A method, apparatus, and equipment for predicting the diameter of silicon single crystals over multiple time intervals.
By constructing an augmented bias state vector and using delayed residual correction, the problem of insufficient accuracy and generalization in multi-time interval prediction of silicon single crystal diameter across furnace batches was solved, achieving stable and reliable prediction in new furnace batches.
Patent Information
- Application Number
- CN202611342487.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-01
- Publication Date
- 2026-11-17
AI Technical Summary
Under cross-furnace conditions, existing technologies are insufficient in terms of accuracy and generalization for multi-time interval prediction of silicon single crystal diameter, the mechanism model parameters are difficult to obtain continuously and accurately, and data-driven methods are prone to generalization degradation on new furnaces.
By constructing an augmented deviation state vector, the basic diameter prediction is generated using the kinematic relationship of the crystal growth interface, and the diameter prediction across furnace cycles is achieved through correction by delay residuals and trend confidence.
It significantly improves the reliability and engineering practicality of multi-time interval prediction in actual deployment across furnace cycles, avoids systematic bias and unreliable trend extrapolation caused by parameter changes, and improves prediction stability.
Smart Images

Figure CN123120146A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of crystal growth prediction, specifically relating to a method, apparatus, and equipment for predicting the diameter of silicon single crystals over multiple time intervals. Background Technology
[0002] The Czochralski method is the main industrial method for preparing semiconductor-grade silicon single crystals. Its growth process involves multi-physics coupling, including heat conduction, melt flow, interfacial phase transformation, and crystal pulling. Crystal diameter is a key variable characterizing growth stability and ingot geometric quality. Especially during the constant-diameter growth stage, diameter fluctuations directly affect the effective utilization rate of the ingot and the consistency of subsequent slicing. Accurate diameter prediction at multiple time intervals (30 seconds, 60 seconds, and 120 seconds) provides a basis for proactively adjusting heating power and pulling speed, which is a crucial foundation for achieving high-quality automated growth.
[0003] However, the CZ silicon single crystal growth process is characterized by strong nonlinearity, large inertia, strong coupling, and time-varying parameters, resulting in a significant dynamic lag between control input and diameter response. While the mechanistic model possesses clear physical meaning and interpretability, its predictive accuracy is highly dependent on model assumptions and parameter calibration. In actual production, parameters such as equilibrium angle, surface tension, and equivalent thermal resistance are difficult to obtain consistently and accurately, and may change with each furnace cycle, inevitably leading to parameter mismatch and unmodeled dynamics in the mechanistic model. Data-driven methods learn the mapping relationship between process variables and diameter using historical data; however, semiconductor-grade silicon single crystal production exhibits typical batch characteristics, with significant differences in raw material conditions, thermal field distribution, and equipment conditions between different furnace cycles. Offline-trained models are prone to significant generalization degradation on new furnace cycles. Therefore, balancing the physical consistency of the mechanistic model with the online adaptability of data-driven methods under cross-furnace conditions becomes the core challenge of multi-time-spacing prediction. Summary of the Invention
[0004] To address the shortcomings in accuracy and generalization of diameter prediction over multiple time intervals under cross-furnace conditions in existing technologies, this invention provides a method, apparatus, and device for predicting the diameter of silicon single crystals over multiple time intervals.
[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting the diameter of a silicon single crystal over multiple time intervals, the method comprising: Obtain the measured diameter of the secondary equal-diameter growth stage in the silicon single crystal growth furnace; Based on the measured diameter at the current moment, the basic diameter prediction at each prediction time interval is generated using the kinematic relationship of the crystal growth interface. For each prediction time interval, an augmented bias state vector matching the discrete prediction step number of that prediction time interval is constructed. The augmented bias state vector uses the discrete prediction step number as its dimension and stores the system bias estimate of the current moment and the system bias estimates of each historical moment tracing back the discrete prediction step number in the same vector. When the actual diameter observation corresponding to the historical prediction arrives, the prediction residual of the historical prediction is used as the delay residual, and the delay residual is used as the observation input to update the bias state component in the augmented bias state vector corresponding to the discrete prediction step number. The system bias estimate of the current moment in the augmented bias state vector is extracted and superimposed on the basic diameter prediction to obtain the bias-corrected prediction result. Based on the measured diameter, the current trend and its estimation uncertainty are estimated. The trend confidence is determined according to the relative magnitude of the estimated value of the current trend and the estimation uncertainty. When the trend confidence reaches a preset condition, the correction amount of the corresponding trend is superimposed on the prediction result after the deviation correction; otherwise, it is not superimposed. The result after superimposing or not superimposing the trend correction amount is convexly combined with the continuous prediction using the current measured diameter as the prediction value to generate the final diameter prediction result for each prediction time interval.
[0006] Optionally, the step of generating the base diameter prediction for each prediction time interval using the kinematic relationship of the crystal growth interface includes: Calculate the angular deviation relative to the constant diameter growth equilibrium angle based on the crystal radius, axial crystal growth rate, and actual interface angle at the current moment. The crystal diameter change rate is calculated based on the product relationship between the angular deviation and the axial crystal growth rate. The diameter change rate is discretely accumulated over the prediction time interval to obtain the diameter change predicted by the mechanism model; The predicted base diameter is obtained by superimposing the diameter change onto the measured diameter at the current moment.
[0007] Optionally, constructing an augmented bias state vector matching the number of discrete prediction steps for each prediction time interval includes: Using the discrete prediction steps as the dimensional basis of the augmented deviation state vector, the augmented deviation state vector includes the system deviation state component at the current moment and each historical deviation state component that traces back the discrete prediction steps. An augmented state transition matrix is established such that, during the time recursion process of the augmented deviation state vector at each sampling time, each deviation state component in the augmented deviation state vector of the previous time is shifted and stored sequentially, and the deviation state component of the previous time at the current time is recursively used as the first component of the current time. An observation matrix is established that matches the augmented bias state vector, such that when the delay residual is used as the observation input for updating, the delay residual only forms an observation relationship with the historical bias state component in the augmented bias state vector corresponding to the discrete prediction step.
