• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    In a sampling method with a sampling rate decision scheme, a mean absolute percentage error (MAPE) and a maximum absolute percentage error (MaxErr) of virtual measurements of all workpieces in a set of determinative samples (DS) and various index values representing various state changes of a machining tool (such as maintenance operation , Part replacement, parameter adjustment, etc.) and / or informational anomalies of the machining tool (such as abnormal process data, parameter drift / offset, abnormal measurement data, etc.) that occur in a manufacturing process, applied to an Automated Sampling Decision (ASD). To develop a sample sampling rate to further reduce VM sample accuracy.
    公开号:AT517250A2
    申请号:T50480/2016
    申请日:2016-05-27
    公开日:2016-12-15
    发明作者:Fan-Tien Cheng；Chung-Fang Chen；Jhao-Rong Lyu；Yao-Sheng Hsieh
    申请人:Univ Nat Cheng Kung；
    IPC主号:
    专利说明:

    BACKGROUND Field of the Invention
    The present invention relates to a sampling method with a sampling rate decision scheme. More specifically, the present invention relates to a sampling method for reducing and automatically adjusting a workpiece sampling rate.
    Description of the Prior Art
    Nowadays, most of the semiconductor and TFT LCD factories use spot checks to test the quality of each product or workpiece (ie, "wafers" in IC foundries or "glass substrates" in TFT LCD factories) that is machined by machining tools to monitor. In general, after each N (for example, 25) workpieces have been machined by the machining tool, a manufacturing system selects the Nth workpiece from each N workpieces as a (scheduled) workpiece expected for measurement, ie, the sampling rate is 1 / N determined. The workpiece expected to be measured will then be sent to a measuring tool for measurement, thereby checking the manufacturing quality of the machining tool. This conventional sampling method is based on the assumption that no unusual circumstances will occur suddenly during the manufacturing process and therefore the measurement result of the sampled product or workpiece can be used to determine the quality of the workpieces in the same lot with the sample taken at random. The measuring time and the tool required by the actual workpiece measurement lead to an increase in the cycle time and the manufacturing costs. Therefore, decreasing the sampling rate to the lowest possible value is an important issue for manufacturers to reduce cycle time and production costs. Incidentally, the value of conventional
    Workpiece sampling rate 1 / N selected only according to an empirical value of the production system without other technical basis and therefore can not be effectively adjusted.
    On the other hand, a virtual measurement (VM) can be used to reduce the frequency of actual measurement on the workpiece to reduce the sampling rate. However, if a production deviation occurs at the workpiece that is not scheduled for measurement, no actual measurement is available during this period for updating the VM models, resulting in a low accuracy of the VM prediction. Therefore, the predictive accuracy of VM models is influenced by how a suitable workpiece is sampled and obtained on a timely basis.
    Therefore, there is a need to provide a sampling method for overcoming the above-mentioned drawbacks of the conventional arts.
    SHORT VERSION
    An object of the present invention is to provide a sampling method for automatically adjusting and decreasing a sampling rate with respect to a measurement for workpieces.
    Another object of the present invention is to provide a sampling method for timely providing an actual measurement value of one for recalibrating or re-learning a VM model to thereby ensure the accuracy of the VM.
    In accordance with the above-mentioned objects, a sampling method is provided. In the sampling methods, a plurality of sets of historical machining data, which are used by a machining tool for machining a plurality of historical workpieces, and a plurality of historical measured values of the historical workpieces are collected according to the sets of historical process data. Then, a model generation step is performed and comprises: forming a conjecture model in accordance with a presumption algorithm by using the sets of historical process data and the historical measurements. Thereafter, a workpiece sampling rate represented by 1 / N is initialized, wherein the workpiece sampling rate is directed to select the Nth workpiece as a workpiece expected to be measured from every N workpieces processed by one machining workpiece. Then a workpiece is added to a set of determinative samples. Then, a set of process data used by the machining tool to machine the workpiece and an actual measurement of the workpiece corresponding to the set of process data are collected. Thereafter, the set of process data of the workpiece is input to the conjecture model, whereby a virtual measurement of the workpiece is calculated. Then, an absolute percentage error of the virtual measurement of the workpiece is calculated, and a step is performed to determine if the absolute percentage error of the virtual measurement is greater than a specification of a maximum virtual measurement error defined for the machining tool, thereby producing a first result is obtained. If the first result is true, an OOS (Out of Spec) counter is incremented by one. If the first result is false, a Mean Absolute Percentage Error (MAPE) of virtual measurements of all workpieces in the set of determinative samples is calculated, and a step is taken to determine if the mean absolute percentage error is equal to an upper control limit of MAPE or greater, thereby obtaining a second result. If the second result is true, the workpiece sampling rate is increased by decreasing N, the set of determinative samples is cleared and the OOS counter is set to zero. If the second result is false, a step is taken to determine if the number of workpieces in the set of determinative samples equals or exceeds a threshold of the number of determinate samples, thereby obtaining a third result.
    If the third result is false, the workpiece sampling rate is kept unchanged. If the third result is true, a maximum Absolute Percent Error (MaxErr) of virtual measurements of all workpieces in the set of determinative samples is calculated, and a step is taken to determine if the maximum absolute percentage error is less than an upper control limit of MaxErr, which gives a fourth result. If the fourth result is true, the sample sampling rate is decreased by incrementing N, the set of determinative samples is cleared and the OOS counter is set to zero. If the fourth result is false, an oldest workpiece in the set of determinative samples is discarded and the workpiece sampling rate is kept unchanged.
    According to an embodiment of the present invention, if the first result is true, a step is performed to determine whether the OOS counter is equal to or greater than an OOS threshold, thereby obtaining a fifth result. If the fifth result is true, the sample sampling rate is increased by decreasing N, the set of determinative samples is cleared and the OOS counter is set to 0.
    According to an embodiment of the present invention, a minimum workpiece sample extraction rate represented by 1 / Nmax and a predetermined workpiece sample extraction rate represented by 1 / Ndefauit are obtained, where Nmax is a largest value of N and Ndefauit is a default value of N. Then, a third conservative factor is multiplied by Nmax to obtain a test value. A step is performed to determine if the check value is greater than Ndefauit, thereby obtaining a sixth result. If the sixth result is true, Ndefauit is set equal to the check value.
    According to an embodiment of the present invention, a first checking step is performed to check whether a state change of the machining tool occurs, thereby obtaining a first check result. If the first check result is true, the set of determinative samples is cleared and the OOS counter is set to 0. A second check step is performed to check if N is greater than Ndefauit, thereby obtaining a second check result. If the second check result is true, N is set to Ndefauit.
    According to an embodiment of the present invention, in the aforementioned sampling method, the model generating step is further performed to form a DQIx (Process Data Quality Index) model and a GSI (Global Similarity Index) model, and calculate a DQIX threshold and a GSI threshold using historical process data sets. Then, a workpiece sample measuring step is performed. By doing
    For the workpiece sampling step, the set of process data of the workpiece is input to the DQIx model and the GSI model, whereby a DQIX value and a GSI value of the set of process data of the workpiece are obtained. After this, a workpiece counter is incremented by 1. If the first check result is true, a third check step is performed to check if the DQIX value is equal to or lower than the DQIX threshold, thereby obtaining a third check result. If the third verification result is wrong, a measurement of the workpiece is skipped. If the third check result is true, a fourth check step is performed to check if the workpiece counter is greater than or equal to N, thereby obtaining a fourth check result. If the fourth verification result is true, a measurement is made on the workpiece and the workpiece counter is set to 0. If the fourth check result is false, a fifth check step is performed to check if the GSI value of the workpiece is equal to or less than the GSI threshold, thereby obtaining a fifth check result. If the fifth check result is true, the measurement of the workpiece is skipped.
    According to an embodiment of the present invention, in the sampling measurement method, a reference model is formed according to a reference prediction algorithm by using the sets of historical process data and the historical measurement values, the assumption algorithm being different from the reference prediction algorithm. An RI (Receptance Index) threshold is calculated based on a maximum tolerable margin of error defined by errors of virtual measurements obtained from the presumption model. The set of process data of the workpiece is input to the reference model, whereby a reference prediction value of the workpiece is calculated. An overlap area between the statistical distribution of the virtual measurement of the workpiece and the statistical distribution of the reference prediction value of the workpiece is calculated, thereby producing an RI value of the workpiece. If the fourth check result is false, a sixth check step is performed to check if the Rl value of the workpiece is equal to or greater than the Rl threshold, thereby obtaining a sixth check result. If the sixth verification result is true, the measurement of the workpiece is skipped.
    According to an embodiment of the present invention, a measurement is performed on the workpiece and the workpiece counter is set to 0 if the fifth verification result or the sixth verification result is false and the GSI values of k workpieces processed before the workpiece are all greater than that GSI threshold or the RI values of k workpieces that were machined before the workpiece are all smaller than the RI threshold, where k is a positive integer.
    In accordance with the above-mentioned objects, a computer program product stored on a non-volatile, tangible, computer-readable recording medium is provided. When the computer program product is loaded and executed by a computer, the aforementioned sampling methods are performed.
