CN100533677C - A method of fault detection in manufacturing equipment - Google Patents

Abstract

A method of fault identification on a semiconductor manufacturing tool includes monitoring tool sensor output, establishing a fingerprint of tool states based on the plurality of sensors outputs, capturing sensor data indicative of fault conditions, building a library of such fault fingerprints, comparing present tool fingerprint with fault fingerprints to identify a fault condition and estimating the effect of such a fault condition on process output. The fault library is constructed by inducing faults in a systematic way or by adding fingerprints of known faults after they occur.

Description

Error detecting method in the manufacturing equipment

Technical field

The present invention relates to the error detecting method in the manufacturing equipment, more specifically, relate to but be not limited only to use the semiconductor manufacturing facility of plasma-reaction-chamber.

Background technology

The manufacturing of semiconductor integrated circuit is the retrofit of the many complex steps of needs.Typical semiconductor manufacturing factory (or fab, wafer factory) need the instrument of a hundreds of high complexity on silicon substrate or wafer, to make complex apparatus, as microprocessor or storage chip etc.The manufacture process of single-chip often wants more than 200 independent step to finish.These steps comprise that the offset printing pattern of formulating semiconductor wafer to define each equipment, corrodes circuit with the establishment structure, and come blind to create interested electronic equipment with metal or insulator.This process needs several weeks just can finish from start to end.

Mistake may and appear among these tools of production really.Mistake on the single-chip may jeopardize all devices on this wafer, and all later step on this wafer all may be unworthy, and wafer goes out of use.Therefore error checking is very necessary timely and effectively.Fig. 1 has described an example of semiconductor production instrument, shows transducer output 4, the data acquisition interface 5 of a plasma-reaction-chamber 1, the substrate that will process 2, processing input or set-point 3, tool state and a machining state.

The tool of production is very complicated and various mistake may take place, some mistake is that ongoing tool processes is specific, this has just influenced the productivity ratio and the output (under the situation of plasma-reaction-chamber, any preset time of ongoing processing is called as " prescription " in this area) of equipment.For example a kind of mistake that may occur is considered heat chemistry vapor deposition (CVD) instrument, and it is to be used for deposited semiconductor or layer of insulator material in manufacturing equipment.Crudy is by output decision, measures by some and measures, as the uniformity of film, pressure or the like.The quality of output depends on the input of processing, for example air velocity, reactor pressure and the temperature etc. under the situation of heat chemistry vapor deposition (CVD) instrument again.If any one machined parameters produces deviation, then may produce negative influence to the quality of processing output.

Another kind of mistake relates to the skew in the processing self.A lot of examples is arranged, comprise harm vacuum in the reative cell, the change of reactor wall conditioned disjunction chamber hardware, electric arc, perhaps or even the problem of wafer of input.The output that this also can influence the quality of output and then influence instrument.

A universal feature in all these mistakes is that the transducer on the instrument is the variation of indication mechanism state briefly, although this depends on the sensitivity of tool sensor really.Plasma production chamber generally all is equipped with as airometer, manometric tool state sensor and machining state transducer, for example optical detector and impedance monitors.If the input of processing has changed, some tool sensor will write down this variation so usually.If the condition of process chamber has changed, same tool sensor will write down once and change.

The universal method of machining control on the semiconductor manufacturing facility instrument and error detection occurs is statistics machining control method (SPC), if not the processing input all, so also has much to go on record thus, at uncontrolled incident, monitors control diagram.Fig. 2 shows based on a typical SPC chart from the sensing data of semiconductor manufacturing tool.The control restriction is based on impossible deviation on the statistics of data mean value.In Fig. 2, they are shown as a upper control limit (UCL) and a lower control limit (LCL).Generally speaking, these restrictions are set to use 3 to 4 times of standard deviation (σ) of average of the data set of normal distribution model.This control technology has certain limitation.