[0008] Optionally, the update of the deviation state component in the augmented deviation state vector corresponding to the discrete prediction step number adopts a robust update method, including: The difference between the delay residual and the estimated delay residual predicted based on the augmented bias state vector is calculated as the original innovation, and the innovation variance corresponding to the original innovation is calculated. The original information is standardized and then nonlinearly mapped using the Huber function. When the absolute value of the standardized information does not exceed the preset Huber threshold, its original value is maintained. When it exceeds the Huber threshold, it is compressed to the Huber threshold. The mapped value is multiplied by the standard deviation to obtain a robust innovation. This robust innovation is used to replace the original innovation in calculating the update amount. The robust innovation is also used to replace the original innovation in updating the bias state component in the augmented bias state vector corresponding to the discrete prediction step.
[0009] Optionally, the current trend and its estimation uncertainty are estimated based on the measured diameter, including: Calculate the rate of change of the diameter at the current moment relative to the previous moment; The diameter change rate is estimated by using filtering methods with different smoothing coefficients to obtain fast trend estimates and slow trend estimates. Calculate the estimated uncertainties for the fast trend and the slow trend, respectively.
[0010] Optionally, estimating the diameter change rate using filtering methods with different smoothing coefficients includes: The rate of change of the diameter is estimated using a first exponentially weighted moving average to obtain a fast trend estimate; a second exponentially weighted moving average is also used to estimate the rate of change of the diameter to obtain a slow trend estimate; wherein the half-life of the first exponentially weighted moving average is less than the half-life of the second exponentially weighted moving average. The calculation of the estimation uncertainties for the fast trend and the slow trend, respectively, includes: The variances of the fast and slow trends are estimated using an exponentially weighted moving average method, which serves as the estimation uncertainty for the corresponding trends.
[0011] Optionally, determining the trend confidence level based on the relative magnitude of the current trend estimate and the estimation uncertainty includes: Using the standard deviation of the corresponding trend as a measure of the estimated uncertainty, a trend confidence score is constructed based on the estimated value and the estimated uncertainty of the current trend. The expression: ; in, This is the absolute value of the estimate of the current trend. The standard deviation of the corresponding trend. To prevent the use of tiny constants with a denominator of zero, the trend confidence increases with the increase of the estimated value of the current trend and decreases with the increase of the estimated uncertainty. When the trend confidence level is greater than or equal to the preset minimum confidence level threshold, the preset condition is met, and the corresponding trend correction amount is activated; When the trend confidence level is less than the preset minimum confidence level threshold, the preset condition is not met, and the corresponding trend correction amount is not activated.
[0012] Optionally, the trend correction amount is obtained by multiplying the estimated value of the corresponding trend by a preset gain coefficient; the step of convexly combining the result after superimposing or not superimposing the trend correction amount with the persistent prediction using the current measured diameter as the predicted value includes: The fusion weight is determined based on the predicted time interval. The fusion weight is used as the coefficient of the result after superimposed or unsuperimposed trend correction. The difference between the fusion weight and the preset standard number is used as the coefficient of the continuous prediction. The two are weighted and summed to obtain the final diameter prediction result.
[0013] A device for predicting the diameter of a silicon single crystal over multiple time intervals, the device comprising: The acquisition module is used to acquire the measured diameter of the silicon single crystal growth furnace during the secondary equal-diameter growth stage; The generation module is used to generate the basic diameter prediction at each prediction time interval based on the measured diameter at the current time and the kinematic relationship of the crystal growth interface. The first prediction module is used to construct an augmented deviation state vector matching the discrete prediction step number for each prediction time interval. The augmented deviation state vector stores the system deviation estimate at the current moment and the system deviation estimates at each historical moment tracing back the discrete prediction step number in the same vector, with the discrete prediction step number as the dimension. When the actual diameter observation corresponding to the historical prediction arrives, the prediction residual of the historical prediction is used as the delay residual, and the delay residual is used as the observation input to update the deviation state component in the augmented deviation state vector corresponding to the discrete prediction step number. The system deviation estimate at the current moment in the augmented deviation state vector is extracted and superimposed on the basic diameter prediction to obtain the prediction result after deviation correction. The second prediction module is used to estimate the current trend and its estimation uncertainty based on the measured diameter, determine the trend confidence level according to the relative magnitude of the estimated value of the current trend and the estimation uncertainty, and when the trend confidence level reaches a preset condition, add the corresponding trend correction amount to the prediction result after deviation correction; otherwise, do not add it. The result after adding or not adding the trend correction amount is convexly combined with the continuous prediction using the current measured diameter as the prediction value to generate the final diameter prediction result for each prediction time interval.
[0014] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned method for multi-time-distance prediction of silicon single crystal diameter.
[0015] The multi-time-distance prediction method for silicon single crystal diameter provided by this invention has the following beneficial effects: This invention stores current and historical deviation estimates by augmenting the deviation state vector with the deviation step count of the prediction time interval as the dimension. This allows the delayed prediction residuals to be time-aligned by corresponding to the historical position of the deviation step count. The trend confidence is determined based on the relative magnitude of the trend estimate and the uncertainty. Trend correction is only activated when preset conditions are met. This avoids unreliable trend extrapolation from destroying the established prediction stability when the continuous compensation mechanism model suffers from systematic deviations caused by parameter changes in new furnace batches. It effectively overcomes the generalization degradation caused by distribution offset in offline training models and significantly improves the reliability and engineering practicality of multi-time interval prediction in actual deployment across furnace batches. Attached Figure Description
[0016] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a flowchart illustrating a multi-time-distance prediction method for the diameter of a silicon single crystal according to an exemplary embodiment of the present invention.
