FCCSP chip support plate production processing analysis and optimization method and system based on etching process data

Abstract

The invention relates to the field of chip carrier production data processing, in particular to an FCCSP chip carrier production processing analysis and optimization method and system based on etching process data, and the method comprises the steps: synchronously collecting etching equipment technological parameters, AOI images and other multi-source time sequence data; microcosmic, mesoscopic and macroscopic multi-scale feature extraction is carried out; inputting the feature sequence into an improved Transform model, and through multi-scale position coding and a cross-scale cross attention mechanism, realizing real-time prediction of line width and defects and mining recessive causal association of process parameters and quality indexes; dynamically generating a process parameter adjustment strategy by adopting a self-adaptive multi-target optimization algorithm introducing abnormal confidence coefficient weight and constraint projection based on a prediction result; the strategy is issued and executed to complete closed-loop control, upstream and downstream data are fused to calculate a yield risk index, and cross-process early warning and presetting are achieved. And the whole-process intelligence from sensing, analysis, decision making to execution is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of FCCSP chip carrier board production data processing, in particular to a FCCSP chip carrier board production processing analysis and optimization method and system based on etching process data. BACKGROUND

[0002] In the high-density line manufacturing of FCCSP (flip chip ball grid array) chip carrier boards, the etching process is the core link that determines the line width precision, uniformity and final yield. In the prior art, the etching process parameters depend on the fixed formula calibrated by offline experiments, and it is difficult to respond to real-time working condition fluctuations. Although sensors and AOI (automatic optical inspection) systems have been deployed on some mainstream production lines to realize basic data acquisition and post-detection, they mainly focus on single-scale parameters and do not realize multi-scale feature fusion, resulting in one-sided correlation analysis of process parameters and quality indicators.

[0003] At the same time, when the AOI detects line width deviation or defects, unqualified products have already been produced, and it is impossible to realize prior prediction and intervention. Although some studies attempt to introduce machine learning for quality prediction, most of them are "black box" models, and there is no causal explanation for how process parameters affect quality indicators in the etching process. In addition, yield control is mostly limited to the etching process, and does not effectively integrate upstream and downstream process data such as lithography and plating, making it difficult to identify cross-process cumulative risks in advance, resulting in delayed abnormal response and affecting overall production stability and cost control effect. SUMMARY

[0004] To solve the above problems, the application provides a FCCSP chip carrier board production processing analysis and optimization method and system based on etching process data.

[0005] In a first aspect, the application provides a FCCSP chip carrier board production processing analysis and optimization method based on etching process data, comprising the following steps: S1, real-time acquisition of multi-source time series data in the etching process of the FCCSP chip carrier board, and synchronous processing; the multi-source time series data includes process parameters of the etching equipment and line topography image data obtained through an automatic optical inspection (AOI) system; the process parameters at least include spray pressure, etching liquid temperature and ion concentration; S2, multi-scale feature extraction of the synchronized multi-source time series data to generate a multi-scale feature sequence in microscale, mesoscale and macroscale; S3, inputting the multi-scale feature sequence into an improved Transformer model to obtain real-time prediction results of the carrier board line width quality and defect risk; the model mines the implicit causal relationship between the process parameters and the quality indicators through multi-scale position encoding and cross-scale cross-attention layers; S4, based on the real-time prediction result, an adaptive multi-objective optimization algorithm is used to dynamically generate a process parameter adjustment strategy that meets the requirements of line width precision, etching uniformity, production efficiency and cost control; the optimization algorithm introduces an abnormal confidence weight and a constraint projection mechanism; S5, the adjustment strategy is sent to the etching equipment control system for execution to complete the closed-loop control; and based on the real-time prediction result and upstream and downstream process data, a yield risk index is calculated, and pre-adjustment or alarm is performed when the index exceeds a threshold.

[0006] Multi-scale feature extraction, improved Transformer causal prediction and adaptive multi-objective optimization are deeply integrated to solve the pain points of one-sided data analysis, decision lag and neglect of multi-objective conflict in traditional methods. The scheme can realize pre-prediction and intervention of the etching process, improve line width precision and yield, ensure process stability through closed-loop control, and realize cross-process collaborative early warning and optimization through system-level risk index.

[0007] As a further limitation of the technical scheme of the application, the steps of S1 include: S11, real-time acquisition of process parameter data through sensors deployed in the spraying system, circulating pipeline and chamber of the etching equipment; wherein the spraying pressure is obtained through a pressure transmitter installed on the main spraying pipeline of the etching machine; the etching liquid temperature is obtained through a platinum resistance temperature sensor immersed in the etching liquid circulating pipeline; and the ion concentration is obtained through an online ion selective electrode sensor installed in the bypass of the etching liquid circulating pipeline; S12, scanning each board through an automatic optical inspection (AOI) system integrated after the etching process to obtain its line topography image data and generate a corresponding detection result time sequence; S13, for sensor data of different sampling frequencies, linear interpolation or spline interpolation method is used for missing data compensation to make all numerical process parameters continuous at the same time resolution; S14, taking the time when the board enters the etching equipment as the reference time point, a uniform time stamp is stamped on all data streams, and the AOI detection result time sequence is associated and aligned with the process parameter sequence experienced by a specific board batch to generate a multi-source time sequence data set.

[0008] The deployment position and type of the sensor are determined to ensure the accuracy and reliability of the process parameter acquisition; missing data is compensated by linear interpolation / spline interpolation, time stamp alignment and batch association to solve the problem of asynchronous multi-source data, generate a high-quality multi-source time sequence data set, and provide an accurate input basis for subsequent multi-scale feature extraction and prediction model, avoiding analysis errors caused by data deviation.

[0009] As a further limitation of the technical scheme of the application, the steps of S2 include: S21. Microscale feature extraction: The instantaneous change rate and statistical variance of ion concentration in the multi-source time series dataset are calculated using a sliding window to obtain the instantaneous change features of ion concentration. S22. Mesoscale feature extraction: Based on the spatial distribution data of spray pressure and etchant temperature field in the multi-source time series dataset, the surface of the carrier plate is divided into several regular grids. The mean and standard deviation of spray pressure and the mean and standard deviation of temperature are calculated in each grid. The local etching rate is estimated by combining the chemical reaction kinetic model, and then the regional distribution characteristics of etching rate are obtained. S23. Macro-scale feature extraction: Statistical process control analysis is performed on the linewidth measurements of all carrier boards in the same batch in the multi-source time series dataset to extract the batch linewidth mean, range, moving range and batch mean offset to obtain the batch process stability characteristics. S24. The obtained microscale features, mesoscale features and macroscale features are organized into a multiscale feature sequence according to time sequence or spatial location.

[0010] Features are extracted hierarchically at the micro, meso, and macro scales to capture instantaneous parameter fluctuations, regional etching uniformity, and batch stability, forming a hierarchical feature system. Microscopic ion concentration variations reflect the instantaneous state of the process, mesoscopic etching rate distribution characteristics pinpoint regional quality differences, and macroscopic batch stability characteristics control the overall process level. The combination of these three features provides a more comprehensive characterization, offering rich and targeted inputs for subsequent prediction models and improving the accuracy of quality prediction and causal analysis.