First problem is that all SPC charts of monitoring can not be expanded, because each instrument has tens transducers, and work in-process has a hundreds of instrument.The output that second problem is each transducer control range of may drifting about out, and the output of processing is not obviously influenced, and/or the input of processing can remain in the control range, but the output of processing because the change of processing conditions can be drifted about out control range.This is because machining tool is generally very complicated, and their output dependence is in their condition of uniting input and equipment self.For this reason, semiconductor machining adopts the sampling of the conventional quality of production of testing wafer usually, because this at least can forecast production.For example, frequently move testing wafer, to check crudy, as the rete pressure in the CVD processing, or the critical dimension in the corrosion processing (CD).Hence one can see that, and this is a very expensive processing control method, because operation testing wafer and time-out actual production come test process quality can have a negative impact to output of a factory and productivity ratio.The 3rd problem relates to the difficulty that the SPC scope is set on the device senses device.The SPC method is added up, and the data of hypothesis normal distribution.Situation is not so usually.The drift of instrument and transducer, and normal instrument intervention such as preventive maintenance (PM) property are to cause the Non-Gaussian Distribution data set.

Fig. 3 showed in the cycle of about 1100 wafers, from the output parameter 1 of the transducer of oxide etch plasma process instrument and two data flow of 2, at this moment between inherent wafer number 1018 places, detect a pressure mistake.It has the pressure controller of mistake to cause by one.In the interval before mistake, carry out the wet cleaning of two preventive maintenances (PM) of plasma-reaction-chamber.The periodic effect of these PM incidents and reative cell is apparent in initial data.Also as can be seen, these data are unusual normal state not, have autocorrelation and discontinuity.Therefore the SPC method can not handle this data effectively, and critical event may be lost in these data.Certainly in the example of Fig. 3, the mistake that occurs at wafer 1018 places can not use the SPC method to pick out these data.

The multivariate statistics technology has been used for attempting to remedy above-mentioned preceding two problems (for example U.S. Patent No. 5479340).Polytechnics is not only considered each variance of Control Parameter, also considers their covariance.Because the multivariate statistics method can be used for packed data, thereby reduce the quantity of control form and brought the solution that more is added with autgmentability, this has just handled the some shortcomings part of SPC technology.For example, can replace a plurality of sensor data streams with single statistic, as Hotelling T ², it has caught the variance of each transducer and the covariance between transducer.Utilize these technology to significantly reduce the quantity of control diagram, and more can represent the health of whole system with single statistic.

Yet because multivariate method is based on statistically, the 3rd problem is not processed.This is illustrated in Fig. 4, and it has demonstrated a Hotelling T based on the sensing data that comprises data flow shown in Fig. 3 (and being used for the more data flow of the output parameter of multisensor) ²Statistic.As mentioned above, have only an error event, appear at wafer 1018 places in this data centralization.Other all data comprise that the discontinuity of drift and PM all is normal.Yet this list multivariate statistics amount because they have departed from the normal behaviour of statistics, but has been missed real error condition to report out several statistics skews greater than 99% confidence level.The multivariate statistics method has an extra shortcoming.Offset amplitude is difficult to explain, and is same because it is also based on statistics.The a certain significant process quality issue of the perhaps not strict correspondence of a huge deviation statistically, but a little deviation may be indicated a main process excursion once in a while.

When multiplex's tool semiconductor manufacturing place uses statistical method, can cause further problem.In practice, the unsatisfactory coupling of plasma process chamber.The transducer response of a reative cell also is different from, and can be different from very much, the transducer response of another reative cell of same type (promptly being built in same nominal rating), even when the identical prescription of operation.Therefore statistics error detection occurs model can not be transferred to another from a reative cell, and small differences may trigger a mistake alarm in the transducer response.Statistical model need be analyzed between the differential responses chamber.This is the further restriction in this method.

