Hybrid driving quality prediction method for wafer cutting whole process

Abstract

This invention relates to the field of quality prediction and control technology for photovoltaic silicon wafer cutting processes, specifically a hybrid-driven quality prediction method for the entire silicon wafer cutting process. The method includes collecting and preprocessing process parameters, silicon rod properties, environmental parameters, and quality data during the silicon wafer cutting process; constructing a mechanistic simulation model to output a quality benchmark prediction value; constructing an attention-enhanced spatiotemporal network residual correction model to output a residual compensation value; and then adaptively weighted fusion to obtain the final quality prediction result, which is then input into a digital twin closed-loop system. This invention, by setting up data acquisition and preprocessing, hybrid modeling, and digital twin closed-loop optimization mechanisms, integrates the mechanistic simulation model and the residual correction model to achieve real-time prediction of silicon wafer cutting quality. This reduces the impact of operating condition fluctuations on the prediction results and can be used for parameter optimization, anomaly early warning, and closed-loop control, thereby reducing reliance on manual labor and improving production stability and automation levels.

Description

Technical Field

[0001] This invention relates to the field of quality prediction and control technology for photovoltaic silicon wafer cutting process, and in particular to a hybrid-driven quality prediction method for the entire silicon wafer cutting process. Background Technology

[0002] With the photovoltaic industry's continuous pursuit of improved cell conversion efficiency and reduced manufacturing costs, silicon wafer manufacturing is developing towards larger sizes, thinner wafers, and lower losses. Diamond wire cutting technology, due to its advantages of high cutting efficiency, low material loss, and high processing precision, has become the mainstream method for photovoltaic silicon wafer processing. This process uses the high-speed reciprocating motion of a steel wire to drive diamond abrasive grains to grind and cut the silicon rod. The cutting process involves various complex physical phenomena such as steel wire vibration, abrasive grain removal, silicon rod contact stress, cutting heat conduction, coolant flow, and brittle fracture of silicon material. It is a typical multi-physics coupled process involving mechanical, thermal, fluid, and material removal. Because various process parameters influence each other, and because the silicon wafer itself is thin and brittle, the cutting quality is prone to fluctuations.

[0003] In actual production, the factors affecting silicon wafer cutting quality mainly include process parameters such as diamond wire cutting speed, feed rate, wire tension, coolant temperature, coolant flow rate, and wire wear; silicon rod property parameters such as silicon rod diameter, length, resistivity distribution, and crystal orientation; and environmental parameters such as ambient temperature and equipment vibration spectrum. These factors collectively influence key quality indicators such as wafer trace depth, total thickness variation, and fragmentation rate. Currently, many companies rely on manual experience to set and adjust process parameters and judge finished product quality through offline sampling inspection. This approach suffers from problems such as delayed inspection, untimely parameter adjustments, and difficulty in timely warnings of quality anomalies. When changes occur in silicon rod batches, equipment wear conditions, or environmental conditions, traditional experience-based control methods struggle to establish a dynamic correlation between parameters and quality in a timely manner.

[0004] Existing methods for predicting silicon wafer dicing quality mainly include traditional experimental design methods, pure data-driven models, pure mechanistic simulation models, and some digital twin solutions. Traditional orthogonal experiments and response surface methodology are difficult to fully utilize historical production data and cannot comprehensively characterize multi-parameter coupling relationships. While pure data-driven models possess nonlinear fitting capabilities, they have poor extrapolation capabilities and insufficient interpretability. Pure mechanistic models, although physically interpretable, are complex to model and computationally inefficient, making them difficult to meet real-time online prediction requirements. Existing digital twin solutions generally lack dynamic adaptation and closed-loop optimization capabilities across the entire silicon wafer dicing process. Therefore, a hybrid-driven quality prediction method that balances physical mechanisms, data fitting accuracy, real-time prediction capabilities, and operational condition adaptability is urgently needed. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a hybrid-driven quality prediction method for the entire silicon wafer dicing process. The technical solution is as follows: A hybrid-driven quality prediction method for the entire silicon wafer dicing process includes the following steps: S1. Deploy a multi-source heterogeneous data acquisition system on the diamond wire multi-wire cutting production line to simultaneously collect process parameter data, silicon rod attribute data, environmental parameter data and quality index data during the silicon wafer cutting process, and preprocess the collected data to form a standardized time series dataset. S2. Construct the first sub-model, which is a mechanism simulation model based on the physical mechanism of the cutting process. It is used to establish a benchmark mapping relationship between process parameters and cutting quality indicators based on the standardized time series dataset, and output the quality benchmark prediction value. S3. Construct a second sub-model, which is an attention-enhanced spatiotemporal network residual correction model. It is used to learn the residual mapping relationship of unmodeled dynamic error by taking the residual between the quality benchmark prediction value output by the first sub-model and the actual quality measurement value as the training target, and output the residual compensation value. S4. The quality benchmark prediction value output by the first sub-model and the residual compensation value output by the second sub-model are fused using an adaptive weighted fusion strategy to obtain the final prediction value of the silicon wafer cutting quality index. S5. Input the final predicted value into the digital twin closed-loop optimization system, and perform process parameter optimization, anomaly warning, defect tracing and control command output based on the final predicted value to realize online quality prediction and closed-loop control of the silicon wafer cutting process.

