A Big Data Analysis Platform for Wafer Cleaning to Predict the Lifespan of PVA Sponge Rollers

Abstract

This invention belongs to the field of semiconductor cleaning big data technology and discloses a wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers. It constructs a four-dimensional, all-encompassing acquisition system based on PVA multi-source acquisition modules, covering PVA characteristics, process parameters, cleaning effects, and environmental conditions. This system not only covers the physical properties, degradation characteristics, and formulation information of the sponge rollers but also simultaneously collects process operation data, environmental adaptation data, and wafer cleaning quality feedback data. It adds correlation data between defect types and sponge roller wear locations, completely establishing a chain of correlation between sponge roller status, process parameters, and cleaning effects. Utilizing edge node preprocessing, timestamp and process stage dual-dimensional time sequence alignment, and an adaptive acquisition strategy, the acquisition frequency for high-difficulty cleaning tasks during the degradation period and below 7nm is increased to 200Hz, while the frequency for routine tasks is reduced to 50Hz. This ensures data integrity and timeliness while effectively reducing cloud transmission pressure and resource consumption.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor cleaning big data technology, specifically a wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers. Background Technology

[0002] In the semiconductor wafer manufacturing process, the cleaning process is a crucial step in ensuring wafer surface quality and improving the yield of subsequent processes. PVA sponge rollers are a core consumable in wafer cleaning equipment, and their lifespan directly affects the stability of the cleaning effect. Changes in the wear and aging of PVA sponge rollers have an increasingly significant impact on the wafer surface defect rate. Current wafer cleaning and PVA sponge roller management technologies face the following technical challenges: Existing semiconductor cleaning equipment data analysis primarily focuses on monitoring operational parameters such as temperature, pressure, and rotation speed. It fails to achieve multi-dimensional data fusion, encompassing the physical properties of the PVA sponge roller itself (e.g., porosity, water absorption, compressive stress), real-time interactive data from the cleaning process (e.g., friction, residual cleaning fluid), and wafer cleaning quality feedback data (e.g., number of surface defects, cleanliness). These data sources are isolated, making it difficult to construct a correlation chain between sponge roller status, process parameters, and cleaning effect, resulting in a lack of comprehensive data support for lifetime prediction.

[0003] Existing lifetime prediction methods are mostly based on traditional statistical models such as linear regression or single machine learning algorithms such as LSTM, without incorporating professional knowledge in areas such as the degradation mechanism of PVA sponge roller materials and the physical constraints of cleaning processes. Furthermore, the wear patterns of sponge rollers differ depending on the wafer fabrication process and the type of cleaning equipment used, making it difficult for these models to adapt to different scenarios and resulting in prediction accuracy that fails to meet the demands of high-precision production.

[0004] For extreme process conditions such as high temperature and high concentration of cleaning fluid, existing technologies have not specifically modeled the degradation characteristics of sponge rollers, resulting in a significant decrease in the prediction accuracy of the models in such scenarios. In addition, most data-driven models are black box structures, which cannot explain the core influencing factors of the life prediction results and are difficult to support process parameter optimization decisions. Summary of the Invention

[0005] The purpose of this invention is to provide a big data analysis platform for wafer cleaning that predicts the lifespan of PVA sponge rollers, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers, comprising: Preferably, the PVA multi-source acquisition module adopts a four-dimensional, real-time acquisition architecture encompassing PVA characteristics, process parameters, cleaning effect, and environmental conditions, constructing a data acquisition system that combines specificity and completeness. It acquires core physical characteristic parameters such as porosity, water absorption rate, surface wear, and compressive stress through an embedded micro-sensor array, and combines this with high-frequency ultrasonic detection technology to achieve non-contact wear monitoring. Simultaneously, it adds the acquisition of PVA degradation characteristic parameters and synchronous recording of sponge roller formulation information. It also synchronously acquires real-time operating parameters such as equipment rotation speed, cleaning fluid temperature and concentration, spray pressure, and cleaning time, as well as environmental parameters within the cleaning chamber such as humidity and oxygen concentration, adapting to the needs of PVA thermal oxidation degradation kinetic analysis. Furthermore, it acquires the number of wafer surface defects and cleanliness level through machine vision, and combines this with offline detection data from semiconductor testing equipment to add correlation data between defect types and sponge roller wear locations. Machine vision acquisition parameters: An industrial area array camera with a resolution of 20 million pixels and a frame rate of 30fps is selected and installed 30cm above the discharge port in the cleaning chamber. The lens is vertically downward to capture images of the wafer surface. A ring LED light source (wavelength 500-600nm) is used to eliminate reflection interference. Image processing adopts grayscale threshold segmentation + morphological filtering algorithm to identify defects such as scratches and particles on the wafer surface. The defect size recognition accuracy is 0.5m. The cleanliness level is classified into 1-5 according to the SEMI standard.

[0007] Embedded micro-sensor array deployment: 3-5 sets of sensors are evenly distributed on both ends of the PVA sponge roller and on the surface of the roller body. The porosity and water absorption rate sensors are embedded 10mm inside the roller body to avoid direct rinsing by the cleaning fluid. The surface wear and compressive stress sensors are attached to the surface of the roller body and are waterproof and sealed with an IP68 protection rating to resist corrosion from the cleaning fluid. The number of sensors is adapted to the length of the sponge roller. 3 sets are deployed for a length of 500mm and 5 sets are deployed for a length greater than 500mm to ensure full detection coverage.

[0008] High-frequency ultrasonic testing technical parameters: The ultrasonic sensor is installed on the inner wall of the cleaning chamber, with a distance of 5-8mm from the surface of the sponge roller. The installation angle is 30° with the tangent of the roller. A 2-5MHz high-frequency ultrasonic probe is selected, with a sampling point interval of 0.5mm. The wear thickness is calculated by the attenuation of the echo signal, and the signal interference caused by air bubbles in the cleaning fluid is filtered out simultaneously.