As mentioned above, the machining control in the semiconductor industry is the same with the statistical monitoring of device fabrication, uses conventional crudy sampling.Certainly, because output directly depends on crudy, finally this is the most healthy and the strongest method.Yet the crudy of each wafer of measurement is particularly got metric from wafer in each procedure of processing, is forbidding aspect the cost that reduces plant capacity and measuring equipment.U.S. Patent No. 5,926,690 described a kind of based on measuring CD (critical dimension) and controlling the method for machining control of the corrosion tool of processing by change etching time according to measured value.The output of single crudy and CD are by the input that optionally changes single processing, and the etching time of photoresist is controlled.If the coating survey tool is integrated with corrosion tool, CD regulates before each wafer is moved and measured afterwards at any time so.Whether this machining control method depends on the accurate measurement of CD and judges important in a variation on all wafers or on the suitable statistical sample.Yet in the case to the exactly determined dependence of CD, perhaps in the ordinary course of things to the dependence of the tolerance of crudy, it is very expensive that these technical operation are got up.An optional method of accurate measurement that does not need to have the tolerance of crudy will have advantage very much.

In U.S. Patent No. 6,174, another notion of machining control has been described in 450.In this case, single machined parameters also is a direct current biasing, controls by the variation of RF power.This notion is by fixing the input of a particular process, can control the output of particular process better.A problem of this method is, the processing output dependence is in a plurality of inputs, unless and all inputs Be Controlled all, otherwise just can not infer that processing exports.

Independence but related problem is the instrument coupling.Generally speaking, manufactory is assembled into production line, and every production line is used for a specific procedure of processing.For example, processing comprises a lithographic printing line, corrosion line, precipitation line or the like.Wafer is processed according to the process of the equipment of structure by every production line.Each independently production line by one the cover similar equipment collection constitute, each all has a plasma-reaction-chamber at least.A typical processing (fab) can comprise tens similar chamber types, is used for one group of procedure of processing.Each of these procedure of processings all is assigned with independently prescription, when the processing particular device, and the given prescription of operation on all wafers that a plurality of reative cells will be used in manufactory to be processed.Ideally, aspect equipment quality, the prescription that on any given process chamber, moves, the same on other similar reative cell with all, will produce identical output.For example, move a specific etch recipe, ideally, all reative cells have identical cross wafer uniformity or the like with same speed corrosion wafer.Yet as discussed, the difference of similar reacting chamber space may and be certain to occur on the surface, causes a unmatched output collection.This does not match finally can influence the productivity ratio of factory and output.

The mismatch problem of reative cell and reacting chamber space can solve by a plurality of methods at present.At first, attempt having the course of processing of wide action pane, so that the little difference of reative cell and reacting chamber space can be ignored to the influence of processing output with design at every turn.The second, by the device class according to ultimate criterion, a large amount of differences of reative cell output become tolerable.For example, under the situation that microprocessor is made, quicken vanning speed.The 3rd, attempt so that all reative cells are all identical at every turn.This relates to the exchange of repetition test parts, and rectifying inspection widely, and this generally is a very method of effort.

During less than the equipment of 100nm, it is more and more tightr that process window becomes, and worsened the influence of the output difference between the reative cell at the grid length of semiconductor machining (fab) beginning processed transistor and line width.The equipment standard classification is expensive, and the average equipment below has lower market value simultaneously.Finally, it is an interests equation that reduces that exchange by test parts and rectifying inspection make all identical effort of all reative cells, because under many circumstances, a large amount of time and efforts may be consumed on this problem.

Measure the output of reative cell yes the method for determining output difference.Certainly, the conventional crudy inspection of general work in-process use is done these.These quality examinations normally on-the-spot (Ex-situ), and one group of wafer of processing with know whether export difference can influence and certainly exist some time delays between the output.Field monitor is a kind of more and more expensive method, determines that before the scene of output quality is determined the difference of reacting chamber space will have superiority more.

As previously mentioned, the response of the transducer on reative cell may have very big-difference with the transducer response on another reative cell of the operation same recipe of same type.These difference reactions are following some or all:

(a) " really " difference of reacting chamber space will obtain proof in the output of these reative cells;

(b) based on the optimum reacting chamber space difference of the condition of reative cell, structure tolerance, chamber life cycle or the like;

(c) the tiny difference in the output of the set of sensors on each instrument that causes by difference calibration surplus.

Use original sensor data to judge that the problem of top (a) is that these data are obscured by (b) with (c).