[0011] As a further limitation of the technical solution of the present invention, step S3 includes: S31. The multi-scale feature sequences are grouped according to their respective scale categories and normalized to obtain the grouped feature sequences, including batch process stability feature sequences, mesoscale etching rate regional distribution feature sequences, and microscale ion concentration instantaneous change feature sequences. S32. Inject positional information into the grouped feature sequences respectively; among them, low-frequency positional encoding is used for macroscopic features that reflect long-period changes, and high-frequency positional encoding is used for microscopic features that reflect instantaneous fluctuations. S33. Input the encoded feature sequence into the encoding layer of the improved Transformer model; wherein the attention mechanism in the encoding layer is constructed as a cross-scale attention layer, wherein the macro-scale feature sequence is used as the input of the query vector, the meso-scale feature sequence is used as the input of the key vector, and the micro-scale feature sequence is used as the input of the value vector. Through the calculation of the cross-scale attention layer, an attention weight matrix representing the dependency between macro-scale and meso-scale features is generated, and the output context vector of the cross-scale attention layer is calculated based on the attention weight matrix and the value vector. S34. The context vector output by the cross-scale cross-attention layer is processed by the feedforward neural network and then input into the regression prediction head and the classification prediction head in parallel to obtain the real-time prediction value of linewidth quality and the probability of defect risk, respectively. S35. Simultaneously, the attention weight matrix is ​​merged and analyzed to identify the weight matrix row corresponding to the query vector representing the quality anomaly; in the weight matrix row, the key vector with the highest weight score is located, which points to the carrier plate region where the uniformity problem occurs; by analyzing the micro-features corresponding to the value vector component that contributes the most to the generation of the context vector of the region in the attention calculation path, the instantaneous process parameter fluctuation that causes the uniformity problem of the region is located, thereby obtaining the implicit causal relationship between process parameters and quality indicators.

[0012] By employing multi-scale location encoding and cross-scale attention mechanisms, the model not only makes accurate predictions but also reveals the causal path between macroscopic batch bias, mesoscopic regional inhomogeneity, and microscopic parameter fluctuations. Transforming the model from a "black box" to a "gray box" enhances the interpretability and engineering credibility of the results, enabling process engineers to trust and make decisions based on the model's findings.

[0013] As a further limitation of the technical solution of this invention, the calculation formula for the cross-scale attention layer is as follows:

[0014] in, For macroscopic scale feature sequences, For mesoscale feature sequences, For microscale feature sequences, This is the output of the cross-scale attention layer, i.e., the context vector.

[0015] The calculation formula for the cross-scale attention layer clearly defines the interaction logic of macroscopic, mesoscopic, and microscopic features, and achieves the orderly fusion of features at different scales through mathematical expression. This formula ensures that the attention weight matrix can accurately quantify the degree of influence of features at each scale on quality indicators, which not only provides a rigorous calculation basis for improving the Transformer model, but also makes the analysis of implicit causal relationships between process parameters and quality indicators interpretable, thereby improving the efficiency of problem investigation.

[0016] As a further limitation of the technical solution of the present invention, step S4 includes: S41. Construct a dynamic multi-objective optimization problem. The optimization objectives should include at least the linewidth accuracy loss, etching uniformity deviation function, production efficiency loss, and cost control increment. S42. Calculate the overall uncertainty of the current prediction based on the variance of the linewidth prediction value output by the regression prediction head and the entropy of the defect risk probability distribution output by the classification prediction head; dynamically adjust the weights of the optimization objectives based on the overall uncertainty: when the overall uncertainty is higher than the upper threshold, increase the weights of the linewidth accuracy and etching uniformity deviation function; when the overall uncertainty is lower than the lower threshold, increase the weights of production efficiency and cost control. S43. An improved multi-objective optimization algorithm is used to solve the Pareto optimal solution set; during the iteration process, the abnormal confidence weight is used to guide the search direction, and a constraint projection mechanism is used to handle process constraints. S44. From the Pareto optimal solution set, combined with the current real-time prediction results and the abnormal confidence level, select the most suitable combination of process parameters for the current production situation and generate the final process parameter adjustment strategy.

[0017] The dynamic multi-objective optimization problem is constructed to comprehensively address the multi-dimensional needs of linewidth accuracy, defect risk, production efficiency, and cost control. By integrating uncertainty calculation and dynamic weight adjustment mechanisms, the optimization priority adapts to the reliability of prediction results, avoiding optimization imbalances caused by static weights. The improved multi-objective optimization algorithm combines anomaly confidence weight guidance and constraint projection mechanisms to improve search efficiency while ensuring that the solution meets equipment process constraints. It quickly generates the optimal process parameter adjustment strategy adapted to the current operating conditions, balancing optimization effectiveness and engineering feasibility.

[0018] As a further limitation of the technical solution of the present invention, step S5 includes: S51. Convert the generated process parameter adjustment strategy into an industrial standard communication protocol data packet and send it to the programmable logic controller of the etching equipment for execution. S52. The programmable logic controller drives the actuator to adjust the process parameters and collects the actual values ​​of the adjusted process parameters in real time to confirm that the closed-loop control action is completed. S53. Based on the real-time prediction results of the current etching process output by S3, and combined with real-time quality data from upstream processes, calculate the system-level yield risk index. ; The yield risk index The calculation formula is:

[0019] in: This is the risk component of the current etching process calculated based on the predicted linewidth quality and defect risk probability output by S3. This is a risk component calculated based on upstream process quality data; These are the weighting coefficients, and ; S54, the yield risk index R Compare with a preset threshold: like R If the threshold is exceeded, a pre-adjustment strategy is generated for the downstream process; like R If the threshold is exceeded, an alarm message is sent to the upstream process and production management system.

[0020] Industrial standard communication protocol conversion ensures seamless integration of adjustment strategies with equipment PLCs, enabling efficient linkage between optimization decisions and equipment control; closed-loop control forms a complete closed loop from prediction, optimization, control to feedback by providing real-time feedback of actual process parameters, ensuring the stability of adjustment effects; the system-level yield risk index integrates etching and upstream process data to achieve early identification of cross-process risks, and graded threshold response makes risk handling more targeted, effectively preventing the spread of anomalies and improving the overall production process's anti-interference capability.

[0021] As a further limitation of the technical solution of this invention, the risk component of the current etching process... The calculation method is as follows: The normalized deviation between the predicted linewidth quality value output by S3 and its target value is weighted and fused with the defect risk probability, i.e.:

[0022] in, These are the weighting coefficients, and , This is a real-time predicted value for line width quality. The preset line width target value, This represents the probability of defect risk.

[0023] The risk component of the etching process is calculated by weighting and fusing the normalized deviation of the predicted linewidth with the probability of defect risk, thus taking into account both linewidth accuracy and defect risk, two core quality indicators. The flexible setting of the weighting coefficients can be adapted to the quality requirements of different production scenarios, ensuring that the risk component accurately reflects the quality status of the etching process.