    Thus, with the application of the embodiments of the present invention, the workpiece sampling rate can be automatically adjusted and significantly lowered, and the VM accuracy can be effectively ensured.
    It should be understood that both the foregoing general description and the following detailed description are exemplary and are intended to further illustrate the claimed invention.
    BRIEF DESCRIPTION OF THE DRAWINGS
    The invention may be more fully understood by reading the following detailed description of the embodiment with reference to the accompanying drawings, in which:
    FIG. 1 is a schematic diagram for explaining a reliability index (RI) according to some embodiments of the present invention; FIG.
    FIG. 2 is a schematic diagram for defining an RI threshold (RIT) according to some embodiments of the present invention; FIG.
    3 is a schematic flow diagram illustrating a sample sampling method with a sampling rate decision scheme according to various embodiments of the present invention;
    FIGS. 4A to 4C are schematic diagrams used to explain a step of adjusting a workpiece sampling rate according to some embodiments of the present invention;
    FIGS. 5A and 5B are schematic flow diagrams showing a step of adjusting a workpiece sampling rate according to various embodiments of the present invention;
    FIG. 6 is a schematic flow diagram illustrating a
    Workpiece sample measuring step according to various embodiments of the present invention; and
    7 is a schematic flowchart showing a virtual measurement method according to various embodiments of the present invention.
    DETAILED DESCRIPTION
    Reference will now be made in detail to the embodiments of the present invention, examples of which are illustrated in the accompanying drawings. Wherever possible, the same reference numbers have been used in the drawings and the description to designate the same or similar parts.
    Embodiments of the present invention are directed to a sampling method that applies a mean absolute percent error (MAPE) and a maximum absolute percent error (MaxErr) of virtual measurements of all workpieces in a set of deterministic samples (DS) to set a sample sampling rate. The sampling method also combines various index values which can detect various state changes of a machining tool (such as maintenance operation, part change, parameter setting, etc.) and / or informational abnormalities of the machining tool (such as abnormal process data, parameter drift / offset, abnormal measurement data, etc.). which appear in a manufacturing process are used to develop an automated sampling decision (ASD) scheme to reduce a sampling rate while maintaining VM accuracy. The indices include a reliability index (RI), a global similarity index (GSI), a process data quality index (DQIX), and a metrics data quality index (DQIy). The R1 value, the GSI value, the DQIx value and the DQIy value used in the embodiments of the present invention may be related to U.S. Patent No. 8,095,484 B2.
    Embodiments of the present invention may be combined with the VM system disclosed by U.S. Patent No. 8,095,484 B2. U.S. Patent No. 8,095,484 B2 is hereby incorporated by reference. The RI value is designed to calibrate the reliability level of a virtual metric. The GSI value is used to evaluate the degree of similarity between the current set of input process data and all sets of process data used to form and train a presumption model. The GSI value is provided to help the RI value calibrate the reliability level of the VM system. The DQIX value is used to evaluate whether a set of process data used to make a workpiece is abnormal, and the DQIy value is used to evaluate whether the measurement data of the workpiece is abnormal.
    In the following, theoretical basics regarding the RI value (RI model), the GSI value (GSI model), the DQIX value (DQIx model), and the DQIy value (DQlyModel) are explained first. RI and the GSI are used to learn in real time if the VM value is reliable. The RI model is used to calculate an RI value between 0 and 1 by analyzing the process data of the machining tool, thereby determining whether the virtual measurement result can be trusted. The GSI model is used to calculate the GSI value for the process. The GSI value is defined as the degree of similarity between the current set of input process data and all sets of process data used to formulate or teach the models.
    Referring to Table 1, it is assumed that there are n sets of historical data including process data (.Xi, i = 1, 2,..., N) and corresponding actual
    Measured values (yz ·, ί = 1,2, ···, η) are collected, each set of process data p containing individual parameters (from parameter 1 to parameter p), in particular Xi = [xu, jc.2, .. ., xp] r. In addition, (m-n) sets of process data are also collected in actual production, but no actual measurements beyond yn + i are available. That is, only the first among (m-n) pieces of the products is selected and actually measured. In current manufacturing practice, the actual measurement yn + i obtained is used to derive and evaluate the quality of the (m-n-1) pieces of the products.
    Table 1:
    As shown in Table 1, y1, y2r -, yn are historical measurements and yn + 1 is the actual measurement of the first piece of manufactured products. in the
    In general, a set of actual measurements {yu ΐ = 1,2, ···, η) is a normal distribution with a mean value μ and a standard deviation σ, i.e.,
    All the actual readings can be standardized with respect to the mean and standard deviation of the sample set {yt, i = 1,2, ···, n). Their standardized values (also referred to as Z-score or standardized random variables), z z ···, z are thus derived, each Z-score having an average (expected value) 0 and a standard deviation (variance) 1, i.e.,
    Zy. ~ N (0, l). With respect to the actual measurements, a corresponding zy near 0 indicates that the actual reading approaches the mean of the specification. The equations for standardization are listed below:
    (1)
    (2)
    (3) where yt is the i-th actual measurement, zyi is the standardized i-th actual measurement, y is the average of all actual measurements, and is the standard deviation of all actual measurements.
    The discussion herein adopts a neural network (NN) algorithm as the presumption algorithm for establishing the presumptive model performing a virtual measurement and uses something like a multi-regression (MR) algorithm to obtain the reference algorithm for establishing the reference model, which is used as a basis for comparison the conjecture model serves to be. However, the present invention may also employ other algorithms to be the presumption algorithm or reference algorithm, such as a backpropagation neural network (BPNN) algorithm, a general neural network regression algorithm (e.g. GRNN) algorithm, a Radial Basis Function Neural Network (RBFNN) algorithm, a Simple Recurrent Network (SRN) algorithm, a Support Vector Data Description (SVDD) Vector Algorithm, an assist vector machine (SVM). Support Vector Machine) algorithm, multi-regression (MR) algorithm, partial least squares (PLS), non-linear iterative partial least squares (NIPALS) or generalized linear models (GLM), as long as the reference algorithm is different from the courage such as an SVM algorithm and other related algorithms, and thus the present invention is not limited thereto.
    If the NN and MR algorithms are used, if their convergence conditions are both that SSE (Sum of Square Error) is minimized with n -► °°, their standardized predictive measures (defined as resp. ) be the same as the standardized actual
    Measured value zy .. In other words, if n-> °°, represent cy. = Z ^ = Z ^. all the standardized actual metric, but have different names because they have different purposes and different estimation models.
    Therefore, type zym ~ afj and Z_ ,. , ~ w (// z ^, σ | J that Zym and Z, v., share the same statistical distribution, but it is due to the existence of different estimation models that the estimates of mean and standard deviation differ between these two prediction algorithms the estimation equation of the standardized mean [fiZy = ZßNJ and the estimation equation of the standard deviation (σζ = σζ_) with respect to the NN conjecture model of the estimation equation of the standardized
    Mean value ψζ = Ζλ) and the estimation equation of the standard deviation 'z /' ri '[σζ ^ = σζ. ) with respect to the MR reference model.
    The RI is designed to calibrate the reliability level of the virtual metric. The RI should therefore estimate the degree of similarity between the statistical distribution Zp of the virtual metric and the statistical distribution
    Take into account the actual measured value. However, if the virtual measurement is applied, no actual metric can be used to verify the credibility of the virtual metric. (In particular, a virtual measurement becomes unnecessary when actual measurements are obtained.) Instead, the present invention adopts the estimated statistical distribution Zpr by the reference algorithm, which is one such as the MR algorithm, to replace z. The reference algorithm can also be such as a
    Time series algorithm and other related algorithms, and thus the present invention is not limited thereto.
    Referring to Fig. 1, Fig. 1 is a schematic diagram for explaining the reliability index (RI) according to the preferred embodiment of the present invention. The RI of the present invention is defined as the overlap area value (overlap area A) between the statistical distribution Zp of the virtual
    Measured value from the conjecture model (formed by such as the NN algorithm) and the statistical distribution Z ^ of the reference prediction value from the
    Reference model (formed by such as the MR algorithm) defined. As such, the RI equation is listed below:
    (4) with μ = Ζβχ for Z.M <Z.ri μ = Ζ. ffir Ζ, ή <Z, ni where σ is set to 1.
    The RI increases with increasing overlap area A. This phenomenon indicates that the result obtained by the presumption model is closer to that obtained from the reference model and thus the corresponding virtual measurement is more reliable. On the other hand, the reliability of the corresponding measured value decreases with decreasing RI. If the distribution estimated from ZyN completely overlaps with the distribution estimated from Zyr, then according to
    Distribution Theory of Statistics the Rl value equals 1; and when these two distributions are largely separated, the Rl value approaches zero.
    The following explains the method of calculating the statistical distribution of the virtual measurements {Z-n_ and σζ) from the presumption model.