The difference of isolating in real time reacting chamber space provides about the definite information from the output of the crudy of given machining production line for process operation person.After isolating the very poor reative cell of coupling, next step be make this reative cell turn back to production line on the state of matched.As mentioned above, this method often is exactly a repetition test, comprises parts swap and calibration, up to reative cell output coupling.The real-time classification of the basic reason of chamber difference has superiority more.

U.S. Patent No. 6,586 has been recognized the mismatch problem of reative cell in 265, and has been disclosed a kind of based on the method that is used to optimize work flow of selecting optimum machining path by one group of processing line.This method has solved not matching of reative cell easily, and uses the very poor reative cell of coupling as few as possible.

On the proceeding in March, 2003 of European advanced machining control discussion, disclosed a kind of method of in the instrument Computer-Assisted Design, Manufacture And Test, isolating chamber difference.This method is collected the data of all the sensors relevant with each Processing Room on the given instrument, a principal component model (PCA) of structure sensor data set.PCA obtains the variance of all processing effectively from the relevant multivariate data collection (transducer) of one group of incoherent principal component, each all is the linear combination of original set.First principal component is represented the deviation of initial data as much as possible, and second principal component is represented remaining deviation as far as possible, and not relevant with first principal component etc.Generally can find, particularly when sensor data set be correlated with just as on machining tool the time, the major part of deviation is obtained by first principal component of minority.Therefore, in the PCA space, draw sensing data and make the user can observe sensor bias at an easy rate, and obtain the difference between the reative cell.Yet, the mixture of (output influence) chamber difference, benign chamber differences and sensor device difference when observed deviation keeps really in PCA.And the basic reason for difference does not provide classification.

Therefore an object of the present invention is to provide the improved error detecting method in the manufacturing equipment, more specifically, but be not limited to use the semiconductor manufacturing facility of plasma-reaction-chamber, it can be used to avoid or reduce the problem of above-mentioned machining control and reative cell coupling.

Summary of the invention

Therefore, the invention provides the method for the error checking in a kind of manufacturing equipment, wherein said equipment has transducer, and transducer has the output of the current state of indicating equipment, and this method comprises the following steps:

(a) set up fault fingerprint, this fault fingerprint comprises the sensing data of equipment state under the expression error condition;

(b) with these storage in a fault fingerprint storehouse;

(c) use transducer to determine the current state of equipment;

(d), check mistake based on the comparison between the fault fingerprint in current sensor data and the fault fingerprint storehouse.

Description of drawings

Below with reference to the accompanying drawings, by the case description embodiments of the invention, wherein:

Fig. 1 has described a typical semiconductor manufacturing tool, has the transducer output of input setting and indicating equipment state.

Fig. 2 shows one based on one statistics machining control chart in the transducer output;

Fig. 3 shows not processed sensing data in one period, comprises the mistake of preventive maintenance incident and a reality;

Fig. 4 shows a polynary Hotelling T based on a selection of tool sensor output ²The machining control chart;

Fig. 5 shows the transducer output response as the function of some typical process input;

Fig. 6 shows an example with the correlation of processing the transducer output of importing;

Fig. 7 shows a typical error fingerprint that is made of 15 sensor parameters;

Fig. 8 shows an example with processing output to the correlation of processing input;

Fig. 9 is the flow chart of first embodiment of the present invention;

Figure 10 shows the sensing data from 3 different process chamber;

Figure 11 is the flow chart of the second embodiment of the present invention;

Figure 12 shows the result of the method for second embodiment that is applied to 3 test reaction chambers;

Figure 13 is one and shows the form of inducing variation that generates result shown in Figure 12.

Embodiment

In the first embodiment of the present invention, the processing control method that is used for semiconductor manufacturing facility comprises, and at first uses the instrument of this method to determine a tool profile thereon for each.In this embodiment, the profile of instrument is to be made up by a plurality of tool sensor data.These sensing datas can be from the multidimensional data of single-sensor or from the data of one group of transducer, but under any circumstance these data must be change sensitivities to tool state and machining state.Important criterion is that sensing data has enough dimensions to allow for a plurality of different error condition definition a plurality of different fingerprint separately.As used herein, " fingerprint " is the one group of sensing data that has defined the particular state of equipment, and like this, a fault fingerprint means the one group of sensing data that has defined the equipment state under the error condition.