[0024] As a further limitation of the technical solution of the present invention, the risk component is calculated based on the quality data of the upstream process. The calculation method is as follows: Based on the quality data of the upstream photolithography process obtained by the manufacturing execution system, the normalized deviation of its key parameters is calculated; the key parameters include at least the uniformity of photoresist thickness and the alignment accuracy.

[0025] in, This represents the normalized deviation for the uniformity of photoresist thickness. This is the normalized deviation for alignment accuracy; These are the weighting coefficients, and .

[0026] Risk components are calculated for key parameters in the upstream lithography process to accurately capture fluctuations in upstream processes that significantly impact etching quality. Normalization ensures the comparability of risk contributions for different parameters, and weighting coefficients are adapted to the actual impact of each parameter, ensuring that the calculation of upstream process risk components is scientific and reasonable. This provides comprehensive and accurate input for cross-process system-level risk assessment, avoiding yield losses caused by overlooking fluctuations in upstream processes.

[0027] Secondly, the present invention also provides an FCCSP chip carrier board production processing analysis and optimization system based on etching process data, comprising: The data acquisition and synchronization module is used to acquire multi-source timing data in the etching process of FCCSP chip carrier in real time and perform synchronous processing; the multi-source timing data includes process parameters of the etching equipment and circuit morphology image data obtained by an automated optical inspection (AOI) system; the process parameters include at least spray pressure, etching solution temperature and ion concentration. The multi-scale feature extraction module is used to extract multi-scale features at the micro, meso, and macro scales from synchronized multi-source time-series data, generating multi-scale feature sequences. The intelligent prediction and causal analysis module is used to input the multi-scale feature sequence into the improved Transformer model to obtain real-time prediction results of the carrier board linewidth quality and defect risk, and to explore the implicit causal relationship between process parameters and quality indicators; the model is implemented through multi-scale position encoding and cross-scale cross attention layer. An adaptive optimization decision module is used to dynamically generate process parameter adjustment strategies that meet the requirements of linewidth accuracy, etching uniformity, production efficiency, and cost control based on the real-time prediction results and an adaptive multi-objective optimization algorithm. The optimization algorithm introduces anomaly confidence weights and constraint projection mechanisms. The closed-loop execution and risk monitoring module is used to send the adjustment strategy to the etching equipment control system for execution to complete the closed-loop control; and to calculate the yield risk index based on the real-time prediction results and upstream and downstream process data, and to perform pre-adjustment or alarm when the threshold is exceeded.

[0028] As can be seen from the above technical solutions, this application has the following advantages: Through a full-link design that integrates multi-source time-series data synchronization, multi-scale feature extraction, improved Transformer prediction, and adaptive multi-objective optimization, the etching process is transformed from offline experience-driven to real-time data-driven intelligence. Multi-scale feature fusion overcomes the limitations of single-parameter monitoring; improved Transformer models uncover implicit causal relationships to enhance prediction accuracy; adaptive optimization algorithms dynamically balance multi-objective requirements; and closed-loop control combined with cross-process risk warning improves carrier board linewidth accuracy and reduces defect rates, while simultaneously considering production efficiency and cost control, thus comprehensively improving the stability and yield of FCCSP carrier board production. Attached Figure Description

[0029] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention.

[0031] Figure 2 A block diagram of a system provided in an embodiment of the present invention. Detailed Implementation

[0032] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0034] like Figure 1 As shown, this embodiment of the invention provides a method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data, including the following steps: S1. Real-time acquisition of multi-source timing data during the etching process of the FCCSP chip carrier board, and synchronous processing; the multi-source timing data includes process parameters of the etching equipment and circuit morphology image data obtained by an automated optical inspection (AOI) system; the process parameters include at least spray pressure, etching solution temperature and ion concentration. The spray pressure is acquired in real time by a pressure transmitter installed on the main spray pipe, with a sampling frequency of 10Hz. The temperature of the etching solution is acquired by a platinum resistance temperature sensor immersed in the etching solution circulation pipeline. The copper ion concentration is monitored in real time by an online ion-selective electrode integrated in the circulation pipeline bypass.

[0035] Specifically, the steps in S1 include: S11. Process parameter data are collected in real time by sensors deployed in the spray system, circulation pipeline and chamber of the etching equipment; wherein, the spray pressure is obtained by a pressure transmitter installed on the main spray pipeline of the etching machine; the etching solution temperature is obtained by a platinum resistance temperature sensor immersed in the etching solution circulation pipeline; and the ion concentration is obtained by an online ion selective electrode sensor installed in the bypass of the etching solution circulation pipeline. S12. By using an automated optical inspection (AOI) system integrated after the etching process, each carrier board is scanned to obtain its circuit morphology image data and generate a corresponding time sequence of inspection results. S13. For sensor data with different sampling frequencies, linear interpolation or spline interpolation methods are used to compensate for missing data, so that all numerical process parameters are continuous at the same time resolution. In the data processing server, all process parameter time-series data are resampled using 1Hz as the baseline time axis. For high-frequency data (pressure, temperature), spline interpolation is used to fit a smooth curve using known data points, and then estimated values ​​are taken at the 1Hz time point. This method better preserves the dynamic characteristics of the data. For random missing data caused by brief communication interruptions, computationally simple linear interpolation is used to fill in the gaps. Ultimately, this ensures that all numerical process parameters are continuous and complete at a 1Hz time resolution.

[0036] S14. Using the time when the carrier board enters the etching equipment as the reference time point, a unified timestamp is applied to all data streams from all sources. The time sequence of AOI detection results is associated and aligned with the process parameter sequence experienced by a specific batch of carrier boards to generate a multi-source time sequence dataset.

[0037] When the carrier board enters the etching equipment via RFID or barcode scanning, the manufacturing execution system records the carrier board's unique batch / serial number and precise entry timestamp. Tenter .by Tenter Starting from the conveyor belt speed, the time window for the carrier plate in each process chamber of the etching equipment is calculated.

[0038] From the process parameter database with a unified time axis processed by S13, all process parameter sequences (pressure, temperature, ion concentration) within the specified time window are extracted. The corresponding detection results for each carrier are then matched with the carrier serial number from the AOI system.

[0039] The extracted process parameter time series and the final quality inspection results of the carrier board are stored as a complete data record or data slice in a dedicated analysis database. Thus, each data record fully describes all the process conditions experienced by a specific carrier board during etching and its final output quality, forming a multi-source time series dataset that can be used for model training and real-time analysis.

[0040] S2. Multi-scale feature extraction at the micro, meso, and macro scales is performed on the synchronized multi-source time-series data to generate multi-scale feature sequences. This step involves in-depth processing of the synchronized dataset, extracting features representing the state of the production process from different physical scales.

[0041] Microscale feature extraction: For high-frequency changing parameters such as ion concentration, a sliding window with a duration of 30 seconds is used to calculate the instantaneous rate of change and statistical variance of the data within the window, thereby capturing the rapid fluctuation characteristics of the chemical activity of the etching solution.