[0006] As a further aspect of the present invention, the process parameter data includes at least one of diamond wire speed, feed rate, wire tension, coolant temperature, coolant flow rate, wire wear, and coolant viscosity. The silicon rod attribute data includes at least one of silicon rod diameter, silicon rod length, silicon rod resistivity distribution, and crystal orientation; The environmental parameter data includes at least one of ambient temperature and vibration spectrum; The quality indicator data includes at least one of line mark depth, total thickness change (TTV), and fragmentation rate; The preprocessing includes missing value imputation, outlier removal, data standardization or normalization, and timestamp alignment.

[0007] As a further embodiment of the present invention, the first sub-model includes a wire mesh dynamics model, a silicon rod-steel wire contact mechanics model, and a thermo-fluid-solid multiphysics coupling model. The wire mesh dynamics model is used to calculate the transverse vibration amplitude and vibration frequency of the diamond wire during the cutting process, and to establish the mapping relationship between the transverse vibration amplitude and vibration frequency and the depth of the wire mark on the silicon wafer surface. The silicon rod-steel wire contact mechanics model is used to calculate the cutting force distribution and the internal stress field of the silicon wafer, and to assess the risk of silicon wafer breakage based on the maximum principal stress criterion and fracture mechanics model. The thermo-fluid-solid multiphysics coupling model is used to calculate the temperature distribution in the cutting area and couples the temperature field to the contact mechanics model to correct the influence of thermal expansion on stress distribution and cutting quality.

[0008] As a further aspect of the present invention, the wire network dynamics model is established based on the following dynamic equations: Where M is the net mass matrix, C is the damping matrix, K is the stiffness matrix, Ft is the external force vector acting on the net, and x t For wire mesh displacement response; The silicon rod-steel wire contact mechanical model is based on constitutive relations and follows: The internal stress field of the silicon wafer is calculated, and the fragmentation risk is calculated based on the Weibull fracture risk model. For stress, For the elasticity matrix, In response to the situation; The heat-fluid-solid multiphysics coupling model is based on the heat conduction equation: Calculate the temperature distribution in the cutting region, where, Where is the material density, c is the specific heat capacity, T is the temperature, k is the thermal conductivity, and q is the cutting heat source term.

[0009] As a further embodiment of the present invention, the second sub-model includes a stacked enhanced autoencoder layer, an attention mechanism module, a bidirectional long short-term memory network layer, and a fully connected regression layer connected in sequence. The stacked enhanced autoencoder layer is used to perform hierarchical feature extraction on multi-scale process parameters, silicon rod property parameters, environmental parameters, and the output value of the first sub-model. The attention mechanism module is used to dynamically weight the features of historical samples, the features of current samples, and the temporal features at different times; The bidirectional long short-term memory network layer is used to capture the forward and backward time dependencies in the process parameter sequence of the cutting process; The fully connected regression layer is used to map deep spatiotemporal features into residual compensation values.

[0010] As a further aspect of the present invention, the training objective of the second sub-model is: Where e is the residual, Y 真实 Y represents the actual measured mass. 机理 The quality baseline prediction value output by the first sub-model; The second sub-model learns the nonlinear mapping relationship between the multi-source input parameters and the residuals under the current operating conditions, and outputs a residual compensation value ê, which is used to compensate for the nonlinear error, time-varying error and equipment state drift error that the first sub-model failed to represent.

[0011] As a further aspect of the present invention, the second sub-model employs an on-the-fly learning fine-tuning strategy, specifically including: For each new sample to be predicted, calculate the similarity between the new sample and historical samples in the historical database; Select the most similar nearest neighbor historical samples to form a localized training sample set; The second sub-model is locally fine-tuned or its parameters are updated based on the localized training sample set. The fine-tuned second sub-model outputs the residual compensation value corresponding to the current sample to adapt to the operating condition drift under different silicon rod batches, different equipment wear conditions, and different environmental disturbance conditions.

