Operation method and operation device of failure detection and classification model

Abstract

A method and device for operating a failure detection and classification model. The method for operating the failure detection and classification model includes the following steps: Continuously obtain a plurality of detection curve graphs. Determine whether these detection curve graphs have changed from a first waveform to a second waveform. If these detection curve graphs have changed from a first waveform to a second waveform, determine whether at least N of these detection curve graphs have changed to the second waveform. If at least N of these detection curve graphs have changed to the second waveform, automatically cut these detection curve graphs that have changed to the second waveform to obtain a plurality of windows. Automatically set an algorithm for each window. Obtain an indicator data of each window through each algorithm. Retrain the failure detection and classification model based on these indicator data.

Description

Technical Field

[0001] The present invention relates to an operating method and an operating device, and more particularly to an operating method and an operating device for a failure detection and classification model. Background Art

[0002] With the rapid advancement of semiconductor technology, process complexity and precision continue to increase. In semiconductor manufacturing, various instrumentation information is analyzed and then used through analytical models for prognostic and health management (PHM) or virtual metrology (VM). If the predicted or measured information from a process instrument is found to be suboptimal, adjustments must be made promptly to avoid the production of large quantities of defective products.

[0003] Traditionally, human effort has been used to monitor analytical models to determine if they require adjustment or retraining with new datasets. However, this approach consumes considerable manpower. Furthermore, with increasing product complexity and process precision, the need for analytical model adjustment or further retraining is increasing. Without the ability to promptly modify analytical models, accurate process monitoring is impossible. Summary of the Invention

[0004] The present invention relates to an operating method and operating device for a failure detection and classification model (FDC model). If the failure detection and classification model is found to be insufficiently trained, the failure detection and classification model can be retrained. If the prediction results of the failure detection and classification model are found to be offset, the failure detection and classification model can be adjusted.

[0005] According to a first aspect of the present invention, a method for operating a failure detection and classification model is proposed. The method for operating a failure detection and classification model includes the following steps. Continuously obtain a plurality of detection curve graphs. Determine whether these detection curve graphs have changed from a first waveform to a second waveform. If these detection curve graphs have changed from a first waveform to a second waveform, determine whether at least N of these detection curve graphs have changed to the second waveform. If at least N of these detection curve graphs have changed to the second waveform, automatically cut these detection curve graphs that have changed to the second waveform to obtain a plurality of windows. Automatically set an algorithm for each window. Obtain an indicator data for each window through each algorithm. Retrain the failure detection and classification model based on these indicator data.

[0006] According to a second aspect of the present invention, an operating device for a failure detection and classification model is proposed. The operating device for the failure detection and classification model includes a data acquisition unit, a management unit, a window cutting unit, an indicator calculation unit, a failure detection and classification model, and a training unit. The data acquisition unit is used to continuously obtain a plurality of detection curve graphs. The management unit is used to determine whether these detection curve graphs have changed from a first waveform to a second waveform, and to determine whether at least N of these detection curve graphs have changed to the second waveform. If at least N of these detection curve graphs have changed to the second waveform, the cutting unit automatically cuts these detection curve graphs that have changed to the second waveform to obtain a plurality of windows. The management unit further automatically sets an algorithm for each window. The indicator calculation unit is used to obtain an indicator data for each window through each algorithm. The training unit retrains the failure detection and classification model based on these indicator data.

[0007] In order to better understand the above and other aspects of the present invention, embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 FIG. 1 is a schematic diagram illustrating process monitoring according to an embodiment.

[0009] Figure 2 FIG. 4 is a block diagram illustrating an apparatus for operating a failure detection and classification model according to an embodiment.

[0010] Figure 3 FIG. 4 illustrates a model retraining process and a model adjustment process according to one embodiment.

[0011] Figure 4 A flowchart illustrating how the failure detection and classification model operates.

[0012] Figure 5 This example illustrates detecting changes in a graph.

