Automatic Window Generation for Process Tracing

Abstract

Automatic definition of windows for trace analysis. For each process step, the trace data is aligned to both the beginning of the process step and the end of the process step, and statistics including rates of change are calculated from both the beginning of the process step and the end of the process step. A window is generated based on an analysis of the calculated statistics.

Description

Technical Field

[0001] This application relates to the use of process trace analysis for detecting and classifying semiconductor device failures, and more particularly, to a machine-based method for automatically defining windows for process trace analysis.

Background Art

[0002] The detection of device failures by monitoring the time-series traces of device sensors has long been recognized but is a very difficult problem in semiconductor manufacturing. Typically, methods for failure detection and classification ("FDC") begin by splitting complex traces into logical "windows" and then calculating statistics (often called metrics or key numbers) for the trace data within the windows. Metrics can be monitored using statistical process control ("SPC") techniques, mainly based on engineering knowledge, to identify features or anomalies and can be used as inputs for predictive models and root cause analysis. However, the quality of the metrics determines the value of all subsequent analyses: high-quality metrics require high-quality window definitions.

[0003] In the conventional approach, window definitions are mostly manual and are one of the key limitations and the largest cost in the use of FDC techniques. Furthermore, there are existing automatic windowing algorithms, but they typically require extensive manual intervention to generate high-quality windows. Therefore, it would be desirable to find improved techniques for defining and using trace windows in FDC analysis methods.

Brief Description of the Drawings

[0004]

Figure 1A

Figure 1B

Figure 2

Figure 3A

Figure 3B

Figure 4A

Figure 4B

Figure 5A

Figure 5B

Figure 6

Figure 7A

Figure 7B

Figure 7C

Figure 7D

DETAILED DESCRIPTION OF THE INVENTION

[0005] As used herein, the term "sensor trace" refers to time series data that periodically measures important physical quantities, such as sample values of physical sensors at each point in time, during the operation of a device. Note that the sampling rate may vary and the period between samples is not necessarily the same. The term "trace" or "device trace" refers to the set of sensor traces for all important sensors identified for a particular processing instance. The term "step" refers to a distinct device processing period, such as one of the steps in a process recipe.

[0006] Referring to FIG. 1A, an exemplary graphical plot 100 of trace data is presented that represents data taken from 400 individual traces, i.e., time series values from distinct sensors taken during six individual steps I-VI of a semiconductor manufacturing process. Sensor values are plotted on the y-axis against the actual elapsed time measured in seconds on the x-axis starting from the start of step I. Each step of the process is represented as a unique icon in FIG. 1A, which better illustrates and differentiates sensor behavior, especially during transition periods. Note that process steps typically start at a particular point in time, but the length of a step can vary during different executions of the process.

[0007] From a simple visual observation of the sensor data over time in FIG. 1A, the problems associated with the calculation of basic statistical metrics should be apparent, i.e., especially when the trace values are changing rapidly, the statistical measures simply cannot provide a complete picture of the sensor activity. For example, in step IV, clearly a lot is going on and the basic statistical measures cannot fully explain that activity.

[0008] Step I is executed with a nominal sensor value from t = 0 to approximately t = 10 seconds. Step II is executed from approximately t = 10 to t = 13, and the sensor value surges and then drops rapidly. Step III is executed from approximately t = 13 to t = 25, and the sensor value first drops and then increases, with a period of stability at the nominal sensor value between approximately t = 17 and t = 22.

[0009] Step IV has the longest duration of any of these steps, lasting approximately from t = 25 to t = 100. However, an important transition occurs from the start to the end of this step, and some sensor traces begin to decline at approximately t = 75. The end of the sensor trace declines at t > 100, and the sensor values eventually spread even more widely due to fluctuations in the step time. Thus, a clear distinction in time is evident between one group of sensor traces that decline between approximately t = 75 and t = 85 and a second group of sensor traces that decline between approximately t = 90 and t = 100. There is also a long period of stability between approximately t = 45 and t = 75.