    In the NN conjecture model, if the convergence condition is to minimize SSE, then it can be assumed that "for given Z Z ^
    The distribution with the mean value μζ ^ and the standard deviation σζ is ", ie, for given Zx .. Z ^ ~. / V (pZy, ozJ, where for μζ ^ the NN estimation equation £ iZy = Z ^ N crf ^ the NN estimation equation σ2ζ = <rz.
    Before the NN conjecture model is constructed, the process data must be standardized. The equations for standardizing the process data are listed below: (5)
    "J
    (6)
    (7) where xi: j is the jth process parameter in the ith set of process data, zx is the standardized jth process parameter in the ith set of
    Process data is; xj is the mean value of the jth process data; σ is the standard deviation of the jth process data. j
    The n sets of standardized process data zx, i = 1, 2, ..., n; j = 1,2, ..., p) and the n standardized actual measured values (zy, i = J, 2, ... , n) are exploited to form the NN conjecture model. The m sets of standardized process data [zx .., i = l, 2, ..., m; j = 1,2, ..., pj are then input to the NN conjecture model to obtain the corresponding standardized virtual measures: Zp, ZpN2, ··· 'zyN' 'z9n' + i 'Zy'Nm
    Accordingly, the estimate of μζ ^ (i.e., μζ =) and the estimate of σζ (i.e., σζ = σζ.Ν) can be calculated as follows:
    (8th)
    (9)
    (10) where Z- is the average of the standardized virtual measurements.
    y N
    The following will explain the method of calculating the reference prediction values (Z ~ and σ7) from the MR model.
    The basic assumption of the MR is that "for given Zx Z, the distribution i, j sfi with the mean μζy_ and the standard deviation is σζ", ie for given Zx we have Z ,, ~ Ν μζ, ol), where for μ the MR estimation equation μ = z0 holds and for / r / X "yi / Zyi w Zyi sn o2Zy the MR estimation equation a2Zy is = σ |. applies.
    The MR relationship between the n sets of standardized process data px i = l, 2, - ·, n; j = 1,2, -, p) and the n standardized actual measured values [zy, i = l, 2, ···, «), the weighting factors ßr = [ßrO'ßri'ßr2 '-' ßrpY ' which correspond to these P parameters can be defined by using the MR analysis. The relationship between Z and Zx is therefore constructed as i, j follows:
    (11)
    It was
    (12) and
    (13)
    The least squares method may use the estimation equation of βr,
    received as
    (14)
    Therefore, the MR reference model can be obtained as
    / = 1,2, ..., n, n + 1, ..., m
    Therefore, during the presumption phase, after inputting a set of process data, its MR estimate corresponding thereto can be obtained through the equation (15). The MR estimation equation of the standard deviation σ is σζ zy yr with
    (16)
    (17)
    After obtaining the NN estimation equations (Z ~ and σζ.) And the MR
    yNi yN
    Estimation equations (Z ^ and σζ.) Can be applied to their normal distribution curves, as shown in FIG. Therefore, the RI of each virtual measurement can be derived by calculating the overlap area value (overlap area A).
    After receiving the RI, the RI threshold (RIT) must be defined. If RI> RIT, the reliability level of the virtual measure is acceptable. A systematic approach for determining R Ij will be described below.
    Before determining the RIT, it is necessary to define a maximum tolerable error limit (EL). The virtual metric error is an absolute percentage of the difference between the actual metric yt and yNi derived from the NN
    Assumption model, divided by the average of all the actual measured values y.
    (18)
    The value EL may then be specified based on the error defined in equation (18) and the accuracy specification of the virtual measurement (VM). As a result, RIT is defined as the value of R1 corresponding to the EL as shown in FIG. ie .:
    (19) where μ and σ are defined in equation (4), and
    (20) where Oy is specified in equation (3).
    The following explains a method of forming a GSI model. As mentioned above, when the virtual measurement is applied, no actual measurement is available to verify the accuracy of the virtual measurement. Therefore, instead of the standardized actual measurement value z, the standardized MR prediction value Z% is assumed to calculate the RI. These
    Replacement can cause unavoidable calibration errors in the RI. To compensate for this unavoidable replacement, a global similarity index (GSI) is provided to help the RI calibrate the reliability level of the virtual measurement and to identify the key process parameters with large deviations (z-score values or values of the standardized random variables). ,
    The GSI evaluates the degree of similarity between each set of process data and the model set of process data. This model set is derived from all of the historical process data sets used to form the presumption model.
    The present invention may use a statistical distance measurement, such as the Mahalanobis distance, to quantify the degree of similarity. The Mahalanobis distance is a distance measurement performed by P.C. Mahalonobis was introduced in 1936. This measurement is based on a correlation between variables to identify and analyze different patterns of sampling sets. The Mahalonobis distance is a useful way to determine a similarity of an unknown sample set to a known one. This method takes into account the correlation of the data set and is scale-invariant, that is, it is independent of the scale of measurements. If the data set has a high similarity, the calculated Mahalanobis distance will become comparatively small.
    The present invention uses the calculated GSI size (using the Mahalonobis distance) to determine if the newly entered set of process data is similar to the model set of process data. If the calculated GSI is small, the newly entered sentence is comparatively similar to the model set. Therefore, the virtual measurement of the newly entered (highly similar) sentence is comparatively accurate. Conversely, if the calculated GSI is too large, the newly entered sentence is slightly different from the model set. As a result, the estimated virtual measure estimated in accordance with the newly entered (slightly similar) set has a low level of confidence in accuracy.
    The equations for computing the standardized process data Zx of the presumption model are shown in Equations (5), (6) and (7). First, the model set of process parameters is defined as XM = [xm, i> xm, 2> - 'xm, pY, where xMJ is 3c'j, j = 1,2, -, p such that each element in the model set after standardization (also referred to as the standardized model parameter ZMJ) has a value of 0. In other words, all of the elements in ZM = [zM1, ZM2, ..., ZMJ are 0. Thereafter, the correlation coefficients between the standardized model parameters are calculated.
    Assuming that the correlation efficiency between the s-th parameter and the t-th parameter is rst and that k records exist, then
    (21)
    After calculating the correlation coefficients between the standardized model parameters, the matrix of correlation coefficients can be obtained as
    (22)
    Assuming that the inverse matrix (if;) of R is defined as A, holds
    (23)
    Thus, the equation for calculating the Mahalanobis distance (öf) between the standardized λ-th set of process data (zj and the standardized model set of process data (zM) is as follows:
    (24)
    Finally, equation (25) is obtained.
    (25)
    The GSI of the standardized λ-th set of process data is then equal to Df / p.
    After receiving the GSI, the GSI threshold (GSIT) should be defined as follows:
    (26)
    In the so-called LOO (Leave-One-Out) method, a sample data set from all sets of process data used to form models is selected as a simulated running test sample set, and then the remaining sets of process data are used to generate a GSI Model and hereafter the GSI model is used to calculate a GSI value for the test sample set, ie GSILoo. The above steps are repeated on all of the sample data sets (process data) used to form models, thus calculating all the GSI | _oo values of the respective sample data sets. Thus, for example, the GSILOo shown in equation (26) stands for the 90% trimmed average of all GSI | _oo values computed by the sample data sets, respectively. The "a" shown in equation (26) is between 2 and 3, and can be slightly adjusted in accordance with actual conditions, where "a" is predetermined to be 3.
    A method of constructing a DQIx model is described below. Assume that n sets of historical process data are received to construct the first DQIx model, with each set of historical process data constructed from p parameters. These n sets of historical process data are used to generate p eigenvectors with p eigenvalues (^ ^ ^ ··· ^ λρ) in descending order by principal component analysis (PCA). Then, a set of k significant eigenvalues (with λ> 1) is selected for the construction of a feature extraction matrix M which is expressed as:
    (27)
    The procedure for calculating the DQIX values is explained as follows.
    First, an equation (28) is applied to transform the i-th input set of process data X to k data feature variables A. = [aI, a2, ..., alc].
    (28)
    Then these k data feature variables are transformed to% A = -zai> za2> -> zak 1, which is then converted by the Euclid's Distance (ED) algorithm into a consolidated index, ie the DQIx -Value:
    (29) where i represents the ith input set of process data;
    Zaj: The mean of the jth standardized variables of the training samples.
    Theoretically, the value of Za] is equal to zero, and for this the equation (29) can be simplified as:
    (30)
    Meanwhile, the cross validation exit method (LOO) is used to determine the process data quality threshold (DQIXt) as:
    (31)
    In the so-called cross-validation exhaustion (LOO) method, a sample record is selected from all sets of process data used to form models as a simulated running test sample set, and then the remaining sets of historical process data are used to construct a DQIx model, and hereafter the newly formed DQIx model is used to calculate a DQIX value for the simulated tracking test sample set, ie, DQIxloo. The above steps are repeated on all the check records (process data) used to form models, thereby calculating all the DQIxLoo values of the respective sample records. Thus, for example, the DQ ^ xLOO shown in equation (31) stands for the 90% trimmed average of all the DQIxloo values computed by the sample data sets, respectively. The "a" shown in equation (31) is between 2 and 3, and can be slightly adjusted in accordance with the actual conditions, with "a" preset to be 3.
