Virtual metrology method with convolutional autoencoder and transfer learning and system thereof

Abstract

The present application provides a virtual measurement method and system with convolutional autoencoder and transfer learning. The modeling operation of the virtual measurement method with convolutional autoencoder and transfer learning includes classification of paired data and unpaired process data; a pre-training model is established using the unpaired process data, and then the paired data is input into the pre-training model to establish a virtual measurement model based on convolutional autoencoder, which includes a convolutional neural network model. In addition, the calculation operation includes one of the prediction stage and the transfer learning stage according to whether the actual measurement value is obtained, to calculate one of the first stage virtual measurement value and the second stage virtual measurement value. In this way, fast modeling can be achieved without collecting a large amount of paired data, and online relearning can be performed when the working conditions change to maintain the prediction accuracy.

Description

Technical Field

[0001] This invention relates to a virtual measurement method and system, and more particularly to a virtual measurement method and system incorporating a convolutional autoencoder and transfer learning. Background Technology

[0002] To ensure stable processes and high-yield output, factories (such as semiconductor and TFT-LCD plants) require offline quality sampling inspections. However, this method only measures a small number of samples and has the drawback of time delay, failing to achieve real-time monitoring and online full inspection. Automatic Virtual Metrology (AVM) collects information from production equipment, inputs it into an AVM prediction model for prediction, and outputs virtual measurement values, thus achieving real-time monitoring and online full inspection. As technology advances and processes become increasingly precise, the accuracy requirements for virtual measurement are even higher. This necessitates the integration of Convolutional Neural Network (CNN) algorithms into the AVM architecture (AVM). CNN A new method has also been proposed, which not only improves the accuracy of the original AVM but also performs better in predicting extreme values. However, in practical applications, AVM... CNN Two problems remain to be overcome: first, how to quickly build a model when there is a lack of measurement data; and second, the model lacks self-learning ability, making it unable to maintain prediction accuracy when conditions change. Therefore, it is clear that the market currently lacks a virtual measurement method and system that can quickly build models without requiring large amounts of paired data, can perform online relearning, and maintain prediction accuracy. Consequently, relevant industry players are seeking solutions. Summary of the Invention

[0003] Therefore, the purpose of this invention is to provide a virtual measurement method and system with convolutional autoencoders and transfer learning. This method integrates the modeling and prediction mechanisms of both convolutional autoencoders and transfer learning, overcoming the limitation of known deep learning-based algorithms requiring the collection of large amounts of pairwise data for modeling. Furthermore, by constructing a two-stage mechanism based on transfer learning through different learning strategies, it can continue online relearning even when conditions change, maintaining prediction accuracy and thus enabling more widespread application of virtual measurement in increasingly sophisticated semiconductor processes in the future.

[0004] According to one embodiment of the method of the present invention, a virtual measurement method with convolutional autoencoder and transfer learning is provided, comprising the following steps: obtaining multiple sets of process data, wherein these sets of process data are used or generated when processing multiple workpieces on a production machine, and these sets of process data are mapped one-to-one to these workpieces. Each set of process data includes the values ​​of multiple parameters, each of the parameter values ​​being multiple sets of time-series data respectively corresponding to these workpieces, and each set of time-series data having a data length; performing a data alignment operation on these sets of process data, the data alignment operation including performing a data length adjustment operation, such that after the last data point of each set of time-series data for each of these parameters, at least one data point having the value of the last data point is repeatedly added until the data length of each set of time-series data for each of these parameters is equal to the longest data length of these sets of process data; obtaining multiple actual measurement values ​​of these workpieces; and performing a modeling operation, the modeling operation including dividing the sets of process data and these actual measurement values ​​into multiple pairs of data. The system includes at least one set of unpaired process data, each of which contains one set of process data and a corresponding actual measurement value; it also includes using the at least one set of unpaired process data to build at least one pre-trained model, then inputting the pair of data into the at least one pre-trained model to build a virtual measurement model based on a convolutional autoencoder, wherein the virtual measurement model based on the convolutional autoencoder includes at least one convolutional neural network model; and it performs a computational operation, which includes obtaining at least one of another set of process data and another actual measurement value for another workpiece, and performing either a prediction stage or a transfer learning stage depending on whether the other actual measurement value is obtained, to calculate either a first-stage virtual measurement value or a second-stage virtual measurement value for the other workpiece. The prediction stage calculates the first-stage virtual measurement value based on the other set of process data using the virtual measurement model based on the convolutional autoencoder, and the transfer learning stage calculates the second-stage virtual measurement value for the other workpiece based on the other set of process data and the other actual measurement value using the virtual measurement model based on the convolutional autoencoder.

[0005] Therefore, the virtual measurement method of the present invention, which integrates convolutional autoencoders and transfer learning, combines the modeling and prediction mechanisms of both convolutional autoencoders and transfer learning. It enables rapid modeling without requiring large amounts of pairwise data, overcoming the limitation of known deep learning-based algorithms that require the collection of large amounts of pairwise data for modeling. Furthermore, by constructing a two-stage mechanism based on transfer learning through different learning strategies, it can continue online relearning even when conditions change, thus maintaining prediction accuracy.

[0006] Other embodiments of the aforementioned implementation are as follows: In the aforementioned calculation operation, when the other actual measurement value is not obtained, a prediction phase is executed to calculate the first stage virtual measurement value of the other workpiece; when the other actual measurement value is obtained, a transfer learning phase is executed to calculate the second stage virtual measurement value of the other workpiece.

[0007] Other embodiments of the aforementioned implementation are as follows: The aforementioned prediction stage includes performing a data integration operation on the other set of process data for the other workpiece, and then inputting the other set of process data for the other workpiece into the virtual measurement model based on the convolutional autoencoder to calculate the first stage virtual measurement value of the other workpiece.

[0008] Other embodiments of the aforementioned implementation are as follows: The aforementioned transfer learning stage includes a strategy selection confirmation step for the other actual measurement value of the other workpiece to generate a confirmation result, and executing one of the first strategy step and the second strategy step according to the confirmation result to update the virtual measurement model based on the convolutional autoencoder; and inputting the other set of process data of the other workpiece into the updated virtual measurement model based on the convolutional autoencoder to calculate the second stage virtual measurement value of the other workpiece.

[0009] Other embodiments of the aforementioned implementation are as follows: The aforementioned strategy selection confirmation step includes confirming whether the parts of the production machine have been maintained or replaced. When the confirmation result is that the parts of the production machine have been maintained or replaced, the transfer learning phase executes the first strategy step; when the confirmation result is that the parts of the production machine have not been maintained or replaced, the transfer learning phase executes the second strategy step.

[0010] Other embodiments of the aforementioned implementation are as follows: In the aforementioned modeling operation, the number of the at least one convolutional neural network model is multiple, and the virtual measurement model based on the convolutional autoencoder includes these convolutional neural network models and an estimation model. The multiple inputs of these convolutional neural network models are the pairs of data, and the multiple outputs of these convolutional neural network models are the inputs of the estimation model.