[0042] Mesoscale feature extraction: The surface of a single carrier plate is divided into... A regular grid is constructed. For each grid, the mean and standard deviation of the spray pressure and the mean of the temperature field data are calculated. Combining chemical reaction kinetic models such as Faraday's law and the Arrhenius equation, the local etching rate of each grid region is estimated using the temperature and pressure data of each grid, thereby obtaining the spatial distribution characteristics of the etching rate on the carrier surface.

[0043] Macro-scale feature extraction: Taking a production batch as a unit, statistical process control analysis is performed on the AOI linewidth measurements of all carrier boards in that batch. The mean, range, moving range, and batch mean offset relative to the target value of the linewidth for that batch are calculated to evaluate the process stability of the entire batch production.

[0044] The features extracted from the three scales are concatenated in chronological order to form a comprehensive multi-scale feature sequence, which is then used as the input for subsequent models.

[0045] S3. Input the multi-scale feature sequence into the improved Transformer model to obtain real-time prediction results of the carrier board linewidth quality and defect risk; the model mines the implicit causal relationship between process parameters and quality indicators through multi-scale position encoding and cross-scale cross-attention layers. This step groups the multi-scale feature sequences obtained in S2 into three categories: microscopic, mesoscopic, and macroscopic, and performs Z-score normalization on each category. Positional encoding is then performed on each grouped feature sequence. To emphasize the temporal characteristics of different scales, slow-changing macroscopic features are encoded using low-frequency sine waves, while rapidly fluctuating microscopic features are encoded using high-frequency sine waves. In the Transformer encoding layer, instead of using the traditional self-attention mechanism, a mandatory attention mechanism is employed: The query vector is derived from the batch stability features at a macro scale.

[0046] The bond vectors originate from the regional homogeneity characteristics at the mesoscale.

[0047] The value vector originates from the instantaneous fluctuation characteristics at the microscale.

[0048] By calculating Attention (Q=macro, K=meso, V=micro), the model is forced to learn the reasoning path that identifies which regions' inhomogeneities (meso, K) cause batch-specific quality problems (macro Q) and further traces back to which instantaneous process parameter fluctuations (micro, V) are the root cause.

[0049] The output of the attention layer is fed into two output heads in parallel via a feedforward neural network: a regression head that outputs continuous linewidth predictions and a classification head that outputs the probability of defects occurring.

[0050] Simultaneously, the attention weight matrix is ​​analyzed. For example, when a macroscopic query vector representing a batch with a large linewidth exhibits high attention weight to the mesoscopic bond vector representing the edge region of the substrate, it can be determined that uneven edge etching is the main cause of the problem in this batch, and the value vector can be used to correlate it with specific microscopic parameters such as insufficient spray pressure.

[0051] S4. Based on the real-time prediction results, an adaptive multi-objective optimization algorithm is adopted to dynamically generate process parameter adjustment strategies that meet the requirements of linewidth accuracy, etching uniformity, production efficiency, and cost control; the optimization algorithm introduces anomaly confidence weights and constraint projection mechanisms. This step transforms the prediction results into optimal process decisions. Four optimization objectives are defined: linewidth accuracy loss, etching uniformity deviation function, production efficiency loss, and cost control increment. The weights of these objectives are dynamically adjusted based on the prediction uncertainty provided by the S3 model. When uncertainty is high, the system tends to be conservative, increasing the weights of linewidth accuracy and defect risk objectives; in this embodiment, the sum of these two can be greater than 0.7, prioritizing quality assurance. When uncertainty is low, the system tends to be economical, increasing the weights of efficiency and cost objectives.

[0052] An improved NSGA-III algorithm was used for the solution. The optimization variables were spray pressure, etchant temperature, and ion concentration. During the algorithm iteration, a constraint projection mechanism was employed to directly project solutions exceeding the equipment's safe range to the boundary values, ensuring that all solutions were feasible. Finally, from the obtained Pareto optimal solution set, the optimal combination of process parameters was selected as the adjustment strategy based on the current weights.

[0053] S5. The adjustment strategy is sent to the etching equipment control system for execution to complete closed-loop control; and the yield risk index is calculated based on the real-time prediction results and upstream and downstream process data, and pre-adjustment or alarm is performed when the threshold is exceeded.

[0054] This step ensures that the optimization strategy is implemented and prevents systemic risks.

[0055] The adjustment strategy generated by S4, for example, increasing the spray pressure setpoint from 2.0 Bar to 2.2 Bar, is encapsulated into an OPC UA command and sent to the PLC of the etching equipment via industrial Ethernet. The PLC drives the actuators (pressure regulating valve, temperature control unit, replenishment pump) to complete the parameter adjustment and feeds back the adjusted actual value to the system, forming a closed loop.

[0056] By combining the real-time prediction results of the current etching process with the upstream lithography process data from MES, a yield risk index R between 0 and 1 is calculated using a weighted formula.

[0057] Two preset threshold levels are set at 0.3 and 0.6. When 0.3 ≤ R < 0.6, the system determines that the risk is controllable but requires attention, and will generate a pre-adjustment strategy for the downstream electroplating process, generally suggesting adjusting the electroplating current to compensate for possible slight linewidth deviations. When R ≥ 0.6, the system determines that there is a high risk and immediately sends an alarm to the upstream process and production administrator, prompting them to conduct root cause investigation and intervention.

[0058] In some embodiments, step S2 includes: S21. Microscale feature extraction: The instantaneous change rate and statistical variance of ion concentration in the multi-source time series dataset are calculated using a sliding window to obtain the instantaneous change features of ion concentration. This step aims to capture rapid, subtle fluctuations in the chemical activity of the etching solution, which are the microscopic sources affecting etching uniformity and linewidth accuracy.

[0059] From the multi-source time-series dataset generated in S14, ion concentration time-series data for a single substrate throughout its entire etching cycle are extracted. This data is a continuous sequence with a frequency of 1 Hz after interpolation processing in S13. A sliding window of 30 seconds is used, with each window containing 30 data points. The window's sliding step size is set to 1 second to achieve high temporal resolution feature capture.

[0060] Within each window, two key features are computed: Instantaneous rate of change: Calculate the first difference between adjacent data points within the window, and then calculate their average value to characterize the average trend of ion concentration change within that time period, including acceleration or deceleration.

[0061] Statistical variance: Calculates the variance of all data points within the window to characterize the degree of fluctuation in ion concentration during that time period.

[0062] As the window slides, a set of instantaneous rate of change and statistical variance sequences over time are generated for this substrate. These two sequences together constitute the microscale characteristics of the substrate.