[0009] Edge computing nodes are deployed at the cleaning equipment end to preprocess high-frequency sensor data in real time to reduce the transmission pressure on the cloud. A dual-dimensional synchronization mechanism of timestamp and process stage is adopted to achieve time sequence alignment of multi-source data. An adaptive acquisition strategy is designed based on the PVA degradation stage and wafer process type to accelerate the degradation period and high-difficulty cleaning tasks below 7nm by increasing the acquisition frequency to 200Hz, while reducing it to 50Hz for regular tasks, thus optimizing resource consumption while ensuring data integrity.

[0010] Specific process of edge node preprocessing: ① The noise reduction process uses a median filtering algorithm with a window size of 5 to filter high-frequency jitter noise from the sensor; ② Downsampling uses mean sampling to compress high-frequency data of 100Hz and above by an average of time windows, ensuring that the data retains the core trend of change after downsampling; ③ The format is standardized and uniformly encapsulated in JSON format. The fields include the acquisition timestamp, process stage identifier, parameter type, parameter value, sensor number, and data quality mark. The data quality mark is labeled in four levels: excellent, good, medium, and poor. It is based on the sensor signal strength, with 80dB being excellent, 60-79dB being good, 40-59dB being medium, and <40dB being poor.

[0011] Preferably, the knowledge graph processing module overcomes the limitations of general semiconductor knowledge graphs by constructing a dedicated knowledge graph for PVA sponge roller-wafer cleaning. This dedicated knowledge graph integrates four core knowledge bases: a PVA material properties and degradation mechanism base containing thermal oxidation and hydrolytic degradation laws of sponge rollers with different formulations and degradation rate constant correlation data based on the Arrhenius equation; a cleaning process specification base containing standard parameter ranges and extreme working condition thresholds corresponding to each process; a fault case base containing the correlation between wear anomalies and defect types, and the correspondence between temperature and humidity fluctuations and degradation acceleration; and a data quality rule base containing abnormal data judgment standards based on PVA degradation kinetics. Data sources and update mechanisms for the four core knowledge bases: ① PVA Material Properties and Degradation Mechanism Library: The basic data comes from laboratory accelerated aging tests (degradation monitoring under temperature 40-80℃ and humidity 50%-90%) and authoritative industry literature. The actual factory operation data is added and updated every quarter. ② Cleaning process specification library: Based on semiconductor industry standards and combined with the process optimization experience of various partner factories, threshold calibration is performed by collecting factory process adjustment feedback every month. ③ Fault Case Library: Real-time collection of equipment fault handling records and wafer defect tracing reports from various factories, updated weekly after manual annotation of correlations; ④ Data quality rule base: Initial rules are established based on PVA degradation kinetic test data, and updated every six months in combination with model prediction error feedback. Rule optimization is triggered when the error is >15%.

[0012] Based on the dedicated knowledge graph, three core functions are realized: intelligent noise identification and removal, cross-device data semantic standardization, and multi-dimensional data quality assessment and repair. Intelligent noise identification and removal accurately identifies abnormal data caused by sensor failure and environmental interference through dedicated association rules in the graph. Cross-device data semantic standardization unifies synonymous parameters into cleaning roller linear speed based on the ontology matching mechanism of the knowledge graph and establishes a unified dimensional standard for PVA characteristic parameters. The ontology matching mechanism is implemented as follows: The ontology of the dedicated knowledge graph is constructed according to a three-level architecture: top-level ontology, domain ontology, and instance ontology. The top-level ontology contains core elements such as entities, attributes, and relationships. The domain ontology is subdivided into five sub-domains: PVA sponge roller, cleaning equipment, process parameters, environmental conditions, and cleaning effect. The instance ontology corresponds to specific sponge roller models, equipment parameter values, etc. The matching process consists of two steps: the first step uses a string similarity algorithm to initially match synonymous parameters, and the second step verifies the rationality of the matching through semantic association rules in the domain ontology (such as the linear relationship between equipment rotation speed and cleaning roller linear speed). Finally, a unified and standardized parameter name and dimension are output to ensure semantic consistency of data across devices.

[0013] Multi-dimensional data quality assessment and repair uses evaluation indicators for completeness, consistency, accuracy, and timeliness, combined with the PVA degradation process threshold in the knowledge graph to evaluate the data quality level in real time. For low-quality data, a missing value completion method based on generative adversarial networks is used to finally output standardized data.

[0014] Details of missing value completion using Generative Adversarial Network (GAN): The generator employs a 3-layer fully connected network. The input layer dimension is the dimension of the associated parameters of the missing data, such as temperature, humidity, wear, and other related parameters. The activation function of the hidden layer is ReLU, and the output layer is the target parameters after completion. The discriminator also employs a 3-layer fully connected network with the activation function Sigmoid. During training, the normal data distribution in the knowledge graph is used as a constraint, the number of iterations is set to 500, and the learning rate is 0.001 to ensure that the completed data conforms to the PVA degradation pattern.

[0015] The core-specific correlation rules have been supplemented: when the cleaning fluid concentration exceeds 15%, the PVA hydrolysis degradation rate increases by 3 times; when the wear on the sponge roller surface is >0.8mm, the wafer scratch defect rate increases by 40%; when the ambient humidity is >85% and the temperature is >55℃, the thermal oxidation degradation is accelerated by 2.5 times; when the spray pressure fluctuation range exceeds 0.3MPa, the local wear deviation of the sponge roller is >0.5mm; when the wafer surface particle defects are concentrated in the edge area, the probability of the wear at both ends of the corresponding sponge roller exceeding the standard reaches 80%. These 10 core correlation rules cover the main fault and parameter correlation scenarios.

