Overlay error real-time correction system based on image recognition and dynamic fitting

Abstract

The invention discloses an overlay error real-time correction system based on image recognition and dynamic fitting, and belongs to the field of semiconductor advanced packaging and manufacturing. The system comprises an image acquisition module, an error decoupling module and a dynamic compensation module to form closed-loop control. The image acquisition module adopts a high-resolution linear array CCD (Charge Coupled Device) and an AI (Artificial Intelligence) algorithm to realize high-speed and high-precision mark identification and error quantization; the error decoupling module performs decoupling and trend prediction on system errors and random errors by using PCA, a region weight matrix, LSTM and the like; and the dynamic compensation module realizes mu s-level partition graph compensation by a DMD optical system based on an error thermodynamic diagram through heterogeneous platform operation. The system supports the real-time processing of more than 100Hz, the average overlay error is less than or equal to 0.3 mu m and the response time is less than 5 seconds in the FOPLP and other processes, the problems of lag, low precision and the like of the traditional method are effectively solved, and the manufacturing yield and the production line efficiency are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of advanced semiconductor packaging manufacturing technology, specifically relating to a real-time overlay error correction system based on image recognition and dynamic fitting. Background Technology

[0002] With the widespread application of ABF (Ajinomoto Build-up Film) substrates in advanced packaging, their line widths have gradually approached submicron and even nanometer levels from the 10 μm level, while multilayer interconnect structures (such as RDLs with more than 32 layers) are gradually becoming mainstream. This trend poses unprecedented challenges to the overlay alignment accuracy in photolithography processes. Traditional overlay error correction methods mainly rely on offline measurement and static compensation strategies, which are no longer sufficient to meet the high-precision manufacturing requirements under dynamic and complex deformations, specifically in the following aspects:

[0003] Offline calibration suffers from lag and severely insufficient response speed: Current mainstream error correction processes still rely on high-precision offline equipment such as interferometers and white-light interferometry systems to periodically measure alignment errors. For example, the technical process disclosed in patent CN101424283B includes multiple steps such as error measurement, model building, parameter updating, and compensation uploading, with an overall cycle typically ranging from 30 minutes to 1 hour. On high-density interconnect (HDI) production lines, this lag directly leads to a disconnect between the production line and measurement, especially in scenarios with frequent process fluctuations or large batch-to-batch temperature differences. The high failure rate of compensation data significantly reduces yield and affects production line throughput.

[0004] The applicability of global correction models is declining: some existing systems attempt to compensate the substrate as a whole using partitioned affine models or global geometric transformations (US6331937B1), but the core logic of these methods is still based on the assumption of "uniform deformation of the entire substrate." In real-world production, ABF substrates often generate nonlinear local stresses during lamination, exposure, and curing, leading to different expansion rates or contraction trends between quadrants. For example, the central region of the substrate is heated uniformly, while the edge regions dissipate heat quickly and experience stress concentration, easily forming irregular deformations, manifested as overlay error gradients in the "center-edge" or "top left-bottom right" directions. In this situation, global affine models cannot accurately fit the local error distribution and instead introduce new error accumulation.

[0005] Lack of a real-time response mechanism for sudden deformation: ABF materials belong to a polymer resin system, and their coefficient of thermal expansion is much higher than that of inorganic carriers, making them prone to rapid deformation under heat or stress concentration. For example, factors such as deviations in pressing pressure, fluctuations in heat input during laser drilling, and local temperature rises can all trigger "sudden deformation events" in a short period of time, manifesting as a sudden increase in overlay error from 0.5 μm to over 2 μm in a certain area. However, traditional static templates and correction models are often based on statistical averages and lack a real-time dynamic update mechanism. Once the process conditions deviate from the reference state, errors cannot be identified and corrected in time, ultimately leading to batch defects or failure risks such as interlayer short circuits and broken lines. Summary of the Invention

[0006] Based on the shortcomings of the existing technologies, this invention provides a real-time overlay error correction system based on image recognition and dynamic fitting. Through a closed-loop architecture of "image acquisition-error decoupling-dynamic compensation", it integrates core technologies such as multimodal AI recognition, multi-source error modeling, and partitioned dynamic fitting to solve the bottleneck problems of lag, low precision, and weak adaptability of traditional correction methods, and achieve sub-micron level overlay alignment control.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] Image acquisition module: Adopting a "high-resolution linear CCD + multimodal AI recognition" architecture, it achieves high-precision, high-adaptability, and high-frame-rate online acquisition of alignment marks.