[0012] As a further aspect of the present invention, the adaptive weighted fusion strategy dynamically adjusts the fusion weights based on the real-time prediction errors of the first sub-model and the second sub-model under the current operating conditions. When the real-time prediction error of the first sub-model is less than that of the second sub-model, the weight of the first sub-model is increased. When the second sub-model learns significant residual compensation information or the real-time prediction error of the first sub-model increases, the residual compensation weight of the second sub-model is increased. The final predicted value is obtained as follows: Among them, Y 预测 Y represents the final predicted value of the silicon wafer dicing quality index. 机理 ê is the quality baseline prediction value output by the first sub-model, ê is the residual compensation value output by the second sub-model, and α(t) is the residual compensation weight coefficient that dynamically changes with the model prediction error under the current operating conditions.

[0013] A hybrid-driven quality prediction system for the entire silicon wafer dicing process, the system comprising: The data acquisition module is used to simultaneously collect process parameter data, silicon rod property data, environmental parameter data, and quality index data on the diamond wire multi-wire cutting production line. The data preprocessing module is used to perform missing value imputation, outlier removal, data standardization or normalization, and timestamp alignment on the data collected by the data acquisition module to form a standardized time series dataset. The mechanism simulation module is used to construct the first sub-model based on the physical mechanism of the cutting process and output the quality benchmark prediction value; The residual correction module is used to construct the attention-enhanced spatiotemporal network residual correction model and output the residual compensation value based on the residual between the predicted quality benchmark value and the actual quality measurement value. An adaptive fusion module is used to dynamically adjust the fusion weights based on the real-time prediction accuracy of the mechanism simulation module and the residual correction module, and output the final predicted value of the silicon wafer cutting quality index. The digital twin optimization module is used to construct a virtual quality evaluator based on the final predicted value, and to perform real-time optimization of key process parameters, anomaly warning, defect tracing, and control command generation. The control execution module is used to send the optimization control instructions generated by the digital twin optimization module to the programmable logic control system to realize online self-optimization and closed-loop control of process parameters.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention organically combines multi-source heterogeneous data acquisition, physical mechanism modeling, residual correction modeling, and adaptive fusion prediction by setting up a data acquisition and preprocessing layer, a hybrid modeling layer, and a quality prediction output layer. This forms a hybrid-driven quality prediction scheme for the entire silicon wafer cutting process. It utilizes process parameters, silicon rod attribute parameters, environmental parameters, and quality inspection data from the production site, and provides a benchmark prediction with physical constraints through a mechanism simulation model. The residual correction model compensates for complex nonlinear factors and unmodeled dynamic errors, thereby balancing the interpretability, generalization ability, and prediction accuracy of the model.

[0015] 2. This invention, by setting up an instant learning fine-tuning mechanism, an adaptive weighted fusion mechanism, and a digital twin closed-loop optimization mechanism, enables the quality prediction model to dynamically adjust according to different silicon rod batches, equipment operating status, and environmental disturbances. The prediction results are then used for process parameter optimization, anomaly early warning, defect tracing, and control execution. This not only enhances the model's adaptability to different production conditions and reduces the impact of condition changes on prediction results, but also forms a closed-loop operation process from real-time data acquisition, quality prediction, optimization decision-making to control feedback. This reduces reliance on human experience, shortens the process debugging cycle, and improves the stability and automation level of the production process. Attached Figure Description

[0016] Figure 1 : Overall flowchart of the hybrid-driven quality prediction method of this invention; Figure 2 System architecture diagram for quality prediction and closed-loop optimization based on hybrid-driven digital twin; Figure 3 : Structural diagram of the first sub-model (mechanism simulation model); Figure 4 : Deep learning network architecture diagram of the second sub-model (residual correction model). Detailed Implementation

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. All equivalent substitutions, simple modifications, or combinations made based on the concept of the present invention should fall within the scope of protection of the present invention.

[0018] This invention provides a hybrid-driven quality prediction method for the entire silicon wafer dicing process. It simultaneously collects process parameters, silicon rod properties, environmental parameters, and quality index data on a diamond wire multi-wire dicing production line. A first sub-model based on physical mechanisms and a second sub-model based on attention-enhanced spatiotemporal networks are constructed. Through adaptive weighted fusion, the final predicted values ​​of quality indicators such as wafer trace depth, total thickness variation (TTV), and fragmentation rate are obtained. Furthermore, the prediction results are input into a digital twin closed-loop optimization system to achieve real-time optimization of process parameters, anomaly warning, defect tracing, and PLC control execution.