[0016] Preferably, the PINN prediction module is based on the standardized data output by the knowledge graph processing module. It constructs a physical information neural network prediction model that integrates the PVA-specific degradation mechanism. This model embeds the PVA sponge roller-specific degradation physical equation into the neural network loss function. The model prediction results are guaranteed by physical constraints. At the same time, a graph neural network is used to extract the spatial correlation features of multi-source data, and a temporal convolutional network is combined to capture the long and short-term temporal features of the data. The two features are then fused and input into the PINN model to improve the prediction accuracy. Feature fusion method: The spatial correlation features extracted by the graph neural network and the long and short-term temporal features extracted by the temporal convolutional network adopt a dynamic weighted fusion strategy. The weights are adaptively learned during the model training process, and the initial weight ratio is set to 1:2. During fusion, the features are concatenated through a fully connected layer and then compressed in dimension, finally outputting a 256-dimensional fused feature to ensure effective coupling of the two features. SHAP value feature attribution output format: The output results include two parts. The first is a parameter contribution ranking table, which is arranged from largest to smallest in terms of absolute contribution value, and marks the positive / negative impact of each parameter on the lifetime prediction results. The second is a visual beehive diagram, which intuitively shows the contribution distribution of each parameter under different values. The core influencing factor screening criterion is an absolute contribution value of 0.1, to ensure the pertinence of process optimization decisions.

[0017] Graph Neural Network Node and Edge Definitions: Graph nodes include two categories. The first category is the key monitoring areas of the sponge roller, such as the middle, both ends, and high-wear areas of the roller body. The second category is the core monitoring parameters, including physical characteristics, process parameters, and environmental parameters. The edge weights represent the correlation strength between nodes. For example, the edge weights of the nodes in the middle area of ​​the sponge roller and the defect rate parameter node at the center of the wafer are calculated based on the correlation of historical data. The spatial correlation between nodes is quantified through the adjacency matrix to achieve the fusion of spatial features of multi-dimensional data.

[0018] Key parameters of the temporal convolutional network: the kernel size is set to 3-5, the dilation coefficients are increased in the order of 1, 2, 4 to adapt to the capture of long and short-term temporal features, the number of network layers is 4, the activation function is GELU, and layer normalization is used to avoid gradient vanishing and ensure accurate capture of the changing trend of degradation rate stages.

[0019] Incremental learning fine-tuning strategy: Fine-tuning is performed only on the output layer and the last two hidden layers of the PINN model, while freezing the parameters of the previous feature extraction network; the learning rate is initially set to 0.0005, decaying by 50% every 100 iterations, and fine-tuning stops when the model loss function converges to below 0.01; a fine-tuning is triggered when the amount of new data reaches 20% of the existing training data to ensure that the model dynamically adapts to process changes.

[0020] The model adds an extreme condition feature extraction layer and introduces the entropy change feature analysis approach related to molecular dynamics simulation. It optimizes the prediction performance under extreme conditions by using phonon entropy change feature sequences. The model adopts an incremental learning mechanism to achieve dynamic adaptive updates. When new data or significant changes in process parameters occur, there is no need to retrain. Optimization can be completed by simply fine-tuning the network parameters. At the same time, a feature attribution module based on SHAP values ​​is introduced to analyze the contribution of each parameter to the lifetime prediction results to identify core influencing factors such as temperature fluctuations and wear. The output remaining lifetime prediction results of the sponge roller are simultaneously sent to the digital twin calibration module for simulation verification and provide early warning judgment basis for the edge intelligent monitoring module.

[0021] Details of extreme operating condition entropy change feature processing: Phonon entropy change features were acquired using a laser Raman spectrometer with a detection wavelength of 532 nm and a scanning range of 200-3000 cm⁻¹. -1 The vibrational peak shift of PVA molecules in the spectrum is extracted as an entropy change correlation index; the feature sequence length is set to 100 consecutive sampling points, and the sampling interval is consistent with the acquisition module, 200Hz under extreme conditions, i.e. 5ms / point; the mean, variance, and trend slope of the sequence are calculated by the sliding window method to form a three-dimensional feature vector, which is input into the extreme condition feature extraction layer to improve the model's sensitivity to extreme scenarios.

[0022] Preferably, the digital twin calibration module constructs a full-element digital twin model encompassing microstructure, macrostate, and process. This full-element digital twin model is built upon the raw data from the PVA multi-source acquisition module and standardized data processed by the knowledge graph. It includes core elements such as the microstructure of the PVA sponge roller, macroscopic operating status, cleaning equipment operating status, and wafer surface characteristics, intuitively displaying the real-time status of the sponge roller. The digital twin calibration module also possesses multi-scenario simulation verification capabilities, including visualization verification of lifetime prediction results, extreme condition simulation, and microscopic degradation evolution simulation. Full-element digital twin modeling tools and data interaction interfaces: A macroscopic scene (cleaning equipment, sponge roller, wafer) visualization model was built using Unity3D, and a microscopic structure (pores, fibers) finite element simulation model was built using ANSYS. Data linkage between the two models was achieved through Python scripts. The MQTT protocol is used to realize real-time communication between digital twins and physical entities and prediction models. The interface fields include entity ID, parameter type, parameter value, timestamp, and calibration status. The data transmission latency is controlled within 10ms. Microscopic visualization uses volume rendering technology, setting different color gradients according to porosity and crystallinity distribution; porosity >80% is blue, 60%-80% is green, 40%-60% is yellow, and <40% is red; crystallinity >60% is dark, 30%-60% is medium, and <30% is light.

[0023] Microstructure modeling details: The finite element method is used for microscopic modeling, with a mesh size of 10m x 10m x 10m to match the pore scale. The pore distribution is digitally mapped based on the actual microscopic images of the sponge roller obtained by CT scan. The fiber strength parameters are fitted by the material tensile test data. The microscopic degradation evolution simulation adopts a coupled molecular dynamics and finite element method to simulate the effect of molecular chain breakage on macroscopic compressive stress and porosity. The simulation step size is set to 1e-12 seconds.