[0018] Figure 2 This invention provides an overall framework for multi-time-distance prediction of the diameter of semiconductor silicon single crystals based on mechanism anchoring and fixed hysteresis robust Kalman correction, according to an exemplary embodiment.
[0019] Figure 3 This is a block diagram of a silicon single crystal diameter multi-time interval prediction device provided by the present invention according to an exemplary embodiment. Detailed Implementation
[0020] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.
[0021] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0022] First, this invention provides a method for predicting the diameter of silicon single crystals over multiple time intervals, specifically as follows: Figure 1 As shown, it includes the following steps: S101. Obtain measured data of the secondary equal diameter growth stage in the silicon single crystal growth furnace.
[0023] The measured data should include at least the measured diameter of the crystal, the main heating power, the crystal pulling speed, and the interface state variables.
[0024] S102. Based on the measured diameter at the current moment, generate the basic diameter prediction for each prediction time interval using the kinematic relationship of the crystal growth interface.
[0025] In this step, based on the measured diameter and interface state variables at the current moment, a mechanism model predictor is constructed using the kinematic relationship of the crystal growth interface to generate the basic diameter prediction at each prediction time interval.
[0026] For example, based on the crystal radius, axial crystal growth rate, and actual interface angle at the current moment, the angular deviation relative to the constant diameter growth equilibrium angle is calculated; the crystal diameter change rate is calculated based on the product relationship between the angular deviation and the axial crystal growth rate; the diameter change rate is discretely accumulated over the prediction time interval to obtain the diameter change predicted by the mechanism model; the diameter change is superimposed on the measured diameter at the current moment to obtain the basic diameter prediction.
[0027] S103. For each prediction time interval, construct an augmented deviation state vector that matches the number of discrete prediction steps for that prediction time interval. When the actual diameter observation corresponding to the historical prediction arrives, update the deviation state component in the augmented deviation state vector corresponding to the number of discrete prediction steps. Extract the current time system deviation estimate from the augmented deviation state vector and superimpose it onto the basic diameter prediction to obtain the prediction result after deviation correction.
[0028] In this step, for each prediction time interval, an augmented bias state vector matching the discrete prediction step number of that prediction time interval is constructed. The augmented bias state vector uses the discrete prediction step number as its dimension and stores the system bias estimate of the current moment and the system bias estimates of each historical moment tracing back the discrete prediction step number in the same vector. When the actual diameter observation corresponding to the historical prediction arrives, the prediction residual of the historical prediction is used as the delay residual, and the delay residual is used as the observation input to update the bias state component in the augmented bias state vector corresponding to the discrete prediction step number. The system bias estimate of the current moment in the augmented bias state vector is extracted and superimposed on the basic diameter prediction to obtain the prediction result after bias correction.
[0029] For example, for the current prediction time interval, its discrete prediction steps are used as the dimension base of the augmented deviation state vector, so that the augmented deviation state vector contains the system deviation state component at the current time and the historical deviation state components at the previous discrete prediction steps. An augmented state transition matrix is established so that during the time recursion process at each sampling time, the deviation state components in the augmented deviation state vector at the previous time are shifted and stored sequentially, and the deviation state component at the current time of the previous time is recursively used as the first component at the current time. An observation matrix matching the augmented deviation state vector is established so that the currently arriving delay residual only forms an observation relationship with the historical deviation state components in the augmented deviation state vector corresponding to the discrete prediction steps, so that the time of the delay prediction residual is aligned with the historical deviation state. The difference between the delay residual and the estimated delay residual predicted based on the augmented bias state vector is used as the original innovation, and the corresponding innovation variance is calculated. The original innovation is standardized and then nonlinearly mapped using the Huber function. When the absolute value of the standardized innovation is less than or equal to a preset Huber threshold, its original value is maintained; when it is greater than the Huber threshold, it is compressed to the Huber threshold. The mapped value is multiplied by the standard deviation to obtain the robust innovation. The robust innovation is used to replace the original innovation to calculate the update amount, and the robust innovation is used to replace the original innovation to update the bias state component in the augmented bias state vector corresponding to the discrete prediction step.
[0030] S104. Based on the measured diameter, estimate the current trend and its estimation uncertainty. Determine the trend confidence level according to the relative magnitude of the estimated value of the current trend and the estimation uncertainty. When the trend confidence level reaches the preset condition, add the corresponding trend correction amount to the prediction result after deviation correction; otherwise, do not add it. Combine the result after adding or not adding the trend correction amount with the continuous prediction using the current measured diameter as the prediction value in a convex combination to generate the final diameter prediction result for each prediction time interval.
[0031] In one embodiment, the rate of change of the diameter at the current moment relative to the previous moment is calculated; filtering methods with different smoothing coefficients are used to estimate the rate of change of the diameter to obtain fast trend estimates and slow trend estimates; and the estimation uncertainties of the fast trend and slow trend are calculated respectively.
[0032] For example, the rate of change of diameter is estimated using a first exponentially weighted moving average to obtain a fast trend estimate; a second exponentially weighted moving average is used to estimate the rate of change of diameter to obtain a slow trend estimate; wherein the half-life of the first exponentially weighted moving average is less than the half-life of the second exponentially weighted moving average; the variances of the fast and slow trends are estimated using exponentially weighted moving averages respectively, which are used as the estimation uncertainties of the corresponding trends.
[0033] In this step, the diameter change rate relative to the previous time step is determined based on the measured diameter sequence. Filtering methods with different smoothing coefficients are used to estimate the diameter change rate, obtaining fast trend estimates and slow trend estimates. The estimation uncertainty of each trend is determined, and the trend confidence is determined based on the relative magnitude of the estimated value and the estimation uncertainty of each trend. The correction amount for each trend is determined based on the estimated value of each trend, and whether to enable the correction amount for each trend is determined based on the comparison result of the trend confidence and the preset minimum confidence threshold. The filtered fast trend correction amount and slow trend correction amount are superimposed on the prediction result after deviation correction to obtain trend enhancement candidate predictions. The trend enhancement candidate predictions are convexly combined with the persistent prediction using the current measured diameter as the prediction value to generate the final diameter prediction result for each prediction time interval.