[0063] S22. Mesoscale feature extraction: Based on the spatial distribution data of spray pressure and etchant temperature field in the multi-source time series dataset, the surface of the carrier plate is divided into several regular grids. The mean and standard deviation of spray pressure and the mean and standard deviation of temperature are calculated in each grid. The local etching rate is estimated by combining the chemical reaction kinetic model, and then the regional distribution characteristics of etching rate are obtained. This step aims to quantify the spatial uniformity of the etching process on the carrier surface and reveal localized issues. Specifically, the effective area of ​​the FCCSP carrier is divided into a... The algorithm uses a regular grid with 100 cells. Each cell represents an independent analysis area. Based on the fixed position and orientation of the carrier plate in the etching equipment, the spray pressure distribution model and temperature field data collected by the infrared thermal imager are mapped onto the corresponding grid. For example, the grid located at the edge of the carrier plate will have a lower spray pressure value by default than the central area. For each grid... Within its corresponding etching time window, the mean and standard deviation of the spray pressure and the mean and standard deviation of the temperature are calculated. Estimation is performed using a chemical reaction kinetics model. A simplified implementation model is as follows: Local estimation rate

[0064] in, It is a grid average temperature It is the average ion concentration. It is a grid average pressure, These are model constants obtained by fitting historical data. Using this model, a local etching rate is calculated for each grid cell. The etching rates of all 100 grids are combined into a vector, and the statistical characteristics of this vector are calculated to serve as a mesoscopic feature describing the etching uniformity of this substrate.

[0065] S23. Macro-scale feature extraction: Statistical process control analysis is performed on the linewidth measurements of all carrier boards in the same batch in the multi-source time series dataset to extract the batch linewidth mean, range, moving range and batch mean offset to obtain the batch process stability characteristics. Extract the final linewidth measurements {CD1, CD2, ..., CD} of all 50 carrier boards from the same batch from the dataset. 50}

[0066] Batch line width average: Calculated as (CD1 + CD2 + ... + CD) 50 ) / 50.

[0067] Range: Calculate Max(CD1, CD2, ..., CD) 50 )-Min(CD1, CD2, ..., CD 50 ).

[0068] Moving range: Calculate the absolute values ​​of the linewidth differences between adjacent carrier plates |CD2-CD1|, |CD3-CD2|, ..., and then calculate the average of these absolute values.

[0069] Batch mean offset: Calculate |batch line width mean - target line width value|.

[0070] The four statistics mentioned above together constitute the macro-scale characteristics of this production batch, reflecting the overall central tendency, dispersion, and whether systematic drift has occurred.

[0071] S24. The obtained microscale features, mesoscale features and macroscale features are organized into a multiscale feature sequence according to time sequence or spatial location.

[0072] For the analysis of a single carrier board, all features about that carrier board extracted in S21, S22, and S23 are treated as a feature vector. The number of such feature vectors is equal to the number of carrier boards in a batch, generated sequentially over time, forming a feature sequence.

[0073] When detailed analysis of the interior of the carrier plate is required, the features of each grid in S22 can be combined with the micro-features covering the area of ​​that grid to form a feature vector based on spatial location.

[0074] In some embodiments, step S3 includes: S31. The multi-scale feature sequences are grouped according to their respective scale categories and normalized to obtain the grouped feature sequences, including batch process stability feature sequences, mesoscale etching rate regional distribution feature sequences, and microscale ion concentration instantaneous change feature sequences. S32. Inject positional information into the grouped feature sequences respectively; among them, low-frequency positional encoding is used for macroscopic features that reflect long-period changes, and high-frequency positional encoding is used for microscopic features that reflect instantaneous fluctuations. S33. Input the encoded feature sequence into the encoding layer of the improved Transformer model; wherein the attention mechanism in the encoding layer is constructed as a cross-scale attention layer, wherein the macro-scale feature sequence is used as the input of the query vector, the meso-scale feature sequence is used as the input of the key vector, and the micro-scale feature sequence is used as the input of the value vector. Through the calculation of the cross-scale attention layer, an attention weight matrix representing the dependency between macro-scale and meso-scale features is generated, and the output context vector of the cross-scale attention layer is calculated based on the attention weight matrix and the value vector. The calculation formula for cross-scale attention layers is as follows:

[0075] in, For macroscopic scale feature sequences, For mesoscale feature sequences, For microscale feature sequences, This is the output of the cross-scale attention layer, i.e., the context vector.

[0076] Perform matrix multiplication between the macroscopic query sequence and the mesoscopic scale feature sequence; An attention weight matrix is ​​generated. After normalization, each row corresponds to a macroscopic feature, and the sum of the weights in each row is 1. The higher the weight value, the stronger the causal relationship between the macroscopic problem and the mesoscopic problem in that region. This is the attention weight matrix to be analyzed and recorded in S35.

[0077] The calculation process of the cross-scale attention layer is as follows: First, the dot product of the macroscopic query matrix and the mesoscopic key matrix is ​​calculated to obtain the initial attention score; then, the score matrix is ​​scaled and Softmax normalized to generate the attention weight matrix; finally, the weight matrix is ​​multiplied by the microscopic value matrix to obtain the weighted context vector output. This calculation process forces the model to establish an information flow path from macroscopic stability to mesoscopic homogeneity, and then to microscopic fluctuations.

[0078] S34. The context vector output by the cross-scale attention layer is processed by a feedforward neural network and then input in parallel to the regression prediction head and the classification prediction head to obtain the real-time predicted value of linewidth quality and the probability of defect risk, respectively. It should be noted that the regression prediction head consists of a linear layer, which maps the final feature representation to a continuous numerical value, i.e., the real-time predicted value of linewidth quality. The classification prediction head consists of a linear layer and a sigmoid activation function, mapping the feature representation to a probability between 0 and 1, i.e., the defect risk probability. .

[0079] S35. Simultaneously, the attention weight matrix is ​​merged and analyzed to identify the weight matrix row corresponding to the query vector representing the quality anomaly; in the weight matrix row, the key vector with the highest weight score is located, which points to the carrier plate region where the uniformity problem occurs; by analyzing the micro-features corresponding to the value vector component that contributes the most to the generation of the context vector of the region in the attention calculation path, the instantaneous process parameter fluctuation that causes the uniformity problem of the region is located, thereby obtaining the implicit causal relationship between process parameters and quality indicators.

[0080] In the attention weight matrix generated by S33, find the row corresponding to a known macroscopic feature of quality anomaly (in this embodiment, this includes a severely out-of-tolerance average batch linewidth), i.e., a specific macroscopic query vector. Within that row, find the key vector with the highest weight score. This key vector corresponds to a specific mesoscopic feature, thereby locating the carrier plate region where uniformity issues occur.

[0081] In attention calculations, high-weighted key vectors extract more information from the value vector (V, derived from micro-features). By analyzing the micro-feature dimensions that dominate this attention path, the instantaneous process parameter fluctuations that cause uniformity issues in the region can be located.