[0024] Two-way closed-loop calibration interaction frequency: The data interaction cycle between the digital twin and the physical entity is 10 seconds / time (synchronizing real-time status data). The parameter calibration cycle between the digital twin and the PINN prediction model is after one batch of cleaning tasks is completed, that is, after each batch of wafer cleaning is completed, the physical constraint parameters of the model are calibrated based on the actual wear and degradation data. Under extreme conditions such as temperature exceeding the process threshold, the interaction and calibration frequency is increased to 5 seconds / time to ensure that the model can adapt to abnormal scenarios in a timely manner.

[0025] A two-way closed-loop calibration mechanism is established for physical entities, digital twins, and prediction models. The model parameters are dynamically calibrated through real-time data interaction between the digital twin and the physical entity to improve simulation accuracy. At the same time, the microscopic degradation simulation results of the digital twin are fed back to the PINN prediction module to optimize the physical constraint parameters of the model. The cross-condition degradation law data generated by the digital twin model are synchronously sent to the federated learning sharing module to achieve cross-plant reuse.

[0026] Two-way closed-loop calibration process: Under normal operating conditions, the trigger is activated every 10 seconds; under extreme operating conditions (parameters exceeding process thresholds), the trigger is activated every 5 seconds. Upload real-time status data (wear, temperature, defect rate, etc.) and the digital twin model updates the virtual entity's status; Based on the comparison between the degradation trend of virtual simulation and the actual trend of physical entity, the deviation value is calculated. When the deviation is >8%, calibration is triggered to adjust the physical constraint parameters of PINN model. The optimized prediction results are fed back to update the simulation boundary conditions of the digital twin, forming a closed loop; Cross-condition degradation data format: The data is stored in tabular form, including four core fields: condition dimension, sponge roller formulation type, degradation index, and confidence level. Each condition contains more than 1,000 sets of sample data to ensure the effectiveness of cross-plant reuse. The condition dimension includes temperature, humidity, cleaning fluid concentration, and wafer process. The degradation index includes wear rate, crystallinity change rate, and remaining life.

[0027] Preferably, the federated learning sharing module takes cross-plant data from the PVA multi-source acquisition module, standardized knowledge from the knowledge graph processing module, and simulation patterns from the digital twin calibration module as core inputs, and uses an improved federated averaging algorithm to construct a cloud-edge two-level federated learning framework. This framework first clusters the plant nodes participating in federated learning according to the PVA sponge roller formulation type, and prioritizes parameter aggregation within the same formulation cluster. Details of the improved federated average algorithm: To address the slow convergence and poor adaptation to heterogeneous data issues of the traditional federated average algorithm, two improvements have been added: ① Gradient filtering mechanism: Outlier detection is performed on the parameter gradients uploaded by each factory. The three principles are adopted to remove gradient data that exceed three times the standard deviation of the mean, so as to avoid outlier data from interfering with the aggregation results. ② Adaptive learning rate: The learning rate is dynamically adjusted according to the proportion of data in each factory. The learning rate coefficient is set to 1.2 for factories with a data proportion >30%, 1.0 for 10%-30%, and 0.8 for <10%, which accelerates model convergence. The weight calculation cycle of gradient weighted average during aggregation is one training round, and each round contains 1000 training samples.

[0028] Each factory, acting as an edge node, trains model parameters locally and only uploads the parameter gradients to the cloud server for aggregation and updates. There is no need to share the original production data to ensure data privacy and security. At the same time, the private process knowledge of each factory is represented in a federated manner through a PVA-specific knowledge graph. The framework also supports the dynamic addition and removal of factory nodes, ensures the security of federated learning through a permission management mechanism, and adopts a weighted aggregation strategy for the differences in data distribution among different factories. The optimized model parameters and shared knowledge generated by federated learning will be back-synchronized to the knowledge graph processing module and PINN prediction module of each plant.

[0029] Preferably, the edge intelligent monitoring module takes the remaining lifespan result output by the PINN prediction module, the real-time status data of the PVA multi-source acquisition module, and the simulation early warning information of the digital twin calibration module as inputs, and deploys the lightweight PINN lifespan prediction model to the edge computing node to realize real-time inference of the state of the sponge roller during the cleaning process. Based on lifetime prediction results, real-time status data, and PVA degradation thresholds, a three-level early warning mechanism is set up, including warning, alarm, and emergency shutdown. In the early warning stage, warning information and key influencing factors are pushed out. In the alarm stage, audible and visual alarms are triggered and maintenance solutions such as reducing cleaning load and regular ultrasonic cleaning are automatically recommended. In the emergency shutdown stage, a shutdown signal is automatically sent to the cleaning equipment, and a maintenance work order is generated in the MES system. It supports seamless integration with the control system of wafer cleaning equipment and the factory MES system to achieve automated linkage between early warning and disposal. Linkage communication protocol and early warning push method: It interfaces with the cleaning equipment control system using the Modbus TCP protocol to transmit shutdown signals and parameter adjustment instructions; it interacts with the MES system using a RESTful API interface, and the work order data format follows the factory's MES system standard, including fields such as warning level, sponge roller ID, equipment number, suggested handling method, and deadline for handling. During the early warning phase, notifications are pushed through the local display screen of the equipment and pop-up windows of the factory production management platform; during the alarm phase, SMS notifications are added and pushed to the mobile phone number of the equipment maintenance personnel; during the emergency shutdown phase, the factory workshop's audible and visual alarms are triggered simultaneously, with an alarm volume of 85dB and a light flashing frequency of 2Hz, and the emergency work order is displayed at the top of the MES system.