[0034] When the trend confidence level is lower than the preset minimum confidence level threshold, the corresponding trend correction amount is not added up, and the trend enhancement candidate prediction is equal to the prediction result after bias correction.
[0035] For example, a fast trend estimate is obtained by estimating the diameter change rate using a first exponentially weighted moving average, and a slow trend estimate is obtained by estimating the diameter change rate using a second exponentially weighted moving average, where the half-life of the first exponentially weighted moving average is shorter than that of the second exponentially weighted moving average. The variances of the fast and slow trends are estimated using exponentially weighted moving averages, respectively, as the estimation uncertainty of the corresponding trends. For each type of trend in both fast and slow trends, the absolute value of the trend estimate is divided by the sum of the absolute value of the trend estimate and the standard deviation of the corresponding trend, yielding a trend confidence level between 0 and 1. The trend confidence level increases with the magnitude of the trend estimate and decreases with the estimation uncertainty. When the trend confidence level is greater than or equal to a preset minimum confidence threshold, the correction amount for that trend is enabled; when the trend confidence level is less than the preset minimum confidence threshold, the correction amount for that trend is disabled. The fast trend estimate and the slow trend estimate are multiplied by preset gain coefficients corresponding to the prediction time interval and the corresponding trend type, respectively, to obtain the fast trend correction amount and the slow trend correction amount. The fusion weight is determined based on the prediction time interval. The fusion weight is used as the coefficient of the result after superimposed or unsuperimposed trend correction. The difference between the fusion weight and the preset standard number is used as the coefficient of the continuous prediction. The two are weighted and summed to obtain the final diameter prediction result. For example, if the fusion weight is between 0 and 1, 1 is used as the difference between the preset standard number and the fusion weight as the coefficient of the continuous prediction. The two are weighted and summed to obtain the final diameter prediction result.
[0036] Using the above method, the current and historical bias estimates are stored in the bias state vector with the deviation step number of the prediction time interval as the dimension. This allows the delayed prediction residuals to be time-aligned by corresponding to the historical position of the deviation step number. The trend confidence is determined based on the relative magnitude of the trend estimate and the uncertainty. Trend correction is only activated when preset conditions are met. In this way, under the condition of continuous compensation mechanism model caused by the systematic bias due to parameter changes in new furnaces, unreliable trend extrapolation is avoided from destroying the established prediction stability. This effectively overcomes the generalization degradation caused by the distribution offset of the offline training model and significantly improves the reliability and engineering practicality of multi-time interval prediction in the actual deployment across furnaces.
[0037] Based on the above steps, the present invention also provides an embodiment for illustration.
[0038] Let the measured diameter sequence of the constant diameter growth stage of CZ silicon single crystal be: ; in, Indicates time Crystal diameter, The number of valid samples is [number], and the sampling interval is [number]. The prediction time interval set considered in this invention is as follows:
[0039] ; For any prediction time interval The corresponding number of discrete prediction steps is: ; At the present moment Utilizing information sets Predicting future moments Crystal diameter: ; in, This includes current and historical crystal diameters, main heating power, crystal pulling speed, interface state, and historical prediction residuals already reached. Time intervals are not used in the prediction process. Any subsequent measured information is used to ensure the causality of online predictions.
[0040] Based on the above problem definition, this invention constructs as follows: Figure 2 The three-stage cross-furnace diameter multi-time interval prediction framework is shown. M1 provides a physically constrained basic prediction using the kinematic relationship of the crystal growth interface; M2, through delayed residual time alignment and robust Kalman update online estimation and correction of system bias, is the core stage of the framework to achieve cross-furnace adaptation; M3, based on M2, appropriately refines the credible local dynamics according to fast and slow trends and their uncertainties, and actively reduces the correction amplitude when the trend credibility is low. The three stages perform independent rolling predictions for prediction time intervals of 30 s, 60 s, and 120 s, respectively.
[0041] like Figure 2 As shown, at the current time t, M1 calculates the future diameter change based on the current measured diameter and the relevant process state, forming a mechanism-anchored prediction. For the prediction time interval h, a time-aligned prediction residual can only be constructed after the actual observation corresponding to the historical prediction has arrived. M2 uses this delayed residual as an observation and recursively estimates the current system bias through fixed-lag augmented states and Huber robust innovation, which is the core component of the proposed framework to achieve online adaptation.
[0042] After M2 completes the systematic bias correction, M3 further extracts the fast and slow timescale trends in the diameter sequence and adjusts the local correction intensity based on the trend uncertainty. When the trend is stable and reliable, M3 appropriately refines the output of M2; when the trend reliability is low, the prediction results actively shrink towards M2 and the persistence prediction to avoid error amplification caused by unreliable trend extrapolation. Therefore, M3 mainly undertakes the role of prediction refinement and stability maintenance, rather than replacing the core bias correction function of M2.
[0043] The three prediction stages can be uniformly represented as: ; ; as well as ; in, The diameter change predicted by the mechanistic model. For predicting time interval System bias estimation under the following conditions This is a dual-trend correction amount. For weight fusion.
[0044] During the growth of CZ silicon single crystals, the change in crystal radius is determined by both the axial growth rate and the geometric state near the solid-liquid interface. Let... Where is the crystal radius, This represents the axial crystal growth rate. For the actual interface angle, To maintain the equilibrium angle during constant-diameter growth of the crystal, the angular deviation relative to the constant-diameter state can be expressed as:
[0045] ; The rate of change of crystal radius is expressed as: ; when When, the crystal radius remains unchanged; when At this time, the crystal radius increases or decreases with changes in the interface state. Because the crystal diameter satisfies:
[0046] ; Therefore, the rate of change of diameter is: ; For the predicted time interval The mechanistic model gives the following future diameter change: ; in, and They are time points The future growth rate and interface angle are estimated based on the current process state.