[0082] In some embodiments, step S4 includes: S41. Construct a dynamic multi-objective optimization problem, where the optimization objectives include at least linewidth accuracy loss, etching uniformity deviation function, production efficiency loss, and cost control increment; specifically including: S411. Constructing the linewidth accuracy loss function ,in This is the real-time predicted value of linewidth quality output by S3. The preset target line width value; S412, Constructing the Etching Uniformity Deviation Function ,in This represents the standard deviation of the etching rate in each grid region at the mesoscale. This represents the average etching rate for each grid region; S413. Constructing the production efficiency loss function ,in For standard etching time, To optimize variables Calculated estimated etching time; S414. Constructing the cost control increment function ,in To optimize variables The corresponding estimated process cost is the weighted sum of etching solution consumption, equipment energy consumption, and maintenance costs; Based on the baseline process cost; S415. Combine the objective functions from S411-S414 to form a dynamic multi-objective optimization problem:

[0083] Where Ω represents the feasible region defined by the physical limits of the equipment and the process specifications. x =[ p , t , c [ ] Represent spray pressure, etching solution temperature, and ion concentration, respectively.

[0084] S42. Calculate the overall uncertainty of the current prediction based on the variance of the predicted linewidth value output by the regression prediction head and the entropy of the defect risk probability distribution output by the classification prediction head; dynamically adjust the weights of the optimization objectives based on the overall uncertainty: when the overall uncertainty is higher than the upper threshold, increase the weights of linewidth accuracy and etching uniformity deviation function; when the overall uncertainty is lower than the lower threshold, increase the weights of production efficiency and cost control; specifically including: S421. Extract the variance estimate corresponding to the current real-time prediction from the regression prediction head output. The linewidth uncertainty component is then normalized to obtain the linewidth uncertainty component. :

[0085] in, This represents the maximum variance of the linewidth prediction based on historical data statistics.

[0086] S422. Defect risk probability based on the output of the classification prediction head in S3. Calculate its probability distribution entropy As a measure of uncertainty, and after normalization, the defect uncertainty component is obtained. : ,

[0087] in, This represents the maximum entropy of the Bernoulli distribution.

[0088] S423, Calculating Comprehensive Uncertainty ,in These are the preset weighting coefficients; according to U The value of is used to classify production scenarios: High uncertainty scenarios: when hour; Low uncertainty scenario: when hour; and This is the preset threshold.

[0089] S424. Assign weight vectors to the optimization objective according to different scenarios. These correspond to the targets for line width accuracy, defect risk, production efficiency, and cost control, respectively: In situations of high uncertainty, allocation ; In a low-uncertainty situation, allocation .

[0090] S43. An improved multi-objective optimization algorithm is used to solve the Pareto optimal solution set; during the iteration process, the abnormal confidence weight is used to guide the search direction, and a constraint projection mechanism is used to handle process constraints; the abnormal confidence weight is a dynamic weight used to bias the search direction in the optimization algorithm, calculated based on the comprehensive uncertainty of the prediction model. S43, which employs an improved multi-objective optimization algorithm, specifically includes the following steps: S431. Initialize the population by combining process parameters, including spray pressure, etching solution temperature, and ion concentration, and encoding them as individuals in the population. S432. In the selection operation of each generation, a roulette wheel selection method based on the weight of abnormal confidence is introduced; the selection probability of an individual is positively correlated with its fitness performance on high-weight targets. When the weight of line width and defect targets is high, individuals that perform well on these targets are given priority. In S432, the specific implementation of the weight adaptive selection operation is as follows: Based on the obtained dynamic weight vector Calculate the weighted overall fitness of each individual F :

[0091] in For the first i The result of normalizing the objective function values; Individual choice probability Its weighted overall fitness F Proportional.

[0092] S433. For new individuals generated after crossover and mutation operations, a constrained projection mechanism is used to map them to the process feasible region; the constrained projection mechanism is as follows: if the process parameters of a certain sample are... Beyond its feasible scope Then it is corrected to a boundary value:

[0093] S434. Employ an elite retention strategy to preserve outstanding individuals from the parent generation to the offspring generation. When the improvement rate of the objective function of the population is lower than the threshold for several consecutive generations or the maximum number of iterations is reached, stop the iteration and output the final Pareto optimal solution set.

[0094] S44. From the Pareto optimal solution set, combined with the current real-time prediction results and anomaly confidence levels, the most suitable combination of process parameters for the current production situation is selected, generating the final process parameter adjustment strategy. Specifically, this includes: S441. Based on the current real-time prediction results and the abnormal confidence level, define a dynamic ideal point. ; When in a high uncertainty situation, the linewidth accuracy will be lost. With Etching Uniformity Deviation Function The ideal value is set to zero, that is... ; When in a low-uncertainty situation, the production efficiency will be lost. With cost control increment The ideal value is set to zero, that is... .

[0095] S442. For each candidate solution x in the Pareto optimal solution set, calculate its objective vector f(x) to the dynamic ideal point. Weighted Euclidean distance d (x);

[0096] in, This is a dynamic weight vector.

[0097] S443. Select the weighted Euclidean distance from the Pareto optimal solution set. d (x) The smallest candidate solution This is the combination of process parameters that best suits the current production situation.

[0098] S444, Optimal solution The code is decoded into specific process parameter settings, generating adjustment strategies including target spray pressure, target etchant temperature, and target ion concentration, and encapsulated into an instruction format that can be issued via the OPC UA protocol.

[0099] In some embodiments, S5 completes closed-loop control and risk monitoring, specifically including the following steps: S51. Convert the generated process parameter adjustment strategy into an industrial standard communication protocol data packet and send it to the programmable logic controller of the etching equipment for execution. S52. The programmable logic controller drives the actuator to adjust the process parameters and collects the actual values ​​of the adjusted process parameters in real time to confirm that the closed-loop control action is completed. S53. Based on the real-time prediction results of the current etching process output by S3, and combined with real-time quality data from upstream processes, calculate the system-level yield risk index. R ; The yield risk index R The calculation formula is:

[0100] in: This is the risk component of the current etching process calculated based on the predicted linewidth quality and defect risk probability output by S3. This is a risk component calculated based on upstream process quality data; These are the weighting coefficients, and In FCCSP substrate production, the etching process is the decisive step in forming the circuit pattern. Defects occurring at this stage are extremely difficult to repair in subsequent processes. Therefore, the real-time status of the current etching process has the highest weight in terms of final yield. Based on historical data analysis and expert experience, the risk weight of the current etching process is... The value is typically set between 0.6 and 0.8, with 0.7 being the recommended value. The quality of the upstream photolithography process is an input condition for the etching process. Its weight... and This is complementary, reflecting the proportion of responsibility for upstream issues in the final yield loss. It is typically set between 0.2 and 0.4, with a recommended value of 0.3. The ratio is calculated by statistically analyzing the historical proportion of yield loss caused by etching issues and the proportion caused by problems with incoming photolithography materials. and The initial value. Later, feedback and optimization can be performed using system runtime data.

[0101] S54, the yield risk index R Compare with a preset threshold: like R If the threshold is exceeded, a pre-adjustment strategy is generated for the downstream process; like R If the threshold is exceeded, an alarm message is sent to the upstream process and production management system.

[0102] Risk component of the current etching process The calculation method is as follows: The normalized deviation between the predicted linewidth quality value output by S3 and its target value is weighted and fused with the defect risk probability, i.e.:

[0103] in, These are the weighting coefficients, and , This is a real-time predicted value for line width quality. The preset line width target value, This represents the probability of defect risk.