[0030] The data on the effectiveness of early warning and response recorded by the edge intelligent monitoring module will be fed back to the PVA multi-source acquisition module.

[0031] The beneficial effects of this invention are as follows: 1. This invention constructs a four-dimensional, all-dimensional acquisition system for PVA characteristics, process parameters, cleaning effects, and environmental conditions through a PVA multi-source acquisition module. This system not only covers the physical properties, degradation characteristics, and formulation information of the sponge roller, but also simultaneously acquires process operation data, environmental adaptation data, and wafer cleaning quality feedback data. It adds correlation data between defect types and sponge roller wear locations, and completely establishes the correlation chain between sponge roller status, process parameters, and cleaning effects. By utilizing edge node preprocessing, time stamp and process stage dual-dimensional time sequence alignment, and combined with an adaptive acquisition strategy, the acquisition frequency for high-difficulty cleaning tasks during the degradation period and below 7nm is increased to 200Hz, while the frequency for regular tasks is reduced to 50Hz. This ensures both data integrity and timeliness, while effectively reducing cloud transmission pressure and resource consumption.

[0032] 2. The PINN prediction module of this invention embeds the PVA-specific degradation physical equation into the model loss function, using physical constraints to ensure that the prediction results conform to the actual degradation law. Simultaneously, it combines graph neural networks to extract spatial correlation features from multi-source data, and temporal convolutional networks to capture long-term and short-term temporal features, improving prediction accuracy through feature fusion. For extreme scenarios such as high temperature and high-concentration cleaning solutions, an extreme condition feature extraction layer is added, introducing entropy change feature analysis related to molecular dynamics simulation to optimize prediction performance under extreme conditions. The model adopts an incremental learning mechanism; when faced with new data or significant changes in process parameters, it does not require retraining, but only fine-tuning the network parameters to complete the optimization.

[0033] 3. The federated learning sharing module of this invention constructs a cloud-edge two-level framework, clustering and aggregating parameters according to the sponge roller formula. Each factory only uploads the parameter gradient, without sharing the original production data. Under the premise of ensuring data privacy, it realizes cross-factory knowledge security sharing and model collaborative optimization, improving model robustness. The edge intelligent monitoring module deploys the lightweight PINN model to the edge computing node to realize real-time inference of the sponge roller status. Based on the remaining life result, real-time status data and PVA degradation threshold, it establishes a three-level mechanism of early warning, alarm, and emergency shutdown. In the early warning stage, it pushes information and core influencing factors. In the alarm stage, it triggers audible and visual alarms and recommends maintenance solutions. In the emergency shutdown stage, it automatically sends a shutdown signal and generates a maintenance work order. It also seamlessly connects with the equipment control system and MES system to realize automated linkage. Attached Figure Description

[0034] Figure 1 This is a flowchart of the wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers, as described in this invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] like Figure 1 As shown, this embodiment of the invention provides a wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers, including: The PVA multi-source acquisition module adopts a four-dimensional real-time acquisition architecture that includes PVA characteristics, process parameters, cleaning effect, and environmental conditions. This breaks through the limitations of traditional methods that only collect equipment operating parameters and constructs a data acquisition system that combines specificity and completeness. Specifically, core physical property parameters such as porosity, water absorption rate, surface wear, and compressive stress are collected through an embedded micro-sensor array and combined with high-frequency ultrasonic detection technology to achieve non-contact wear monitoring. At the same time, new PVA degradation characteristic parameters are collected (including the degree of molecular chain breakage detected by near-infrared spectroscopy and changes in crystallinity detected by X-ray diffraction) and sponge roller formulation information (such as crosslinking density and filler ratio) are recorded synchronously. Real-time operating parameters such as equipment speed, cleaning fluid temperature and concentration, spray pressure, and cleaning time are collected simultaneously, as well as environmental parameters such as humidity and oxygen concentration in the cleaning chamber to meet the needs of PVA thermal oxidation degradation kinetic analysis. The number of wafer surface defects and cleanliness level are collected through machine vision and combined with offline detection data from semiconductor testing equipment to add correlation data between defect type and sponge roller wear location.

[0037] In addition, the module deploys edge computing nodes on the cleaning equipment to perform real-time preprocessing (denoising, downsampling, and format standardization) on high-frequency sensor data (such as 100Hz wear data) to reduce cloud transmission pressure. It adopts a dual-dimensional synchronization mechanism of timestamp and process stage to achieve accurate time-series alignment of multi-source data. Based on the PVA degradation stage (initial, stable, and accelerated degradation) and wafer process type, it designs an adaptive acquisition strategy. During the accelerated degradation period and high-difficulty cleaning tasks below 7nm, the acquisition frequency is increased to 200Hz, while it is reduced to 50Hz for regular tasks, thus optimizing resource consumption while ensuring data integrity.

[0038] The timing alignment is implemented as follows: the timestamp adopts the UTC time format with millisecond precision, and the clock deviation of each sensor is ensured to be 1ms through the equipment clock synchronization protocol; the process is divided into 6 stages: material preparation, pre-cleaning, main cleaning, rinsing, drying, and unloading. Each stage is triggered by the equipment PLC signal to write the stage identifier. The timing alignment of multi-source data is completed by matching the timestamp 1ms and binding the process stage identifier, ensuring accurate association of data under the same process action.

[0039] To fully leverage the value of the PVA multi-source acquisition module's output data, eliminate noise interference, and achieve data standardization, the knowledge graph processing module overcomes the limitations of general semiconductor knowledge graphs by constructing a dedicated knowledge graph for PVA sponge roller-wafer cleaning. This dedicated knowledge graph integrates four core knowledge bases: a PVA material properties and degradation mechanism library containing data on the thermal oxidation and hydrolytic degradation laws of sponge rollers with different formulations and the correlation data of degradation rate constants based on the Arrhenius equation; a cleaning process specification library containing standard parameter ranges and extreme operating condition thresholds corresponding to each process; a fault case library containing the correlation between wear anomalies and defect types, and the correspondence between temperature and humidity fluctuations and degradation acceleration; and a data quality rule library containing abnormal data judgment standards based on PVA degradation kinetics.