[0011] Other embodiments of the aforementioned implementation are as follows: In the aforementioned transfer learning stage, the first strategy step includes inputting the time-series data of multiple parameters of the other set of process data and the other actual measurement values ​​of the other set of workpieces into the convolutional neural network models; and inputting the outputs of the convolutional neural network models into the estimation model to update the virtual measurement model based on the convolutional autoencoder.

[0012] Other embodiments of the foregoing implementation are as follows: In the aforementioned transfer learning phase, the second strategy step includes inputting the outputs of these convolutional neural network models into the estimation model to update part of this virtual measurement model based on convolutional autoencoders.

[0013] Other embodiments of the aforementioned implementation are as follows: The aforementioned data integration operation further includes performing a frequency distribution calculation on the data length of each of the groups of time series data for each of these parameters to obtain a distribution of the number of data occurrences as a function of data length. The data length with the highest frequency among the groups of time series data for each of these parameters is a reference data length; the time series data with the reference data length among the groups of time series data for each of these parameters are averaged to obtain a set of reference time series data for each of these parameters; a Dynamic Time Warping (DTW) algorithm is used to calculate the distance between each of the groups of time series data for each of these parameters and its corresponding reference time series data; a distance threshold is set; and when the distance is greater than the distance threshold, the process data corresponding to that distance is deleted.

[0014] Other embodiments of the aforementioned implementation are as follows: either the first-stage virtual measurement value produced in the aforementioned prediction stage or the second-stage virtual measurement value produced in the transfer learning stage corresponds to the critical dimension (CD) of the dry etch shallow trench isolation (Dry Etching STI) process applied to the production equipment in semiconductor manufacturing.

[0015] According to one embodiment of the structural method of the present invention, a virtual measurement system with convolutional autoencoder and transfer learning is provided, comprising memory and a processor. The memory stores multiple sets of process data and multiple actual measurement values. These sets of process data are used or generated when a production machine processes multiple workpieces. These sets of process data are mapped one-to-one to these workpieces. Each set of process data contains the values ​​of multiple parameters, each of which represents multiple sets of timing data corresponding to these workpieces. Each set of timing data has a data length. The processor is electrically connected to memory. The processor receives these sets of process data and these actual measurements and configures them to perform an operation comprising the following steps: performing a data merging operation on these sets of process data, which includes performing a data length adjustment operation by repeatedly adding at least one data point with the value of the last data point after the last data point of each set of timing data for each of these parameters, until the data length of each set of timing data for each of these parameters is equal to the longest data length of these sets of process data; performing a modeling operation, which includes dividing these sets of process data and these actual measurements into multiple pairs of data and at least one unpaired set of process data. Each of the following includes one set of process data and a corresponding actual measurement value; and uses at least one unpaired set of process data to build at least one pre-trained model, then inputs these paired data into the at least one pre-trained model to build the virtual measurement model based on a convolutional autoencoder, wherein the virtual measurement model based on a convolutional autoencoder includes at least one convolutional neural network model; and performs a computational operation, the computational operation including obtaining at least one of another set of process data and another actual measurement value for another workpiece, and performing one of a prediction stage and a transfer learning stage depending on whether the other actual measurement value is obtained, to calculate one of a first-stage virtual measurement value and a second-stage virtual measurement value for the other workpiece. The prediction stage calculates the first-stage virtual measurement value based on the other set of process data using the virtual measurement model based on the convolutional autoencoder, and the transfer learning stage calculates the second-stage virtual measurement value for the other workpiece based on the other set of process data and the other actual measurement value using the virtual measurement model based on the convolutional autoencoder.

[0016] Therefore, the virtual measurement system of the present invention, which integrates convolutional autoencoders and transfer learning, combines the modeling and prediction mechanisms of both convolutional autoencoders and transfer learning. It requires only a small amount of paired data for rapid modeling, overcoming the limitation of known deep learning-based algorithms that require large amounts of paired data for modeling. Furthermore, by constructing a two-stage mechanism based on transfer learning through different learning strategies, it can continue online relearning even when conditions change, thus maintaining prediction accuracy.

[0017] Other embodiments of the aforementioned implementation are as follows: In the aforementioned calculation operation, when the other actual measurement value is not obtained, a prediction phase is executed to calculate the first stage virtual measurement value of the other workpiece; when the other actual measurement value is obtained, a transfer learning phase is executed to calculate the second stage virtual measurement value of the other workpiece.

[0018] Other embodiments of the aforementioned implementation are as follows: The aforementioned prediction stage includes performing a data integration operation on the other set of process data for the other workpiece, and then inputting the other set of process data for the other workpiece into the virtual measurement model based on the convolutional autoencoder to calculate the first stage virtual measurement value of the other workpiece.

[0019] Other embodiments of the aforementioned implementation are as follows: The aforementioned transfer learning stage includes a strategy selection confirmation step for the other actual measurement value of the other workpiece to generate a confirmation result, and executing one of the first strategy step and the second strategy step according to the confirmation result to update the virtual measurement model based on the convolutional autoencoder; and inputting the other set of process data of the other workpiece into the updated virtual measurement model based on the convolutional autoencoder to calculate the second stage virtual measurement value of the other workpiece.

[0020] Other embodiments of the aforementioned implementation are as follows: The aforementioned strategy selection confirmation step includes confirming whether the parts of the production machine have been maintained or replaced; when the confirmation result is that the parts of the production machine have been maintained or replaced, the transfer learning stage executes the first strategy step; when the confirmation result is that the parts of the production machine have not been maintained or replaced, the transfer learning stage executes the second strategy step.

[0021] Other embodiments of the aforementioned implementation are as follows: In the aforementioned modeling operation, the number of the at least one convolutional neural network model is multiple, and the virtual measurement model based on the convolutional autoencoder includes these convolutional neural network models and the estimation model. The multiple inputs of these convolutional neural network models are the pairs of data, and the multiple outputs of these convolutional neural network models are the inputs of the estimation model.

[0022] Other embodiments of the aforementioned implementation are as follows: In the aforementioned transfer learning stage, the first strategy step includes inputting the time-series data of multiple parameters of the other set of process data and the other actual measurement values ​​of the other set of workpieces into the convolutional neural network models; and inputting the outputs of the convolutional neural network models into the estimation model to update the virtual measurement model based on the convolutional autoencoder.

[0023] Other embodiments of the foregoing implementation are as follows: In the aforementioned transfer learning phase, the second strategy step includes inputting the outputs of these convolutional neural network models into the estimation model to update part of this virtual measurement model based on convolutional autoencoders.

[0024] Other embodiments of the aforementioned implementation are as follows: The aforementioned data integration operation further includes calculating the frequency distribution of the data length of each of the time series data sets for each of these parameters, to obtain a distribution of the number of data occurrences as a function of data length, wherein the data length that occurs most frequently in the time series data sets for each of these parameters is a reference data length; averaging the time series data sets with the reference data length in the time series data sets for each of these parameters, to obtain a set of reference time series data for each of these parameters; using a Dynamic Time Warping (DTW) algorithm to calculate the distance between each of the time series data sets for each of these parameters and its corresponding reference time series data; setting a distance threshold; and deleting the process data corresponding to the distance when the distance is greater than the distance threshold.