[0013] Figure 6 A detection curve diagram with a first waveform and a detection curve diagram with a second waveform according to an embodiment are shown. DETAILED DESCRIPTION

[0014] Please refer to Figure 1, which illustrates a schematic diagram of process monitoring according to one embodiment. In semiconductor processing, various semiconductor tools 900 are used to perform various processes on wafers, such as deposition, etching, and annealing. Sensors 910 are installed on semiconductor tools 900 to sense values ​​such as pressure, temperature, and concentration. Sensors 910 typically perform continuous sensing to obtain a raw trace RT. The raw trace RT can be transmitted to a remote operating device 100 via a network 800. Operating device 100 can be, for example, a computer, a server, a cluster computing center, or an edge computing center. Operating device 100 can input the raw trace RT into a fault detection and classification model (FDC model) MD to output a prediction result RS. Prediction result RS can be, for example, a process yield prediction for prognostics and health management (PHM). Alternatively, prediction result RS can be, for example, an inferred thickness or linewidth of another wafer to achieve virtual metrology (VM).

[0015] As product complexity and process precision increase, the need for adjustment or further retraining of the failure detection and classification model MD is increasing. If the failure detection and classification model MD cannot be corrected in real time, accurate process monitoring and analysis will be impossible.

[0016] Please refer to Figure 2 , which illustrates a block diagram of an operating device 100 for a failure detection and classification model MD according to one embodiment. In addition to performing functions such as prognostic health management (PHM) and virtual measurement (VM) using the failure detection and classification model MD, the operating device 100 is also capable of adaptively retraining and adjusting the failure detection and classification model MD.

[0017] In addition to the failure detection and classification model MD, the operating device 100 also includes a data acquisition unit 110, a management unit 120, a window cutting unit 130, an indicator calculation unit 140, a training unit 150, a user interface 160, an adjustment unit 170, and a database 180. The data acquisition unit 110, the management unit 120, the window cutting unit 130, the indicator calculation unit 140, the training unit 150, the adjustment unit 170, and / or the failure detection and classification model MD may be, for example, a circuit, a chip, a circuit board, program code, or a storage device storing program code. The user interface 160 may be, for example, a display screen, a touch screen, or an operation screen. The database 180 may be, for example, a hard drive, a memory, or a cloud data center. The management unit 120 includes a template selection unit 121, a waveform determination unit 122, and a counting unit 123. The template selection unit 121 , the waveform determination unit 122 and the counting unit 123 are, for example, a circuit, a chip, a circuit board, a program code or a storage device storing a program code.

[0018] After the operation device 100 obtains the detection curve RT via the data acquisition unit 110, if it finds that the failure detection and classification model MD is undertrained, it can retrain the failure detection and classification model MD through the operations of the management unit 120, the window cutting unit 130, the indicator calculation unit 140, and the training unit 150. Furthermore, after the operation device 100 obtains the detection curve RT via the data acquisition unit 110, if it finds that the prediction result RS of the failure detection and classification model MD has deviated, it can adjust the failure detection and classification model MD through the operations of the adjustment unit 170. The operations of each of these components are further described below using a flowchart.

[0019] Please refer to Figures 3-4 , Figure 3 The model retraining process ST1 and the model adjustment process ST2 according to one embodiment are shown. Figure 4 A flowchart illustrating the operation method of the failure detection and classification model is shown. The model retraining process ST1 includes steps S120-S190, and the model adjustment process ST2 includes steps S210-S220. In step S110, the data acquisition unit 110 is used to continuously obtain several detection curves RT. These detection curves RT represent the values ​​detected by the same sensor. Under the same process, the same equipment, and the same parameter settings, the waveforms of these detection curves RT should be similar.

[0020] Next, in step S120, the waveform determination unit 122 of the management unit 120 determines whether the detection curve RT has changed from the first waveform to the second waveform. If the detection curve RT has changed from the first waveform to the second waveform, the process proceeds to step S130; if the detection curve RT remains at the first waveform, the process proceeds to step S210. If a change in the waveform of the detection curve RT is detected, it may indicate a change in the process, machine, or parameter settings, necessitating a model retraining step ST1.

[0021] Then, in step S130, the counting unit 123 of the management unit 120 determines whether at least N images in the detection curve RT have changed to the second waveform. N is, for example, 4 or 5. If at least N images in the detection curve RT have changed to the second waveform, the process proceeds to step S140; if not, the process returns to step S120. For example, please refer to Figure 5 , which illustrates the changes in the detection curve graphs RT1 to RT8. The detection curve graphs RT1 to RT4 have a first waveform, and as time changes, the detection curve graph RT5 has changed to a second waveform. When only one detection curve graph RT5 has been accumulated and changed to the second waveform, the process will return to step S120 to continue to determine whether the next detection curve graph RT6 has also changed to the second waveform. The process will not enter step S140 until four detection curve graphs RT5 to RT8 have been accumulated and changed to the second waveform (taking N=4 as an example). Through step S130, it can be confirmed that the detection curve graphs RT5 to RT8 have indeed undergone continuous changes and will indeed have an impact on subsequent analysis.