[0019] Example 1, see Figure 1The method of this invention generally includes steps such as data acquisition and preprocessing, mechanism simulation modeling, residual machine learning modeling, adaptive weighted fusion, quality index prediction output, and digital twin closed-loop optimization. Specifically, firstly, a multi-source heterogeneous data acquisition system acquires data of the entire silicon wafer cutting process, which is then preprocessed by edge computing nodes and input into a hybrid modeling layer. Subsequently, the first sub-model outputs the quality benchmark prediction value, and the second sub-model outputs the residual compensation value. Finally, the final quality prediction result is obtained through an adaptive weighted fusion strategy, and this result is used for parameter optimization and control command generation in the digital twin system. Implementation environment setup: In this embodiment, a mass production line for diamond wire multi-wire cutting of photovoltaic 160μm monocrystalline silicon wafers is selected as the implementation object. The production line includes a diamond wire multi-wire cutting machine, AOI online inspection equipment, downstream thickness inspection equipment, and a programmable logic control system (PLC). A multi-source heterogeneous data acquisition system is deployed on a diamond wire cutting machine. The acquisition system includes a high-precision tension sensor, a patch-type temperature sensor, a laser thickness gauge, an industrial high-definition vision camera, a wire loss detector, a vibration sensor, a coolant flow sensor, and an environmental temperature and humidity acquisition device. Among them, the high-precision tension sensor has a measurement accuracy of ±0.1N, the patch-type temperature sensor has a measurement accuracy of ±0.1℃, the laser thickness gauge has a measurement accuracy of ±0.01μm, and the industrial high-definition vision camera has 20 million pixels. Industrial control computers are configured as edge computing nodes on the production site. These computers are equipped with i7 processors and 32GB of RAM to perform real-time data access, data preprocessing, model inference calculations, and data interaction with the digital twin data platform. Meanwhile, a digital twin data platform and a digital twin simulation system are built on a server or local industrial control platform to achieve real-time mapping between the physical production line and the virtual model. The digital twin system communicates with the PLC control system to send optimized process parameter combinations to the production equipment for online adjustment. See Figure 2 The system of this invention can be divided into a physical production line layer, a digital twin layer, and an intelligent decision-making layer. The physical production line layer is used to complete the data acquisition and process parameter execution of the silicon wafer cutting process; the digital twin layer is used to construct a virtual model that is synchronously mapped with the physical production line and carries the hybrid drive quality prediction model; the intelligent decision-making layer is used to perform anomaly warning, defect tracing, process parameter optimization, and PLC control instruction generation based on the quality prediction results.

[0020] Data sample collection and preprocessing: In this embodiment, production line measured data were continuously collected for 30 days, resulting in a total of 150,000 production samples. The production samples covered silicon wafer cutting quality data under different diamond wire speeds, feed speeds, wire tension, coolant temperature, coolant flow rate, wire wear, silicon rod length, silicon rod diameter, silicon rod resistivity distribution, crystal orientation, ambient temperature, and vibration spectrum conditions. The process parameter data includes diamond wire speed, feed rate, wire tension, coolant temperature, coolant flow rate, wire wear, and coolant viscosity; silicon rod attribute data includes silicon rod diameter, silicon rod length, silicon rod resistivity distribution, and crystal orientation; environmental parameter data includes ambient temperature and vibration spectrum; and quality index data includes wire mark depth, total thickness variation (TTV), and fragmentation rate. The quality index data is provided by AOI online inspection equipment, laser thickness gauges, and downstream inspection equipment. During the specific data acquisition process, the diamond wire speed was controlled at 1100–1500 m / min, the feed rate was controlled at 0.8–1.4 mm / min, the wire tension was controlled at 10–16 N, the silicon rod length was 120–180 mm, the processing temperature was 20–45 ℃, and the data sampling frequency was 100 ms / time. Edge computing nodes preprocess the collected raw data, including: Linear interpolation is used to supplement missing values ​​caused by short-term sensor loss or intermittent data acquisition. The 3σ criterion is used to eliminate abnormal noise data and data points that deviate significantly from normal operating conditions from the sensor. The Z-score normalization algorithm is used to standardize the scale of process parameters, silicon rod property parameters and environmental parameters of different dimensions. Based on a unified timestamp alignment rule, data from different sampling frequencies and sources are timestamped to form a standardized time series dataset; After preprocessing, the standardized time series dataset is divided into training, validation, and test sets in an 8:1:1 ratio for subsequent first sub-model initialization, second sub-model training, adaptive fusion parameter calibration, and prediction performance verification.