[0009] Hardware configuration: It uses a Sony ILX554B high-resolution linear CCD (5μm pixel size, scanning width ≥610mm), paired with a real-time triggered sampling control circuit, with a sampling frequency of over 100Hz, supporting full-coverage scanning of large-size panels of 610mm×457mm; it adopts a non-contact optical imaging solution to avoid disturbing the substrate during the acquisition process.

[0010] AI recognition algorithm:

[0011] (1) Based on the improved U-Net convolutional neural network (introducing the attention mechanism CBAM), the alignment mark area is precisely segmented at the pixel level with a segmentation accuracy of ≤0.1μm, effectively shielding noise interference such as scratches and stains on the substrate surface;

[0012] (2) By integrating Siamese Network and ORB feature matching algorithm, a dual verification mechanism of "structure matching + feature matching" is constructed to quantify the deviation between the design drawing and the actual image. The deviation calculation accuracy is ≤0.05μm.

[0013] (3) Built-in marker shape adaptive library supports various standard / customized alignment markers such as circles, crosses, hexagons, and rings, which can adapt to different packaging process requirements without changing the algorithm model.

[0014] Data output: Synchronously outputs the actual coordinates (X_actual, Y_actual), design coordinates (X_design, Y_design), deviation values ​​(Δx, Δy), and confidence scores (≥0.95 is considered valid data) for each marker point, providing a highly reliable data source for subsequent error modeling.

[0015] II. Error Decoupling Module

[0016] A mechanism of "multi-source error decoupling + spatiotemporal fusion modeling" is constructed to achieve accurate separation of systematic errors and random errors, and dynamic prediction of error trends.

[0017] Error classification and decoupling:

[0018] Systematic errors: such as laser drilling offset, exposure stage error, mask pattern distortion, etc.

[0019] Random errors include nonlinear deformation caused by thermal expansion, stress release, and local material inhomogeneity.

[0020] Composite error coupling: for example, the synergistic deformation caused by thermal stress and mechanical loading;

[0021] Decoupling method: Based on the joint algorithm of principal component analysis (PCA) and independent component analysis (ICA), the error vector E=[Δx, Δy] is first reduced in dimension by PCA to extract the principal components (corresponding to systematic errors) with a cumulative contribution rate of ≥95%; then, the residual error is separated by ICA to separate the independent random error components, with a decoupling accuracy of ≥98%.

[0022] Regionalized error modeling:

[0023] (1) An adaptive meshing algorithm is adopted to divide the substrate into dynamic blocks of 10mm×10mm~100mm×100mm according to the error gradient distribution (the block size is automatically reduced in areas with large error gradients).

[0024] (2) Construct an independent error weight matrix for each block. The weight coefficients are determined by the density of marker points, error confidence, and deformation complexity within the block, so as to achieve accurate modeling of "where the error is large, the correction is dense".

[0025] Spatiotemporal fusion prediction model:

[0026] (1) Spatial modeling: A locally weighted quadratic polynomial fitting model is used for each block, and the fitting function is:

[0027] In the formula: (x,y) are the coordinates of any point within the block, and wi is the weighting coefficient of the sampling point (calculated based on the Gaussian kernel function: w) i =exp(-((x i -x0)²+(y i -y0)²) / (2σ²)), where σ is the spatial influence radius), and a0~a5 are the fitting coefficients;

[0028] (2) Time prediction: A bidirectional LSTM network is introduced on the basis of the ARIMA model to construct a hybrid time series prediction model; the input is the error principal component data of the past 10 frames (after PCA compression), and the output is the error trend prediction value of the next 1~2 frames; the model is driven by the AI ​​acceleration module built into the FPGA platform (computing power ≥1TOPS), and the prediction delay is ≤1ms.

[0029] (3) Fusion strategy: The error of the current frame adopts the spatial fitting result, and the error of the next frame adopts the weighted fusion result of "spatial fitting value × 0.7 + temporal prediction value × 0.3", which takes into account both real-time performance and foresight.

[0030] Dynamic error mapping: Real-time generation of error heatmaps across the entire substrate, presenting the error distribution in a pseudo-color format. The heatmap update frequency is synchronized with the sampling frequency (≥100Hz), providing accurate error location and amplitude information for the dynamic compensation module.