[0104] It's important to note that defects in FCCSP chip carrier boards render them absolute rejects, while linewidth deviations are acceptable within a certain range. Therefore, defect risk carries a veto power in risk assessment. When the model predicts a high defect risk, a high-level alert must be triggered immediately. This refers to the probability weighting of defect risk. The value is usually set between 0.5 and 0.7, with a recommended value of 0.6.

[0105] Linewidth deviation affects the consistency of product performance, but it is a parametric deviation. Linewidth accuracy deviation weighting. and Complementarity is used to capture process states that have deviated from the optimization goal but have not yet caused fatal defects. It is usually set between 0.3 and 0.5, with a recommended value of 0.4.

[0106] Risk component of upstream process The calculation method is as follows: Based on the quality data of the upstream photolithography process obtained by the manufacturing execution system, the normalized deviation of its key parameters is calculated; the key parameters include at least the uniformity of photoresist thickness and the alignment accuracy.

[0107] in, This represents the normalized deviation for the uniformity of photoresist thickness. This is the normalized deviation for alignment accuracy; These are the weighting coefficients, and .

[0108] The normalized deviation of photoresist thickness uniformity is obtained from photoresist thickness measurement data acquired from upstream MES. In this embodiment of the invention, it is the standard deviation of multi-point measurements. Divided by its process control upper limit Normalization yields, i.e. The higher the value, the more uneven the adhesive thickness, and the greater the potential impact on etching uniformity.

[0109] The normalized deviation of the alignment accuracy is the actual alignment offset in this embodiment of the invention, obtained from the exposure alignment accuracy measurement data acquired by the upstream MES. Divide by its specification tolerance Normalization yields, i.e. The larger the value, the closer the alignment error is to or beyond the tolerance, resulting in a higher risk of line bridging or disconnection.

[0110] In FCCSP micro-circuit fabrication, alignment accuracy is paramount. Even minute alignment errors can directly lead to circuit bridging or breakage, resulting in fatal and irreparable defects. Its impact is structural; therefore, alignment accuracy deviation has the highest weighting. A value of 0.6 is recommended.

[0111] Uneven photoresist thickness can affect local etching rates, primarily leading to poorer linewidth uniformity. It is a performance-impacting factor rather than a directly fatal one. The weight of photoresist thickness uniformity deviation is typically lower than that of alignment accuracy, with a recommended value of 0.4.

[0112] Process simulation or historical data correlation analysis is employed. By simulating the impact of different alignment errors and resist thickness variations on the linewidth and defects after etching, their importance weights can be deduced.

[0113] like Figure 2 As shown, this embodiment of the invention also provides an FCCSP chip carrier production processing analysis and optimization system based on etching process data, including: The data acquisition and synchronization module is used to acquire multi-source timing data in the etching process of FCCSP chip carrier in real time and perform synchronous processing; the multi-source timing data includes process parameters of the etching equipment and circuit morphology image data obtained by an automated optical inspection (AOI) system; the process parameters include at least spray pressure, etching solution temperature and ion concentration. The multi-scale feature extraction module is used to extract multi-scale features at the micro, meso, and macro scales from synchronized multi-source time-series data, generating multi-scale feature sequences. The intelligent prediction and causal analysis module is used to input the multi-scale feature sequence into the improved Transformer model to obtain real-time prediction results of the carrier board linewidth quality and defect risk, and to explore the implicit causal relationship between process parameters and quality indicators; the model is implemented through multi-scale position encoding and cross-scale cross attention layer. An adaptive optimization decision module is used to dynamically generate process parameter adjustment strategies that meet the requirements of linewidth accuracy, etching uniformity, production efficiency, and cost control based on the real-time prediction results and an adaptive multi-objective optimization algorithm. The optimization algorithm introduces anomaly confidence weights and constraint projection mechanisms. The closed-loop execution and risk monitoring module is used to send the adjustment strategy to the etching equipment control system for execution to complete the closed-loop control; and to calculate the yield risk index based on the real-time prediction results and upstream and downstream process data, and to perform pre-adjustment or alarm when the threshold is exceeded.

[0114] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data, characterized in that, Includes the following steps: S1. Real-time acquisition of multi-source timing data during the etching process of the FCCSP chip carrier board, and synchronous processing; the multi-source timing data includes process parameters of the etching equipment and circuit morphology image data obtained by an automated optical inspection (AOI) system; the process parameters include at least spray pressure, etching solution temperature and ion concentration. S2. Multi-scale feature extraction at the micro, meso, and macro scales is performed on the synchronized multi-source time-series data to generate multi-scale feature sequences. S3. Input the multi-scale feature sequence into the improved Transformer model to obtain real-time prediction results of the carrier board linewidth quality and defect risk; the model mines the implicit causal relationship between process parameters and quality indicators through multi-scale position encoding and cross-scale cross-attention layers. S4. Based on the real-time prediction results, an adaptive multi-objective optimization algorithm is adopted to dynamically generate process parameter adjustment strategies that meet the requirements of linewidth accuracy, etching uniformity, production efficiency, and cost control; the optimization algorithm introduces anomaly confidence weights and constraint projection mechanisms. S5. The adjustment strategy is sent to the etching equipment control system for execution to complete closed-loop control; and the yield risk index is calculated based on the real-time prediction results and upstream and downstream process data, and pre-adjustment or alarm is performed when the threshold is exceeded.

2. The method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data according to claim 1, characterized in that, The steps in S1 include: S11. Process parameter data are collected in real time by sensors deployed in the spray system, circulation pipeline and chamber of the etching equipment; wherein, the spray pressure is obtained by a pressure transmitter installed on the main spray pipeline of the etching machine; the etching solution temperature is obtained by a platinum resistance temperature sensor immersed in the etching solution circulation pipeline; and the ion concentration is obtained by an online ion selective electrode sensor installed in the bypass of the etching solution circulation pipeline. S12. By using an automated optical inspection (AOI) system integrated after the etching process, each carrier board is scanned to obtain its circuit morphology image data and generate a corresponding time sequence of inspection results. S13. For sensor data with different sampling frequencies, linear interpolation or spline interpolation methods are used to compensate for missing data, so that all numerical process parameters are continuous at the same time resolution. S14. Using the time when the carrier board enters the etching equipment as the reference time point, a unified timestamp is applied to all data streams from all sources. The time sequence of AOI detection results is associated and aligned with the process parameter sequence experienced by a specific batch of carrier boards to generate a multi-source time sequence dataset.