[0025] Other embodiments of the aforementioned implementation are as follows: either the first-stage virtual measurement value produced in the aforementioned prediction stage or the second-stage virtual measurement value produced in the transfer learning stage corresponds to the critical dimension (CD) of the dry etch shallow trench isolation (Dry Etching STI) process applied to the production equipment in semiconductor manufacturing. Attached Figure Description

[0026] To gain a more complete understanding of the embodiments and their advantages, the following description, taken in conjunction with the accompanying drawings, is provided, in which...

[0027] Figure 1 This is a block diagram illustrating a virtual measurement system with convolutional autoencoder and transfer learning according to a first embodiment of the present invention;

[0028] Figure 2 This is a flowchart illustrating a virtual measurement method with convolutional autoencoder and transfer learning according to a second embodiment of the present invention;

[0029] Figure 3 It is a drawing Figure 2 A schematic diagram of the modeling operation based on a convolutional autoencoder;

[0030] Figure 4 It is a drawing Figure 3 A schematic diagram of a convolutional autoencoder network;

[0031] Figure 5 It is a drawing Figure 3 A schematic diagram of a virtual measurement model based on a convolutional autoencoder;

[0032] Figure 6 It is a drawing Figure 2 A flowchart illustrating the two-stage mechanism of computational operations based on transfer learning;

[0033] Figure 7 It is a drawing Figure 6 A diagram illustrating the first strategic step;

[0034] Figure 8 It is a drawing Figure 6 A diagram illustrating the second strategy step;

[0035] Figure 9 This illustrates the prediction results of key dimensions for different strategy combinations in application examples of the present invention; and

[0036] Figure 10 This illustrates the prediction results of key dimensions for different prediction algorithms in application examples of the present invention.

[0037] [Symbol Explanation]

[0038] 100: Virtual Measurement System with Convolutional Autoencoder and Transfer Learning

[0039] 1002: Memory

[0040] 1004: Processor

[0041] 102: Process Data

[0042] 104: Measurement data

[0043] 106: Process Data Preprocessing Module

[0044] 108: ADAS

[0045] 110: Virtual Measurement Model with CAE and TL

[0046] 112: Measurement Data Preprocessing Module

[0047] 114: CAE-based Virtual Measurement Model

[0048] 116: Two-phase mechanism based on TL

[0049] 120:RI model

[0050] 130: GSI model

[0051] 200: Virtual Measurement Methods with Convolutional Autoencoders and Transfer Learning

[0052] 212: Unpaired process data

[0053] 214: Paired Data

[0054] 3E_1, 3E_2, 3E_p: CNN model encoder

[0055] 3D_1, 3D_2, 3D_p: CNN model decoders

[0056] 310: Convolutional layer

[0057] 320: Pooling layer

[0058] 330: Leveling layer

[0059] 340: Discard Layer

[0060] 350: Fully Connected Neural Network

[0061] 400: Estimation Model

[0062] 500: Forecasting Phase

[0063] 600: Transfer Learning Stage

[0064] 710, 720, 730, 810, 820, 830, 840, 850: Curves

[0065] 740,860: Group

[0066] S02, S04, S042, S06, S0612, S0614, S062, S064, S0642, S0644, S08, S082, 502, 504, 506, 508, 510, 512, 602, 604, 606, 608, 610, 612, 614, 616, 618, 620, 622, 624, 626, 628, 630: Steps Detailed Implementation

[0067] Several embodiments of the present invention will now be described with reference to the accompanying drawings. For clarity, many practical details will be set forth in the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the invention, these practical details are not essential. Furthermore, for the sake of simplicity in the drawings, some known and conventional structures and elements will be illustrated in a simple schematic manner; and repeated elements may be denoted by the same reference numerals.

[0068] Furthermore, in this document, when a component (or unit or module, etc.) is "connected" to another component, it can mean that the component is directly connected to the other component, or that the component is indirectly connected to the other component, meaning that there is another component between the component and the other component. Only when it is explicitly stated that a component is "directly connected" to another component does it indicate that there is no other component between the component and the other component. The terms "first," "second," and "third" are only used to describe different components and do not limit the components themselves; therefore, "first component" can also be referred to as "second component." Moreover, the combinations of components / units / circuits in this document are not combinations generally known, conventional, or known in this field. Whether the components / units / circuits themselves are known cannot be used to determine whether their combination relationships are easily performed by someone of ordinary skill in the art.

[0069] Please refer to the following: Figure 1 and Figure 2 ,in Figure 1 This is a block diagram illustrating a virtual measurement system 100 with a convolutional autoencoder (CAE) and transfer learning (TL) according to a first embodiment of the present invention; and Figure 2 This is a flowchart illustrating a virtual measurement method 200 with convolutional autoencoder and transfer learning according to a second embodiment of the present invention. The virtual measurement system 100 with convolutional autoencoder and transfer learning includes a memory 1002 and a processor 1004.

[0070] Memory 1002 stores multiple sets of process data 102 and multiple sets of measurement data 104 (i.e., multiple actual measurement values). These sets of process data 102 are used or generated when the production equipment processes the workpiece, and they are mapped one-to-one to the workpiece. Each set of process data 102 contains the values ​​of multiple parameters, each parameter value corresponding to multiple sets of timing data for the workpiece, and each set of timing data has a data length. In other words, memory 1002 obtains the aforementioned multiple sets of process data 102 for historical workpieces from the production equipment. Furthermore, these measurement data 104 (these actual measurement values) are obtained by measuring the quality items of the workpiece using a measurement equipment. For wafer fabrication (semiconductor fabrication), the production equipment is a wafer fabrication equipment, such as an etching machine, deposition machine, sputtering machine, etc.; the actual measurement values ​​are film thickness, etching depth, etch sidewall angle, critical dimension (CD), etc.; the process data 102 includes temperature. For wafer dicing, the production equipment is a wafer dicing machine; the actual measurement is wafer chipping; process data 102 includes blade clogging, coolant flow rate, spindle speed, feed rate, wafer conditions (such as thickness, coating, etc.), and / or kerf width. For machine tool machining, the production equipment is a machine tool; the actual measurement includes surface roughness, straightness, angle, perpendicularity, parallelism, and / or roundness; process data 102 includes operating current, vibration and / or audio data obtained by sensors mounted on the machine tool (e.g., a triaxial accelerometer and / or an acoustic emission (AE) sensor).