[0010] In Step V, the sensor traces transition to the nominal value in two different time groups, and in Step VI, the sensor traces are stable at the nominal value in two different time groups.

[0011] In a conventional approach, technical staff simply attempt to manually establish windows based on a visual review of the graph results, roughly defining the windows where (i) the trace data is stable and (ii) the rate of change is the same. For example, given those objectives, Figure 1B shows an example of a window manually imposed over regions where either the trace data is stable or the rate of change is the same.

[0012] Figure 2 shows one embodiment of method 200 for automatically defining a window useful for evaluating trace data. At item 202, the trace data for each process step is aligned, in the first instance, from the start of the process step and, in the second instance, from the end of the process step. To "align" the data at the start of the process step, the shortest time for each sensor trace within that process step is subtracted from each value of each sensor trace within that process step. Next, the time variables of all sensor traces are interpolated to a constant time step, which is called "StepTime". To "align" the data at the end of the process step, the maximum value of the step time for each sensor trace is subtracted from all values of the step time for that sensor trace, which is called "remStepTime".

[0013] At item 204, statistics are calculated from both the start and end of each process step. Specifically, a new definition of stability is included in the calculated statistics. At item 206, a window is calculated by analyzing the statistics from both the start and end of the step. Each item of method 200 will now be described in more detail.

[0014] For example, FIGS. 3A and 3B show the results of method item 202. In FIG. 3A, the sensor data from step IV for two different recipes is still plotted on the y-axis, but all values are aligned in time at the start of the step. Similarly, in FIG. 3B, the sensor data values from step IV are aligned in time from the end of the step. The advantage of analyzing the window from both the start and end of the trace is apparent when observing the aligned data. Since the processing time or length of the step is variable, it should be clear from the examples of FIGS. 3A and 3B that the statistics for the first part of the trace can be calculated best from the start of the trace and, similarly, the statistics for the last part of the trace can be calculated best from the end of the trace.

[0015] The second key to an effective automatic window definition is to calculate the statistic in item 204 for each point in time of each process step from both the start and the end of the step. By constructing the summary statistic in both directions, it is possible to identify the optimal window even if the steps vary by trace in time. A number of statistics can be seen to be useful, including the median, mean, standard deviation (including robust estimation), and estimated gradient.

[0016] One new additional statistic that is a key to constructing an effective window is the estimation of the rate of change with respect to the trace at each point in time. For the purposes of the present disclosure, this estimation of the rate of change is referred to herein as "stability" and is the best indicator for separating the transition window from the stable window. For example, FIG. 4A illustrates trace data that is aligned and plotted from the start of the step, except that here the sensor values on the y-axis are scaled from 0 to 1 and the rate of change, i.e., stability, is plotted against the step time or the length of the step such that it is equal among sensors that vary in various magnitudes. This scaled sensor value is then used to calculate all the statistics, facilitating the evaluation of those statistics without considering the original range of the sensor. Based on the data in FIG. 4A, the rate of change for each trace is estimated, and FIG. 4B shows the corresponding plot of the stability estimated over the time frame aligned with the same step, aligned at the start of the step. Similarly, FIG. 5A illustrates the scaled trace data for step IV plotted aligned from the end of the step, and FIG. 5B shows the corresponding stability plot over the frame aligned with the same step, aligned at the end of the step.

[0017] When the sensor trace changes smoothly, the rate of change is simply the maximum absolute value of the difference in the sensor values with scale between the current time value and the values before and after, divided by the standard time step. The stability can be the rate of change or some monotonic transformation of the rate of change. If the trace does not change smoothly due to repeated data points or inherent noise, the calculation of the rate of change may require a more complex algorithm based on specific requirements.