[0047] Under discrete-time conditions, the above equation can be written as: ; To reduce the impact of absolute state drift in the mechanistic model on the prediction results, this invention does not directly use the absolute diameter output by the mechanistic model. Instead, it superimposes the diameter change calculated by the mechanistic model onto the current measured diameter. Therefore, the prediction result for M1 is:
[0048] ; This process uses the measured diameter to determine the prediction starting point and only uses a mechanistic model to describe the direction and magnitude of future diameter changes, thus avoiding the cumulative drift caused by long-term integration of the absolute state. However, the equilibrium angle, interfacial thermal state, and equivalent heat transfer conditions may change with each furnace cycle, and M1 may still be consistently overestimated or underestimated, thus requiring further online deviation correction.
[0049] For the predicted time interval ,time The generated M1 prediction is: ; At that moment Once the actual diameter is reached, the residuals corresponding to the historical predictions can be calculated: ; The residual corresponds to The predictions are generated 100 times prior to the current sampling step, not at the current time. Since the sampling interval is 2 seconds, the prediction intervals of 30 seconds, 60 seconds, and 120 seconds correspond to fixed lags of 15, 30, and 60 sampling steps, respectively. Therefore, there is a fixed time lag between the prediction residuals and the bias state that matches the prediction interval.
[0050] To describe the bias of slowly varying systems in mechanistic predictions, a random walk state model is established for each prediction interval: ; in, Indicates time Predicted time interval Systematic prediction bias under The noise is zero-mean process noise, and satisfies: ; To achieve time alignment between the delayed residual and the historical bias state, an augmented state vector is constructed: ; The corresponding state equation is: ; in, ; Due to the currently obtained prediction residuals Corresponding to The deviation state before each sampling step is expressed by the observation equation as follows: ; in, ; And the observation noise satisfies: .
[0051] The prior estimate of the augmented bias state is: ; The corresponding prior covariance is: ; Calculate the observational information based on the currently arriving historical prediction residuals: ; Its new information variance is: ; Traditional Kalman updates are highly sensitive to anomalous information. To reduce the impact of local oscillations, measurement glitches, and short-term anomalous disturbances on state updates, the information is first standardized:
[0052] ; The normalized information is constrained using Huber functions: ; in, For predicting time interval The corresponding Huber threshold. Robust innovation is:
[0053] .
[0054] The Kalman gain is calculated as follows: ; The posterior estimate of the augmented state is: ; To improve numerical stability, the posterior covariance is updated using the Joseph form: .
[0055] The first component of the augmented state vector represents the system bias estimate at the current time step: ; Therefore, the prediction result for M2 is: .
[0056] M2 updates the bias state using only the actual observations that have arrived at the current time; future predictions that have not yet obtained the actual values do not participate in the current state update, thus satisfying the causality requirement of online prediction.
[0057] M2 is primarily used to compensate for slowly varying system biases in the mechanistic model, but the crystal diameter sequence may also contain rapid changes caused by local process adjustments and slow drifts caused by thermal field evolution. To characterize the dynamic features at different time scales, the current diameter change rate is defined as:
[0058] ; The fast and slow trends are estimated separately using exponentially weighted moving averages: ; in, and These represent the fast and slow trends, respectively, with the forgetting coefficient determined by the corresponding half-life: ; And satisfy: ; To measure the stability of the trend estimate, the trend uncertainty is estimated using exponentially weighted variance: ; Based on the relative relationship between trend magnitude and estimation uncertainty, the trend confidence level is defined as follows: ; in, This is the absolute value of the estimate of the current trend. The standard deviation of the corresponding trend. To prevent the use of tiny constants with a denominator of zero, and given the following: ; Further define the confidence-gated variables: ; in, This is the minimum confidence threshold. When the trend fluctuates significantly or the estimate is unstable, the corresponding trend correction is automatically turned off.
[0059] For the predicted time interval , No. The diameter correction amount generated by the trend-like pattern is: ; in, For trend correction gain, This is the confidence level adjustment index. The total bi-trend adjustment is...
[0060] ; Based on the M2 forecast, the candidate forecasts for enhanced trend are: .
[0061] Considering the strong local persistence of crystal diameter during the constant-diameter growth stage, relying solely on trend-enhancing predictions may lead to over-extrapolation under oscillating conditions. Therefore, a convex combination of trend-enhancing candidate predictions and persistence predictions is used:
[0062] ; in, ; when At that time, M3 degenerates into a persistent predictor: ; when In this case, M3 fully utilizes the candidate predictions of M2 after bi-trend correction. This fusion structure enables the model to enhance dynamic tracking capabilities when the trend is stable and to proactively reduce the correction magnitude when the trend confidence is low.
[0063] Figure 2 The overall information flow and progressive relationship between M1, M2, and M3 are presented. This section further explains the mechanism forward calculation, bias state update, and trend refinement processes under different prediction time intervals, following the online execution order. For each prediction time interval h, M2 maintains independent augmented bias states and covariance matrices, while M3 sets the corresponding trend parameters and fusion weights. The specific steps are as follows:
[0064] (1) Read the crystal diameter and related process variables at the current time (t). When the actual diameter corresponding to the historical prediction is reached, calculate the prediction residual with time alignment and update the augmented bias state of M2 at the prediction time interval (h).
[0065] (2) Using the current process state to drive the mechanism model forward calculation, the mechanism diameter change and M1 prediction at prediction time intervals of 30s, 60s and 120s are obtained respectively.
[0066] (3) Add the current system bias estimate to the output of M1 to obtain the prediction of M2.