[0040] Based on this dedicated knowledge graph, the module can realize three core functions: intelligent noise identification and removal, cross-device data semantic standardization, and multi-dimensional data quality assessment and repair. Among them, intelligent noise identification and removal accurately identifies abnormal data caused by sensor failure and environmental interference through dedicated association rules in the graph, such as abnormal fluctuations in spray pressure leading to increased local wear of sponge rollers and temperatures exceeding 60°C leading to a doubling of PVA degradation rate. Cross-device data semantic standardization, based on the ontology matching mechanism of the knowledge graph, unifies synonymous parameters such as rotation speed and linear speed of different devices into cleaning roller linear speed and establishes a unified dimensional standard for PVA characteristic parameters.

[0041] Multi-dimensional data quality assessment and repair uses evaluation indicators for completeness, consistency, accuracy, and timeliness, combined with the PVA degradation process threshold in the knowledge graph to evaluate the data quality level in real time. For low-quality data, a missing value completion method based on generative adversarial networks is used to complete the missing values, and finally output high-quality standardized data.

[0042] The PINN prediction module is based on high-quality standardized data output by the knowledge graph processing module. By constructing a Physical Information Neural Network (PINN) prediction model that integrates the PVA-specific degradation mechanism, the PINN prediction module overcomes the problem of insufficient generalization ability of existing general PINN models. This model embeds the PVA sponge roller-specific degradation physical equations (including the Arrhenius equation for thermal oxidation degradation and the first-order kinetic equation for hydrolytic degradation) into the neural network loss function. Through physical constraints, it ensures that the model prediction results conform to the actual degradation law, thereby improving the cross-scenario generalization ability. At the same time, it uses graph neural networks to extract spatial correlation features of multi-source data, such as the spatial correspondence between sponge roller wear and wafer defect rate, and the correlation features between PVA crystallinity and compressive stress. It combines temporal convolutional networks to capture the long-term and short-term temporal features of the data, such as the stage change trend of degradation rate. The two features are then fused and input into the PINN model to improve the prediction accuracy.

[0043] For extreme scenarios such as high temperature and high concentration of cleaning fluid, the model adds an extreme condition feature extraction layer and introduces the entropy change feature analysis approach related to molecular dynamics simulation. The prediction performance under extreme conditions is optimized by using phonon entropy change feature sequences. In addition, the model adopts an incremental learning mechanism to achieve dynamic adaptive updates. When new data or significant changes in process parameters are added, there is no need to retrain. Optimization can be completed by simply fine-tuning the network parameters. At the same time, a feature attribution module based on SHAP value is introduced to analyze the contribution of each parameter to the lifetime prediction results to clarify the core influencing factors such as temperature fluctuation and wear, and solve the problem of insufficient model interpretability.

[0044] The predicted remaining life of the sponge roller will be simultaneously transmitted to the digital twin calibration module for simulation verification, and will also provide early warning judgment basis for the edge intelligent monitoring module.

[0045] To verify the accuracy of the PINN prediction module's output and further optimize prediction precision, the digital twin calibration module overcomes the limitations of existing digital twins that only focus on macroscopic states by constructing a full-element digital twin model encompassing microstructure, macroscopic state, and process. This full-element digital twin model is built based on the raw data from the PVA multi-source acquisition module and standardized data processed from the knowledge graph. It includes core elements such as the microstructure of the PVA sponge roller (pore distribution, fiber strength, crystallinity distribution), macroscopic operating state (wear distribution, degradation degree), cleaning equipment operating state, and wafer surface characteristics, which can intuitively display the real-time state of the sponge roller. Simultaneously, the module has multi-scenario simulation verification functions, specifically including visualization verification of lifetime prediction results (comparing the wear and degradation data of the virtual model with that of the physical entity), extreme condition simulation (simulating the degradation process of the sponge roller under abnormally high temperatures, high concentrations of cleaning fluid, high humidity, etc., and extrapolating the impact of extreme conditions on the remaining lifetime to provide boundary constraints for process parameter optimization), and microscopic degradation evolution simulation. Based on the PVA degradation kinetics model, it simulates the impact of molecular chain breakage and crystallinity changes on macroscopic performance.

[0046] In addition, the module establishes a two-way closed-loop calibration mechanism for physical entities, digital twins, and prediction models. It dynamically calibrates model parameters through real-time data interaction between digital twins and physical entities to improve simulation accuracy. At the same time, it feeds back the microscopic degradation simulation results of the digital twins to the PINN prediction module to optimize the physical constraint parameters of the model and further improve prediction accuracy. Meanwhile, the cross-condition degradation law data generated by the digital twin model can be used as core knowledge and synchronously transmitted to the federated learning sharing module to achieve cross-plant reuse.

[0047] To integrate the experience of PVA sponge roller life prediction from multiple factories and improve the adaptability and robustness of models from different factories, the federated learning sharing module takes cross-factory data from the PVA multi-source acquisition module, standardized knowledge from the knowledge graph processing module, and simulation rules from the digital twin calibration module as core inputs. It uses an improved federated averaging algorithm to construct a cloud-edge two-level federated learning framework. This framework first clusters the factory nodes participating in federated learning according to the PVA sponge roller formulation type (such as high crystallinity, low crystallinity, and nanofiller modification). Within the same formulation cluster, parameter aggregation is prioritized to avoid model bias caused by interference from different formulation data.