[0071] Processor 1004 is electrically connected to memory 1002. Processor 1004 receives sets of process data 102 and measurement data 104 (these actual measurements) and configures them to implement a virtual measurement method 200 with convolutional autoencoder and transfer learning. Specifically, processor 1004 includes a process data preprocessing module 106, a virtual measurement model 110 with convolutional autoencoder (CAE) and transfer learning (TL), a measurement data preprocessing module 112, a Reliance Index (RI) model 120, and a Global Similarity Index (GSI) model 130. Process data preprocessing module 106 receives multiple sets of process data 102. Process data preprocessing module 106 performs data alignment operations on the process data 102 according to an Automated Data Alignment Scheme (ADAS) 108 to delete process data 102 with dissimilar time distribution shapes and to make the data lengths of the process data 102 the same. Before or after data reconciliation, the process data preprocessing module 106 can determine the process data quality index (DQI) based on the process data quality index. X The model is used to assess the data quality of process data 102, and the raw process data 102 from the production equipment is organized and standardized. The measurement data preprocessing module 112 uses the Measurement Data Quality Index (DQI) to... yThe model is used to evaluate the data quality of historical workpiece measurement data 104 (i.e., the aforementioned multiple actual measurement values), remove outliers, and standardize the measurement data 104. Then, using the historical workpiece measurement data 104 and the adjusted process data 102 as modeling samples, virtual measurement models 110, RI model 120, and GSI model 130 with convolutional autoencoders and transfer learning are established according to the convolutional neural network algorithm. The virtual measurement model 110 with convolutional autoencoders and transfer learning includes a virtual measurement model 114 based on convolutional autoencoders and a two-stage mechanism 116 based on transfer learning. After the virtual measurement models 110, RI model 120, and GSI model 130 with convolutional autoencoders and transfer learning are modeled, a virtual measurement model 114 based on convolutional autoencoders is generated. Then, virtual measurement calculations can be performed on subsequent workpieces according to the two-stage mechanism 116 based on transfer learning. The two-stage mechanism 116 based on transfer learning includes a prediction stage (Phase-I) and a transfer learning stage (Phase-II). In the prediction stage (Phase-I), after obtaining a set of process data 102 for a workpiece, the process data preprocessing module 106 performs data integration operations and / or other data preprocessing according to the automatic data integration mechanism (ADAS) 108. Then, the processed process data 102 is input to the virtual measurement models 110, RI model 120, and GSI model 130 with convolutional autoencoders and transfer learning to calculate the first-stage virtual measurement value (VM) of the workpiece. I The process data 102 and measurement data 104 of the workpiece are used to retrain or calibrate the virtual measurement model 110, RI model 120, and GSI model 130 with convolutional autoencoder and transfer learning, after the workpiece has been processed by the production machine and its quality items have been measured by the measurement machine to obtain its measurement data 104 (i.e., actual measurement value). Then, the process data 102 and measurement data 104 of the workpiece can be used to calculate the second-stage virtual measurement value (VM) of the workpiece. II Its confidence index (RI) and overall similarity index (GSI) are used to measure its confidence index (RI) and overall similarity index (GSI).

[0072] The aforementioned RI model 120, GSI model 130, and DQI X Model and DQI yThe model can be referenced to U.S. Patent No. 8,095,484B2. That is, embodiments of the present invention incorporate the relevant provisions of this U.S. Patent No. 8,095,484B2. Furthermore, the memory 1002 described above may include random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions for execution by the processor 1004. The processor 1004 may include any type of processor, microprocessor, or field-programmable gate array (FPGA) capable of compiling and executing instructions. The processor 1004 may include a single device (e.g., a single core) or a group of devices (e.g., multiple cores).

[0073] Therefore, the virtual measurement system 100 of the present invention, which integrates convolutional autoencoders and transfer learning, combines modeling and prediction using both mechanisms, thus overcoming the limitation of known deep learning-based algorithms requiring the collection of large amounts of pairwise data for modeling. Furthermore, a two-stage mechanism 116 based on transfer learning is constructed through different learning strategies, enabling online relearning even when conditions change, thereby maintaining prediction accuracy.

[0074] Figure 2 The virtual measurement method 200 with convolutional autoencoder and transfer learning includes multiple steps S02, S04, S06, and S08. Step S02 includes a data collection operation to obtain multiple sets of process data 102. These sets of process data 102 are used or generated when the production machine processes multiple workpieces. These sets of process data 102 are mapped to these workpieces in a one-to-one manner. Each set of process data 102 contains the values ​​of multiple parameters. The value of each parameter is a set of time-series data corresponding to these workpieces. Each set of time-series data has a data length.

[0075] Step S04 includes performing a data integration operation on these sets of process data 102. This data integration operation involves calculating the frequency distribution of the data lengths across multiple sets of time-series data for each parameter, obtaining the distribution of data occurrences over data length. The data length with the highest frequency across the multiple sets of time-series data for each parameter is designated as a reference data length. Then, a reference time-series data acquisition operation is performed by averaging the time-series data with the aforementioned reference data length across the multiple sets of time-series data for each parameter, resulting in a set of reference time-series data for each parameter. After obtaining a set of reference time-series data for each parameter, a data distance calculation operation is performed using a Dynamic Time Warping (DTW) algorithm to calculate the distance between each set of time-series data for each parameter and its corresponding reference time-series data. DTW is a method for measuring the similarity between two time-series data sets, widely used in speech recognition to determine whether two speech segments represent the same word. Finally, a distance threshold value is set. When the distance (data distance) between the time series data of a parameter and its corresponding reference time series data is greater than a distance threshold, the process data 102 corresponding to this distance is deleted. That is, this set of process data 102 is too dissimilar to other sets of process data 102 and cannot be used for modeling. Next, a data length upper limit is set. Then, step S042 is executed, which includes performing a data length adjustment operation to repeatedly add at least one data point with the value of the last data point after the last data point of each set of time series data for each parameter, until the data length of each set of time series data for each parameter is equal to the longest data length of all process data 102.

[0076] After performing the data length adjustment operation (step S042), multiple measurement data 104 (i.e., multiple actual measurement values) of these workpieces are obtained, and then step S06 is executed. Step S06 includes performing a modeling operation, which includes steps S062 and S064. Step S062 includes dividing these sets of process data 102 and these actual measurement values ​​into multiple pairs of data and at least one unpaired process data, each pair of data including one of the process data 102 and one of the corresponding actual measurement values. Step S064 includes using this at least one unpaired process data to build at least one pre-trained model, and then inputting these pairs of data into this at least one pre-trained model to build a virtual measurement model 114 based on a convolutional autoencoder, wherein this virtual measurement model 114 based on a convolutional autoencoder includes at least one convolutional neural network model.

[0077] Step S08 performs a calculation operation, which includes step S082, to obtain at least one of another set of process data 102 for another workpiece and another actual measurement value, and, depending on whether the other actual measurement value is obtained, executes either a prediction phase (Phase-I) or a transfer learning phase (Phase-II) to calculate the first-stage virtual measurement value (VM) for the other workpiece. I ) and the second-stage virtual measurement (VM) II One of them. The prediction phase (Phase-I) is based on another set of process data 102, which calculates the first-stage virtual measurement value (VM) according to the virtual measurement model 114 based on this convolutional autoencoder. I The transfer learning phase (Phase II) involves using another set of process data 102 and another actual measurement value to calculate the second-stage virtual measurement value (VM) of the other workpiece based on the virtual measurement model 114 using the convolutional autoencoder-based virtual measurement model. II The two-stage mechanism 116 based on transfer learning of the present invention is implemented through a prediction stage (Phase-I) and a transfer learning stage (Phase-II). The following detailed embodiments illustrate the modeling operation in step S06 and the computation operation in step S08.

