases by about 15%~25%. Typical values ​​are approximately 0.01~0.02 / %RH.

[0078] The coordinate drift is: , The radial and angular drift components in polar coordinates are converted into x- and y-direction compensations in Cartesian coordinates. Multiplying these by an environmental correction factor gives the final coordinate drift.

[0079] In step (3), the method for online calibration of the parameters of the spatiotemporal prediction model specifically includes three sub-steps.

[0080] First, maintain a capacity of A time-sliding window is used to collect measured offset data for each grain within that window; the time-sliding window refers to retaining the most recent data. The offset data of each grain is used for parameter updates. Typical values ​​range from 50 to 200 grains. A sliding window ensures sufficient data for parameter estimation to avoid overfitting while also reflecting the latest changes in creep characteristics. During die bonding, the die bonder's vision positioning system precisely locates the target grain before picking it up. The difference between the predicted coordinates and the actual coordinates generated during this positioning process is the grain picking offset, which can be obtained through the die bonder's data output port.

[0081] Second, the parameters of the spatiotemporal prediction model are adaptively updated using a recursive parameter estimation algorithm. The recursive parameter estimation algorithm is a recursive least squares algorithm with a forgetting factor, and the parameter update formula is as follows: ,in, Here is the gain matrix. The observation matrix is ​​composed of the Jacobian matrices of the model with respect to the parameters. For the first The updated model parameter set. For the first The updated model parameter set. The residual is the difference between the measured crystal offset and the predicted offset. Forgetting factor, forgetting factor Typical values ​​are 0.95 to 0.99, used to enhance sensitivity to recent data. The smaller the value, the greater the weight the algorithm places on recent data, and the faster the parameters are updated, but the weaker its resistance to noise; conversely, the larger the value, the greater the weight the algorithm places on recent data, and the faster the parameters are updated, but the weaker its resistance to noise. A larger value indicates more stable parameter updates but a slower response to sudden changes. In practical applications, an appropriate forgetting factor can be selected based on the stability of the production environment.

[0082] Third, anomaly detection is performed on the model prediction residuals. When the prediction residuals of a preset number of consecutive grains exceed the residual threshold, it is determined to be a model mismatch and local recalibration is triggered. The calibration points near the mismatch area are quickly scanned and updated.

[0083] For example, if the residual threshold is set to 30 μm, a model mismatch is determined when the residuals of n consecutive grains (n≥5) exceed the threshold. In this case, it is not necessary to rescan the entire wafer; instead, only the calibration points near the mismatch area are quickly scanned and updated, shortening the anomaly response time. Model mismatch may be caused by sudden environmental changes, local defects in the film material, or other unforeseen factors.

[0084] In addition, this embodiment provides a method for initializing the parameters of the spatiotemporal prediction model, which is performed through one of the following methods:

[0085] Historical data-driven initialization: This method retrieves historical records of the same film material type and die expansion parameters as the current batch from the historical production database, extracting the statistically averaged parameters as initial values. The database is organized using a three-dimensional index: "film material type + die expansion intensity level + wafer size." The advantage of historical data-driven initialization is that it eliminates the need for additional calibration time, allowing direct entry into die bonding production. This approach is suitable for mature production lines that have accumulated a substantial amount of historical data.

[0086] Quick calibration initialization: within a preset time window after initial coordinate data creation (e.g., ... +5min and Within two time points (+15 minutes), multiple pre-set calibration dies on the wafer are rapidly repositioned to obtain measured offsets. The initial parameters of the spatiotemporal prediction model are then solved using the least squares method. Rapid calibration initialization is suitable for scenarios involving the first use of new film materials or a lack of historical data. In rapid calibration initialization, the selection strategy for calibration grains is as follows: the wafer is divided into multiple rings with equal radial spacing, and multiple calibration points are distributed at equal angles within each ring, with a total of no fewer than 12 calibration grains. Specifically, the wafer is divided into 3-4 rings based on radial distance (the inner ring's radial distance is approximately 30% of the wafer radius, the middle ring approximately 60%, and the outer ring approximately 90%), with 4-6 points distributed at equal angles within each ring. This distribution strategy ensures that the calibration points are spatially uniformly distributed on the film surface, enabling sufficient sampling of creep characteristics in both radial and angular dimensions, providing sufficient information for model parameter fitting.

[0087] Example 3

[0088] This embodiment further illustrates the steps for evaluating the compensation effect, based on Embodiment 1.

[0089] Furthermore, the coordinate compensation method also includes a compensation effect evaluation step: continuously monitoring the crystal retrieval success rate, average offset, standard deviation of model prediction residuals, and parameter drift rate, and recording the compensation data and model parameters into the production database.