[0048] Specific steps for node clustering: The core characteristics of the sponge roller formulation were extracted, including three key parameters: crosslinking density, filler ratio, and molecular weight distribution. Standardize the features and unify their dimensions; K-means clustering was used, and the number of clusters K was determined by the elbow method. Based on the variance contribution of each factory's formula data, K was set to a value of 3-5. Factory nodes with 85% similarity in formula features are grouped into the same cluster to ensure consistent data distribution within the same cluster. The parameter gradient data is encrypted using the AES-256 symmetric encryption algorithm. The key is generated by the cloud server and each factory node through negotiation using an asymmetric encryption algorithm, and the key is automatically updated every 24 hours. The uploaded data is transmitted in fragments, and the fragment verification is performed through hash value verification to ensure the security and integrity of data transmission.

[0049] Each factory, acting as an edge node, trains its model parameters locally, uploading only the parameter gradients to the cloud server for aggregation and updates. This eliminates the need to share raw production data, ensuring data privacy and security. Simultaneously, each factory's private process knowledge (such as customized cleaning experience and formula-specific degradation patterns) is represented in a federated manner using a PVA-specific knowledge graph, enabling secure sharing and fusion of knowledge across factories. The framework also supports dynamic addition and removal of factory nodes, ensuring the security of federated learning through a permission management mechanism. Furthermore, it employs a weighted aggregation strategy to address the differences in data distribution across different factories, enhancing the adaptability of the aggregation model to various factory scenarios.

[0050] Weighted aggregation strategy weight calculation: The weight consists of three parts: data volume weight, accounting for 40%, calculated based on the number of valid data entries of participating nodes; data quality weight, accounting for 30%, calculated based on the completeness and accuracy scores of the knowledge graph evaluation; and model local performance weight, accounting for 30%, calculated based on the prediction error of the node's local model. Total weight = 0.4 data volume ratio + 0.3 data quality score + 0.3 (1 - local prediction error), ensuring that the weight allocation takes into account both data scale and quality.

[0051] Permissions are divided into three levels: administrator permissions (responsible for configuring the federated framework and approving nodes), participating node permissions (uploading parameter gradients and obtaining optimization models), and visitor permissions (only viewing the public part of the shared knowledge graph). To join a node, you need to submit materials such as factory qualifications and sponge roller formula type. After the administrator approves the application, participation permissions will be assigned. When a node leaves, its parameter cache in the cloud will be automatically deleted, while the common rules in the shared knowledge graph will be retained to protect data privacy.

[0052] Ultimately, the optimized model parameters and shared knowledge generated by federated learning will be back-synchronized to the knowledge graph processing module and PINN prediction module of each plant.

[0053] The edge intelligent monitoring module takes the remaining lifespan result output by the PINN prediction module, the real-time status data of the PVA multi-source acquisition module, and the simulation early warning information of the digital twin calibration module as inputs. It deploys the lightweight PINN lifespan prediction model (including the interpretability module) to the edge computing node to realize real-time inference of the state of the sponge roller during the cleaning process. The inference delay is 50ms, avoiding the problem of untimely early warning caused by cloud transmission delay.

[0054] The lightweight PINN model implementation method is as follows: It adopts a combination of model pruning (pruning ratio of 40%, removing connections with absolute weight values ​​less than 0.001) and INT8 quantization (converting 32-bit floating-point parameters into 8-bit integers) to reduce the amount of inference computation. It also reduces the inference computation by using operator fusion technology (merging convolution, activation, and normalization operations into a single operator). During deployment, it uses the TensorRT framework for optimization to ensure an inference latency of 50ms.

[0055] Specific thresholds for Level 3 early warning: ① Warning stage: 10 batches of remaining service life, or a single parameter exceeds the process standard range by 5%, such as temperature exceeding the set value by 3℃ or cumulative wear reaching 0.3mm; ② Alarm stage: 5 batches of wafers remain on the lifespan, or local wear > 0.6 mm, or wafer defect rate increases by 30% compared to the average; ③ Emergency shutdown phase: 2 batches with remaining lifespan, or wear > 0.8 mm, or a sudden increase of more than 2 times in thermal oxidation degradation rate, or a predicted wafer defect rate of 5%, with a risk of batch defects; The wear threshold of sponge rollers with different formulations can be adjusted according to the formulation type. The threshold is reduced by 10% for high crystallinity formulations and increased by 15% for nanofiller modified formulations.

[0056] The module sets up a three-level early warning mechanism—early warning, alarm, and emergency shutdown—based on lifetime prediction results, real-time status data, and PVA degradation thresholds. In the early warning stage (less than 10 batches remaining), it pushes warning information and key influencing factors, such as accelerated degradation due to high temperature. In the alarm stage (less than 5 batches remaining or excessive local wear), it triggers audible and visual alarms and automatically recommends maintenance solutions such as reducing cleaning load and regular ultrasonic cleaning. In the emergency shutdown stage (extreme degradation risk or potential wafer batch defects), it automatically sends a shutdown signal to the cleaning equipment and generates a maintenance work order in the MES system. The module supports seamless integration with the wafer cleaning equipment's control system and the factory's MES system, achieving automated linkage between early warning and response, reducing manual intervention to ensure stable operation of the cleaning process.

[0057] Maintenance plan details: Reduce the equipment speed by 10%-20%, extend the cleaning time by 5%-10%, and keep the spray pressure unchanged; Perform this process once after every two batches of cleaning, with ultrasonic power set to 300-500W, cleaning time 5-8 minutes, cleaning medium being deionized water, and temperature 25-30℃. Prioritize replacing the local components of the sponge roller corresponding to the worn area, such as detachable roller sleeves, instead of replacing the entire roller; The handling effect record includes eight fields: warning level, handling time, handling method, pre-handling parameters (wear amount, temperature, defect rate), post-handling parameters (same dimension), wafer batch pass rate, and maintenance time. The recorded data is automatically associated with the sponge roller ID and equipment number, serving as the core basis for model iteration and data acquisition strategy optimization.