[0022] Next, in step S140 , the management unit 120 triggers an alarm signal to notify the operator.

[0023] Then, in step S150, the management unit 120 or the operator determines whether the event is solely an emergency. If so, the process proceeds to step S180; if not, the process proceeds to step S160. Examples of emergencies include unexpected power outages, abnormal pressure leaks, and unstable voltage. These emergencies do not occur continuously, so this step eliminates the emergency and eliminates the need to continue with the model retraining process ST1.

[0024] In step S160, the window cutting unit 130 automatically cuts the detection curve graph RT, which has been changed to the second waveform, to obtain a plurality of windows WD*. In this step, the template selection unit 121 of the management unit 120 first selects a golden sample GS from the detection curve graph RT, which has been changed to the second waveform. After the sample GS is selected, it is passed to the window cutting unit 130, which cuts the sample GS according to the trace type. These trace types include a constant trace, a fluctuating trace, a zero-point trace, a process trace, an ascending trace, a descending trace, a regional peak trace, and a regional valley trace. Different trace types are cut into different windows WD*.

[0025] Please refer to Figure 6 , which depicts a detection curve graph RT61 having a first waveform and a detection curve graph RT62 having a second waveform according to one embodiment. The detection curve graph RT61 having the first waveform is cut out to form a window WD. The detection curve graph RT61 having the second waveform is then re-cut out to form a window WD*.

[0026] Then, in step S170, the indicator calculation unit 140 automatically sets an algorithm for each window WD*. The management unit 120 can set different algorithms based on different track types, such as calculating the average, slope, standard deviation, peak, or wave number.

[0027] Next, in step S180 , the indicator calculation unit 140 obtains indicator data IC* of each window WD* through various algorithms.

[0028] Then, in step S190, the training unit 150 retrains the failure detection and classification model MD according to the indicator data IC* to obtain an updated failure detection and classification model MD* (shown in FIG. Figure 3 ).

[0029] The model retraining procedure ST1 is completed through the above steps S120 to S190. Figure 3 As shown, if the failure detection and classification model MD is found to be undertrained, new windows WD* can be automatically cut out and indicator data IC* calculated based on these windows WD*. Then, the indicator data IC* is used for retraining to obtain an updated failure detection and classification model MD*.

[0030] The model adjustment procedure ST2 includes steps S210 to S220. In step S210, it is confirmed that the detection curve graph RT maintains the first waveform, and the adjustment unit 170 determines whether the prediction result RS of the failure detection and classification model MD deviates. If the prediction result RS of the failure detection and classification model MD deviates, the process proceeds to step S220; if the prediction result RS of the failure detection and classification model MD does not deviate, the process proceeds to step S180. In this step, the predicted line width and predicted thickness of the prediction result RS can be confirmed by electron microscope (SEM) images to see whether there are deviations such as being too low or too high.

[0031] Next, in step S220, the adjustment unit 170 adjusts the prediction result RS of the failure detection and classification model MD to the prediction result RS*. In this step, the adjustment unit 170 performs a translation adjustment on the prediction result RS to obtain the prediction result RS*. In this way, if the prediction result RS of the failure detection and classification model MD is found to be offset, the prediction result RS can be adjusted without updating the failure detection and classification model MD to obtain an accurate prediction result RS*.

[0032] In summary, although the present invention has been disclosed above with reference to the embodiments, these are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations can be made 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 appended claims.

Claims

1. A method for operating a failure detection and classification model, used in semiconductor processing, wherein various processes are performed on wafers using various semiconductor tools, and sensors are disposed on the semiconductor tools. The method comprises: The sensor is used to continuously sense and obtain a plurality of detection curve graphs, and the failure detection and classification model performs failure detection and classification on these detection curve graphs; determining whether the detection curve graphs have changed from a first waveform to a second waveform; If the detection curves have been changed from the first waveform to the second waveform, determining whether at least N of the detection curves have been changed to the second waveform; If at least N of the detection curve graphs are changed to the second waveform, the detection curve graphs changed to the second waveform are automatically cut according to multiple track types to obtain multiple windows; Automatically setting an algorithm for each window according to the plurality of trajectory types; Obtaining index data of each window through each algorithm, each index data being an average value, a slope, a standard deviation, a peak value, or a wave number; and The failure detection and classification model is retrained based on the indicator data.