    It is noted that the feature extraction matrix M and the DQIX value constitute a DQIX model, and the DQIx model will be updated in accordance with a relearn or calibrate condition (in the model update procedure).
    After that, the Z-score values of the historical process data are calculated. Then a DQIy model is generated, where the DQly model is composed of m similar patterns.
    In the present embodiment, the m-like patterns {Pi, P2, ..., Pm} are sorted from Z score values of these n sets of historical process data by applying the adaptive resonance theory 2 (ART2) with p = 0.98.
    The method of calculating DQly values will be described below. First, when a new actual metric yj is collected, the Z-score values become ZXG. in accordance with the actual measured value yj by the adaptive resonance theory 2 (ART2), to obtain from the similar patterns {Pi, P2, ..., Pm} the most similar pattern P ^ = [X ; 1, X9; 2, ... , X9; V] to look for. Then the v-samples (v> 2) within P = [X 1, X 2, ..., X V] with their corresponding actual measurements Yq = [y9> i, y9> 2, ..., y9> v] and this new actual measure y, used to calculate the DQI and threshold (DQlyr) of the DQIy. The DQIy of y is obtained as normalized variability (NV):
    (32)
    (33) with yq: the mean of all yq, i; v: the number of samples within the pattern Pq.
    The DQIyr of a particular pattern Pq is defined to be the maximum tolerable variance of the Pq. It is assumed that y, the maximum tolerable
    Measured value having the maximum tolerable variance in Pq, then yt of this paragraph can be represented as:
    (34) where Rmax is the maximum tolerable variance;
    (35) where Rp, i = 1, 2, m, is the area in the pattern P. and m is the total number of all groups of similar patterns.
    By adding yt to the similar pattern Pq, DQlyr can be obtained as:
    (36)
    If, after receiving DQIy and DQlyi, DQI> DQlyr is true, it means that the new actual metric is abnormal; otherwise, the actual reading is normal.
    The above-mentioned PCA, LOO, ART2, Z-score and ED algorithms are all known to those skilled in the art and, therefore, the details thereof are not described herein.
    Referring to FIG. 3, FIG. 3 is a schematic flow diagram showing a sampling method with a sampling rate decision scheme according to various embodiments of the present invention. First, a workpiece sampling rate 1 / N is initialized (step 110), wherein the workpiece sampling rate is directed to select from every N workpieces processed by a machining tool the Nth workpiece as a workpiece expected to be measured. For example, a conventional workpiece sampling rate is defined such that out of every 25 workpieces, the 25th workpiece (workpiece counter = 25) is selected as a workpiece to be measured after every 25 workpieces have been processed by the machining tool. Embodiments of the present invention assume a variable "workpiece counter" to allow the application of the workpiece sampling rate, whose initial value is zero. In other words, the workpiece counter is the number of workpieces that have been processed by the machining workpiece but are not subjected to measurement after the last workpiece has been measured. In step 110, N may be initialized to 1 or a number determined by a conventional manufacturing system. Theoretically, the workpiece with the workpiece counter, which is equal to N, will be selected as the workpiece expected to be measured. After that, a plurality of sets of historical process data used by the machining tool for machining a plurality of historical workpieces are collected, and a plurality of historical measured values of the historical workpieces corresponding to the sets of historical process data are collected (step 120). Then, a model generation step 130 is performed to form an RI model (a presumption model and a reference model), a DQIx model, a DQIy model, and a GSI model, and a DQIX threshold, a DQly threshold, and a GSI threshold using the conjecture model to calculate virtual measurements of workpieces. The details of the model generation step 130 are explained above. After the model generation step 130 is completed, a step 140 of adjusting the workpiece sampling rate is performed to adjust the value of N of the workpiece sampling rate. After the value of N is automatically adjusted, a workpiece sample measuring step 150 may be performed to further reduce the workpiece sampling rate. The step 140 of adjusting the workpiece sampling rate is based on the virtual measurement (VM) accuracy for adjusting the
    Workpiece sampling rate, wherein if the VM accuracy is poor, the workpiece sampling rate is increased (by decreasing N); and, if the VM accuracy is good, the workpiece sampling rate is decreased or maintained (by increasing or leaving N). Thus, after step 140 of adjusting the workpiece sampling rate, the workpiece sampling rate has been appropriately adjusted. Then, a user may use the above-mentioned workpiece sampling rate to perform the workpiece sampling measurement step 150.
    Hereinafter, technical bases of the workpiece sampling rate adjusting step 140 will be explained. In embodiments of the present invention, each workpiece that is fed to the machining tool is added to a set of deterministic samples (hereinafter referred to as a DS set) and becomes a mean absolute percent error (MAPE) and a maximum absolute percentage error (MaxErr) of virtual measurements of all workpieces in the DS set, referred to as MAPEds and MaxErrDs, for determining the VM accuracy, where the mean absolute percentage error (MAPE) and the maximum absolute percentage error (MaxErr) are defined as:
    (37) (38) representing a VM value; y, represents an actual measurement; y represents a setpoint; and n represents the number of workpieces in the DS set.
    The DS set must collect a sufficient number of workpieces to be representative. However, if there are too many workpieces in the DS set, much time will be spent. Therefore, a threshold of the number of determinative samples (referred to as TDs) must have an appropriate value. The closer MAPE and MaxErr are to zero, the better VM accuracy is achieved. Embodiments of the present invention use a specification of a virtual maximum measurement error for the machining tool (referred to as SPECMax), an upper control limit of MaxErr (referred to as UCLMax), and an upper control limit of MAPE (referred to as UCLmape) to determine if the workpiece sampling rate is adjusting with the equations of UCLMax and UCLmape set out below:
    (39) where SPECMax varies due to the physical properties of the machining tool; α is a first conservative factor; β is a second conservative factor; 0> a, ß <1; and a> ß. The aforementioned Tos, α, β can be determined by sensitivity analysis. For an example of Plasma Enhanced Chemical Vapor Deposition (PECVD), a tool may be Tds 5; SPECMax can be 1.08%; can be α 0.85; and can be .65, and thus has 0.92% UCLm3x and 0.70% UCLmape.
    Referring to FIGS. 4A to 4C, FIGS. 4A to 4C are schematic diagrams used to explain the sample sampling rate adjusting step 140 according to some embodiments of the present invention, where "o" is used for MAPEDs (a mean absolute percent error of virtual measurements of all workpieces in the DS set); for MaxErrDs (a maximum absolute percentage error of virtual measurements of all workpieces in the DS set); "1-3" in the "o" and (the DS set) for the first workpiece to the third workpiece, which are successively supplied to the machining tool stands; "2-6" in the "o" and (the DS set) for the second workpiece up to the sixth workpiece, the
    Processing tool are supplied in succession, stands; "3-7" in the "o" and (the DS set) for the third workpiece to the seventh workpiece, which are fed to the machining tool sequentially, and "©" for a 00S- (Out of Spec - outside the specification) event is.
    As shown in Fig. 4A, the cases are shown in which the workpiece sampling rate is left unchanged. During a process of collecting determinative samples (workpieces), while the number of determinative samples collected in the DS set is less than Tds, MAPEds is used to check whether the current one
    Workpiece sampling rate 1 / N needs to be changed or not. In this example, TDs are set to 5 for illustrative purposes. As shown in the left half of Fig. 4A, when the MAPEds of each of the DS sets formed of the first four workpieces ("1", "1-2", "1-3", "1-4") less than UCLMApe is left, the sample sampling rate unchanged. When the number of determinative samples collected in the DS set reaches TDs (5), not only MAPE dsj but also MaxErros is used to calculate the
    Determine workpiece sampling rate 1 / N. As shown in the right half of FIG. 4A, when the MaxErrDS of the DS set ("1-5") composed of the first workpiece to the fifth workpiece is larger than UCLMax and the MAPEds of the DS sentence (" 1-5 ") is still smaller than UCLmape, which means that the VM accuracy is moderate, leaving the sample sampling rate unchanged. After that, the oldest sample in the DS sentence is discarded (step 274), thereby preventing too many workpieces in the DS sentence from spending too much time.
    As shown in Fig. 4B, the cases are shown in which the workpiece sampling rate 1 / N is decreased (N is increased). In this example, TDs are set to 5 for illustrative purposes. If the number of determinative samples collected in the DS set ("1-5") is equal to TDS (5), if the MAPEds of the DS set ("1-5") is smaller than UCLmape and the MaxErrDs of the DS set ("1-5") is also smaller than UCLmape, which means that the VM accuracy is good and the VM values are accurate enough to reflect the actual readings that will decrease the sample sampling rate 1 / N. Then, the DS set is cleared and an OOS counter is reset to 0 (step 286), the function of the OOS counter will be explained later. The deletion of the DS set means to start a fresh new collection of determinative samples, that is, the number of workpieces in the DS set (referred to as SIZEDs) is reset to 0; and resetting the OOS counter to 0 means that the frequency of occurrence of OOS is counted again.