[0058] In addition, the early warning and response effect data recorded by the module will be fed back to the PVA multi-source acquisition module as supplementary data for subsequent data acquisition strategy optimization and model iteration.

[0059] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers, characterized in that, include: PVA Multi-Source Acquisition Module: Constructs a four-dimensional real-time acquisition system for PVA characteristics, process parameters, cleaning effect, and environmental conditions. It collects sponge roller body characteristics, degradation features, formula information, process operation, cleaning feedback, and environmental adaptation data. After edge node preprocessing and time sequence alignment, it outputs full-dimensional raw data in combination with an adaptive acquisition strategy. Knowledge graph processing module: Constructs a dedicated knowledge graph for PVA sponge roller-wafer cleaning, integrating core knowledge bases such as material degradation mechanism, process specifications, fault cases and data quality rules, to achieve abnormal data identification, cross-device data semantic standardization and data quality optimization, and output standardized data; PINN Prediction Module: Based on standardized data, a physical information neural network prediction model is constructed that integrates the degradation mechanism of PVA. Physical equations are embedded to form physical constraints. Multiple networks are integrated to extract spatiotemporal correlation features. Extreme working condition adaptation, incremental learning and feature attribution mechanisms are added to output the remaining life of the sponge roller. Digital twin calibration module: Constructs a full-element digital twin model of microstructure, macrostate and process, realizes lifetime prediction verification, extreme condition simulation and micro-degradation simulation, optimizes prediction model parameters through a two-way closed-loop calibration mechanism, and generates degradation law data across operating conditions; Federated Learning Sharing Module: Taking cross-plant data, standardized knowledge, and simulation rules as input, it constructs a cloud-edge two-level federated learning framework, clusters and aggregates parameters according to sponge roller formula, and realizes cross-plant knowledge security sharing and model collaborative optimization. Edge intelligent monitoring module: Deploys lightweight predictive models to achieve real-time inference, builds a three-level early warning mechanism, links devices and MES system to achieve automated handling, and records handling effect data for reverse feedback.

2. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 1, characterized in that, The four-dimensional acquisition system of the PVA multi-source acquisition module includes the core physical properties, degradation characteristics, and formulation information of the sponge roller, equipment operating parameters, wafer cleaning effect data, and environmental parameters adapted to PVA degradation analysis. Edge computing nodes are deployed at the cleaning equipment end for data preprocessing, and a time-series alignment is achieved by using a dual-dimensional synchronization mechanism of timestamps and process stages. An adaptive acquisition strategy is designed based on the PVA degradation stage and wafer process type, and the acquisition frequency is adjusted differentially to balance data integrity and resource consumption.

3. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 2, characterized in that, The dedicated knowledge graph processing module integrates four core knowledge bases: a PVA material properties and degradation mechanism database, a cleaning process specification database, a fault case database, and a data quality rule database. Based on this dedicated knowledge graph, it realizes intelligent noise identification and removal, cross-device data semantic standardization, and multi-dimensional data quality assessment and repair functions. It identifies abnormal data through dedicated association rules, unifies parameter semantics and dimensions based on ontology matching mechanism, and assesses data quality and repairs low-quality data by combining PVA degradation process thresholds.

4. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 3, characterized in that, The PINN prediction module embeds the PVA-specific degradation physical equation into the neural network loss function to form physical constraints. It uses graph neural networks and temporal convolutional networks to extract spatial correlation features and long-term and short-term temporal features respectively and fuses them into the input model. An extreme condition feature extraction layer is added to optimize the prediction performance in extreme scenarios. An incremental learning mechanism is used to realize the dynamic adaptive update of the model. A feature attribution module is introduced to analyze the contribution of each parameter to the lifetime prediction result. The output remaining lifetime result is synchronously sent to the digital twin calibration module and the edge intelligent monitoring module.

5. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 4, characterized in that, The full-element digital twin model of the digital twin calibration module is built based on raw and standardized data, covering core elements such as PVA sponge rollers, cleaning equipment, and wafer surface. It has functions such as lifetime prediction visualization verification, extreme working condition simulation, and micro-degradation evolution simulation. It establishes a two-way closed-loop calibration mechanism of physical entity, digital twin, and prediction model, dynamically optimizes model parameters, and sends the generated cross-working condition degradation law data to the federated learning sharing module.

6. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 5, characterized in that, The federated learning sharing module adopts a two-level federated learning framework, clustering factory nodes according to PVA sponge roller formulation type, and prioritizing parameter aggregation for clusters with the same formulation; each factory node trains the model locally, only uploading parameter gradients to the cloud for aggregation and updates, realizing the federated representation of private process knowledge through knowledge graph, supporting dynamic node addition and removal and weighted aggregation strategies, and optimizing model parameters to be synchronized to the knowledge graph processing module and PINN prediction module.

7. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 6, characterized in that, The edge intelligent monitoring module deploys a lightweight PINN model on edge nodes to achieve real-time inference. Based on the remaining lifetime result, real-time status data, and PVA degradation threshold, it sets a three-level early warning mechanism, links the cleaning equipment control system and MES system to achieve automated early warning and handling, and the recorded handling effect data is fed back to the PVA multi-source acquisition module.

8. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 7, characterized in that, The number of embedded micro-sensor arrays deployed is adapted to the length of the sponge roller: 3 sets are deployed for lengths of 500mm and 5 sets are deployed for lengths greater than 500mm.

9. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 8, characterized in that, The dedicated knowledge graph contains core association rules covering major fault and parameter correlation scenarios.

10. The wafer cleaning big data analysis platform for predicting the lifespan of PVA sponge rollers according to claim 9, characterized in that, The three-tiered early warning mechanism adapts to different push methods, including display screens, SMS messages, and audible and visual alarms, based on the early warning level.

Images