[0090] The system continuously monitors the following indicators: crystal success rate (target ≥99.5%), average deviation (target <10μm), standard deviation of model prediction residuals (reflecting the stability of model accuracy), and parameter drift rate (reflecting the rate of change of creep characteristics). All compensation data and model parameters are recorded in the production database in time series. This provides data support for driving initialization with historical data from subsequent batches, making model initialization increasingly accurate; it can also be used for long-term trend analysis to discover regular differences in creep characteristics under different membrane material types and seasonal environmental conditions, providing data basis for process optimization.

[0091] After adopting the method of this embodiment, the chip misalignment remains at a low level throughout the entire 8-hour production shift, while the misalignment continuously increases over time in the uncompensated scheme. Comparative data using a 4-inch Mini LED wafer as an example shows that the misalignment of the uncompensated scheme can reach approximately 98μm in the 8th hour, the misalignment of the timed rescan scheme (rescanned every 2 hours) is approximately 8μm but requires 3 rescans, while the method of this invention achieves a chip success rate of 99.7% in 8 hours of continuous production, with a total misalignment of no more than approximately 7μm and no need for rescanning, saving approximately 90 minutes of effective production time, as detailed in Table 1.

[0092] Table 1 Comparison of data for 4-inch Mini LED wafers

[0093]

[0094] Example 4

[0095] This embodiment provides an alternative to the analytical model in Embodiment 2, that is, the spatiotemporal prediction model is constructed using a neural network model;

[0096] The neural network model takes the radial distance, angular position, time difference from the initial scan, current temperature and current humidity of the grain as input features and outputs the coordinate drift of the corresponding grain. The neural network model is pre-trained offline using historical production data and continuously fine-tuned through incremental learning during online production.

[0097] The neural network model employs a lightweight multilayer perceptron (MLP) structure. The network input features are 5-dimensional vectors: radial distance... Angular position Time difference Current temperature Current humidity The network structure is a three-layer fully connected network with hidden layer neurons numbered 32-16-8 respectively, using ReLU as the activation function, and a 2D output layer. , Neural network models do not require assumptions about specific mechanical models and are more adaptable to complex nonlinear deformations, but they require more historical data to support them.

[0098] During the offline pre-training phase, a large amount of actual drift data from historical batches (real coordinate differences obtained from multiple scans by AOI) is used as training samples to enable the neural network to learn the general rules of film surface creep. During online production, measured offset data from visual feedback from the die bonder are used to continuously fine-tune the network weights through incremental learning with a small learning rate, allowing the model to adapt to the specific creep characteristics of the current batch.

[0099] Example 5

[0100] An AOI device, wherein a computer program / instructions are running in the AOI device, and when the computer program / instructions are executed by a processor, the steps of any of the above methods are implemented.

[0101] AOI (Action of Information) devices comprise computer programs or instruction sets designed to perform specific tasks or achieve specific functions. These programs or instructions are designed to be executed by a processor to implement a series of predefined steps or operations. The program product may be stored in various forms of computer storage media, such as memory, hard disks, solid-state drives, optical discs, or other forms of digital storage devices. It may exist in the form of compiled binary code or in the form of scripts or bytecode executable by an interpreter. Through carefully designed algorithms and logical instructions, the program product enables the processor to process data in a specific order and manner, performing various functions such as data analysis, user interaction, and device control.

[0102] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment / mode or example is included in at least one embodiment / mode or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0103] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0104] Those skilled in the art should understand that the above embodiments are merely for illustrating the present invention and are not intended to limit the scope of the invention. Those skilled in the art can make other changes or modifications based on the above invention, and these changes or modifications still fall within the scope of the present invention.

Claims

1. A method for dynamic compensation of Mini / Micro LED die expansion coordinates, characterized in that, Includes the following steps: Perform an initial scan and record-keeping on the expanded wafer to obtain the initial time. The set of physical coordinates of each grain ,in This represents the total number of effective grains on the wafer, and also records the ambient temperature at the initial scan time. and humidity ,in For the first The initial coordinates of each grain; A parameterized spatiotemporal prediction model for the deformation of the expanded crystal film is established. The spatiotemporal prediction model is used to predict the positional drift of the grain relative to the initial coordinates based on the spatial position of the grain and the current time. During the die bonding process, the die-taking offset data generated by the die bonding machine's visual positioning system is acquired, and the die-taking offset data is used as a feedback signal to calibrate the parameters of the spatiotemporal prediction model online. For the current grain to be extracted The coordinate drift of the grain at the current moment is calculated based on the calibrated spatiotemporal prediction model. The coordinate drift is then superimposed onto the initial physical coordinates to obtain the compensated coordinates. The compensation coordinates are output to the die bonder control system; The spatiotemporal prediction model is constructed using a neural network model. The neural network model takes the radial distance, angular position, time difference from the initial scan, current temperature and current humidity of the grain as input features, and outputs the coordinate drift of the corresponding grain. The neural network model is pre-trained offline using historical production data and continuously fine-tuned through incremental learning during online production.