[0067] (4) Update the fast trend, slow trend and uncertainty of M3 according to the current diameter change, and calculate the corresponding trend confidence and gating state.
[0068] (5) Calculate the dual-trend correction based on the confidence gating results, and merge the trend enhancement prediction with the persistence prediction to obtain the final prediction of M3.
[0069] Mechanism model parameters, process noise variance Observation noise variance Huber threshold The fast and slow trend half-life, confidence threshold, trend correction gain, and fusion weight were all determined through verification using an inner-layer test batch. The outer-layer test batches did not participate in any parameter identification, hyperparameter selection, or data standardization process.
[0070] The entire online prediction process uses only the information obtained at the current moment and in the past, without recalculating the current prediction using future observations. Therefore, it can meet the real-time and causal requirements of multi-time-distance rolling prediction of CZ silicon single crystal growth process.
[0071] Based on the steps of the above embodiments, the present invention has also been experimentally verified.
[0072] The experimental data came from an industrial Czochralski silicon single crystal growth system of a semiconductor company, comprising five independent production runs, denoted as Run 1–Run 5. Crystal diameter and related process variables were recorded synchronously at a fixed sampling interval of 2 seconds, corresponding to a sampling frequency of 0.5 Hz. All data were extracted from the constant-diameter crystal growth stage, with each run containing several hours of continuous industrial operation records. Crystal diameter was used as the prediction target, while process variables such as main heating power, crucible position, crystal pulling speed, and interface contact angle were used for mechanistic model calculations or data-driven model training.
[0073] To illustrate the data scale and diameter distribution characteristics of different industrial furnace runs, Table 1 summarizes the number of valid samples, runtime, and diameter statistics for each furnace run after data cleaning and continuous segment filtering. This table is only used to describe the characteristics of the dataset and is not used for model parameter identification, hyperparameter selection, or test furnace standardization.
[0074] Table 1. Overview of industrial furnace data for the equal diameter growth stage of semiconductor-grade silicon single crystals Run Cleaned samples Validsamples Excludedsamples (%) Continuoussegments Valid evaluation duration (min) Diameter mean ±SD (mm) Diameter range (mm) Run 1 6560 6496 64 (1.0) 2 216.5 305.95 ± 4.22 293.18–320.95 Run 2 8074 8005 69 (0.9) 4 266.8 303.29 ± 2.33 290.47–319.53 Run 3 8822 8822 0 (0.0) 1 294.1 305.35 ± 1.46 301.94–308.11 Run 4 9497 9497 0 (0.0) 1 316.6 305.90 ± 0.55 304.84–307.00 Run 5 8354 8354 0 (0.0) 1 278.5 306.26 ± 0.45 305.07–307.20 Table 1 shows that the five industrial furnace runs exhibit significant differences in data scale, continuity, and diameter fluctuation characteristics. Each furnace run contains 6496–9497 valid samples, with an effective evaluation duration of 216.5–316.6 min. Run 1 and Run 2 contain 2 and 4 continuous data segments, respectively, with sample exclusion rates of 1.0% and 0.9%. Run 3–Run 5 each contain a single continuous data segment with no invalid samples. The standard deviation of diameter varies from 0.45 to 4.22 mm across different furnace runs, with significant differences in the diameter range. The intra-furnace fluctuation in Run 1 and Run 2 is significantly higher than that in Run 4 and Run 5. This heterogeneity among furnace runs provides a representative industrial scenario for evaluating the model's generalization ability under conditions of varying data continuity, differences in diameter fluctuation, and absence of furnace runs.
[0075] Let the original data sampling interval be... For the predicted time interval:
[0076] ; The corresponding number of discrete prediction steps is: .
[0077] Data preprocessing first extracts constant-diameter growth data for each furnace based on production stage identifiers, eliminating data from stages such as crystal pulling, shoulder formation, shoulder turning, finishing, and equipment shutdown. Then, the samples are sorted by timestamp, and discontinuous data segments are identified based on the difference between adjacent timestamps. When the timestamp difference between adjacent samples exceeds the normal 2-second sampling interval, indicating missing sampling points, the corresponding records are divided into different continuous segments to prevent the model input window or prediction target from crossing data breakpoints.
[0078] Samples exhibiting equipment restarts, sensor failures, missing key process variables, or significant measurement glitches are excluded from model training and performance evaluation. However, actual local oscillations, slow drifts, and control adjustments occurring during normal production processes are retained to verify the model's predictive stability under complex industrial conditions.
[0079] For any prediction time interval A valid prediction sample is constructed only when the historical input window and the future target interval are both located within the same furnace batch and the same continuous data segment. Let the required historical window length for the model be... Then, a valid sample must satisfy:
[0080] ; It does not cross furnace boundary, data breakpoint, or invalid production stage.
[0081] To ensure fairness in comparisons between different methods, all methods use the same common valid sample mask to calculate evaluation metrics under the same outer layer test furnace number and prediction time interval. No method may selectively change the test sample range due to its own prediction deficiencies.
[0082] To avoid introducing test information during data preprocessing, standardized parameters are calculated only from the calibration furnace in the current outer layer tradeoff. For process variables... The standardized form is:
[0083] ; in, and These represent the variables in the calibration furnace. The mean and standard deviation. The same parameters were then used for the inner layer validation furnaces and the outer layer test furnaces; the statistical information of the test furnaces was not included in the calculation of the standardized parameters.
[0084] M1 operates directly using process variables with actual physical dimensions. For data-driven models requiring standardized input, the predictions are restored to actual diameter dimensions before evaluation, and performance metrics are calculated on a uniform, common, valid sample.
[0085] CZ silicon single crystal production exhibits significant batch characteristics, with strong temporal correlations between adjacent samples within the same furnace. If all time-series samples are randomly divided into training and testing sets, adjacent time segments within the same furnace may simultaneously enter model training and performance evaluation, resulting in leakage of temporal information and overestimation of the model's generalization performance on unseen furnaces.