Fig. 5 shows the part of the typical tool profile of the plasma-reaction-chamber that is used to move particular etch recipe.For the variation that each instrument of two instrument inputs is imported, show the response A1-A15 of 15 transducers outputs.In the case, transducer output is magnitude of voltage, current value and the phase value that encourages 5 RF harmonic waves that produce by the RF of plasma, and the instrument input is RF power and tonnage.Can see that the output valve of each transducer changes according to the variation of processing input.For example, output valve A8 reduces with the increase of the RF power of carrying, but increases with the increase of pressure.Like this, particularly when considering all transducer output, the variation of tonnage will be different from and be different from the variation of the RF power of conveying.If a plurality of equipment inputs are changed in the experiment that designs, so just can set up a complete tool profile that constitutes by the transducer response of one group of processing input.In fact, the variation that shows among Fig. 5 is the variation with respect to the mean value that has moved transducer several times, even because for single reative cell, to identical prescription sensor values slight variation can be arranged, although far away not as good as the deviation between the reative cell.

The present invention relies on the reliability of transducer output to come the prediction processing input value, is independent of the nominal setting of input value.Fig. 6 shows under the RF power of this situation, and one section typical process input value is to the curve chart of the input value estimated from the output of tool sensor (as the response curve of the RF transducer of the plasma etcher of Fig. 5).As can be seen, in the case, exist typical good correlation between actual input and the predicted value based on this input of transducer dateout.Like this, tool sensor at least one equipment output that can be used for calculating to a nicety.Therefore, for example under certain error condition, even the operator is a nominal value with the RF power setting, compare with the equipment of the conveying of controlling RF power, sensing data still can provide the more reliable measurement of the RF power of conveying.

As in U.S. Patent No. 6,441, described in 620, the profile of instrument can be used for adding signature in specific input.After this, if transducer output changes, and these change to change with expectation from one group of response curve of having learnt and are complementary, and so just can classify to the basic reason of mistake immediately.Yet, U.S. Patent No. 6,441,620 is just useful aspect the diagnosis mistake after mistake is detected by (for example by test products after processing); It can not be when mistake takes place or detects mistake when taking place probably.

As will be explained below, in the method, before running into mistake, fault fingerprint to be classified, this process has guaranteed that this method is very healthy and strong aspect these mistakes of detection.

In case set up the profile of instrument, just can pass through or the simulation mistake, for example by the output variation of input of change equipment and measuring transducer by force, by understanding fault fingerprint when new mistake occurs; Perhaps, generate the storehouse of known errors fingerprint by importing the fault fingerprint data from other instrument.Last option is the method that has superiority very much, because it has avoided knowing the time that model spends for each instrument in manufactory.In current context, instrument mistake is the deviation in the tool state, and in fact it can produce unacceptable injurious effects to the quality by the product of this instrument manufacturing, if perhaps do not note, produces probably and more seriously influences.

In this embodiment, fault fingerprint is stored as sensor output value and difference for the tool profile of the sensor output value of the processing input value of nominal.Compare with the data that come from the identical transducer that does not have mistake, Fig. 7 is the visual representation of the typical change of expression fault fingerprint in the sensing data.For example, a mistake, as the RF power of its calibration value that drifted about can be expressed as one and comprise negative value A8 and on the occasion of the fault fingerprint of A9 etc.The vector representation of these differences is because each all has size (arrow length among Fig. 7) and direction (plus or minus).

For each instrument, tool profile all is essential, and as mentioned above, the absolute value of transducer output all is different usually to each other, although these instruments are of the same type, and moves identical prescription.Yet preferred embodiment is based on the following fact, and when moving identical nominal recipe, the rate of change (slope among Fig. 5) of transducer output is all fully identical to another from an instrument of same type at least.This just means fault fingerprint at same type and to move the Tool Room of same nominal prescription fully identical, and makes fault fingerprint effectively to transplant at Tool Room.Also mean simultaneously, the variation of transducer output, i.e. slope among Fig. 5 only needs to be identified on a good tool of the given prescription of operation of given type.After confirming these slopes, the tool profile of other instrument of the prescription that the operation of same type is identical may include only the output valve to the transducer of nominal process input value.