[0031] III. Dynamic Compensation Module

[0032] Based on a heterogeneous processing platform and partitioned dynamic fitting technology, μs-level high-precision real-time compensation is achieved:

[0033] Hardware architecture: It adopts a heterogeneous processing platform of "CPU+FPGA+AI accelerator". The CPU is responsible for system scheduling and data management, the FPGA is responsible for real-time logic operation and control signal generation, and the AI ​​accelerator is responsible for parallel computing of fitting model and prediction model. The overall operation latency is ≤500μs.

[0034] Zone compensation mechanism:

[0035] (1) Based on the error thermal map, the substrate is divided into multiple independent compensation blocks, each block corresponds to a set of exclusive compensation parameters, and the minimum compensation granularity can be ≤10mm;

[0036] (2) Each block adopts the compensation logic of "distortion inverse mapping + optical fine-tuning": based on the error data of the local polynomial fitting model, the inverse compensation amount (Δx_comp, Δy_comp) is calculated and converted into the pixel offset control signal of the digital micromirror array (DMD).

[0037] Optical fine-tuning execution:

[0038] A high frame rate digital micromirror array (DMD, pixel count ≥ 1920×1080, switching speed ≤ 1μs) is selected. By adjusting the deflection angle of the micromirrors, dynamic offset correction of the exposure pattern is achieved. The compensation parameters are written to the DMD controller through a high-speed LVDS interface with an update frequency ≥ 100Hz, ensuring that each layer of exposure pattern is corrected in real time based on the latest error data.

[0039] Closed-loop verification: After compensation is executed, the image acquisition module collects the deviation of the corrected marker points in real time and feeds it back to the error decoupling module for evaluation of the compensation effect; if the deviation is ≥0.1μm, secondary compensation is triggered until the error meets the process requirements (≤0.3μm).

[0040] IV. Anomaly Response Module

[0041] A full-process mechanism of "detection-early warning-correction-traceability" is designed to address sudden deformations, thereby improving system robustness.

[0042] Anomaly detection: Set dual anomaly judgment criteria: 1) Single frame marker error Δ ≥ 3 times the global error standard deviation σ; 2) The error change rate of markers in 3 consecutive frames ≥ 0.5μm / frame; If either criterion is met, it is judged as an abnormal area.

[0043] Emergency correction:

[0044] (1) Start the “maximum error envelope fitting model”, use cubic spline curves to fit the envelope of the abnormal region, and quickly generate emergency compensation parameters. The compensation response time is ≤1ms.

[0045] (2) Temporarily increase the sampling frequency (from 100Hz to 200Hz) and compensation update frequency (from 100Hz to 200Hz) in the abnormal area to achieve high-frequency dynamic tracking correction;

[0046] Anomaly tracing: Automatically marks the location, time, error amplitude, correction process, and other data of the abnormal area to generate an "anomaly handling report"; at the same time, it marks the edge of the corresponding area on the substrate with a "suboptimal exposure area" mark to facilitate subsequent quality inspection and tracing.

[0047] The technological advantages include:

[0048] (1) Ultra-high real-time performance and response speed: The “acquisition-modeling-compensation” closed-loop control system constructed by this invention supports real-time processing frequencies above 100Hz. In the application example, for the RDL-3 layer pattern of FOPLP package, it achieves rapid correction with a response time of less than 5 seconds. Compared with the traditional process with a compensation cycle of up to 30 minutes, it greatly improves the real-time processing capability of the production line, effectively solves the problems of traditional offline correction lag and disconnection between the production line and measurement, and significantly improves production efficiency and production line throughput.

[0049] (2) Submicron-level high-precision alignment: Image recognition technology combining U-Net convolutional neural network and Siamese network is used to achieve high-precision extraction and error calculation of various shape markers; combined with principal component analysis (PCA) to effectively decouple systematic and random errors, and the fusion application of regional weighted multinomial fitting model and time series prediction model, the average overlay error of FOPLP packaging process is ≤0.3μm when about 400 sampling points are deployed on a 610mm×457mm panel, achieving submicron-level alignment control. This breaks through the accuracy bottleneck of traditional global correction model when dealing with non-uniform deformation, reduces error accumulation, and significantly improves the alignment accuracy and manufacturing yield of photolithography process.