[0086] To simulate the actual application scenario of deploying the model to a new production furnace, this invention adopts a nested leave-one-run-out (LORO) verification scheme. This scheme consists of two independent processes: an outer layer of cross-furnace performance evaluation and an inner layer of model selection.
[0087] Let the set of all furnace batches be: ; In the The outer layer compromises, the complete furnace cycle Designated as the test set: ; The remaining four furnace cycles constitute the calibration set: ; Five furnace runs were sequentially designated as completely unseen outer test runs, forming a total of five outer test layers. These outer test runs did not participate in model structure determination, parameter identification, hyperparameter selection, data standardization, or early shutdown decisions.
[0088] This classification method corresponds to the actual industrial deployment process, which involves using existing production furnaces to complete model development and then directly applying the model to new furnaces that have not participated in the development.
[0089] For each outer layer fold, leave-one-batch validation is performed within the remaining four calibration batches. Each time, one calibration batch is selected as the inner layer validation set, and the remaining three batches are selected as the inner layer training set, so each outer layer fold contains four inner layer validation folds.
[0090] Let the prediction time interval be The following are the candidate parameters: ; in, The candidate parameter space is represented by parameters such as mechanistic model parameters, Kalman process noise variance and observation noise variance, Huber threshold, fast and slow trend half-life, confidence threshold, trend correction gain and fusion weights.
[0091] The objective function for candidate parameters in inner layer validation is defined as: ; in, ; Indicates the candidate parameter in the th... Internal layer verification furnace number and prediction time interval The RMSE is as follows. The optimal parameters are:
[0092] .
[0093] After selecting the inner layer model, the optimal parameters are used to rebuild the model on all four calibration furnaces of the current outer layer fold, and a one-time evaluation is performed on the corresponding outer layer test furnaces. The outer layer test results are not fed back into any parameter selection process.
[0094] For the three prediction time intervals of 30 s, 60 s, and 120 s, M2 maintains independent bias state and noise parameters, while M3 determines the corresponding trend correction parameters and fusion weights. The selection of parameters for all prediction time intervals is strictly limited to the inner layer validation process.
[0095] Secondly, the present invention also provides a device for predicting the diameter of a silicon single crystal over multiple time intervals, such as... Figure 3 As shown, it includes: The acquisition module 201 is used to acquire the measured diameter of the silicon single crystal growth furnace during the secondary equal-diameter growth stage.
[0096] The generation module 202 is used to generate the basic diameter prediction at each prediction time interval based on the measured diameter at the current time and the kinematic relationship of the crystal growth interface.
[0097] The first prediction module 203 is used to construct an augmented deviation state vector that matches the discrete prediction step number for each prediction time interval. The augmented deviation state vector stores the system deviation estimate at the current moment and the system deviation estimates at each historical moment that trace back the discrete prediction step number in the same vector, with the discrete prediction step number as the dimension. When the actual diameter observation corresponding to the historical prediction arrives, the prediction residual of the historical prediction is used as the delay residual, and the delay residual is used as the observation input to update the deviation state component in the augmented deviation state vector corresponding to the discrete prediction step number. The system deviation estimate at the current moment in the augmented deviation state vector is extracted and superimposed on the basic diameter prediction to obtain the prediction result after deviation correction.
[0098] The second prediction module 204 is used to estimate the current trend and its estimation uncertainty based on the measured diameter, determine the trend confidence level according to the relative magnitude of the estimated value of the current trend and the estimation uncertainty, and when the trend confidence level reaches a preset condition, add the corresponding trend correction amount to the prediction result after deviation correction; otherwise, do not add it. The result after adding or not adding the trend correction amount is convexly combined with the continuous prediction using the current measured diameter as the prediction value to generate the final diameter prediction result for each prediction time interval.
[0099] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The steps of a method for predicting the diameter of a silicon single crystal over multiple time intervals are provided.
[0100] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above-mentioned functions. Figure 1 The steps of a method for predicting the diameter of a silicon single crystal over multiple time intervals are provided.
[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0105] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the patent of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for predicting the diameter of a silicon single crystal over multiple time intervals, characterized in that, The method includes: Obtain the measured diameter of the secondary equal-diameter growth stage in the silicon single crystal growth furnace; Based on the measured diameter at the current moment, the basic diameter prediction at each prediction time interval is generated using the kinematic relationship of the crystal growth interface. For each prediction time interval, an augmented bias state vector matching the discrete prediction step number of that prediction time interval is constructed. The augmented bias state vector uses the discrete prediction step number as its dimension and stores the system bias estimate of the current moment and the system bias estimates of each historical moment tracing back the discrete prediction step number in the same vector. When the actual diameter observation corresponding to the historical prediction arrives, the prediction residual of the historical prediction is used as the delay residual, and the delay residual is used as the observation input to update the bias state component in the augmented bias state vector corresponding to the discrete prediction step number. The system bias estimate of the current moment in the augmented bias state vector is extracted and superimposed on the basic diameter prediction to obtain the bias-corrected prediction result. Based on the measured diameter, the current trend and its estimation uncertainty are estimated. The trend confidence is determined according to the relative magnitude of the estimated value of the current trend and the estimation uncertainty. When the trend confidence reaches a preset condition, the correction amount of the corresponding trend is superimposed on the prediction result after the deviation correction; otherwise, it is not superimposed. The result after superimposing or not superimposing the trend correction amount is convexly combined with the continuous prediction using the current measured diameter as the prediction value to generate the final diameter prediction result for each prediction time interval.
2. The method according to claim 1, characterized in that, The method of generating the base diameter prediction for each prediction time interval using the kinematic relationship of the crystal growth interface includes: Calculate the angular deviation relative to the constant diameter growth equilibrium angle based on the crystal radius, axial crystal growth rate, and actual interface angle at the current moment. The crystal diameter change rate is calculated based on the product relationship between the angular deviation and the axial crystal growth rate. The diameter change rate is discretely accumulated over the prediction time interval to obtain the diameter change predicted by the mechanism model; The predicted base diameter is obtained by superimposing the diameter change onto the measured diameter at the current moment.
