Você está na página 1de 5

Lean Randomization and Exception Handling

Jan Rothe, Robert Barlovi , Andreas Becker, Raymond Goss, Dirk Grossmann, Wolfgang Jckel, Sebastian Knappe, Thomas Kowtsch, Stephan Melzig, Mark Reiche, Uwe Schulze, Heiko Wagner, Hermann Weindl, Daniel Zschbitz GLOBALFOUNDRIES Wilschdorfer Landstrae 101 01109 Dresden, GERMANY
Abstract In recent years, randomization of process order as well as randomization of the wafer map in the cassettes delivering the wafers was increasingly used to be able to correlate yield variations with specific segments of the process flow (see [1] and [2]), as well as to reduce yield influences on specific wafer positions such as top and bottom of the wafer cassette. Many randomization efforts relied on wafer sorter based randomization, which costs cycle time and wafer sorter capacity. In this paper, we describe how randomization of both process sequence as well as wafer map changes can be implemented on process equipment as part of normal processing. This kind of randomization does not add cycle time and does not require wafer sorters and will therefore be called lean randomization. We will also describe some of the most important exceptions that can occur when implementing lean randomization and we will discuss how a fab can deal with those exceptions.
I. INTRODUCTION

of yield signals. However, such a step leaves wafer order inside the cassette unchanged and therefore can lead to yield loss (e.g. on the top or bottom wafers). It also reduces the ability to identify slot specific yield signals. The alternative to logical randomization utilizes the process equipment for changing the wafer map inside the cassette, which mimics the physical randomization that is usually implemented using wafer sorters. In this paper we will describe all the potential randomization methods and compare their benefits and disadvantages. We will show how randomization on process equipment can be implemented using existing standards and we will discuss the exceptions that can occur in such a scenario, as well as the potential exception handling.
II. THE DIFFERENT KINDS OF RANDOMIZATION

In order to eliminate wafer position influences on product quality, it has become standard practice to perform randomization steps as part of normal lot movement through a manufacturing line. This way it can be assured that for instance the topmost position in a wafer cassette is populated by different wafers during the life of a lot, so that different wafers are exposed to outgases or environmental light. Changing the wafer map in a cassette is typically done using wafer sorters, which come equipped with their own randomization algorithms. In the standard scenario, different wafer positions in the cassette also lead to different process order. The resulting process order information can later be used to correlate yield data for single or multiple lots to specific process segments, such as e.g. a specific implant or metal layer, thus generating a yield signal. Consequently, this capability allows for quick identification of problematic process sequences, or specific tools and chambers. Finding problems quicker ultimately sped up yield learning. As every fab tries to further shorten learning cycles, there is a natural desire to increase the number of randomization steps so that smaller process segments can be separated based on process order. Utilizing wafer sorters for randomization requires adequate sorter capacity (and thus floor space), and impacts cycle time. As a result, at the same time the fab wants to see more randomization in the process flow, yet less usage of wafer sorters, especially for randomization purposes. Randomization of the process order at process equipment (logical randomization) allows wafer processing to be decoupled from wafer order in the cassette, which again allows using the process order information for the generation

In this section we will discuss several different kinds of randomization in semiconductor manufacturing, together with anticipated or real advantages or drawbacks. A. Randomization on wafer sorter The classical randomization approach inserts wafer sorter operations into the production flow. These wafer sorter steps change the wafer order in the cassette in a random or predetermined way. Through this method, manufacturing equipment succeeding the wafer sorter step processes the wafers in a different order than the equipment preceding the wafer sorter step when processing the wafers in slot order. One advantage of utilizing wafer sorters is that they can verify the wafer IDs, and thus establish the confidence that slot based job setup specifies the right settings for the right wafers. Unfortunately, this approach is quite expensive in terms of wafer sorter capacity, floor space and cycle time, especially when a larger number of randomization steps are inserted into the manufacturing flow. B. On-equipment randomization without changing the wafer order in the cassette A different and potentially complementary version of randomization is to decouple process order from wafer order in the FOUP. In such a scenario, the wafer order in the cassette remains unchanged, while the process order on the specific operation is different than wafer order. This way, it becomes possible to process wafers differently on succeeding operations.