[0021] Construction and implementation of the first sub-model: The first sub-model is a mechanism simulation model based on the physical mechanism of the cutting process. It is used to establish the benchmark mapping relationship between process parameters and cutting quality indicators, and output the quality benchmark prediction value. The first sub-model includes the wire mesh dynamics model, the silicon rod-steel wire contact mechanics model, and the thermo-fluid-solid multiphysics coupling model. See Figure 3The first sub-model takes process parameters, silicon rod property parameters and environmental parameters as inputs, and performs physical mechanism calculations through wire mesh dynamics model, silicon rod-steel wire contact mechanics model and thermal-fluid-solid multi-physics coupling model respectively to obtain intermediate physical quantities such as steel wire transverse vibration amplitude, vibration frequency, internal stress distribution of silicon wafer, and temperature distribution of cutting area. It further maps the wire mark depth, TTV and fragmentation rate to obtain quality benchmark prediction values. Specifically, the mechanism simulation model was constructed using ABAQUS finite element simulation software. The elastic modulus of the silicon wafer was set to 130 GPa and the Poisson's ratio to 0.28. The characteristic fracture strength and Weibull modulus of the silicon rod were calibrated based on measured data. In the wire mesh dynamics model, the wire mesh system formed by the diamond wire rewinding between the guide wheels is regarded as a multibody dynamics system, and the following dynamic equations are established: Where M is the wire mesh mass matrix, C is the damping matrix, K is the wire mesh stiffness matrix, F(t) is the external force vector acting on the wire mesh, x(t) is the displacement response of the wire mesh system, the damping matrix includes steel wire-silicon rod contact damping and air damping, the stiffness matrix considers the geometric stiffness effect generated by steel wire tension, and the external force vector mainly includes cutting force and coolant scouring force. By solving this dynamic equation, the transverse vibration amplitude and vibration frequency of the steel wire during the cutting process are obtained, and the mapping relationship between vibration parameters and the depth of wire marks on the silicon wafer surface is established. In the silicon rod-steel wire contact mechanics model, the contact mechanics relationship between the silicon rod and the diamond wire is established, and the internal stress field of the silicon wafer is calculated using the following constitutive relation: Where σ is stress, C is elastic matrix, and ε is strain. Combining the finite element numerical simulation method, the stress distribution inside the silicon wafer under different process parameters is calculated, high stress concentration areas are identified, and the risk of silicon wafer fragmentation is assessed based on the maximum principal stress criterion and the Weibull fracture risk model. In the thermo-fluid-solid multiphysics coupling model, a temperature field model is established that considers the heat generated by friction between diamond abrasive grains and silicon rods, the heat carried away by coolant, and the internal heat conduction of silicon wafer material. The temperature distribution in the cutting region is calculated based on the following heat conduction equation: Where ρ is the material density, c is the specific heat capacity, T is the temperature, k is the thermal conductivity, and q is the cutting heat source term. The temperature field calculation results are further coupled to the contact mechanics model to correct the influence of thermal expansion on stress distribution, line mark depth, TTV and fragmentation risk. Through the joint simulation of the above three physical mechanism sub-modules, the first sub-model outputs the silicon wafer cutting quality benchmark prediction value Y mechanism. This quality benchmark prediction value has a clear physical meaning and can reflect the quality change trend under ideal working conditions and the constraints of major physical factors. The second sub-model is the attention-enhanced spatiotemporal network residual correction model, which is used to learn the residual mapping relationship between the output value of the first sub-model and the actual quality measurement value, and output the residual compensation value. See Figure 4 The second sub-model consists of a stacked enhanced autoencoder layer, an attention mechanism module, a bidirectional long short-term memory network layer, and a fully connected regression layer connected in sequence. The stacked enhanced autoencoder layer is used to extract hierarchical features from multi-scale process parameters, silicon rod property parameters, environmental parameters, and the output values ​​of the first sub-model. The attention mechanism module is used to dynamically weight historical sample features, current sample features, and temporal features at different times. The bidirectional long short-term memory network layer is used to capture the forward and backward temporal dependencies in the process parameter sequence of the dicing process. The fully connected regression layer is used to map deep spatiotemporal features into residual compensation values. In this embodiment, the input to the second sub-model includes eight disturbance features such as vibration amplitude, equipment operating time, ambient humidity, and silicon rod resistivity distribution, as well as the mass benchmark prediction value Y mechanism output by the first sub-model. The training objective of the second sub-model is: Where e is the residual, Y 真实 Y represents the actual measured mass. 机理 The first sub-model outputs the quality baseline prediction value. The second sub-model learns the nonlinear mapping relationship between the multi-source input parameters and the residuals under the current operating conditions and outputs the residual compensation value ê, which is used to compensate for the nonlinear error, time-varying error and equipment state drift error that the first sub-model failed to fully represent. In terms of network structure, the stacked enhanced autoencoder layer has three hidden layers to extract process parameter features at different scales layer by layer, and alleviates information loss during the layer-by-layer feature extraction process through layer-by-layer reconstruction. The attention mechanism module adopts a soft attention strategy to dynamically allocate weights for different time steps, different parameter features, and the contribution of historical samples. The bidirectional long short-term memory network layer has two hidden layers, with each hidden layer having 24 neurons, to simultaneously extract historical time-series information and subsequent trend information, reflecting the cumulative effect on the line mark depth, TTV, and fragmentation rate during the cutting process. The fully connected regression layer maps the extracted deep spatiotemporal features to residual compensation values. During training, the Adam optimization algorithm was used to iteratively train the second sub-model. The number of iterations was set to 1000, and the learning rate was set to 0.001. When the mean residual of the validation set was lower than 0.2%, the training of the residual correction model was considered complete. To adapt to the drift of operating conditions under different silicon rod batches, different equipment wear states, and different environmental disturbance conditions, the second sub-model further embeds an instant learning fine-tuning strategy. For each new sample to be predicted, the system first calculates the similarity between the new sample and historical samples in the historical database, selects the 100 nearest neighbor samples with the highest similarity to form a localized training sample set, and performs local fine-tuning or parameter updates on the second sub-model based on this localized training sample set. In this way, the second sub-model can maintain good residual compensation capability under different production batches and different equipment states.