[0018] The sensor measurements appear to be stable (low values) at the start of the step (Figure 4B) from approximately t = 3 to t = 18 and from t = 38 to t = 100, but looking at the data from the end of the step (Figure 5B), it is clear from the observation of the stability plot that the sensor measurements appear to be stable from approximately t = 80 to t = 140.

[0019] Following the calculation of the statistics in item 204, the window is calculated in step 206 by analyzing the statistics from both the start and end of each step. Figure 6 illustrates one method 600 for calculating the window using the statistics, and the corresponding graphical results are shown in Figures 7A - 7D. In Figures 7A - 7D, 28 traces showing abnormal behavior were added to the data represented in Figure 1A to demonstrate the ability of the automatic windowing process to handle abnormal traces within the training data set.

[0020] In item 602, the transition periods at the start and end of each step, i.e., the regions where the trace changes rapidly, are determined. This can be done by analyzing the stability and median calculated in item 204 of method 200. Figure 7A shows the analysis results of the statistics for the trace data, in which five windows for the transition regions are indicated by rectangular icons and six windows for the non - transition regions are indicated by circular icons.

[0021] All regions during the transition period are then clustered at item 604 to group adjacent points with similar stability values. This item identifies internal transitions. For example, FIG. 7B shows the result of clustering adjacent data points based on similar stability values. This item results in some splitting of the additional transition and non-transition windows as compared to FIG. 7A.

[0022] Ideally, the transition window should be a window with rapidly changing data between windows with relatively stable data. At item 606, the transition window is extended by a relatively small amount to approach that goal, as shown in FIG. 7C, by including additional points having a small difference in, for example, rate of change / stability factor.

[0023] At item 608, similar windows and very short windows are merged with adjacent windows to reduce the number of windows, and at item 610, the final number and type of windows are estimated based on the statistics of the points within the windows. The result with 11 windows of three separate types is shown in FIG. 7D: 4 transition windows are defined by rectangular icons, 3 stable windows are indicated by circular icons, and 4 general windows (neither stable nor transitional) are shown by triangular icons.

[0024] The window types assigned at item 610 allow the statistical metrics to be customized such that they maximize the quality of the calculated metrics and minimize the number of metrics generated.

[0025] The automatic generation of trace windows is facilitated by the emergence of parallel processing architectures and the development of machine learning algorithms, which enable users to model problems, gain insights, and make predictions using large amounts of data at high speed, making such an approach appropriate and realistic. Machine learning is a field of artificial intelligence involving the construction and study of systems that can learn from data. These types of algorithms, along with parallel processing capabilities, enable much larger data sets to be processed and are particularly well-suited for multivariate analysis.

[0026] The creation and use of a processor-based model for an automatic window hanger can be done on a desktop-based, i.e., stand-alone, or as part of a network system, but given the large amount of information that is processed and presented with some interactivity, the processor capabilities (CPU, RAM, etc.) should be of current state-of-the-art technology in order to maximize effectiveness. In a semiconductor manufacturing plant environment, the Exensio® analysis platform is a useful option for constructing an interactive GUI template. In one embodiment, the coding of the processing routine can be done using Spotfire® analysis software version 7.11 or higher, which is compatible with the Python object-oriented programming language, which is mainly used for coding machine language models.

[0027] The foregoing description has been presented for purposes of illustration only - it is not intended to be exhaustive or to limit the disclosure to the precise form described. Many modifications and variations are possible in light of the above teachings.

Claims

1. A method for automatic window definition for the analysis of semiconductor device trace data, comprising: receiving a plurality of sensor traces from a plurality of device sensors during a plurality of process steps of a semiconductor process; for each respective step of the plurality of process steps, aligning each sensor trace at the start of the respective step; calculating a first rate of change for each sensor trace aligned at the start of the respective step; aligning each sensor trace at the end of the respective step; and calculating a second rate of change for each sensor trace aligned at the end of the respective step; performing; defining a plurality of windows for delimiting, and analyzing the sensor traces within each window based primarily on the first and second rates of change calculated for each of the plurality of sensor traces; A method comprising.