Finally, error condition is determined by following comparison between the two, one is according to the sensor output value that measures and by the current tool state of the deviation between the indicated nominal value of the tool profile among Fig. 5, another is the fingerprint of any known errors state among Fig. 7.Expression is stored in each vector set of the mistake in the fault library, relevant with the vector set of corresponding current tool state, if current tool state and expression are stored in when having important coupling between the defined tool state of vector set of a mistake in the storehouse, then detect a mistake.If a fingerprint matching in deviation and the fault library, it will be labeled out so.

As can be seen from Figure 3, the deviation in each transducer output may be relatively very big between each time operation; Yet unless the tight coupling of the deviation on each standalone sensor and known deviation pattern (also i.e. a fault fingerprint) in the method, it is left in the basket.If sensing data has a lot of dimensions, the possibility of coupling can be ignored so by mistake.Like this, machining control in this method, by relatively when precondition and error condition, rather than normal condition, and carrying out, this makes this technology very healthy and strong.

It can also be seen that from above in order to detect a mistake, tool profile only need comprise the sensor output value at the nominal process input value.Yet, as the back will be explained, need to determine the influence of mistake, so that understand the rate of change of transducer output with respect to the processing input to processing output.

Present embodiment can be used for cognitive any emerging fingerprint, and they are joined in the fault library.When a new mistake occurred, a plurality of tool sensor changed report condition.When taking place for the first time, the fingerprint that in fault library, does not match, this mistake can not be classified.When new mistake was for example confirmed by metrology by oneself, the fingerprint of new mistake was added into.Afterwards, if this mistake occurs, it is classified immediately.Therefore this method allows constantly cognitive and expands fault library.

As mentioned above, in order to quicken cognition, these variations of expression typical fault conditions can be concluded.For example the integrality of hardware and process can be traded off wittingly, so that record and comprise these features.It is wrong or the like that example can be that the omission of the escape of air of concluding, hardware component or improper, wafer are placed.

In this embodiment, mark the error condition of reative cell after, next procedure is to judge whether this mistake can be influential to processing output.

As can be seen, the response curve of Fig. 5 has been set up the size of transducer output variation and has been processed the relation of importing between the size that changes.

Fig. 8 shows the sets of curves of processing output to the correlation of processing input.In typical case, be known for these correlations of given fabrication tool.Horizontal dotted line is to corresponding to one " window ", and each tolerance must be positioned at this window, so that make product meet its target criteria.In this case, in corrosion processing, target criteria requires post-etching CD (critical dimension) between 101nm-103nm.Therefore if said method demonstrates a mistake and takes place, and for example, this mistake be from set-point 130sccm deviation the HBr stream of 15sccm, the influence to CD is to produce the CD of being wider than the tolerable limit of target criteria so.Therefore, this mistake is labeled and stops to process.Now, operating personnel have known the position of mistake, and s/he just can carry out error repair at once.

Like this, not only can dope certain fault and take place, and because the size of mistake, promptly the variation in the processing output that is caused by the deviation in the processing input can be determined out, and this can be used to estimate the influence to the quality of processing output then.

With reference now to Fig. 9,, it illustrates in greater detail first embodiment, the tool profile data are stored as a plurality of response curves as shown in Fig. 5 that tool state (as RF power, air-flow) is associated with transducer output (as voltage, electric current, phase place), step 20.By the forcibly changing to the processing input, by when mistake takes place extra fault fingerprint data being joined in the storehouse, perhaps as mentioned above by importing the fault fingerprint data from other instrument, the error condition data are hunted down and are added into fault library, step 22.This last option makes fault library to be increased fast.For example each fingerprint among Fig. 7 can be labeled as the process variations of specific size.