[0022] Adaptive weighted fusion and final quality prediction: After completing the construction of the first and second sub-models, this invention integrates the outputs of the two sub-models through an adaptive weighted fusion strategy to obtain the final quality prediction value. Specifically, the first sub-model outputs the quality baseline prediction value Y, the second sub-model outputs the residual compensation value ê, and the adaptive fusion module dynamically adjusts the fusion weights based on the real-time prediction errors of the first and second sub-models under the current operating conditions. The final prediction value is obtained as follows: Among them, Y 预测 Y represents the final predicted value of the silicon wafer dicing quality index. 机理 ê is the quality baseline prediction value output by the first sub-model, ê is the residual compensation value output by the second sub-model, and α(t) is the residual compensation weight coefficient that dynamically changes with the model prediction error under the current working condition. When the first sub-model has a small prediction error and high reliability of physical mechanism prediction under the current operating conditions, the residual compensation weight is reduced so that the final prediction value depends more on the output of the physical mechanism model; when equipment fluctuations, environmental disturbances or changes in silicon rod batches cause the prediction error of the first sub-model to increase, the residual compensation weight is increased so that the second sub-model can more fully compensate for the unmodeled dynamic errors. In this embodiment, the prediction error calculation logic and weight dynamic adjustment threshold are calibrated based on the validation set data under different operating conditions. Under stable production conditions, the weight of the mechanism model can be assigned a value of 0.7, and the weight of the residual model can be assigned a value of 0.3. When equipment fluctuations or environmental disturbances increase, the system can automatically adjust the weight of the mechanism model to 0.4 and the weight of the residual model to 0.6 to reduce the overall prediction error. Through the aforementioned adaptive fusion strategy, this invention can balance the interpretability of the mechanistic model and the nonlinear compensation capability of the residual correction model, thereby achieving stable prediction of line mark depth, TTV, and fragmentation rate under different working conditions.

[0023] Digital twin closed-loop optimization and control execution: See Figure 2This invention embeds a hybrid-driven quality prediction model into a digital twin system to construct a closed-loop optimization architecture consisting of a physical production line layer, a digital twin layer, and an intelligent decision-making layer. The digital twin system receives multi-source sensor data from the physical production line layer in real time and maps it to the virtual production line model. The hybrid drive quality prediction model serves as a virtual quality evaluator in the digital twin system, outputting real-time predicted values ​​of quality indicators such as silicon wafer trace depth, TTV, and breakage rate. It also outputs the contribution analysis and counterfactual analysis results of each key parameter to the quality indicators. When the predicted quality indicators exceed the preset threshold, the system automatically triggers an anomaly warning. For example, when the line mark depth, TTV, or fragmentation rate exceeds the set upper limit, the system generates an warning message and, combined with the parameter contribution analysis results, locates the cause of the anomaly. For line marks exceeding the standard, the focus can be on checking the vibration amplitude of the diamond wire, the tension of the steel wire, and the condition of the wire loss; for TTV anomalies, the focus can be on checking the feed rate, the condition of the coolant, and the clamping condition of the silicon rod; for fragmentation rate anomalies, the focus can be on checking the stress concentration area inside the silicon wafer, the temperature of the coolant, and the vibration condition of the equipment. In terms of parameter optimization, the digital twin system uses a hybrid-driven quality prediction model as a virtual quality evaluator and employs the particle swarm optimization algorithm (PSO) to optimize key process parameters in real time. The optimization targets include wire mark depth ≤ 0.5 μm, TTV ≤ 0.7 μm, and fragmentation rate ≤ 0.5%. Under the premise of meeting equipment safety constraints and process boundary conditions, the system selects the Pareto optimal parameter combination and pushes the optimal parameter combination to the PLC control system. After receiving the optimization command, the PLC control system adjusts the adjustable parameters such as diamond wire speed, feed rate, wire tension, and coolant flow rate online, thereby forming a real-time self-optimization closed loop of process parameters.