3. The method according to claim 1, characterized in that, The step of constructing an augmented bias state vector that matches the number of discrete prediction steps for each prediction time interval includes: Using the discrete prediction steps as the dimensional basis of the augmented deviation state vector, the augmented deviation state vector includes the system deviation state component at the current moment and each historical deviation state component that traces back the discrete prediction steps. An augmented state transition matrix is established such that, during the time recursion process of the augmented deviation state vector at each sampling time, each deviation state component in the augmented deviation state vector of the previous time is shifted and stored sequentially, and the deviation state component of the previous time at the current time is recursively used as the first component of the current time. An observation matrix is established that matches the augmented bias state vector, such that when the delay residual is used as the observation input for updating, the delay residual only forms an observation relationship with the historical bias state component in the augmented bias state vector corresponding to the discrete prediction step.
4. The method according to claim 1, characterized in that, The update of the deviation state component in the augmented deviation state vector corresponding to the discrete prediction step number adopts a robust update method, including: The difference between the delay residual and the estimated delay residual predicted based on the augmented bias state vector is calculated as the original innovation, and the innovation variance corresponding to the original innovation is calculated. The original information is standardized and then nonlinearly mapped using the Huber function. When the absolute value of the standardized information is less than or equal to a preset Huber threshold, its original value is maintained; when it is greater than the Huber threshold, it is compressed to the Huber threshold. The mapped value is multiplied by the standard deviation to obtain a robust innovation. This robust innovation is used to replace the original innovation in calculating the update amount. The robust innovation is also used to replace the original innovation in updating the bias state component in the augmented bias state vector corresponding to the discrete prediction step.
5. The method according to claim 1, characterized in that, Based on the measured diameter, the current trend and its estimation uncertainty include: Calculate the rate of change of the diameter at the current moment relative to the previous moment; The diameter change rate is estimated by using filtering methods with different smoothing coefficients to obtain fast trend estimates and slow trend estimates. Calculate the estimated uncertainties for the fast trend and the slow trend, respectively.
6. The method according to claim 5, characterized in that, The estimation of the diameter change rate using filtering methods with different smoothing coefficients includes: The rate of change of the diameter is estimated using a first exponentially weighted moving average to obtain a fast trend estimate; a second exponentially weighted moving average is also used to estimate the rate of change of the diameter to obtain a slow trend estimate; wherein the half-life of the first exponentially weighted moving average is less than the half-life of the second exponentially weighted moving average. The calculation of the estimation uncertainties for the fast trend and the slow trend, respectively, includes: The variances of the fast and slow trends are estimated using an exponentially weighted moving average method, which serves as the estimation uncertainty for the corresponding trends.
7. The method according to claim 1, characterized in that, The determination of trend confidence based on the relative magnitude of the current trend estimate and the estimate uncertainty includes: Using the standard deviation of the corresponding trend as a measure of the estimated uncertainty, a trend confidence score is constructed based on the estimated value and the estimated uncertainty of the current trend. The expression: ; in, This is the absolute value of the estimate of the current trend. The standard deviation of the corresponding trend. To prevent the use of tiny constants with a denominator of zero, the trend confidence increases with the increase of the estimated value of the current trend and decreases with the increase of the estimated uncertainty. When the trend confidence level is greater than or equal to the preset minimum confidence level threshold, the preset condition is met, and the corresponding trend correction amount is activated; When the trend confidence level is less than the preset minimum confidence level threshold, the preset condition is not met, and the corresponding trend correction amount is not activated.
8. The method according to claim 1, characterized in that, The trend correction amount is obtained by multiplying the estimated value of the corresponding trend by a preset gain coefficient; the convex combination of the result after superimposing or not superimposing the trend correction amount with the persistent prediction using the current measured diameter as the prediction value includes: The fusion weight is determined based on the predicted time interval. The fusion weight is used as the coefficient of the result after superimposed or unsuperimposed trend correction. The difference between the fusion weight and the preset standard number is used as the coefficient of the continuous prediction. The two are weighted and summed to obtain the final diameter prediction result.
9. A device for predicting the diameter of a silicon single crystal over multiple time intervals, characterized in that, The device includes: The acquisition module is used to acquire the measured diameter of the silicon single crystal growth furnace during the secondary equal-diameter growth stage; The generation module is used to generate the basic diameter prediction at each prediction time interval based on the measured diameter at the current time and the kinematic relationship of the crystal growth interface. The first prediction module is used to construct an augmented deviation state vector matching the discrete prediction step number for each prediction time interval. The augmented deviation state vector stores the system deviation estimate at the current moment and the system deviation estimates at each historical moment tracing back the discrete prediction step number in the same vector, with the discrete prediction step number as the dimension. When the actual diameter observation corresponding to the historical prediction arrives, the prediction residual of the historical prediction is used as the delay residual, and the delay residual is used as the observation input to update the deviation state component in the augmented deviation state vector corresponding to the discrete prediction step number. The system deviation estimate at the current moment in the augmented deviation state vector is extracted and superimposed on the basic diameter prediction to obtain the prediction result after deviation correction. The second prediction module is used to estimate the current trend and its estimation uncertainty based on the measured diameter, determine the trend confidence level according to the relative magnitude of the estimated value of the current trend and the estimation uncertainty, and when the trend confidence level reaches a preset condition, add the corresponding trend correction amount to the prediction result after deviation correction; otherwise, do not add it. The result after adding or not adding the trend correction amount is convexly combined with the continuous prediction using the current measured diameter as the prediction value to generate the final diameter prediction result for each prediction time interval.
10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 8.