2. The method for dynamic compensation of Mini / Micro LED crystal expansion coordinates according to claim 1, characterized in that, The first Initial coordinates of each grain It is obtained in the following way: The wafer is scanned using the camera of an AOI device, and the positions of N of the dies are identified; Superimpose an identical rectangle on each of the N said grain positions, and obtain the four midpoints of the four sides of each rectangle; Connect the midpoints of adjacent rectangles according to the following requirements: the connecting line segment is the shortest, the difference between the angle between the connecting line segment and the side of the rectangle and 90° is the smallest, and there is exactly one connecting line segment for each midpoint. Based on the connecting line segments, fit N row and column values ​​of the grains to determine the point coordinates of each grain; The initial coordinates of each die are determined based on the position of the point coordinates of each die and the position of the wafer reference mark point. .

3. The method for dynamic compensation of Mini / Micro LED crystal expansion coordinates according to claim 1, characterized in that, The spatiotemporal prediction model is an analytical model. The analytical model decomposes the deformation of the expanded crystal film into radial drift components and angular drift components, and introduces an environmental correction factor to compensate for the influence of environmental temperature and humidity changes on the creep rate of the film material. The coordinate drift is obtained by transforming the radial drift component and the angular drift component into coordinates and multiplying them by the environmental correction factor.

4. The method for dynamic compensation of Mini / Micro LED crystal expansion coordinates according to claim 3, characterized in that, The spatiotemporal prediction model is constructed based on the Burgers viscoelastic mechanical model, and specifically includes: The radial drift component is ,in, For grains radial distance to the center of the wafer The time difference from the initial scan. The delay elasticity coefficient, It is the reciprocal of the delay time constant. The viscosity coefficient; The angular drift component is ,in, For grains Angular position, Here is the anisotropy coefficient. The angular creep attenuation constant; The environmental correction factor is ,in, The temperature sensitivity coefficient, Humidity sensitivity coefficient and These are the current temperature and humidity, respectively. and These are the temperature and humidity at the initial scan time, respectively; The coordinate drift is: , .

5. The method for dynamic compensation of Mini / Micro LED crystal expansion coordinates according to claim 1, characterized in that, The method for online calibration of the parameters of the spatiotemporal prediction model specifically includes: Maintain a capacity of A time-sliding window is used to collect the measured displacement data of each grain within the time-sliding window; The parameters of the spatiotemporal prediction model are adaptively updated using a recursive parameter estimation algorithm. Anomaly detection is performed on the model prediction residuals. When the prediction residuals of a preset number of consecutive grains exceed the residual threshold, it is determined to be a model mismatch and local recalibration is triggered. The calibration points near the mismatch area are quickly scanned and updated.

6. The method for dynamic compensation of Mini / Micro LED crystal expansion coordinates according to claim 5, characterized in that, The recursive parameter estimation algorithm is a recursive least squares algorithm with a forgetting factor, and the parameter update formula is: ,in, Here is the gain matrix. The observation matrix is ​​composed of the Jacobian matrices of the model with respect to the parameters. Forgetting factor, For the first The updated model parameter set. For the first The updated model parameter set. The residual is the difference between the measured crystal offset and the predicted offset.

7. The method for dynamic compensation of Mini / Micro LED crystal expansion coordinates according to claim 1, characterized in that, The parameters of the spatiotemporal prediction model are initialized in one of the following ways: Historical data-driven initialization: Retrieve historical records of the same film material type and crystal expansion parameters as the current batch from the historical production database, and extract the statistical average parameters as the initial values; Rapid calibration initialization: Within a preset time window after the initial coordinate filing is completed, multiple preset calibration dies on the wafer are rapidly positioned to obtain the measured offset, and the initial parameters of the spatiotemporal prediction model are solved by least squares fitting.

8. The method for dynamic compensation of Mini / Micro LED crystal expansion coordinates according to claim 1, characterized in that, It also includes a compensation effect evaluation step: continuously monitoring the crystal success rate, average deviation, standard deviation of model prediction residuals and parameter drift rate, and recording the compensation data and model parameters into the production database.

9. An AOI device, wherein a computer program / instructions run in the AOI device, characterized in that, When the computer program / instruction is executed by the processor, it implements the Mini / Micro LED expansion coordinate dynamic compensation method as described in any one of claims 1-8.

Images