[0024] Production line deployment and trial operation: After completing model training, fusion module calibration, and digital twin system integration, the entire system will be deployed to the target diamond wire multi-wire cutting production line for trial operation, with a trial operation period of 15 days. During the trial operation, the system continuously collects production data and stores new production samples in the historical database. The second sub-model continuously updates the residual correction capability based on the real-time learning fine-tuning strategy to correct model deviations caused by equipment wear, silicon rod batch differences, and environmental disturbances. At the same time, the system continuously verifies the predicted values ​​of line mark depth, TTV and fragmentation rate based on actual detection results, and updates the error statistics and weight adjustment logic in the adaptive fusion module. During the anomaly warning debugging process, corresponding thresholds are set for different quality indicators. When the predicted quality indicator exceeds the set range, the system automatically triggers an early warning and outputs suggested parameters for adjustment. The system is then debugged by recording indicators such as prediction accuracy, response delay, parameter optimization effect, and early warning accuracy. In this embodiment, the system can achieve a single batch silicon wafer quality prediction delay of ≤200ms and a parameter optimization command issuance delay of ≤500ms, meeting the real-time online prediction and closed-loop control requirements in diamond wire multi-wire cutting production scenarios. After trial operation verification, the system can adapt to different silicon rod batches, different equipment wear states, and different environmental disturbance conditions, and is compatible with monocrystalline silicon wafer and polycrystalline silicon wafer cutting scenarios, realizing the joint prediction of multiple indicators such as wire mark depth, TTV, and fragmentation rate.

[0025] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hybrid-driven quality prediction method for the entire silicon wafer dicing process, characterized in that, Includes the following steps: S1. Deploy a multi-source heterogeneous data acquisition system on the diamond wire multi-wire cutting production line to simultaneously collect process parameter data, silicon rod attribute data, environmental parameter data and quality index data during the silicon wafer cutting process, and preprocess the collected data to form a standardized time series dataset. S2. Construct the first sub-model, which is a mechanism simulation model based on the physical mechanism of the cutting process. It is used to establish a benchmark mapping relationship between process parameters and cutting quality indicators based on the standardized time series dataset, and output the quality benchmark prediction value. S3. Construct a second sub-model, which is an attention-enhanced spatiotemporal network residual correction model. It is used to learn the residual mapping relationship of unmodeled dynamic error by taking the residual between the quality benchmark prediction value output by the first sub-model and the actual quality measurement value as the training target, and output the residual compensation value. S4. The quality benchmark prediction value output by the first sub-model and the residual compensation value output by the second sub-model are fused using an adaptive weighted fusion strategy to obtain the final prediction value of the silicon wafer cutting quality index. S5. Input the final predicted value into the digital twin closed-loop optimization system, and perform process parameter optimization, anomaly warning, defect tracing and control command output based on the final predicted value to realize online quality prediction and closed-loop control of the silicon wafer cutting process.

2. The hybrid-driven quality prediction method for the entire silicon wafer dicing process according to claim 1, characterized in that, The process parameter data includes at least one of the following: diamond wire speed, feed rate, wire tension, coolant temperature, coolant flow rate, wire wear, and coolant viscosity. The silicon rod attribute data includes at least one of silicon rod diameter, silicon rod length, silicon rod resistivity distribution, and crystal orientation; The environmental parameter data includes at least one of ambient temperature and vibration spectrum; The quality indicator data includes at least one of line mark depth, total thickness change (TTV), and fragmentation rate; The preprocessing includes missing value imputation, outlier removal, data standardization or normalization, and timestamp alignment.

3. The hybrid-driven quality prediction method for the entire silicon wafer dicing process according to claim 1, characterized in that, The first sub-model includes a wire mesh dynamics model, a silicon rod-steel wire contact mechanics model, and a thermo-fluid-solid multiphysics coupling model; The wire mesh dynamics model is used to calculate the transverse vibration amplitude and vibration frequency of the diamond wire during the cutting process, and to establish the mapping relationship between the transverse vibration amplitude and vibration frequency and the depth of the wire mark on the silicon wafer surface. The silicon rod-steel wire contact mechanics model is used to calculate the cutting force distribution and the internal stress field of the silicon wafer, and to assess the risk of silicon wafer breakage based on the maximum principal stress criterion and fracture mechanics model. The thermo-fluid-solid multiphysics coupling model is used to calculate the temperature distribution in the cutting area and couples the temperature field to the contact mechanics model to correct the influence of thermal expansion on stress distribution and cutting quality.