In process of production, the product wafer is by a plurality of transducer output monitorings, and constantly compares step 24 with the fingerprint of fault library.The size and Orientation that transducer output departs from the expectation nominal value of this instrument is compared with corresponding each fingerprint value.Although a lot of possible methods are arranged in the present embodiment, should be relatively more relevant based on mathematics.Yet also can use Euclidean distance.Like this, when a correlation surpasses given threshold values or Euclidean distance less than given threshold values, error condition of mark then, step 26.Compare by the size of the mistake that will determine by the tool profile among Fig. 5 and as shown in Figure 8 processing ancillary data, determine the influence of mistake, step 28.If mistake is judged out operation output there are negative or unacceptable influence, step 30, then instrument will be stopped, and revise the problem that identifies, step 32.

Obviously, those of skill in the art can be embodied as the computer program with related sensor numerical value with above-mentioned processing, and sensor values wherein is the conduct input after analog to digital conversion.

Should be realized that design this method is not those variations that take place in response to the variation of associated external input setting in order to the variation of the processing input parameter of detection.More precisely, normally constant although it is that these inputs are provided with maintenance, the variation that takes place by some mistakes in the plasma process.For example, intensive airflow rate transducer may produce a mistake, makes the actual airflow rate that enters reative cell be different from the value of transducer indication; Perhaps matching unit may absorbed power, makes the RF power the carried indicated value less than the power instrument relevant with the RF source.

Said method also can be used for determining can not cause the product mistake, but really the processing that may become the product mistake be exported the variation of influential processing conditions.For example, with reference to figure 8, if the variation of known process input, then precognition processing output variation is in the cards in the output area of hope.For example, method described herein can be used for determining a mistake, for example variation of working power.Can judge, do not change CD is released the standard of wishing, but it may cause wideer CD.Although final products do not suffer disastrous effect, it can indicate a trend, makes operating personnel prediction cause this mistake negative or unacceptable influence to processing output before mistake takes place.

Because mistake is big or small known, this method also can be used for the closed circuit machining control.For example, under the situation in Fig. 3,, can detect the mistake of pressure set-point with this method at wafer 1018 places.Operating personnel or can stop processing and modification problem, perhaps the variation based on prediction select to change pressure, and ignores defective pressure gauge.In addition, in this example, can be used for predicting that processing output changes because pressure changes, operating personnel can change pressure based on the prediction that processing is exported so.

In the above-described embodiments, check that mistake is just based on the fingerprint of discerning the error state of having stored.Also promptly, by being compared, the state of current manufacturing equipment and the state in the unexpected storehouse carry out error detection occurs.Only when current instrument condition was identified as error condition, it just was labeled.Therefore, saying does not in essence have false positive, and error checking and fault identification are synonyms.In addition, when determining current state and an error state and being complementary, the size of mistake is determined, and compares with the mistake limit of the instrument that is used for particular process.Therefore, if error condition does not have adverse effect to processing output, just can ignore it.

Present embodiment has following advantage compared to existing technology:

(a) mistake detects by mode identification method, thus statistically can not trigger false alarm unusually,

(b) the normal intervention carried out of the user in the processing environment as preventive maintenance, can not endanger the robustness of fingerprint recognition,

(c) size of mistake is easy to be explained, and can be with important level to user report,

(d) needn't depend on the accurate in-site measurement that continues to crudy, for example, by the feature such as the CD of measurement products.This method is determined based on the mistake that influence crudy to any meeting, the prediction crudy.

(e) fingerprint base is transplantable, so the expansion between the tool set is possible.

In the second embodiment of the present invention, above-mentioned technology is applied to the coupling of plasma-reaction-chamber.Should be realized that, in first embodiment, when the correlated response chamber is in a kind of known kilter, generate the profile (even for the reative cell of the operation same recipe of same kind, recall between the reative cell of front there are differences) of each instrument.Second embodiment compares any given test reaction chamber and reative cell one Known good (reference) equal type, the operation same recipe, has determined that the reative cell of test also is good, also promptly mates with the reference reaction chamber.

As described in introduce, come the original sensor data of reative cell self-test and reference can not be used for the comparison reative cell, because the variation from the output of the transducer between the reative cell may be very big, to such an extent as to " real " (promptly significant) chamber difference is fallen by optimum chamber difference and the differential shading between transducer.