2. The method for operating a failure detection and classification model according to claim 1, further comprising: If at least N of the detection curves are changed to the second waveform, an alarm signal is triggered.

3. The method for operating a failure detection and classification model according to claim 1, further comprising: A sample graph is selected from the plurality of detection curve graphs that are changed into the second waveform. 4 . The method for operating a failure detection and classification model as claimed in claim 1 , wherein the indicator data are displayed on a user interface.

5. The method for operating a failure detection and classification model according to claim 1 , further comprising: Determining whether the failure detection deviates from the prediction result of the classification model and whether the plurality of detection curve graphs maintain the first waveform; as well as If the prediction result of the failure detection and classification model deviates and the plurality of detection curves remain at the first waveform, the prediction result of the failure detection and classification model is adjusted. 6 . The method for operating a failure detection and classification model as claimed in claim 5 , wherein in the step of adjusting the prediction result of the failure detection and classification model, a translation adjustment is performed on the prediction result. 7 . The method for operating a failure detection and classification model as claimed in claim 1 , wherein the windows have different types of traces. 8 . The method for operating a failure detection and classification model as claimed in claim 7 , wherein the plurality of trajectory types include a constant trajectory, a fluctuating trajectory, a zero-point trajectory, a process trajectory, an ascending trajectory, a descending trajectory, a regional peak trajectory, and a regional valley trajectory.

9. A device for operating a failure detection and classification model, used in semiconductor processing, wherein various semiconductor tools are used to perform various processes on wafers, and sensors are installed on the semiconductor tools. The device for operating a failure detection and classification model comprises: a data acquisition unit for continuously acquiring a plurality of detection curve graphs by performing continuous sensing using the sensor, and the failure detection and classification model performing failure detection and classification on these detection curve graphs; a management unit configured to determine whether the detection curves have changed from the first waveform to the second waveform, and to determine whether at least N of the detection curves have changed to the second waveform; a window cutting unit, wherein if at least N of the detection curve graphs are changed to the second waveform, the window cutting unit automatically cuts the detection curve graphs changed to the second waveform according to a plurality of trajectory types to obtain a plurality of windows, and the management unit further automatically sets an algorithm for each of the windows according to the plurality of trajectory types; An indicator calculation unit, configured to obtain indicator data of each window through each algorithm, wherein each indicator data is an average value, a slope, a standard deviation, a peak value, or a wave number; The failure detection and classification model; as well as The training unit retrains the failure detection and classification model based on these indicator data. 10 . The operating device of the failure detection and classification model as claimed in claim 9 , wherein if at least N of the plurality of detection curves change to the second waveform, the management unit triggers a warning signal. 11 . The operating device of the failure detection and classification model as claimed in claim 9 , wherein the management unit is further configured to select a sample graph from the plurality of detection curve graphs that are transformed into the second waveform.

12. The apparatus for operating a failure detection and classification model according to claim 9, further comprising: User interface for displaying these indicator data.

13. The apparatus for operating a failure detection and classification model according to claim 9, further comprising: An adjustment unit is used to determine whether the prediction result of the failure detection and classification model deviates and whether these detection curve graphs maintain the first waveform. If the prediction result of the failure detection and classification model deviates and the multiple detection curve graphs maintain the first waveform, the adjustment unit adjusts the prediction result of the failure detection and classification model. 14 . The operating device of the failure detection and classification model as claimed in claim 13 , wherein the adjustment unit performs a translation adjustment on the prediction result. 15 . The apparatus for operating a failure detection and classification model as claimed in claim 9 , wherein the windows have different types of traces. 16 . The operating device of the failure detection and classification model according to claim 15 , wherein the plurality of trajectory types include a constant trajectory, a fluctuating trajectory, a zero-point trajectory, a process trajectory, an ascending trajectory, a descending trajectory, a regional peak trajectory, and a regional valley trajectory.

Images