3. The method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data according to claim 1, characterized in that, The steps in S2 include: S21. Microscale feature extraction: The instantaneous change rate and statistical variance of ion concentration in the multi-source time series dataset are calculated using a sliding window to obtain the instantaneous change features of ion concentration. S22. Mesoscale feature extraction: Based on the spatial distribution data of spray pressure and etchant temperature field in the multi-source time series dataset, the surface of the carrier plate is divided into several regular grids. The mean and standard deviation of spray pressure and the mean and standard deviation of temperature are calculated in each grid. The local etching rate is estimated by combining the chemical reaction kinetic model, and then the regional distribution characteristics of etching rate are obtained. S23. Macro-scale feature extraction: Statistical process control analysis is performed on the linewidth measurements of all carrier boards in the same batch in the multi-source time series dataset to extract the batch linewidth mean, range, moving range and batch mean offset to obtain the batch process stability characteristics. S24. The obtained microscale features, mesoscale features and macroscale features are organized into a multiscale feature sequence according to time sequence or spatial location.

4. The method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data according to claim 1, characterized in that, The steps in S3 include: S31. The multi-scale feature sequences are grouped according to their respective scale categories and normalized to obtain the grouped feature sequences, including the instantaneous change feature sequence of ion concentration, the regional distribution feature sequence of etching rate, and the batch process stability feature sequence. S32. Inject positional information into the grouped feature sequences respectively; among them, low-frequency positional encoding is used for macroscopic features that reflect long-period changes, and high-frequency positional encoding is used for microscopic features that reflect instantaneous fluctuations. S33. Input the encoded feature sequence into the encoding layer of the improved Transformer model; wherein the attention mechanism in the encoding layer is constructed as a cross-scale attention layer, wherein the macro-scale feature sequence is used as the input of the query vector, the meso-scale feature sequence is used as the input of the key vector, and the micro-scale feature sequence is used as the input of the value vector. Through the calculation of the cross-scale attention layer, an attention weight matrix representing the dependency between macro-scale and meso-scale features is generated, and the output context vector of the cross-scale attention layer is calculated based on the attention weight matrix and the value vector. S34. The context vector output by the cross-scale cross-attention layer is processed by the feedforward neural network and then input into the regression prediction head and the classification prediction head in parallel to obtain the real-time prediction value of linewidth quality and the probability of defect risk, respectively. S35. Simultaneously, the attention weight matrix is ​​merged and analyzed to identify the weight matrix row corresponding to the query vector representing the quality anomaly; in the weight matrix row, the key vector with the highest weight score is located, which points to the carrier plate region where the uniformity problem occurs; by analyzing the micro-features corresponding to the value vector component that contributes the most to the generation of the context vector of the region in the attention calculation path, the instantaneous process parameter fluctuation that causes the uniformity problem of the region is located, thereby obtaining the implicit causal relationship between process parameters and quality indicators.

5. The method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data according to claim 4, characterized in that, The calculation formula for cross-scale attention layers is as follows: in, For macroscopic scale feature sequences, For mesoscale feature sequences, For microscale feature sequences, This is the output of the cross-scale attention layer, i.e., the context vector.

6. The method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data according to claim 1, characterized in that, The steps in S4 include: S41. Construct a dynamic multi-objective optimization problem. The optimization objectives should include at least the linewidth accuracy loss, etching uniformity deviation function, production efficiency loss, and cost control increment. S42. Calculate the overall uncertainty of the current prediction based on the variance of the linewidth prediction value output by the regression prediction head and the entropy of the defect risk probability distribution output by the classification prediction head; dynamically adjust the weights of the optimization objectives based on the overall uncertainty: when the overall uncertainty is higher than the upper threshold, increase the weights of the linewidth accuracy and etching uniformity deviation function; when the overall uncertainty is lower than the lower threshold, increase the weights of production efficiency and cost control. S43. An improved multi-objective optimization algorithm is used to solve the Pareto optimal solution set; during the iteration process, the abnormal confidence weight is used to guide the search direction, and a constraint projection mechanism is used to handle process constraints. S44. From the Pareto optimal solution set, combined with the current real-time prediction results and the abnormal confidence level, select the most suitable combination of process parameters for the current production situation and generate the final process parameter adjustment strategy.

7. The method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data according to claim 1, characterized in that, The steps in S5 include: S51. Convert the generated process parameter adjustment strategy into an industrial standard communication protocol data packet and send it to the programmable logic controller of the etching equipment for execution. S52. The programmable logic controller drives the actuator to adjust the process parameters and collects the actual values ​​of the adjusted process parameters in real time to confirm that the closed-loop control action is completed. S53. Based on the real-time prediction results of the current etching process output by S3, and combined with real-time quality data from upstream processes, calculate the system-level yield risk index. ; The yield risk index The calculation formula is: in: This is the risk component of the current etching process calculated based on the predicted linewidth quality and defect risk probability output by S3. This is a risk component calculated based on upstream process quality data; These are the weighting coefficients, and ; S54, the yield risk index R Compare with a preset threshold: like R If the threshold is exceeded, a pre-adjustment strategy is generated for the downstream process; like R If the threshold is exceeded, an alarm message is sent to the upstream process and production management system.

8. The method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data according to claim 7, characterized in that, Current Etching Process Risk Components The calculation method is as follows: The normalized deviation between the predicted linewidth quality value output by S3 and its target value is weighted and fused with the defect risk probability, i.e.: in, These are the weighting coefficients, and , This is a real-time predicted value for line width quality. The preset line width target value, This represents the probability of defect risk.

9. The method for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data according to claim 8, characterized in that, Risk component calculated based on upstream process quality data The calculation method is as follows: Based on the quality data of the upstream photolithography process obtained by the manufacturing execution system, the normalized deviation of its key parameters is calculated; the key parameters include at least the uniformity of photoresist thickness and the alignment accuracy. in, This represents the normalized deviation of photoresist thickness uniformity. This is the normalized deviation for alignment accuracy; These are the weighting coefficients, and .

10. A system for analyzing and optimizing the production process of FCCSP chip carrier boards based on etching process data, characterized in that, include: The data acquisition and synchronization module is used to acquire multi-source timing data in the etching process of FCCSP chip carrier board in real time and perform synchronization processing. The multi-source time-series data includes process parameters of the etching equipment and circuit topography image data acquired by an automated optical inspection (AOI) system; the process parameters include at least spray pressure, etching solution temperature, and ion concentration. The multi-scale feature extraction module is used to extract multi-scale features at the micro, meso, and macro scales from synchronized multi-source time-series data, generating multi-scale feature sequences. The intelligent prediction and causal analysis module is used to input the multi-scale feature sequence into the improved Transformer model to obtain real-time prediction results of the carrier board linewidth quality and defect risk, and to explore the implicit causal relationship between process parameters and quality indicators; the model is implemented through multi-scale position encoding and cross-scale cross attention layer. An adaptive optimization decision module is used to dynamically generate process parameter adjustment strategies that meet the requirements of linewidth accuracy, etching uniformity, production efficiency, and cost control based on the real-time prediction results and an adaptive multi-objective optimization algorithm. The optimization algorithm introduces anomaly confidence weights and constraint projection mechanisms. The closed-loop execution and risk monitoring module is used to send the adjustment strategy to the etching equipment control system for execution to complete the closed-loop control; and to calculate the yield risk index based on the real-time prediction results and upstream and downstream process data, and to perform pre-adjustment or alarm when the threshold is exceeded.

Images