    As shown in the left half of FIG. 4C, cases are shown in which the workpiece sampling rate 1 / N is increased (N is decreased). If the number of determinative samples collected in the DS set is less than TDs and the VM value MAPEDs is greater than UCLmape, it means that the VM accuracy is poor and the VM value fails to reflect the actual measurement of the workpiece, and it is therefore necessary to increase the workpiece sampling rate 1 / N. Then, the DS set is cleared and an OOS counter is reset to 0 (step 286).
    When a percentage of an absolute error of a VM value of a workpiece is larger than the SPECMax of the machining tool processing the workpiece, as shown in the right half of FIG. 4C, it means that the VM value has a very large error or the VM Accuracy decreases rapidly, so that an OOS event occurs. In order to avoid the workpiece sampling rate 1 / N being adjusted too frequently or too hastily, the present embodiment introduces the "OOS counter" to count the frequency of occurrence of the OOS error. If the OOS counter is greater than or equal to an OOS threshold (eg, 2, ie, at least two consecutive OOS events have been recorded), then the workpiece sampling rate 1 / N is increased, and then the DS set is cleared and turned on OOS counter is set to 0 (step 286).
    A change in state of the machining tool may occur when tool maintenance, repair, or recipe adjustment is performed. This change of state may be modified by an event sent by the manufacturing execution system. At this moment, the current N value may no longer be suitable for use because the process characteristics have changed. The present embodiment obtains a minimum workpiece sampling rate represented by 1 / Nmax and a predetermined workpiece sampling rate represented by 1 / Ndefauit to redefine the value of N, where Nmax is a maximum value of N and Ndefauit is a default value of N. If (γ xNmax) is greater than Ndefauit, then Ndefaux is set to (γ xNmax), where γ is a third conservative factor and 0 <γ ^ 1. In some embodiments, (γ xNmax) is rounded up to the nearest integer. In the example of the PECVD, γ can be 2/3, and therefore Ndefauit is 4 when Nmax is 6. If a state change of the machining tool occurs and the current workpiece sampling rate 1 / N (eg, N = 5) is smaller than 1 / Ndefauit (N> Ndefauit), the workpiece sampling rate 1 / N is set to 1 / Ndefauit, or (eg, N = 2) the current workpiece sampling rate 1 / N is kept unchanged.
    Next, an operation process of the step 140 (as shown in FIG. 3) of adjusting the workpiece sampling rate will be explained, wherein the step 140 includes a sampling rate adjusting step 200 shown in FIGS. 5A and 5B. In the sampling rate adjusting step 200, first, a set of process data of a workpiece currently being machined by the machining tool and an actual measured value of the workpiece corresponding to the set of process data are obtained, the machining tool using the set of process data to machine the workpiece. Then, optionally, the DQIX value, the DQIy value, the GSI value, or the Rl value of the workpiece may be checked to determine if they meet the requirements of their respective thresholds, i.e., if RI ^ RIt; GSI ^ GSIt; DQIx ^ dqiXt; DQIy ^ DQlyT (step 210). If the result is false ("No"), it means that the set of process data and / or the actual measured value of the workpiece are not reliable and the sampling rate adjusting step 200 can not proceed, and therefore, step 280 is performed to set the workpiece sampling rate 1 / Leave N unchanged (ie, the value of N remains unchanged). If the result of step 210 is true ("yes"), it means that the set of process data and / or the actual measurement of the workpiece are reliable, and step 220 may be performed to set the workpiece to a set of determinate samples (FIG. a DS sentence). Thereafter, the set of process data of the workpiece is input to the guess model generated by the model generation step 130 (shown in FIG. 3), whereby an absolute percentage error of the virtual measurement of the workpiece, a mean absolute percentage error (MAPEds) of virtual measurements of all workpieces in the DS set and a maximum absolute percentage error (MaxErrDs) of virtual measurement errors of all workpieces in the DS sentence are calculated (step 230).
    Then, step 240 is executed to determine whether the absolute percentage error of the virtual measurement of the workpiece is greater than the SPECMax of the machining tool to thereby obtain a first result. If the first result is true ("Yes"), the OOS (out of specification) counter is incremented by one. If the first result is false ("No"), step 260 is executed to calculate and determine whether the mean absolute percent error (MAPEDs) of virtual measurements of all workpieces in the DS set is greater than or equal to UCLmape, whereby a second result is obtained. If the second result is true ("Yes"), step 284 is executed to increase the workpiece sampling rate 1 / N (i.e., decrease the value of N, for example, subtract 1 from N). Next, step 286 is executed to reset the OOS counter to 0 and clear the DS set, that is, set the number of workpieces in the DS set (SIZEds) to zero. If the second result is false ("No"), step 270 is performed to determine if SIZEds is greater than or equal to the threshold of number of determinative samples (TDs) to obtain a third result.
    If the third result is false ("No"), step 280 is performed to leave the workpiece sampling rate 1 / N unchanged. If the third result is true ("Yes"), step 272 is executed to calculate and determine whether the maximum absolute percent error (MaxErros) of virtual measurements of all workpieces in the DS set is less than UCLMax, thereby fourth result is obtained. If the fourth result is true ("yes"), step 282 is executed to decrease the workpiece sampling rate 1 / N (that is, to increase the value of N, for example, to add 1 to N). After that, step 286 is executed to clear the DS set and reset the OOS counter to zero. If the fourth result is false ("No"), step 274 is performed to discard the oldest workpiece in the DS sentence, and step 280 is executed to leave the workpiece sample extraction rate unchanged.
    If the first result (step 240) is true ("yes"), step 250 is performed to determine if the OOS counter is greater than or equal to an OOS threshold (eg, 2), thereby obtaining a fifth result , If the fifth result is true ("Yes"), step 284 is executed to increase the workpiece sampling rate 1 / N (ie, decrease N, for example, subtract 1 from N), and then step 286 is executed to set the DS Delete set and set the OOS counter to 0.
    After step 280 or step 286 has been performed, the minimum sample sampling rate represented by 1 / Nmax and the predetermined sample sampling rate represented by 1 / Ndefauit are obtained, where Nmax is the maximum value of N used by the processing tool in its operating history. Then, step 290 is performed to determine if (y xNmax) is greater than Ndefauit, thereby obtaining a sixth result, where γ is the third conservative factor and 0 <γ ^ 1. In some embodiments, (γ xNmax) is rounded up to the nearest integer. If the sixth result is true ("Yes"), step 292 is executed to set Ndefauit to (γ xNmax). Moreover, in steps 340, 342, 344 and 346 as shown in Fig. 6, when a state change of the machining tool occurs (step 340) and the workpiece sampling rate 1 / N is smaller than 1 / Ndefauit at this moment (N = Ndefauit). (Step 344) set the workpiece sampling rate 1 / N to 1 / Ndefauit (N = Ndefauit) (Step 346).
    Next, five scenarios which are considered by the scheme of the workpiece sample measuring step 150 will be explained.
    Scenario 1: A stable process is assumed. When no state changes of a machining tool occur and all values of RI, GSI, DQIx and DQIy of a workpiece machined by the machining tool are within their respective thresholds in a manufacturing process, that is, RI ^ RIT; GSI ^ GSIT; DQIx ^ dqiXt; DQIy ^ DQlyT, then this process is stable. In this situation, no actual measurement is needed to update the VM model, and therefore the ASD scheme N can be set to a larger number (i.e., a lower predetermined value)
    Sample sampling rate) without affecting the accuracy of the VM models.
    Scenario 2: The state of the machining tool is changed. A possible state change of the machining tool may occur when a tool maintenance, repair or recipe adjustment is performed. In this situation, embodiments of the present invention will require an actual measurement of the workpiece being machined when the state of the machining tool is changed, and reset the workpiece counter to zero. For example, if the workpiece originally expected to be measured is the 25th workpiece and the state of the machining tool is changed when the second workpiece is machined, embodiments of the present invention perform an actual measurement on the second workpiece and the next workpiece expected to be measured is the 27th workpiece.
    Scenario 3: The DQIx value of the workpiece is abnormal (i.e., DQIx> DQIXt). The
    The function of the DQIx value is to check the quality of the set of process data used by the machining tool for machining the workpiece. To prevent abnormal process data from degrading the VM models, the workpiece with an abnormal DQIx value should not be selected for measurement. In other words, the measurement for the workpiece with abnormal DQIX value will be skipped. If the workpiece having abnormal DQIx value is originally expected to be measured (at this moment, the workpiece counter for controlling the workpiece sampling is N), embodiments of the present invention do not perform actual measurement on the workpiece, but instead require an actual measurement on a next workpiece. If the DQIx value of this next workpiece is normal (at this moment, the workpiece counter for controlling the workpiece sampling is larger than N), an actual measurement is made on that next workpiece. If the DQIX value of this next workpiece is still abnormal, the measurement for that next workpiece will be skipped. Afterward, the same steps are used to successively accept a next workpiece.