[0050] (3) Strong adaptability and robustness: Addressing the spatial correlation and local deformation characteristics of error distribution in large-size panels, the system employs a region-weighted fitting mechanism and a hybrid prediction model. This effectively handles nonlinear local stress deformation of the substrate under conventional processes and accurately predicts thermally induced sudden deformation and temporal trends using LSTM or ARIMA models. Simultaneously, the anomaly detection and emergency response mechanism can rapidly activate the "maximum error envelope fitting model" when the error at a marked point exceeds three times the global error standard deviation. This increases the sampling frequency for high-frequency monitoring, effectively responding to sudden events such as thermal shock and mechanical impact. This significantly enhances the system's adaptability and robustness in complex and variable process environments, reducing the risk of batch defects.

[0051] (4) High versatility and flexible expansion: It supports various marking shapes such as circles, crosses, and squares, and can process multiple alignment marks at the same time. It is suitable for various advanced packaging lithography processes such as FOPLP (panel-level fan-out packaging), HDI (high-density interconnect) substrates, and RDL (rewiring layer). By adjusting system parameters (such as region division, fitting model parameters, prediction model weights, etc.), it can adapt to different process requirements and has good versatility and flexible expansion capabilities.

[0052] (5) Intelligent monitoring and quality traceability: The anomaly detection mechanism can mark abnormal areas as "suboptimal exposure areas" in real time. Combined with the error heat map output by the system online, it provides intuitive process quality monitoring data for the production line, which facilitates quality backtracking and process optimization analysis, helps enterprises quickly locate production problems, improve process flow, achieve refined production management, and improve product quality control level. Attached Figure Description

[0053] Figure 1 This is a diagram illustrating the overall system framework of the pattern error correction algorithm described in this invention.

[0054] Figure 2 This is a flowchart of image marker extraction and error calculation based on U-Net and Siamese networks;

[0055] Figure 3 This is a schematic diagram of principal component analysis (PCA) for the error separation module;

[0056] Figure 4 This is a schematic diagram of the local polynomial fitting and region weighted calculation method.

[0057] Figure 5 The block diagram of the LSTM model used for time series error prediction;

[0058] Figure 6 Flowchart for generating error fusion and compensation parameters;

[0059] Figure 7 This is a logic diagram of the anomaly detection and emergency compensation triggering mechanism. Detailed Implementation

[0060] To better understand the technical solution of this invention, the system composition, workflow, and key technologies of this invention are described in detail below with reference to the accompanying drawings and embodiments. This invention proposes a real-time overlay error correction system for advanced packaging lithography processes, applicable to process scenarios such as FOPLP (Fan-out Panel Package), HDI (High-Density Interconnect) substrates, and RDL (Rewiring Layer), achieving sub-micron level alignment control of multilayer circuit patterns.

[0061] I. System Overall Architecture

[0062] like Figure 1 As shown, the overall structure of the correction system of this invention consists of three core modules: an image acquisition module for high-speed marker recognition and error quantification; an error modeling module for error decoupling, principal component extraction, and trend prediction; and a dynamic compensation module for zoned dynamic compensation and μs-level pattern fine-tuning. These three modules work together to form a closed-loop "acquisition-modeling-compensation" path, supporting real-time processing frequencies above 100 Hz.

[0063] II. Image Recognition and Error Extraction

[0064] like Figure 2 As shown, a convolutional neural network with a U-Net structure is used to segment and extract marker points in the mask image and the imaging image. Then, a Siamese network is used to perform structural alignment and error calculation on the matched marker pairs to obtain the positional deviations Δx and Δy of each marker point. The calculation formulas are as follows:

[0065]

[0066] Where: S_actual is the actual marker coordinates obtained from image recognition, S_design is the designed coordinates of that point in the mask image, α is the structural response coefficient, β is the thermal expansion response coefficient, and ∂T / ∂t represents the rate of temperature change per unit time. This module supports various marker shapes (such as circles, crosses, squares, etc.) and can process multiple alignment markers simultaneously, achieving high coverage and high robustness in real-time detection.

[0067] III. Principal Component Analysis and Decoupling of Errors

[0068] To effectively decouple systematic errors from random errors, principal component analysis (PCA) is introduced. Figure 3 (1) Normalize all labeled bias vectors E = [Δx, Δy]; (2) Construct the error covariance matrix; (3) Solve for its eigenvalues ​​and principal component vectors; (4) Extract principal components with a cumulative contribution rate exceeding 95% to construct the system error model; (5) Hand over the random error part to subsequent modules for processing. This module ensures that overfitting and local random fluctuations are avoided in the subsequent fitting process, and improves the stability and versatility of the system.

[0069] IV. Regional Weighted Fitting Mechanism

[0070] Considering the significant spatial correlation and local deformation characteristics of the error distribution in large-size panels, the system employs a locally weighted multinomial fitting model to perform high-precision spatial modeling of the errors. Figure 4 ).