978-1-4244-6519-7/10/$26.00 2010 IEEE

236

ASMC 2010

The big advantage of this approach over the classical wafer sorter based randomization is that it comes free of charge in terms of wafer sorter capacity and cycle time. Unfortunately, this kind of randomization does only decouple the processing on one operation from the previous or succeeding operations. To achieve a complete separation of process order of successive operations, it is necessary to implement this kind of randomization on many different equipment platforms, which leads to significant automation implementation and qualification efforts. However, on-equipment randomization can be useful to decouple some of the operations between wafer sorter randomizations from other operations, and therefore could provide useful results at low effort when e.g. implemented on just some families of deposition equipment. C. Randomization including the change of wafer position in the FOUP. Implementing on-equipment randomization that does change the wafer position in a cassette would be the closest one can get to mimicking wafer sorter based randomization. And there are several ways to implement such a randomization, depending on the timing the randomization occurs: (1) Wafers are processed in slot order, as usually the case, but are being put back into the cassette in a different (random) order. This scenario corresponds to implementing a wafer sorter randomization after the process operation that carries out the
S16F11 //PRJobCreateEnh <L[7/1] <U4 4> <A "JobID"> <B d> <L[1] <L[2] <A CarrierID"> <L[n] <U1 1> <U1 2> : <U1 24> <U1 25> > > > <L[3] <U1 1> <A "RecipeID"> <L[0]> > <B ff> <L [0]> // no instruction // necessary here. >

randomization. (2) Wafers are processed in a random order, but placed back into the cassette in the order of processing, i.e. the first processed wafer is being put into slot 1, the second into slot 2, This scenario corresponds to a wafer sorter randomization before the process operation. (3) Wafers are processed in a random order, and placed into arbitrary slots after processing. This scenario corresponds to a process flow in which a wafer sorter randomization precedes the process operation and another wafer sorter randomization succeeds it. (4) Finally, there is also the possibility to process wafers as usual and use the equipments front-end robot to perform a wafer map randomization after all wafers have finished processing, just like a wafer sorter would perform it. This option however does increase cycle time and it impacts the ability of the equipment to process wafers from other FOUPs (if the robot is a throughput bottleneck of the equipment). We will therefore not consider this option further in this paper.
III. IMPLEMENTING RANDOMIZATION

In this section, we will describe how the different types of randomization from the previous section can be implemented,
S16F11 //PRJobCreateEnh <L[7/1] <U4 4> <A "JobID"> <B d> <L[1] <L[2] <A CarrierID"> <L[n] <U1 17> random <U1 6> slot : order <U1 16> determined <U1 1> by host. > > > <L[3] <U1 1> <A "RecipeID"> <L[0]> > <B ff> <L [1] <L [2] <A PRMtrlOrder> <U1 3> > > >

slot numbers of the wafers that need to be processed

Instructing the job to process the wafers in the order determined above

Figure 1 The ProcessJob creation messages used for classical processing (left) and for processing in hostdetermined random order (right).