    Scenario 4: The GSI value or Rl value of the workpiece is abnormal (GSI> GSIT or RKRIt). The purpose of the RI value is to calibrate the reliability level of a VM value. If the Rl value of the workpiece is abnormal (RI <RIT), it represents that the reliability level of the VM value of the workpiece is poor and an actual measurement of the workpiece is required to recalibrate or relearn the VM models. The purpose of the GSI value is to evaluate deviations from process data. A process data deviation of the workpiece can lead to a deviation of its corresponding actual measured value. As such, the workpiece with the abnormal GSI value must be checked. However, if the abnormal RI value or GSI value occurs only once, it may be a false alarm caused by noise or other factors. In order to confirm that an actual deviation is detected, embodiments of the present invention, when at least a certain number of consecutive workpieces (such as two or four) have abnormal RI or GSI values, perform an actual measurement on the second or fourth workpiece by.
    Scenario 5: The DQly value of the workpiece is abnormal (i.e., DQIy> DQlyT). The
    The function of the DQlyValue is to evaluate the quality of the actual measured value of the workpiece. If the quality of the actual measured value of the workpiece is not good, the actual measured value of the workpiece can not be used to recalibrate or retrain the VM models. Instead, embodiments of the present invention require an actual measurement on a next workpiece in immediate succession.
    Hereinafter, an operation process of the workpiece sample measuring step 150 will be explained, wherein the workpiece sample measurement step 150 includes a sampling step 300 shown in FIG. 6 and a step 301 shown in FIG.
    In the sampling step 300, a workpiece is first provided to a machining tool (step 302), the machining tool having a set of process data used to machine the workpiece. The set of process data of the workpiece is input to the DQIx model and the GSI model formed in the model generation step 130, thereby obtaining a DQIX value, a GSI value, and a RI value of the workpiece (step 304). In step 304, the set of process data of the workpiece is also input to the presumption model formed in the model generation step 130, whereby a virtual measurement (VM) value of the workpiece is calculated; and the set of process data of this workpiece is also input to the reference model formed in the model generation step 130, whereby a reference prediction value of the workpiece is calculated. Thereafter, an overlap area between the statistical distribution of the virtual measurement of the workpiece and the statistical distribution of the reference prediction value of the workpiece is calculated to produce an Rl value of the workpiece, the Rl value increasing with increasing overlap area, representing the corresponding virtual Reading is more reliable.
    After that, step 310 is performed to check if the processing tool was idle for a period of time (eg, idle for more than one day). If the check result of step 310 is true ("Yes"), the workpiece is assumed to be the first workpiece after the idling period, and step 392 must be performed to perform a measurement on the workpiece by using a measuring tool and set a workpiece counter to zero whereby it is confirmed whether the machining tool is normal. If the check result of the step 310 is false ("No"), the workpiece counter is incremented by 1 (step 320). After that, step 340 is executed to check if a state change of the machining tool occurs (for example, when tool maintenance, repair or recipe adjustment, etc. is performed). If the check result of step 340 is true ("Yes"), step 342 is executed to clear the DS set and set the OOS counter to 0, and step 344 is executed to check if the value of N is greater than Ndefauit. If the check result of step 344 is false ("No"), the measuring tool is used to perform a measurement on the workpiece, and the workpiece counter is set to 0 (step 392). Thereafter, step 346 is performed. Step 392 is also performed to confirm whether the machining tool is normal.
    If the check result of step 340 is false ("No"), step 350 is executed to check if the DQIx value of the workpiece is good or bad. If the DQIX value is greater than the DQIx threshold, this represents the quality of the
    Set of process data of the workpiece (the DQIx value) is not good (the test result of step 350 is "bad"). Since the actual measured value of the workpiece, which is generated by using the set of process data having the abnormal DQIx value, is not reliable, the measurement of the workpiece is skipped (step 390). If the DQIX value is less than or equal to the DQIx threshold, this represents that the quality of the set of process data of the workpiece is good (the check result of step 350 is "good"), and step 360 is executed to check whether the workpiece counter is greater than or equal to N. If the check result of step 360 is true ("Yes"), it represents that the workpiece is the expected (planned) workpiece for measurement, a measurement should be performed on the workpiece, and the workpiece counter is set to 0 (step 392).
    If the check result of step 360 is false ("No"), step 370 is executed to check if the GSI value and the Rl value of the workpiece are good or bad. If the GSI value of the workpiece is less than or equal to the GSI threshold, and the Rl value of the workpiece is greater than or equal to the Rl threshold, this represents the virtual measurement suggested by the set of process data of the workpiece is reliable (the test result of step 370 is "good") and therefore the workpiece need not be measured (step 390). If the GSI value of the workpiece is greater than the GSI threshold or the Rl value of the workpiece is less than the Rl threshold, this represents that the virtual measurement of the workpiece as measured by the set of process data is not reliable (the test result from step 370 is "bad") and it could be that the workpiece has to be measured. However, if the abnormal RI value or GSI value occurs only once, it could be a false alarm caused by noise or other factors, and thus embodiments of the present invention will result if at least a certain number of successive workpieces (such as two or more) are used four) have abnormal RI or GSI values, then actual measurement on the second or fourth workpiece. In other words, if the check result of step 370 is "bad", step 380 is executed to check if the GSI values of k workpieces (such as the previous one or three workpieces) processed before the work are all greater than the GSI threshold or the RI values of k workpieces that have been processed before the workpiece are less than the RI threshold, where k is a positive integer. If the check result of the step 380 is true ("Yes"), the measuring tool is used to perform a measurement on the workpiece and the workpiece counter is set to 0 (step 392). If the check result of the step 380 is false ("No"), the measurement of the workpiece is skipped (step 390). It is worth noting that step 370 can only check if the GSI value of the workpiece is good or bad. If the GSI value of the workpiece is too large, this represents that the set of process data of the workpiece with respect to the sets of process data used for modeling has some differences such that the quality of the workpiece is likely to be abnormal and an actual Measurement requires. It can be seen from the foregoing that with the applications of the embodiments of the present invention, a user can wait until the machining tool has machined more workpieces to select a workpiece for measurement, ie, N can be increased to reach the predetermined workpiece sampling rate 1 / N without worrying about skipping the measurement of the abnormal workpiece that should be measured. Therefore, the embodiments of the present invention can effectively lower the predetermined workpiece sampling rate. However, the predetermined workpiece sampling rate may be performed only by performing step 350 (checking the workpiece DQIx value), step 360 (checking whether the workpiece is the workpiece expected to be measured), and step 370 (checking the GSI value and the Rl value of the workpiece or only checking the GSI value of the workpiece) are effectively lowered.
    The first to sixth "results" used to explain Figs. 5A and 5B and the "first check" (first to seventh) "check results" used to explain Fig. 6 are only for convenience of description and claims and are not intended to be in any particular order.
    Incidentally, after the actual measurement is performed on the workpiece, the workpiece sample measuring step 150 is also directed to an evaluation of a DQIy value of a workpiece as shown in step 301 of FIG. 7.
    First, an actual measurement of the workpiece and a set of process data corresponding to the actual measurement are collected. The set of process data is converted into a set of z-scores. The set of z-scores and the actual metric are entered into the DQIy model, which calculates a DQIy value of the actual metric of the workpiece. If the DQIy value of the workpiece is greater than the DQIy threshold, this represents that the actual measurement is abnormal and can not be used to calibrate or modify models. In order to cope with such a shortcoming, another workpiece currently in production must be requested for measurement (i.e., the workpiece counter is set to N).
    It is to be understood that the stitching measuring method of the present invention is performed by the above-mentioned steps. A computer program of the present invention stored in a nonvolatile tangible computer readable recording medium is used to perform the sampling method described above. The above-mentioned embodiments may be provided as a computer program product having a machine-readable medium on which instructions for programming a computer (or other electronic devices) to perform a process based on the embodiments of the present invention may be stored. The machine readable medium may include, but is not limited to, a floppy disk, an optical disk, a compact disk read only memory (CD-ROM), a magneto-optical disk, a read-only memory (ROM ), random access memory (RAM), erasable programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), magnetic or optical card, flash memory, or any other type of media / machine readable medium which is suitable for storing electronic instructions. Moreover, the embodiments of the present invention may also be downloadable as a computer program product, which may be transmitted from a remote computer on a requesting computer by using data signals over a communication link (such as a network connection or the like).
    It is also to be understood that the present invention may be described in the context of a manufacturing system. While the present invention may be implemented in semiconductor fabrication, the present invention is not limited to implementation in semiconductor fabrication, but may be applied to other manufacturing industries, wherein the fabrication system is configured to provide workpieces or products, including, but not limited to may include microprocessors, memory devices, digital signal processors, application specific integrated circuits (ASICs), or other similar devices. The present invention can also be applied to workpieces other than semiconductor devices such as vehicle wheels, screws. The manufacturing system includes one or more processing tools that may be used to form one or more products or portions thereof in or on the workpieces (such as wafers). In the illustrated embodiment, the processing tools are shown as a single unit. However, it should be appreciated by those skilled in the art that the machining tools may be implemented in any number of units of any type, including lithography tools, deposition tools, etch tools, polishing tools, annealing tools, machine tools, and the like. In the embodiments, the manufacturing system also includes one or more measurement tools, such as scatterers, ellipsometers, scanning electron microscopes, and the like.