4. The hybrid driving quality prediction method for the whole process of silicon wafer cutting according to claim 1, wherein, The wire mesh dynamics model is established based on the following dynamic equations: Where M is the net mass matrix, C is the damping matrix, K is the stiffness matrix, Ft is the external force vector acting on the net, and x t For wire mesh displacement response; The silicon rod-steel wire contact mechanical model is based on constitutive relations and follows: The stress field inside the silicon wafer is calculated and the risk of chipping is calculated according to the Weibull fracture risk model, wherein is the stress, is the elastic matrix, is the strain; The heat-fluid-solid multiphysics coupling model is based on the heat conduction equation: The temperature distribution in the cutting zone is calculated, wherein where p is the material density, c is the specific heat capacity, T is the temperature, k is the thermal conductivity, and q is the cutting heat source term.

5. The hybrid-driven quality prediction method for the entire silicon wafer dicing process according to claim 1, characterized in that, The second sub-model includes a stacked augmented autoencoder layer, an attention mechanism module, a bidirectional long short-term memory network layer, and a fully connected regression layer connected in sequence. The stacked enhanced autoencoder layer is used to perform hierarchical feature extraction on multi-scale process parameters, silicon rod property parameters, environmental parameters, and the output value of the first sub-model. The attention mechanism module is used to dynamically weight the features of historical samples, the features of current samples, and the temporal features at different times; The bidirectional long short-term memory network layer is used to capture the forward and backward time dependencies in the process parameter sequence of the cutting process; The fully connected regression layer is used to map deep spatiotemporal features into residual compensation values.

6. The hybrid-driven quality prediction method for the entire silicon wafer dicing process according to claim 1, characterized in that, The training objective of the second sub-model is: Where e is the residual, Y 真实 Y represents the actual measured mass. 机理 The quality baseline prediction value output by the first sub-model; The second sub-model learns the nonlinear mapping relationship between the multi-source input parameters and the residuals under the current operating conditions, and outputs a residual compensation value ê, which is used to compensate for the nonlinear error, time-varying error and equipment state drift error that the first sub-model failed to represent.

7. The hybrid-driven quality prediction method for the entire silicon wafer dicing process according to claim 1, characterized in that, The second sub-model employs an on-the-spot learning fine-tuning strategy, specifically including: For each new sample to be predicted, calculate the similarity between the new sample and historical samples in the historical database; Select the most similar nearest neighbor historical samples to form a localized training sample set; The second sub-model is locally fine-tuned or its parameters are updated based on the localized training sample set. The fine-tuned second sub-model outputs the residual compensation value corresponding to the current sample to adapt to the operating condition drift under different silicon rod batches, different equipment wear conditions, and different environmental disturbance conditions.

8. The hybrid-driven quality prediction method for the entire silicon wafer dicing process according to claim 1, characterized in that, The adaptive weighted fusion strategy dynamically adjusts the fusion weights based on the real-time prediction errors of the first and second sub-models under the current operating conditions. When the real-time prediction error of the first sub-model is less than that of the second sub-model, the weight of the first sub-model is increased. When the second sub-model learns significant residual compensation information or the real-time prediction error of the first sub-model increases, the residual compensation weight of the second sub-model is increased. The final predicted value is obtained as follows: Among them, Y 预测 Y represents the final predicted value of the silicon wafer dicing quality index. 机理 ê is the quality baseline prediction value output by the first sub-model, ê is the residual compensation value output by the second sub-model, and α(t) is the residual compensation weight coefficient that dynamically changes with the model prediction error under the current operating conditions.

9. A hybrid-driven quality prediction system for the entire silicon wafer dicing process, characterized in that, include: The data acquisition module is used to simultaneously collect process parameter data, silicon rod property data, environmental parameter data, and quality index data on the diamond wire multi-wire cutting production line. The data preprocessing module is used to perform missing value imputation, outlier removal, data standardization or normalization, and timestamp alignment on the data collected by the data acquisition module to form a standardized time series dataset. The mechanism simulation module is used to construct the first sub-model based on the physical mechanism of the cutting process and output the quality benchmark prediction value; The residual correction module is used to construct the attention-enhanced spatiotemporal network residual correction model and output the residual compensation value based on the residual between the predicted quality benchmark value and the actual quality measurement value. An adaptive fusion module is used to dynamically adjust the fusion weights based on the real-time prediction accuracy of the mechanism simulation module and the residual correction module, and output the final predicted value of the silicon wafer cutting quality index. The digital twin optimization module is used to construct a virtual quality evaluator based on the final predicted value, and to perform real-time optimization of key process parameters, anomaly warning, defect tracing, and control command generation. The control execution module is used to send the optimization control instructions generated by the digital twin optimization module to the programmable logic control system to realize online self-optimization and closed-loop control of process parameters.

Images