2. The defining step comprises defining a first type of window for delimiting sensor traces within at least a first region where the first rate of change or the second rate of change is increasing or decreasing relative to a threshold; and defining a second type of window for delimiting sensor traces within at least a second region where the first rate of change or the second rate of change does not exceed the threshold; The method according to claim 1, further comprising.

3. For each of the calculating steps, performing a first set of other statistical calculations on each sensor trace aligned at the start of the respective step; and performing a second set of other statistical calculations on each sensor trace aligned at the end of the respective step performing; defining the plurality of windows based on the first and second rates of change calculated for each of the plurality of sensor traces and the first and second sets of other statistical calculations; The method according to claim 1, further comprising.

4. Further comprising scaling each of the plurality of sensor traces prior to the aligning step The method according to claim 1, further comprising.

5. The defining step comprises identifying a plurality of transition periods at the start and end of each process step; During the plurality of transition periods, clustering the input data to group sensor traces having similar rates of change; extending the plurality of transition periods; merging similar windows within and between the transition periods; The method according to claim 3, further comprising.

6. The step of identifying the transition period is identifying regions where the values of the plurality of sensor traces change rapidly The method according to claim 5, further comprising.

7. The step of extending the transition period is incorporating a small portion of adjacent regions where the values of the plurality of sensor traces do not change very rapidly into the transition period The method according to claim 5, further comprising.

8. A method for automatic window definition in the analysis of semiconductor device trace data, comprising: receiving trace data obtained from a plurality of semiconductor device sensors, the trace data being associated with a plurality of process steps within a semiconductor process; aligning the trace data for each process step at the start of the process step in a first instance and at the end of the process step in a second instance; calculating a statistic including a rate of change for a first instance and a second instance of the trace data aligned for each process step; generating a plurality of windows for the trace data from an analysis of the statistics calculated for the aligned trace data; A method comprising.

9. The generating step comprises identifying a plurality of transition periods at the start and end of each process step; clustering the trace data between the plurality of transition periods to group trace data having similar rates of change; extending the plurality of transition periods; merging similar windows; The method according to claim 8, further comprising.

10. The step of identifying the transition period is identifying regions where the plurality of sensor traces change rapidly The method according to claim 9, further comprising.

11. The step of extending the transition period is incorporating a small portion of adjacent regions where the sensor traces do not change very rapidly into the transition period The method according to claim 10, further comprising.

12. scaling the trace data to a registration step The method according to claim 8, further comprising: **Claim 13** A method for automatic window definition in the analysis of semiconductor device trace data, comprising: receiving trace data obtained from a plurality of semiconductor device sensors, the trace data being associated with a plurality of process steps in a semiconductor process; for each step of the semiconductor process, aligning the trace data associated with the step at the start of the step; calculating a statistic including a rate of change for the trace data aligned at the start of the step; aligning the trace data associated with the step at the end of the step; calculating a statistic including a rate of change for the trace data aligned at the end of the step; and analyzing the statistics calculated for the trace data aligned at the start of the step and the trace data aligned at the end of the step performing; calculating a window for the trace data based on an analysis of the calculated statistics; A method comprising: **Claim 14** The step of calculating the window comprises: identifying a plurality of transition periods at the start and end of each process step; clustering the trace data between the plurality of transition periods to group trace data having similar rates of change; extending the plurality of transition periods; merging similar windows; The method according to claim 13, further comprising: **Claim 15** The step of identifying the transition period further comprises: identifying regions where the plurality of sensor traces are changing rapidly The method according to claim 14, further comprising: **Claim 16** The step of extending the transition period further comprises: incorporating small portions of adjacent regions where the sensor traces are not changing very rapidly into the transition period The method according to claim 15, further comprising: **Claim 17** Scaling the trace before the registration step The method according to claim 13, further comprising:

Images