    The manufacturing system comprises a sample selection unit provided with the processing tools and the measurement tools for performing a sampling method with a sampling method
    Sampling rate decision scheme is communicatively coupled in accordance with embodiments of the present invention. Those skilled in the art will be able to configure the manufacturing system to provide the necessary interconnections to establish communicative coupling between the sample selection unit, the machining tools, and the measurement tools. In various alternative embodiments, the sample selection unit may be implemented in a computing device, such as a desktop computer, a laptop computer, and the like. However, those skilled in the art should appreciate that, in alternative embodiments, portions of the sample selection unit may be implemented in any number of devices and / or locations.
    On the other hand, a sampling method according to embodiments of the present invention may also be combined with an automatic virtual measurement (AVM) method as disclosed by U.S. Patent No. 8,095,484 B2. Referring to FIG. 7, FIG. 7 is a schematic flow diagram showing a virtual measurement method according to various embodiments of the present invention. Upon completion of all the steps of a first phase virtual measurement, the sampling step 300 is performed as shown in FIG. 6 to determine if a workpiece needs an actual measurement. After performing all the steps of a second phase virtual measurement, the sampling rate adjustment step 200 is performed as shown in FIGS. 5A and 5B to adjust the value of N for use in step 300. In the virtual measurement of the second phase, when an actual measured value of a specific workpiece is obtained, the DQly value of the workpiece is checked. If the DQIy value of the workpiece is greater than the DQIy threshold, step 301 is performed to request a measurement for another workpiece currently in production (i.e., set the workpiece counter to N). After combining the sampling method according to the embodiments of the present invention with the virtual measuring method, an actual measured value of a workpiece can be obtained in time for calibrating or relearning the VM models, thereby ensuring the VM accuracy.
    It can be seen from the above-mentioned embodiments that by using the ADS schemes constructed by various index values representing state changes or abnormal information of a process tool during a fleshing process, the present invention can effectively adjust the workpiece sampling rate, ensure the VM accuracy, and the Can significantly reduce workpiece sampling rate.
    It will be apparent to those skilled in the art that various modifications and variations can be made to the structure of the present invention without departing from the scope or spirit of the invention. In view of the foregoing, it is intended that the present invention cover modifications and variations of this invention provided they come within the scope of the following claims.
    Innsbruck, June 24, 2016
    权利要求:
    Claims (10)
    [1]
    claims
    A sample measurement method comprising: collecting a plurality of sets of historical process data used by a machining tool for machining a plurality of historical workpieces and a plurality of historical measured values of the historical workpieces in accordance with the sets of historical process data; Performing a model generation step, comprising: forming a presumption model in accordance with a presumption algorithm by using the sets of historical process data and the historical measurements; Initializing a workpiece sampling rate represented by 1 / N, wherein the workpiece sampling rate is directed to select from every N workpieces processed by a machining tool the N-th workpiece workpiece as a workpiece expected to be measured; Adding a workpiece to a set of deterministic samples; Collecting a set of process data used by the machining tool to machine the workpiece and an actual measurement of the workpiece according to the set of process data; Inputting the set of process data of the workpiece into the conjecture model, and thereby calculating a virtual measurement value of the workpiece; Calculating an absolute percentage error of the virtual measurement of the workpiece; Determining whether the absolute percentage error of the virtual measure is greater than a specification of a maximum virtual measure error defined for the edit tool, and thereby obtaining a first result; Incrementing an OOS (out of specification) counter by 1 if the first result is true; if the first result is false, calculating a mean absolute percent error (MAPE) of virtual measurements of all the workpieces in the set of determinative samples and determining if the mean absolute percentage error is greater than or equal to an upper control limit of MAPE, and thereby obtaining a second one result; if the second result is true, increasing the sample sampling rate by decreasing N, clearing the set of determinative samples, and setting the OOS counter to 0; if the second result is false, determining whether the number of workpieces in the set of deterministic samples is greater than or equal to a threshold of the number of determinate samples, and thereby obtaining a third result; if the third result is false, leaving the sample-sampling rate unchanged; if the third result is true, calculating a maximum absolute percentage error (MaxErr) of virtual measurements of all the workpieces in the set of determinative samples and determining if the maximum absolute percentage error is less than an upper control limit of MaxErr, thereby obtaining a fourth result; if the fourth result is true, decreasing the sample-sampling rate by increasing N, clearing the set of determinative samples, and setting the OOS counter to 0; and if the fourth result is false, discarding an oldest workpiece in the set of determinative samples and leaving the sample-sampling rate unchanged.
    [2]
    The sampling method of claim 1, wherein the upper control limit of MaxErr is a first conservative factor multiplied by SPECMax, and the upper control limit of MAPE is a second conservative factor multiplied by SPECMax, and the first conservative factor and the second conservative one Factor greater than 0 and less than or equal to 1, and the first conservative factor is greater than the second conservative factor.
    [3]
    3. The sampling method of claim 1, wherein, if the first result is true, the sampling method further comprises: determining whether the OOS counter is greater than or equal to an OOS threshold, and thereby obtaining a fifth result; and if the fifth result is true, increasing the workpiece sampling rate by decreasing N, clearing the set of determinative samples, and setting the OOS counter to zero.
    [4]
    The sampling method of claim 3, wherein the OOS threshold is 2.
    [5]
    5. The sampling method of claim 1, further comprising: obtaining a minimum workpiece sampling rate represented by 1 / Nmax and a predetermined workpiece sampling rate represented by 1 / Ndefauit, where Nmax is a maximum value of N and Ndefauit is a default value of N; Multiplying a third conservative factor by Nmax and thereby obtaining a check value, wherein the third conservative factor is greater than 0 and less than or equal to 1; Determining if the test value is greater than Ndefauit, and thereby obtaining a sixth result; if the sixth result is true, set ndefauit equal to the check value; Performing a first checking step to check whether a state change of the machining tool occurs, and thereby obtaining a first check result; if the first check result is true, clearing the set of determinative samples, setting the OOS counter to 0, and performing a second check step to check if N is greater than Ndefauit, and thereby obtaining a second check result; and if the second check result is true, setting N to ndefauit
    [6]
    6. The sampling method of claim 5, further comprising: performing a model generation step comprising: forming a DQIx (Process Data Quality Index) model and a GSI (Global Similarity Index) model and calculating a DQIX threshold and a GSI threshold by using the sets of historical process data; and performing a workpiece sample measuring step comprising: inputting the set of process data of the workpiece into the DQIx model and the GSI model, and thereby obtaining a DQIX value and a GSI value of the set of process data of the workpiece; Increase a workpiece counter by 1; if the first check result is true, performing a third check step to check if the DQIX value is less than or equal to the DQIX threshold, and thereby obtaining a third check result; if the third verification result is false, skipping a measurement of the workpiece; if the third check result is true, performing a fourth check step to check whether the workpiece counter is greater than or equal to N, and thereby obtaining a fourth check result; if the fourth check result is true, performing a measurement on the workpiece and setting the workpiece counter to 0; if the fourth check result is false, performing a fifth check step to check whether the GSI value of the workpiece is less than or equal to the GSI threshold value, and thereby obtaining a fifth check result; and if the fifth check result is true, skipping the measurement of the workpiece.
    [7]
    7. The sampling method of claim 6, further comprising: building a reference model in accordance with a reference prediction algorithm by using the sets of historical process data and the historical measurements, the assumption algorithm different from the reference prediction algorithm; Calculating an RI (confidence index) threshold based on a maximum tolerable margin of error defined by virtual measurement errors obtained from the presumption model; Inputting the set of process data of the workpiece into the reference model and thereby calculating a reference prediction value of the workpiece; Calculating an overlapping area between the statistical distribution of the virtual measured value of the workpiece and the statistical distribution of the reference prediction value of the workpiece, and thereby generating an Rl value of the workpiece; if the fourth check result is false, executing a sixth check step to check whether the Rl value of the workpiece is greater than or equal to the Rl threshold, thereby obtaining a sixth check result; and if the sixth verification result is true, skipping the measurement of the workpiece.
    [8]
    8. The sample measurement method of claim 7, further comprising: performing a measurement on the workpiece and setting a workpiece counter to zero if the fifth verification result or the sixth verification result is false and the GSI values of k workpieces processed before the workpiece; all are greater than the GSI threshold, or the RI values of k workpieces processed before the workpiece are all less than the RI threshold, where k is a positive integer.
    [9]
    9. The sampling method of claim 6, further comprising: before a process of increasing the workpiece counter by one is performed, executing a seventh checking step to check whether the processing tool has been idling for a period of time, and thereby obtaining a seventh verification result; and making a measurement on the workpiece and setting the workpiece counter to zero if the seventh verification result is true.