[0078] Please refer to the following: Figure 1 , Figure 2 and Figure 3 ,in Figure 3 It is a drawing Figure 2A flowchart illustrating the modeling operation based on a convolutional autoencoder (step S06) is provided. As shown, step S06 includes sequentially executing steps S0612, S0614, S062, and S064. Step S0612 includes collecting historical workpiece process data 102 and measurement data 104 (i.e., actual measurement values). Step S0614 includes performing data alignment on the process data 102 according to an Automatic Data Alignment (ADAS) mechanism 108 to delete process data 102 with dissimilar time distribution shapes and to ensure that the process data 102 have the same data length. Step S062 includes dividing these sets of process data 102 and these actual measurement values ​​into multiple pairs of data 214 and at least one unpaired pair of process data 212, each pair of data 214 containing one piece of process data 102 and one corresponding actual measurement value. Step S064 includes sequentially executing steps S0642 and S0644. Step S0642 includes using at least one unpaired process data 212 to build at least one pre-trained model. Step S0644 includes inputting the pairwise data 214 into the at least one pre-trained model to build a virtual measurement model 114 based on a convolutional autoencoder, wherein the virtual measurement model 114 based on the convolutional autoencoder includes at least one convolutional neural network (CNN) model.

[0079] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 ,in Figure 4 It is a drawing Figure 3 A schematic diagram of a convolutional autoencoder network; and Figure 5 It is a drawing Figure 3 A schematic diagram of the virtual measurement model 114 based on a convolutional autoencoder is shown. As shown, in step S0642, unpaired process data 212 is used to build a pre-trained model, which is formed by a convolutional autoencoder network. The pre-trained model includes multiple CNN model encoders 3E_1, 3E_2, and 3E_p and multiple CNN model decoders 3D_1, 3D_2, and 3D_p. Among them, the CNN model encoders 3E_1, 3E_2, and 3E_p receive parameters 1, 2, and p respectively and generate outputs y1, y2, and y3 respectively. p The CNN model decoders 3D_1, 3D_2, and 3D_p receive outputs y1, y2, and y3 respectively. pAnd respectively generate parameters 1', 2', and p'. Furthermore, the virtual measurement model 114 based on the convolutional autoencoder includes multiple convolutional neural network models (such as CNN model encoders 3E_1, 3E_2, 3E_p) and an estimation model 400. The multiple inputs of these convolutional neural network models (i.e., parameters 1, 2, ..., p) are these paired data 214, and the multiple outputs of these convolutional neural network models (i.e., output y1, output y2, ..., output y...) are... p The input to the estimation model 400 is shown below. A typical convolutional neural network model includes a convolutional layer 310, a pooling layer 320, a flattened layer 330, a dropout layer 340, and a fully-connected neural network 350, wherein the fully-connected neural network 350 includes at least one hidden layer and an output layer. The estimation model 400 has the same structure as the fully-connected neural network 350. Since convolutional neural network models or algorithms are known to those skilled in the art, their details are not described herein.

[0080] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 and Figure 8 ,in Figure 6 It is a drawing Figure 2 A flowchart illustrating the two-stage mechanism 116 based on transfer learning for the computational operation (step S08). Figure 7 It is a drawing Figure 6 A diagram illustrating the first strategic step (Strategy I); and Figure 8 It is a drawing Figure 6 This is a schematic diagram of the second strategy step (Strategy II). As shown in the diagram, during the calculation operation, when this other actual measurement value is not available, prediction phase 500 is executed to calculate the first-stage virtual measurement value (VM) of this other artifact. I Conversely, when another actual measurement value is obtained, transfer learning phase 600 is executed to calculate the second-stage virtual measurement value (VM) of this other artifact. II ).

[0081] In the prediction phase 500, firstly, step 502 is performed to collect process data 102 for a specific workpiece delivered by a process unit. Next, step 504 is performed to check whether the process data 102 for this workpiece has been collected completely. If the result of step 504 is negative, step 502 continues; if the result of step 504 is positive, step 506 is performed to evaluate the DQI of the process data 102. X If the result of step 506 is abnormal, it indicates that the process data 102 is abnormal data, and a notification is issued (step 508); if the result of step 506 is normal, it indicates that the process data 102 is normal data, and step 510 is performed. Step 510 includes performing data alignment on the process data 102 according to the automatic data alignment mechanism (ADAS) 108. Finally, step 512 is performed to perform data alignment on this other set of process data 102 for this other workpiece, and then this other set of process data 102 for this other workpiece is input into the virtual measurement model 114 based on the convolutional autoencoder to calculate the first-stage virtual measurement value (VM) of this other workpiece. I ).

[0082] In the transfer learning phase 600, firstly, step 602 is performed to define a preset parameter value K. Based on this preset parameter value K, the strategy selection is periodically confirmed (i.e., the strategy selection confirmation step 622 is periodically executed). The preset parameter value K is a positive integer. Next, step 604 is performed to set an initial value of a counting parameter value N to 0. The counting parameter value N is an integer. Then, step 606 is performed to collect actual measurement data 104 for a specific workpiece. Next, step 608 is performed to check whether process data 102 for the same workpiece corresponding to this actual measurement data 104 exists, that is, to check the correlation between the measurement data 104 and the process data 102. Afterwards, step 610 is performed to determine whether the correlation check was successful. If the result of step 610 is negative, step 606 continues; if the result of step 610 is positive, DQI is evaluated. yStep 612 is performed to determine whether the actual measurement data 104 is normal. If the result of step 612 is abnormal, a notification is issued (step 614) and step 606 continues; if the result of step 612 is normal, step 616 is performed to confirm whether the production machine status has changed. If the result of step 616 is yes (the production machine status has changed), step 604 is performed; if the result of step 616 is no (the production machine status has not changed), step 618 is performed to set N = N + 1. Next, step 620 is performed to confirm whether the counting parameter value N is equal to the preset parameter value K. If the result of step 620 is no, step 606 continues; if the result of step 620 is yes, a strategy selection confirmation step 622 is performed on the other actual measurement value of the other workpiece to generate a confirmation result, and one of the first strategy step 624 and the second strategy step 626 is executed according to the confirmation result to update the virtual measurement model 114 based on the convolutional autoencoder. Strategy selection confirmation step 622 includes confirming whether the parts of the production machine have been maintained or replaced. If the confirmation result is that the parts of the production machine have been maintained or replaced, then the first strategy step 624 is executed; otherwise, if the confirmation result is that the parts of the production machine have not been maintained or replaced, then the second strategy step 626 is executed.