[0071] The specific steps are as follows: Divide the exposure area into M×N equally spaced grid regions; select K valid marker points around each region (usually K=10~20); and fit a weighted local polynomial model to each block. Where p i (x, y) is the local fitting function. Specifically, a second-order fitting is used: p i (x,y) = a0 + a1x + a2y + a3x 2 +a4y 2 + a5xy. The weighting function is defined as: wi = exp(-((x i – x0)² + (y i -y0) 2 ) / (2σ 2 )), where (x i , y i (x0, y0) is the sampling point, (x0, y0) is the fitting center, and σ is the spatial influence radius.

[0072] V. Fusion of Time Series Forecasting Models

[0073] To address the relationship between thermally induced sudden deformation and its temporal trend, a hybrid prediction model is introduced into the system. Figure 5 – Figure 6 The error for the current frame is provided by a local multinomial fitting model; the error for the next frame is predicted by a time series model (LSTM or ARIMA). The two are combined as follows: f total (x,y,t)=λ·f poly (x,y,t)+(1-λ)·f pred (x,y,t+1), where λ ∈ [0,1], is an adjustable parameter (default λ=0.7). The LSTM model structure consists of two hidden layers. The input sequence is the principal component error (PCA compressed) within the past 10 frames, and the output is the error trend for the next 1-2 frames. It runs in the AI ​​accelerator within the SoC chip.

[0074] VI. Dynamic Compensation and Graphical Fine-tuning Mechanism

[0075] After the error model is output, the system uses a DMD (Digital Micromirror Device) to dynamically adjust the exposure pattern, achieving real-time error correction. Each fitted region corresponds to independent compensation parameters, updated at the μs level, ensuring micrometer-level compensation accuracy. The parameter writing path is: Error Heatmap → Compensation Function Calculation → Optical Control Command → DMD Array.

[0076] VII. Anomaly Detection and Emergency Response Figure 7 )

[0077] In actual working conditions, situations such as thermal shock or mechanical impact can cause sudden, large shifts in the panel graphics, which conventional models cannot respond to in a timely manner. To address this, the system incorporates an anomaly detection and emergency response mechanism: if the error Δ of a marked point exceeds three times the global error standard deviation σ across three consecutive frames, it is marked as an anomaly; the "maximum error envelope fitting model" is activated, using the envelope curve for rapid correction within the original region; the system temporarily increases the sampling frequency (e.g., from 0.5Hz to 2Hz) for high-frequency monitoring of the region; and the anomaly region is marked as a "suboptimal exposure zone" for production line quality backtracking. This mechanism significantly improves the system's robustness and processing stability.

[0078] VIII. Application Examples

[0079] Taking the RDL-3 layer pattern of FOPLP package as an example, the system deploys about 400 sampling points on a 610 mm × 457 mm panel. Based on the above mechanism, it dynamically tracks and corrects thermal overlay errors, achieving an average error of ≤0.3 μm and a response time of <5 seconds. This is a significant improvement over the traditional 30-minute compensation cycle, resulting in a substantial increase in the yield of large-area substrates.

Claims

1. A real-time overlay error correction system based on image recognition and dynamic fitting, characterized in that, It includes an image acquisition module, an error decoupling module, a dynamic compensation module, and an anomaly response module. The modules interact in real time through a high-speed data bus to form a closed-loop control system. The image acquisition module employs a high-resolution linear CCD and real-time triggered sampling control logic to scan the current exposure area at a frequency of ≥100Hz. It performs pixel-level segmentation of alignment marks on the substrate by introducing an improved U-Net structure with an attention mechanism CBAM. It integrates Siamese Network and ORB feature matching algorithms to match and quantify the deviation of marks in the design drawing and the actual drawing, and outputs raw error information including coordinates, deviation values ​​and confidence levels. It supports various types of alignment marks and non-contact online acquisition. The error decoupling module receives the original error information and decouples the systematic error and random error of the error vector E=[Δx, Δy] using the PCA-ICA joint algorithm. It selects the principal component vector with a cumulative contribution rate of ≥95% to construct the systematic error model and separates the random error components using ICA. An adaptive grid partitioning algorithm is used to divide the substrate into dynamic blocks of 10mm×10mm to 100mm×100mm. An error weight matrix and a local weighted quadratic polynomial fitting model are constructed for each block. Based on the ARIMA-BiLSTM hybrid time series prediction model, multi-step prediction is performed on the historical deformation data (Δx_t, Δy_t) of different regions to dynamically generate a global error heat map. The dynamic compensation module adopts a CPU+FPGA+AI accelerator heterogeneous processing platform to perform real-time calculations on the output parameters of the error decoupling module, with a calculation delay of ≤500μs. Based on the error heat map, the substrate is divided into multiple independent compensation blocks with a compensation granularity of ≤10mm. The reverse compensation amount is calculated through a local polynomial fitting model, driving the digital micromirror array (DMD) to perform exposure pattern fine-tuning at a switching speed of ≤1μs, with a compensation update frequency of ≥100Hz. Secondary compensation iteration is achieved through closed-loop verification. The anomaly response module detects abnormal regions using a dual standard of "single frame error threshold + continuous frame change rate". After an anomaly is triggered, the maximum error envelope fitting model is started to generate emergency compensation parameters, with an emergency response time of ≤1ms. The sampling frequency and compensation update frequency of the abnormal region are temporarily increased to 200Hz, and the abnormal region is marked and a traceability report is generated simultaneously.