    [10]
    10. A computer program product stored on a non-transitory, tangible, computer-readable recording medium which, when executed, performs a sampling method as claimed in any one of claims 1 to 9. Innsbruck, June 24, 2016
    类似技术:
    公开号 | 公开日 | 专利标题
    DE102016109232B4|2018-03-29|Sampling method with sampling rate decision scheme and computer program product thereof
    US9829415B2|2017-11-28|Metrology sampling method and computer program product thereof
    US8862525B2|2014-10-14|Method for screening samples for building prediction model and computer program product thereof
    DE60307310T2|2007-10-18|LIKELY-RESTRICTED OPTIMIZATION FOR CONTROLLING A PRODUCTION LINE
    JP4914457B2|2012-04-11|Automatic virtual measurement system and method
    DE69930501T2|2007-03-01|ULTRASENSIVE MONITORING OF SENSORS AND PROCESSES
    DE102009006887B3|2010-07-15|Method and system for semiconductor process control and monitoring using a data quality measure
    DE102017108497B4|2021-05-20|Manufacturing customization system for customizing the state of manufacture by multiple machines
    DE102018108779A1|2018-10-25|Estimator of an automatic machining error factor
    DE10393394T5|2005-08-11|Intelligent integrated lithography control system based on product construction and yield feedback system
    DE112019000739T5|2020-10-22|TIME SERIES ACQUISITION FOR ANALYZING AND CORRECTING A SYSTEM STATUS
    DE102008021558A1|2009-11-12|Process and system for semiconductor process control and monitoring using PCA models of reduced size
    DE112016006546T5|2018-12-06|QUALITY CONTROL DEVICE, QUALITY CONTROL PROCEDURE AND QUALITY CONTROL PROGRAM
    DE10213285A1|2003-01-23|Method of controlling processing device, e.g. photolithographic process in semiconductor manufacture involves predicting current bias correction value from several of the most recent previous correction values
    EP2442248B1|2016-12-14|Coupling method for non-iterative co-simulation
    DE102017111926A1|2018-12-06|Process control circuit and method for controlling a processing arrangement
    DE112015005427T5|2017-08-17|Quality control machine for complex physical systems
    EP1306736A2|2003-05-02|Method for monitoring of processing plants
    DE112016003235T5|2018-05-03|Output efficiency optimization in production systems
    CN101320258B|2011-08-10|Two-stage virtual measurement method
    DE202018102632U1|2018-05-22|Device for creating a model function for a physical system
    CN101598754A|2009-12-09|The system and method for automatic virtual metrology
    DE102014223810A1|2016-05-25|Method and assistance system for detecting a fault in a system
    CN113204857A|2021-08-03|Method for predicting residual life of electronic device based on extreme gradient lifting tree algorithm
    DE102019106939A1|2019-09-26|WORKING CONDITION ADJUSTMENT DEVICE AND MACHINE LEARNING DEVICE
    同族专利:
    公开号 | 公开日
    TW201642147A|2016-12-01|
    CN106206346A|2016-12-07|
    DE102016109232B4|2018-03-29|
    TWI539298B|2016-06-21|
    DE102016109232A1|2016-12-01|
    US10269660B2|2019-04-23|
    KR101930420B1|2018-12-19|
    JP2016224947A|2016-12-28|
    KR20160140474A|2016-12-07|
    CN106206346B|2018-09-28|
    JP6285494B2|2018-02-28|
    US20160349736A1|2016-12-01|
    AT517250A3|2018-05-15|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    JP3208807B2|1991-11-15|2001-09-17|株式会社日立製作所|Electronic device inspection system and electronic device manufacturing method|
    JPH11129145A|1997-10-31|1999-05-18|Denso Corp|Device and method for diagnosing shape precision of work, and recording medium|
    WO2002023289A2|2000-09-15|2002-03-21|Advanced Micro Devices, Inc.|Adaptive sampling method for improved control in semiconductor manufacturing|
    US6976236B1|2002-04-05|2005-12-13|Procket Networks, Inc.|Method for automatically routing connections between top side conductors and bottom side conductors of an integrated circuit package|
    TWI283817B|2003-05-30|2007-07-11|Tokyo Electron Ltd|Method of operating a process control system and method of operating an advanced process control system|
    TWI222696B|2003-09-12|2004-10-21|Grace Semiconductor Mfg Corp|Defect analysis sampling control system and method|
    TWI267012B|2004-06-03|2006-11-21|Univ Nat Cheng Kung|Quality prognostics system and method for manufacturing processes|
    US7076321B2|2004-10-05|2006-07-11|Advanced Micro Devices, Inc.|Method and system for dynamically adjusting metrology sampling based upon available metrology capacity|
    JP2007213147A|2006-02-07|2007-08-23|Denso Corp|Process monitoring apparatus and process monitoring method|
    TWI315054B|2006-05-10|2009-09-21|Nat Cheng Kung Universit|Method for evaluating reliance level of a virtual metrology system in product manufacturing|
    TWI338916B|2007-06-08|2011-03-11|Univ Nat Cheng Kung|Dual-phase virtual metrology method|
    CN101320258B|2007-06-08|2011-08-10|郑芳田|Two-stage virtual measurement method|
    TWI349867B|2008-05-20|2011-10-01|Univ Nat Cheng Kung|Server and system and method for automatic virtual metrology|
    CN101598754B|2008-06-05|2011-09-21|郑芳田|System and method for automatic virtual metrology|
    JP5383379B2|2008-11-26|2014-01-08|キヤノン株式会社|Developing device and cartridge|
    US8392009B2|2009-03-31|2013-03-05|Taiwan Semiconductor Manufacturing Company, Ltd.|Advanced process control with novel sampling policy|
    US8437870B2|2009-06-05|2013-05-07|Taiwan Semiconductor Manufacturing Company, Ltd.|System and method for implementing a virtual metrology advanced process control platform|
    US8620468B2|2010-01-29|2013-12-31|Applied Materials, Inc.|Method and apparatus for developing, improving and verifying virtual metrology models in a manufacturing system|
    TWI427487B|2010-04-02|2014-02-21|Foresight Technology Company Ltd|Method for sampling workpiece for inspection and computer program product performing the same|
    TWI412906B|2010-04-13|2013-10-21|Univ Nat Cheng Kung|Manufacturing execution system with virtual-metrology capabilities and manufacturing system including the same|
    TWI407325B|2010-05-17|2013-09-01|Nat Univ Tsing Hua|Process quality predicting system and method thereof|
    CN102262188B|2010-05-28|2013-06-05|先知科技股份有限公司|Sampling inspection method for workpieces|
    TWI427722B|2010-08-02|2014-02-21|Univ Nat Cheng Kung|Advanced process control system and method utilizing virtual metrology with reliance index and computer program product thereof|
    US10522427B2|2011-07-06|2019-12-31|Taiwan Semiconductor Manufacturing Company, Ltd.|Techniques providing semiconductor wafer grouping in a feed forward process|
    TWI451336B|2011-12-20|2014-09-01|Univ Nat Cheng Kung|Method for screening samples for building prediction model and computer program product thereof|
    TWI463334B|2012-07-20|2014-12-01|Univ Nat Cheng Kung|Baseline predictive maintenance method for target device and computer program product thereof|
    US9240360B2|2012-07-25|2016-01-19|International Business Machines Corporation|Run-to-run control utilizing virtual metrology in semiconductor manufacturing|
    US9508042B2|2012-11-05|2016-11-29|National Cheng Kung University|Method for predicting machining quality of machine tool|
    TWI521360B|2014-03-26|2016-02-11|國立成功大學|Metrology sampling method and computer program product thereof|JP5334787B2|2009-10-09|2013-11-06|株式会社日立ハイテクノロジーズ|Plasma processing equipment|
    CN108268987B|2016-12-30|2021-08-06|郑芳田|Method for estimating quality of various products|
    CN107014635B|2017-04-10|2019-09-27|武汉轻工大学|Grain uniform sampling method and device|
    CN107203496B|2017-06-01|2020-05-19|武汉轻工大学|Grain distribution sampling method and device|
    CN109993182B|2017-12-29|2021-08-17|中移（杭州）信息技术有限公司|Pattern recognition method and device based on Fuzzy ART|
    TWI676899B|2018-02-21|2019-11-11|Measuring instrument data collecting device and method|
    KR102092379B1|2018-04-13|2020-03-23|김대희|Method, apparatus and recording medium for testing semiconductor wafer|
    CN109598052A|2018-11-29|2019-04-09|武汉大学|Intelligent electric meter life cycle prediction technique and device based on correlation analysis|
    KR102305261B1|2020-12-11|2021-09-28|쿠팡 주식회사|Electronic apparatus and information providing method thereof|
    CN113822257A|2021-11-24|2021-12-21|航天智控监测技术有限公司|Abnormal point detection method based on combination of dimensionless features and virtual samples|
    法律状态:
    优先权:
    申请号 | 申请日 | 专利标题
    TW104117016A|TWI539298B|2015-05-27|2015-05-27|Metrology sampling method with sampling rate decision scheme and computer program product thereof|
    [返回顶部]