[0083] The first strategy step 624 includes inputting the time-series data of multiple parameters of the other set of process data 102 and the other actual measurement values ​​of the other set of workpieces into the convolutional neural network models; and inputting the outputs of the convolutional neural network models into the estimation model 400 to update the virtual measurement model 114 based on the convolutional autoencoder. In other words, both the convolutional neural network models and the estimation model 400 are refreshed. Furthermore, the first strategy step 624 also includes retraining the RI model, GSI model, and DQI model. X Model and DQI y Model. The so-called "retraining" refers to retraining each model using updated historical process data 102 and historical measurement values.

[0084] The second strategy step 626 involves inputting the outputs of these convolutional neural network models into the estimation model 400 to update part of the virtual measurement model 114 based on the convolutional autoencoder. In other words, the convolutional neural network models are frozen instead of refreshed; only the estimation model 400 is refreshed. Furthermore, the second strategy step 626 also includes tuning the RI model, GSI model, and DQI model. X Model and DQI yModel. Before performing retraining or calibration steps, the latest set of process data 102 and actual measurement values ​​are used to replace the oldest data in the historical process data 102 and historical measurement values. "Calibration" refers to adjusting the weight values ​​or parameter values ​​of each mode using the updated historical process data 102 and historical measurement values. The execution time of the first strategy step 624 is less than the execution time of the second strategy step 626.

[0085] After executing either step 624 of the first strategy or step 626 of the second strategy, step 628 is performed to update the model, replacing the original virtual measurement model with a tuned or retrained complete set of virtual measurement models, including the CNN model, RI model, GSI model, and DQI. X Model and DQI y These new virtual measurement models are also provided to steps 506, 512, and 612 to evaluate the quality (DQI) of the process data 102 for the next workpiece. X ), and calculate its first-stage virtual measurement (VM) value. I ) and its accompanying Confidence Index (RI) and Global Similarity Index (GSI); and the Quality Index (DQI) for evaluating the actual measurement data 104 of the next workpiece. y Finally, in step 630, another set of process data 102 for this other workpiece is input into the updated virtual measurement model 114 based on the convolutional autoencoder to calculate the second-stage virtual measurement value (VM) for this other workpiece. II ).

[0086] Therefore, the virtual measurement method 200 of the present invention, which integrates convolutional autoencoders and transfer learning, combines modeling and prediction mechanisms of both convolutional autoencoders and transfer learning. It requires only a small amount of paired data 214 for rapid modeling, overcoming the limitation of known deep learning-based algorithms that require the collection of large amounts of paired data 214 for modeling. Furthermore, a two-stage mechanism 116 based on transfer learning is constructed through different learning strategies, enabling online relearning even when conditions change, thus maintaining prediction accuracy.

[0087] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 6 and Figure 9 ,in Figure 9 This illustrates the prediction results of the critical dimension (CD) for different strategy combinations in an application example of the present invention. As shown in the figure, the first-stage virtual measurement (VM) produced by the prediction stage 500 of the present invention... I The second-stage virtual measurement (VM) produced by the transfer learning phase 600II Any of these corresponds to a critical dimension in the Dry Etching Shallow Trench Isolation (Dry Etching STI) process used in semiconductor manufacturing. Figure 9 In the diagram, the horizontal axis represents sampling, the vertical axis represents critical dimensions, the preset parameter value K is 25, the paired data 214 for performing the modeling operation (step S06) consists of 25 data points, the unpaired process data 212 consists of 300 data points, and the paired data 214 for performing the calculation operation (step S08) consists of 79 data points. Curve 710 represents the prediction result of the critical dimension obtained using only the first strategy step 624; curve 720 represents the prediction result of the critical dimension obtained using only the second strategy step 626; curve 730 represents the prediction result of the critical dimension obtained using the method described in this invention. Figure 6 The prediction results for the strategy combinations are shown in Table 1. Point group 740 (composed of "x") represents the actual measured values. Table 1 shows the prediction accuracy of the critical dimensions for different strategy combinations, expressed as Mean Absolute Error (MAE). Figure 9 As shown in Table 1, curve 720 (the prediction result of the key dimensions obtained only using step 626 of the second strategy) is the worst. This is because its initial model only uses the CAE algorithm to extract feature values, therefore the correlation between the actual measurements and the feature values ​​is not well established. In other words, if in Figure 3 In step S0644, the second strategy step 626 (freezing the weights of the convolutional layer) is applied to fine-tune the AVM. CNN If the network is not properly configured, the model will not be able to achieve better prediction results. Furthermore, for curve 710 (which only uses the prediction results of the key dimensions obtained in the first policy step 624), it undergoes initial transfer learning modeling using the first policy step 624, and then... Figure 6 The method flow shown omits the second strategy step 626 (i.e., only the first strategy step 624 is used) for model fine-tuning. This approach reduces prediction accuracy because the model refresh of the convolutional and pooling layers is slightly disturbed by a small number (K=25) of samples. Finally, the curve 730 of this invention uses... Figure 3 and Figure 6 The complete steps allow the model to be established well. Therefore, compared with curves 710 and 720, the accuracy of curve 730 of the present invention is improved by 6.18% and 15.03%, respectively, as shown in Table 1.

[0088] Table 1

[0089]

[0090] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 6 and Figure 10 ,in Figure 10 This figure illustrates the prediction results of critical dimensions using different prediction algorithms in an application example of the present invention. As shown, the upper half represents the prediction results of critical dimensions, and the lower half represents the results of the production machine status. In the production machine status results, the horizontal axis represents sampling, and the vertical axis represents the production machine status. The production machine status changes at sampling point 13, while the present invention uses… Figure 6 Step 616 sets N=0 when the production machine status changes, and steps 606, 608, 610, and 612 are re-executed. In the prediction results of critical dimensions, the horizontal axis represents sampling, the vertical axis represents critical dimensions, the preset parameter value K is 12, the paired data 214 for performing the modeling operation (step S06) consists of 25 data points, the unpaired process data 212 consists of 275 data points, and the paired data 214 for performing the calculation operation (step S08) consists of 56 data points. Curve 810 (AVM) BPNN The image shows the prediction results of the fully automated virtual measurement (AVM) using the Back Propagation Neural Network (BPNN) algorithm; curve 820 (AVM) BPNN Dual Phase represents the prediction results of AVM using the BPNN algorithm and a two-stage mechanism; curve 830 (AVM) CNN The curve 840 (AVM) represents the prediction results using the Convolutional Neural Network (CNN) algorithm. CNN The Dual Phase curve represents the prediction results of AVM using a Convolutional Neural Network (CNN) algorithm and a two-phase mechanism; curve 850 (AVM) CNN (CAE+TL+Dual Phase) is used in this invention. Figure 3 and Figure 6 The prediction results of AVM with convolutional autoencoder and transfer learning are shown; point group 860 (composed of "x") represents the actual measured value. Table 2 shows the prediction accuracy of critical dimensions for different AVM methods when the production machine status changes, with prediction accuracy expressed as MAE. Figure 10 As shown in Table 2, compared with curves 810, 820, 830, and 840, the accuracy of curve 850 of the present invention is improved by 276.36%, 39.23%, 15.70%, and 9.92%, respectively, as shown in Table 2.