237

ASMC 2010

including SEMI standard messaging and the need to obey certain side-conditions. In this section we will limit ourselves to explaining hostdriven randomization approach utilizing SEMI standard behavior. In these scenarios an external system (the host) determines the process order. We will not discuss approaches in which the equipment gets equipped with a special randomization feature. This section therefore assumes the availability of 300mm SEMI Standards implementations, specifically SEMI E40, Standard for Processing Management [3], and SEMI E94, Specification for Control Job Management [4]. We will only describe on-equipment randomization while randomization on wafer sorters will not be discussed as it is assumed to be common practice. A common requirement for on-equipment randomization scenarios is the need to collect process order data in an appropriate engineering data base, to enable yield correlation to process sequence. This data collection can be performed either once the job is created, after the processing has started at the beginning of lot processing or separately for each wafer, or the data can be collected at the end of processing. In the ideal case either option is viable. However, in exceptional cases, the data collection option chosen will impact the level of data collection clean-up, see Section IV. EXCEPTION HANDLING. B. On-equipment randomization without changing the wafer position in the cassette Decoupling the process order from the wafer order in the FOUP is possible by several means, depending on the way jobs are set up at the equipment. If the user implements one process job per cassette, randomization obviously has to occur within the process job. The job attributes used to influence the process order behavior are PRMtrlNamelist (specifying the order of wafers in the job) and PRMtrlOrder (specifying whether the process job should obey the order in which the wafers have been specified). Figure 1 on the previous page shows the features used for a randomizing process job creation in comparison to a classical process job creation message. Process job based randomization is however only possible, if the process job specifies processing for more than one wafer. If on the other hand the user implements one process job per wafer, e.g. to allow wafer-level run-to-run control, the randomization cannot be achieved through process job internal measures. In this case, the control job managing the process jobs will have to schedule process job execution in the desired order. This can be achieved by setting the appropriate ProcessingCtrlSpec and ProcessOrderMgmt attributes. The ProcessingCtrlSpec attribute has to be used to specify the order in which the process jobs shall be executed, while the ProcessOrderMgmt attribute, when set to value LIST makes sure the control job obeys the order defined in the ProcessingCtrlSpec. These features work similarly to the

<L [1] <L [2] <L [2] <A [4] 'CarrierID' > <L [25] <U1 [1] 19> <U1 [1] 3> <U1 [1] 10> <U1 [1] 2> <U1 [1] 12> : > > <L [2] <A [4] 'CarrierID' > <L [25] <U1 [1] 1> <U1 [1] 2> <U1 [1] 3> <U1 [1] 4> <U1 [1] 5> : > > > >

source slots

destination slots

Figure 2. A ControlJobs MtrlOutSpec attribute set up for changing the wafer order after processing. process job features shown in Figure 1, which is why we omit them from this paper. If the user has to set up process jobs for groups of wafers but not the full lot, a full randomization will not be possible (because execution of wafers from different groups cannot overlap), but some randomization within the groups can be achieved on the process job level, while randomization of group order is possible on the control job level. C. Randomization including the change of wafer position in the FOUP. All three kinds of wafer position changing randomization on process equipment have to be implemented on the Control Job level, utilizing the MtrlOutSpec feature of the control job. While usually this attribute is either left empty, or it instructs the job to place wafers back into their original slots, altering the slot order is easily possible with a MtrlOutSpec as shown in Figure 2. In the example of Figure 2, the wafer originating from slot 19 would be moved to slot 1 after processing, the wafer from slot 3 would be moved to slot 2, etc. If the Process and Control Job(s) in the example above have been set to process the wafers in the order 19,3,10,2,12 the combination with the given MtrlOutSpec results in -equipment based randomization that replaces a wafer sorter based randomization before the process step. On the other hand, if process