2. The real-time overlay error correction system based on image recognition and dynamic fitting according to claim 1, characterized in that, The image acquisition module uses a high-resolution linear CCD with a pixel size of 5μm and a scanning width of ≥610mm, supporting alignment marks in various shapes such as circles, crosses, hexagons, and rings.

3. The real-time overlay error correction system based on image recognition and dynamic fitting according to claim 1, characterized in that, The PCA-ICA joint decoupling algorithm of the error decoupling module has a decoupling accuracy of ≥98%. System errors include exposure stage translation / rotation deviation, mask pattern distortion, laser drilling offset, and optical system distortion. Random errors include thermal expansion / contraction, material stress release, and nonlinear deformation caused by local material inhomogeneity.

4. The real-time overlay error correction system based on image recognition and dynamic fitting according to claim 1, characterized in that, The fitting function of the locally weighted quadratic polynomial fitting model is: where (x,y) are the coordinates of any point within the block, wi is the weighting coefficient of the sampling points calculated based on the Gaussian kernel function, and w i =exp(-((x i -x0)²+(y i -y0)²) / (2σ²)), where σ is the spatial influence radius and a0~a5 are the fitting coefficients.

5. The real-time overlay error correction system based on image recognition and dynamic fitting according to claim 1, characterized in that, The ARIMA-BiLSTM hybrid temporal prediction model is driven by an AI accelerator built into the FPGA platform. The AI ​​accelerator has a computing power of ≥1 TOPS. The input is the principal component error data of the past 10 frames, and the output is the predicted error trend value for the next 1-2 frames. The prediction latency is ≤1ms. The fusion strategy of the hybrid temporal prediction model is: f total (x,y,t)=0.7·f poly (x,y,t)+0.3·f pred (x,y,t+1), where f poly For the spatial fitting result, f pred This represents the time series prediction results.

6. The real-time overlay error correction system based on image recognition and dynamic fitting according to claim 1, characterized in that, In the heterogeneous processing platform, the CPU is responsible for system scheduling and data management, the FPGA is responsible for real-time logic operations and control signal generation, and the AI ​​accelerator is responsible for parallel computation of fitting and prediction models.

7. The real-time overlay error correction system based on image recognition and dynamic fitting according to claim 1, characterized in that, The digital micromirror array (DMD) has a pixel count of ≥1920×1080. The compensation parameters are written to the DMD controller through a high-speed LVDS interface. In the closed-loop verification, if the deviation after correction is ≥0.1μm, secondary compensation is triggered until the error is ≤0.3μm.

8. The real-time overlay error correction system based on image recognition and dynamic fitting according to claim 1, characterized in that, The dual anomaly judgment criteria of the anomaly response module are as follows: (1) The error of a single frame marker point Δ is greater than or equal to three times the global error standard deviation σ; (2) The error change rate of marker points for three consecutive frames is ≥0.5μm / frame; The maximum error envelope fitting model uses cubic spline curves for envelope fitting, and abnormal areas are marked as "suboptimal exposure areas" and their location, time, error amplitude, and correction process data are recorded.

9. The real-time overlay error correction system based on image recognition and dynamic fitting according to claim 1, characterized in that, The system is applicable to photolithography processes for FOPLP panel-level fan-out packaging, HDI high-density interconnect substrates, and RDL multilayer redistribution layers.

Images