[0091] Table 2

[0092]

[0093] It is understood that the virtual measurement method with convolutional autoencoder and transfer learning of the present invention comprises the implementation steps described above, and the computer program product stored in the present invention for measurement sampling is used to complete the measurement sampling method as described above. The order of the implementation steps described in the above embodiments can be rearranged, combined, or omitted as needed. The above embodiments can be implemented using a computer program product, which may include a machine-readable medium storing multiple instructions that can program a computer to perform the steps in the above embodiments. The machine-readable medium may be, but is not limited to, a floppy disk, optical disk, read-only optical disk, magneto-optical disk, read-only memory, random access memory, erasable programmable read-only memory (EPROM), electronically erasable programmable read-only memory (EEPROM), optical card or magnetic card, flash memory, or any machine-readable medium suitable for storing electronic instructions. Furthermore, the embodiments of the present invention can also be downloaded as a computer program product, which can be transferred from a remote computer to a requesting computer by using data signals from a communication connection (such as a network connection).

[0094] It should also be noted that the present invention can also be described in the context of a manufacturing system. While the present invention can be implemented in semiconductor manufacturing, it is not limited to semiconductor manufacturing and can be applied to other manufacturing industries. A manufacturing system is configured to manufacture workpieces or products, which include, but are not limited to, microprocessors, memory devices, digital signal processors, application-specific integrated circuits (ASICs), or other similar devices. The present invention can also be applied to workpieces or products other than semiconductor devices, such as vehicle wheel rims and screws. A manufacturing system includes one or more processing tools that can be used to form one or more products or parts thereof on or in a workpiece (e.g., a wafer, a glass substrate). Those skilled in the art will recognize that the processing tools can be of any number and type, including lithography machines, deposition machines, etching machines, grinding machines, annealing machines, machine tools, and similar tools. In embodiments, the manufacturing system also includes a scatterometer, an ellipsometry, a scanning electron microscope, and similar instruments.

[0095] As can be seen from the above embodiments, the present invention has the following advantages: First, by integrating the modeling and prediction mechanisms of convolutional autoencoders and transfer learning, it can quickly model without requiring a large amount of pairwise data, thus overcoming the limitation of known deep learning-based algorithms that require the collection of large amounts of pairwise data for modeling. Second, by constructing a two-stage mechanism based on transfer learning through different learning strategies, it can still perform online relearning when conditions change, thereby maintaining prediction accuracy and enabling virtual measurement to be more widely used in increasingly sophisticated semiconductor processes in the future. Third, experimental verification shows that compared with known AVM algorithms, the improvement rate of the present invention is at least 9.92%, that is, it has higher prediction accuracy.

[0096] Although the present invention has been disclosed above by way of embodiments, it is not intended to limit the present invention. Any person skilled in the art may make various modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the scope defined in the appended claims.

Claims

1. A virtual measurement method with convolutional autoencoder and transfer learning, characterized in that, Includes the following steps: Multiple sets of process data are obtained, wherein the multiple sets of process data are used or generated when a production machine processes multiple workpieces, the multiple sets of process data are mapped to the multiple workpieces in a one-to-one manner, each of the multiple sets of process data contains the values ​​of multiple parameters, each of the values ​​of the multiple parameters is a set of time-series data respectively corresponding to the multiple workpieces, and each of the multiple sets of time-series data has a data length. Perform a data integration operation on the multiple sets of process data. This data integration operation includes: Perform a data length adjustment operation to repeatedly add at least one data point having the value of the last data point after the last data point of each of the multiple sets of time data of each of the multiple parameters, until the data length of each of the multiple sets of time data of each of the multiple parameters is equal to the longest data length of the multiple sets of process data. Obtain multiple actual measurement values ​​of the multiple workpieces; Perform a modeling operation, which includes: The multiple sets of process data and the multiple actual measurement values ​​are divided into multiple pairs of data and at least one unpaired process data. Each of the multiple pairs of data includes one of the process data and one of the corresponding actual measurement values. and At least one unpaired process data is used to build at least one pre-trained model, and then the plurality of paired data are input into the at least one pre-trained model to build a virtual measurement model based on a convolutional autoencoder, wherein the virtual measurement model based on the convolutional autoencoder includes at least one convolutional neural network model; and Perform a calculation operation, which includes: Obtain at least one of another set of process data and another actual measurement value for another workpiece, and perform one of a prediction stage and a transfer learning stage depending on whether the other actual measurement value is obtained, to calculate one of a first-stage virtual measurement value and a second-stage virtual measurement value for the other workpiece. In the prediction stage, the first stage virtual measurement value is calculated by the other set of process data based on the virtual measurement model based on the convolutional autoencoder. In the transfer learning stage, the second stage virtual measurement value of the other workpiece is calculated by the other set of process data and the other actual measurement value based on the virtual measurement model based on the convolutional autoencoder.

2. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 1, characterized in that, In this calculation operation, When the other actual measurement value is not available, the prediction phase is executed to calculate the first-stage virtual measurement value of the other workpiece. and When the other actual measurement value is obtained, the transfer learning phase is executed to calculate the second-stage virtual measurement value of the other workpiece.

3. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 2, characterized in that, This prediction phase includes: After performing the data integration operation on the other set of process data for the other workpiece, the other set of process data for the other workpiece is input into the virtual measurement model based on the convolutional autoencoder to calculate the first stage virtual measurement value of the other workpiece.

4. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 2, characterized in that, This transfer learning phase includes: A strategy selection and verification step is performed on the other actual measurement value of the other workpiece to generate a verification result. Based on the verification result, one of a first strategy step and a second strategy step is executed to update the virtual measurement model based on the convolutional autoencoder; and Input the other set of process data of the other workpiece into the updated virtual measurement model based on the convolutional autoencoder to calculate the second-stage virtual measurement value of the other workpiece.

5. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 4, characterized in that, The strategy selection confirmation step includes: Confirm whether a part of the production machine has been maintained or replaced; Specifically, when the confirmation result indicates that the part of the production machine has been maintained or replaced, the transfer learning phase executes the first strategy step. Specifically, when the confirmation result indicates that the part of the production machine has not been maintained or replaced, the transfer learning phase executes the second strategy step.

6. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 4, characterized in that, In this modeling operation, The number of at least one convolutional neural network model is multiple. The virtual measurement model based on convolutional autoencoder includes the multiple convolutional neural network models and an estimation model. The multiple inputs of the multiple convolutional neural network models are the multiple pairs of data, and the multiple outputs of the multiple convolutional neural network models are the inputs of the estimation model.

7. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 6, characterized in that, In this transfer learning phase, the first policy step includes: The time-series data of multiple parameters of the other set of process data and the other actual measurement value of the other set of workpieces are input into the multiple convolutional neural network models; and The multiple outputs of the multiple convolutional neural network models are input into the estimation model to update the virtual measurement model based on the convolutional autoencoder.

8. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 6, characterized in that, In this transfer learning phase, the second strategy step includes: The multiple outputs of the multiple convolutional neural network models are input into the estimation model to update part of the virtual measurement model based on the convolutional autoencoder.

9. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 1, characterized in that, The data consolidation operation also includes: A frequency distribution calculation is performed on the data length of each of the multiple sets of time series data for each of the multiple parameters to obtain a distribution of the number of occurrences of data with respect to the data length, wherein the data length that occurs most frequently in the multiple sets of time series data for each of the multiple parameters is a reference data length. The average of the time series data with the reference data length in the multiple sets of time series data for each of the multiple parameters is calculated to obtain a set of reference time series data for each of the multiple parameters; A Dynamic Time Warping (DTW) algorithm is used to calculate a distance between each of the multiple sets of time series data and its corresponding reference time series data for each of the multiple parameters. Set a distance threshold value; and When the distance exceeds the distance threshold, delete the process data corresponding to that distance.

10. The virtual measurement method with convolutional autoencoder and transfer learning as described in claim 1, characterized in that, Either the first-stage virtual measurement produced in the prediction phase or the second-stage virtual measurement produced in the transfer learning phase corresponds to a critical dimension (CD) of a dry etch shallow trench isolation (Dry Etching STI) process applied to a semiconductor process by a production machine.

11. A virtual measurement system with convolutional autoencoder and transfer learning, characterized in that, Include: A memory stores multiple sets of process data and multiple actual measurement values. The multiple sets of process data are used or generated when a production machine processes multiple workpieces. The multiple sets of process data are mapped to the multiple workpieces in a one-to-one manner. Each of the multiple sets of process data contains the values ​​of multiple parameters. Each of the values ​​of the multiple parameters is a set of timing data corresponding to the multiple workpieces. Each of the multiple sets of timing data has a data length. as well as A processor, electrically connected to the memory, receives the plurality of sets of process data and the plurality of actual measurements and configures them to perform operations including the following steps: Perform a data integration operation on the multiple sets of process data. This data integration operation includes: Perform a data length adjustment operation to repeatedly add at least one data point having the value of the last data point after the last data point of each of the multiple sets of time data of each of the multiple parameters, until the data length of each of the multiple sets of time data of each of the multiple parameters is equal to the longest data length of the multiple sets of process data. Perform a modeling operation, which includes: The multiple sets of process data and the multiple actual measurement values ​​are divided into multiple pairs of data and at least one unpaired process data. Each of the multiple pairs of data includes one of the process data and one of the corresponding actual measurement values. and At least one unpaired process data is used to build at least one pre-trained model, and then the plurality of paired data are input into the at least one pre-trained model to build a virtual measurement model based on a convolutional autoencoder, wherein the virtual measurement model based on the convolutional autoencoder includes at least one convolutional neural network model; and Perform a calculation operation, which includes: Obtain at least one of another set of process data and another actual measurement value for another workpiece, and perform one of a prediction stage and a transfer learning stage depending on whether the other actual measurement value is obtained, to calculate one of a first-stage virtual measurement value and a second-stage virtual measurement value for the other workpiece. In the prediction stage, the first stage virtual measurement value is calculated by the other set of process data based on the virtual measurement model based on the convolutional autoencoder. In the transfer learning stage, the second stage virtual measurement value of the other workpiece is calculated by the other set of process data and the other actual measurement value based on the virtual measurement model based on the convolutional autoencoder.

12. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 11, characterized in that, In this calculation operation, When the other actual measurement value is not available, the prediction phase is executed to calculate the first-stage virtual measurement value of the other workpiece. and When the other actual measurement value is obtained, the transfer learning phase is executed to calculate the second-stage virtual measurement value of the other workpiece.

13. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 12, characterized in that, This prediction phase includes: After performing the data integration operation on the other set of process data for the other workpiece, the other set of process data for the other workpiece is input into the virtual measurement model based on the convolutional autoencoder to calculate the first stage virtual measurement value of the other workpiece.

14. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 12, characterized in that, This transfer learning phase includes: A strategy selection and verification step is performed on the other actual measurement value of the other workpiece to generate a verification result. Based on the verification result, one of a first strategy step and a second strategy step is executed to update the virtual measurement model based on the convolutional autoencoder; and Input the other set of process data of the other workpiece into the updated virtual measurement model based on the convolutional autoencoder to calculate the second-stage virtual measurement value of the other workpiece.

15. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 14, characterized in that, The strategy selection confirmation step includes: Confirm whether a part of the production machine has been maintained or replaced; Specifically, when the confirmation result indicates that the part of the production machine has been maintained or replaced, the transfer learning phase executes the first strategy step. Specifically, when the confirmation result indicates that the part of the production machine has not been maintained or replaced, the transfer learning phase executes the second strategy step.

16. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 14, characterized in that, In this modeling operation, The number of at least one convolutional neural network model is multiple. The virtual measurement model based on convolutional autoencoder includes the multiple convolutional neural network models and an estimation model. The multiple inputs of the multiple convolutional neural network models are the multiple pairs of data, and the multiple outputs of the multiple convolutional neural network models are the inputs of the estimation model.

17. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 16, characterized in that, In this transfer learning phase, the first policy step includes: The time-series data of multiple parameters of the other set of process data and the other actual measurement value of the other set of workpieces are input into the multiple convolutional neural network models; and The multiple outputs of the multiple convolutional neural network models are input into the estimation model to update the virtual measurement model based on the convolutional autoencoder.

18. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 16, characterized in that, In this transfer learning phase, the second strategy step includes: The multiple outputs of the multiple convolutional neural network models are input into the estimation model to update part of the virtual measurement model based on the convolutional autoencoder.

19. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 11, characterized in that, The data consolidation operation also includes: A frequency distribution calculation is performed on the data length of each of the multiple sets of time series data for each of the multiple parameters to obtain a distribution of the number of occurrences of data with respect to the data length, wherein the data length that occurs most frequently in the multiple sets of time series data for each of the multiple parameters is a reference data length. The average of the time series data with the reference data length in the multiple sets of time series data for each of the multiple parameters is calculated to obtain a set of reference time series data for each of the multiple parameters; A Dynamic Time Warping (DTW) algorithm is used to calculate a distance between each of the multiple sets of time series data and its corresponding reference time series data for each of the multiple parameters. Set a distance threshold value; and When the distance exceeds the distance threshold, delete the process data corresponding to that distance.

20. The virtual measurement system with convolutional autoencoder and transfer learning as described in claim 11, characterized in that, Either the first-stage virtual measurement produced in the prediction phase or the second-stage virtual measurement produced in the transfer learning phase corresponds to a critical dimension (CD) of a dry etch shallow trench isolation (Dry Etching STI) process applied to a semiconductor process by a production machine.

Images