238

ASMC 2010

order would follow the slot order, the example MtrlOutSpec would lead to a wafer map change after the process operation as if the process operation had been followed by a sorter based randomization. Finally, if the process order specified within process and/or control job was different from incoming and outgoing slot order, the example MtrlOutSpec would replace a preceding and a succeeding sorter based randomization. In this example however the host has to obey an important side-condition: The destination slot for a wafer has to be empty at the time the wafer finishes processing. This is nontrivial on many process tools, as the number of wafers in processing influences the distance between the start of processing of the (first) wafer, and the start of the wafer from the target slot of this first wafer. Consider equipment that keeps 5 wafers in processing. If the first wafer scheduled for processing is scheduled to be placed into slot 19, the wafer originating from slot 19 has to start processing at latest as fifth wafer to make sure that slot 19 is not populated when the first wafer finished processing. Similarly, all other wafers have to be processed in an order that makes sure that no wafer is scheduled to be placed into a target slot that is still populated. This requirement can only be somewhat relaxed if the cassette is not completely filled with wafers, so that the free slots can be used for randomization. On equipment that only contains one wafer at a time, randomization of the outgoing wafer map will therefore only be possible, when the cassette delivering the wafers has at least one free slot. In this case, the first wafer(s) processed could be placed into the free slot(s). The only easy example where at least all source wafer slots can always be guaranteed to be empty when a wafer finishes processing are litho clusters. These tools hold several dozen wafers from several cassettes in processing in parallel, and all cassettes become completely empty during processing (if all wafers are scheduled for processing). This is one of the reasons GLOBALFOUNDRIES decided to first implement wafer map randomization on litho equipment. In addition to mastering the equipment scenario and the data collection to an engineering data base for later yield analysis, it is also crucial to feed the wafer map update to the MES, as succeeding operations typically rely on the wafer position for sampling, rework or run-to-run adjustment, rather than employing a WaferID reader. We will discuss the potential effects of a missing wafer map update and ways to deal with these risks in the next section.
IV. EXCEPTION HANDLING

While the implementation of the scenario and the integration with our manufacturing execution system (MES) and data collection infrastructure requires some effort, the implementation of necessary exception handling played a crucial role in the overall architecture of the final solution. In this section, we describe some of the risks that are introduced when implementing lean randomization, and we

discuss several exception handling options considered during the development of a lean randomization solution. Missing Randomization Data Collection An extremely rare exception is missing data collection about the process order. In such a case, yield attribution to process order will not be possible. As this scenario however occurs only extremely rarely, and only generates a potential loss in analysis capability, dealing with this risk is clearly outweighed by the cost of implementing a sophisticated exception handling. Missing Updates of Wafer Map to MES On the other hand a missing wafer map update to the MES in many cases leads to misprocessing, resulting in unnecessary rework or wafer scrap can occur. For example if a wrong wafer is measured after processing (e.g. to verify chamber performance) due to an outdated wafer-to-slot attribution in the MES, wrong run-to-run feed-back can occur and lead to misprocessing of the subsequent lots at the process operation. On the other hand, if feed-forward run-to-run adjustments are based on either the chamber used, or directly on metrology results, relying on the correct wafer map in the cassette can lead to misprocessing of the lot that did not get the wafer map update to the MES. So even though missing wafer map updates to the MES are very rare, a proper exception handling needs to be implemented to eliminate the related misprocessing risk. This risk could be mitigated through the installation of wafer ID readers on process and metrology equipment, ensuring that the right wafers are being processed or measured. But alone the effort to buy and install such devices is prohibitive for a counter-measure to lean randomization exceptions. Furthermore, the qualification of wafer ID readers can require considerable engineering effort. One could also try to implement post-process wafer ID verification on randomizing equipment. But even this cannot reliably prevent the missing data collection to the MES, because the wafer map update might still be missing in case of process aborts (see below). An alternative to a hardware solution is to prevent processing of the cassette if the randomizing process operation has not updated the wafer map accordingly. This can be done by appropriate flags for a lot in processing. Process Aborts Process Aborts already pose a challenge in the standard operational scenario, where no randomization has occurred. As a result of such an incident it is always necessary to determine (1) which wafers have been fully processed according to specification, (2) which wafers have been partially processed, potentially requiring rework, and (3) which wafers have not been processed at all. In the usual processing scenario, this information is easily accessible from equipment logs, equipment messages to the host. The fact that wafers are being processed consecutively from slot 1 to slot 25, makes it easy to determine the wafers that have to be treated separately, as the wafer map is exactly split into processed, partially processed and unprocessed wafers.

239

ASMC 2010

When a process abort happens during a job that was randomizing, the wafer map typically does not show a nice split between fully, partially and unprocessed wafers, as wafer order and process order are independent from each other. This makes extensive operator training necessary to enable them deciding the correct next steps. Furthermore, a randomization GUI can help to give operators access to the necessary information for exception handling. In the worst case, a process abort can leave the wafer map inside the cassette in an undetermined state, as the equipment might not have been able to perform all of the necessary wafer position changes (e.g. because the destination slot of a partially processed wafer is still populated). Such situations need to be detected by either the operator (in case of manual aborts in Offline mode) or the host (in case of host-triggered aborts). Ultimately however, this situation can be dealt with in the same way as a missing wafer map update to the MES. It is clear that a subsequent verification of the wafer map (on a wafer sorter) has to be performed before an operator can make the decision about rework or restart of the process. The sequence in the right part of Figure 4. Cycle Time savings obtained through implementation of lean randomization in the litho area. below shows that an optional wafer sorter step can be implemented to allow for wafer map verification in case of missing updates to the MES. Post-Process Metrology Findings When post-process metrology finds conspicuous wafers, where process results are different from the fabs expectations, operators typically try to determine the cause of the excursion, e.g. to inhibit a certain chamber for processing. If however metrology was only done for a subset of the lot, such an analysis can require re-measuring more wafers, e.g. to determine a pattern in the distribution of conspicuous wafers. Such a pattern however will not be easily visible if the wafer map does not reflect process order on the randomizing process equipment. Supporting the operators with a user interface that shows process order together with metrology overview of the wafers allows to perform the necessary analysis in case of post-process metrology alerts. Randomization after rework Finally, a fab implementing lean randomization has to decide whether wafers or lots that have been reworked and
Randomization Operation

Cycle Time [h] used for Randomization

5 4 3 2 1 0 Standard Randomize Lean Randomize @ Current Lean Randomize @ Every Mask Randomization Mask Layers Layer Randomization Option

Figure 3. Cycle Time savings obtained through implementation of lean randomization in the litho area. need to be reprocessed on the randomizing operation should receive a randomization treatment again. The left part of Figure 3 shows that in case of the classical sorter-based randomization, this question can be answered by configuring the rework route to re-enter the process flow either before or after (in the picture) the randomizing sorter operation. In the right part of Figure 3, where randomizing as well as process operation have been merged together, it is shown that such a decision has to be implemented through flags or other set up to prevent or enforce randomization of (sub)lots after rework.
V. CONCLUSIONS.

GLOBALFOUNDRIES has implemented lean randomization in the litho area using randomization including the wafer position change inside the FOUP. Through this effort, we have implemented randomization steps that would have otherwise consumed ~30% of the wafer sorter capacity reserved for on-route operations. At the same time, the implementation of an equivalent number of wafer sorter steps would have consumed 5 hours of lot cycle time (see Figure 4). Observe however, that the cycle time used for the randomization does not decrease to zero, as the new exception handling mechanisms require some time in case one of the exceptions described above occurs. Ultimately, the new scenario allows implementing randomization on every mask layer without any additional cost in terms of cycle time or wafer sorter capacity. It has thus become possible to greatly improve the ability to attribute yield signals to certain process segments, tools and chambers, and to decouple process order on litho, deposition, and etch steps of successive mask layers.
REFERENCES Yield improvement using statistical analysis of process dates, F Bergeret, C Le Gall; IEEE transactions on semiconductor manufacturing, vol. 16 no 3 2003 Achieving Rapid Yield Improvement, R. Kittler et.al, Semiconductor International, 7/1/2004 SEMI E40-0709, Standard for Processing Management, available from SEMI via www.semi.org. SEMI E94-0309, Specification for Control Job Management, available from SEMI via www.semi.org.

Randomizing Process Operation

Process Operation
process rework process rework

Wafer map verification in error scenario

[1] [2]

Post-Litho Metrology

[2]
Post-Litho Metrology

[3]

Figure 4. Route flow with classical (left) and lean (right) randomization.

240

ASMC 2010

Você também pode gostar