For example, Figure 10 shows the sensing data from 3 plasma-reaction-chambers.These three reative cells mate in appearance, and each all has identical specification and operation same recipe.All input controls all transfer to standard state.Yet the output of finding in the case, reative cell 2 is different.Particularly, the corrosion rate of chamber 2 is lower than chamber 1 and chamber 3, and is lower than acceptable crudy standard.Sensing data A1-A10 is the multidimensional data from single-sensor, can represent the sensor data set of any reative cell.Figure 10 A shows the sensing data from each reative cell, uses different transducers on each reative cell.In the case, there is not evident difference between good reative cell and the bad reative cell, because " real " difference has been mixed up by sensor differences and benign chamber differences.Figure 10 B shows the data from these 3 identical reative cells, uses identical transducer now on these 3 reative cells.Equally, bad reative cell neither be outstanding especially, and " real " difference has been mixed up by benign differences.

Obviously said from the difficulty of Fig. 3 is same, it show sensing data in time tendency and when the PM incident, experienced great variety.The reative cell at difference place in this cycle, the sensing data that output is very different.These data are classified into the difference between benign chamber, because they do not influence the output of reative cell, are the parts of normal reaction chamber operation.Therefore, original sensor data can not return the information about real chamber difference at an easy rate, although it has comprised these information.

Therefore, in a second embodiment, Figure 11, the tool profile of reference reaction chamber of operation special formulation, by at first by with previous description first embodiment time identical mode determine step 40.Next step, step 42, operation same standard recipe, step 44 then on the test reaction chamber, departing between the size and Orientation of transducer output and the current state nominal value of test reaction chamber, quilt compares with corresponding fault fingerprint value in the fault fingerprint storehouse.Yet, be to be used for determining that the nominal sensor value of current state vector is the tool profile of reference reaction chamber with the difference of first specific embodiment key, rather than the tool profile of test reaction chamber.Also promptly, each current state vector is the difference between the nominal value of the respective sensor in the test feature sketch plan of one currency in the transducer of test reaction chamber and reference reaction chamber.We find, this has greatly eliminated the influence to the comparison of reacting chamber space of benign chamber and transducer differences.This relatively can be by mathematics relevant or Euclidean distance carry out, as previously mentioned.

If do not find fingerprint matching in step 46, then the test reaction chamber just is considered to mate with the reference reaction chamber.Yet,, think that the test reaction chamber is defective if find a fingerprint matching.The reason of mistake is identified in step 48, and this test reaction chamber is repaired to correct this mistake simultaneously.Because different fault fingerprint is relevant with the variance of the deviation of differential responses chamber and/or machined parameters, so the reason of mistake can be according to the particular fingerprint identification of coupling.

Figure 12 shows the sensing data of handling from 3 test reaction chambers.Point 1-28 is the wafer of operation in reative cell 1, and some 29-56 is the wafer of operation in reative cell 2, and some 57-84 is the wafer of operation in reative cell 3.In the set of per 28 points, change being forced through this 3 reative cells, shown in form among Figure 13.In Figure 12 A, the correlation between the fault fingerprint of a current tool state and a known variable power is tested.In the case, the y axle is represented the variable power size of an expectation.Clearly, variable power is identified rightly on the suitable wafer in each reative cell.In Figure 12 B, the correlation between the fault fingerprint of a cup state and a processing temperature is tested.The variation that is provided with by the temperature of concluding wafer substrates, and learn fingerprint, the fault fingerprint of processing temperature is before to have learned and joined in the fault fingerprint storehouse.Notice how each reative cell shows the size of the difference of the processing temperature of mating with all wafers.This indicates reative cell not fine coupling aspect temperature.Like this, the underlying cause difference is identified, and is repaired subsequently.In Figure 12 B, it can also be seen that the variable power when a variations in temperature in the test also is labeled out.This is desired, because the variation of the plasma power on this tool types will change the temperature of wafer.

As under the situation of first embodiment, can recognize that top processing can be embodied as a computer program by those skilled in the art, this program has the related sensor value, after analog-to-digital conversion, as input.

The present invention is not limited to embodiment described in the literary composition, and it can be modified without departing from the scope of the invention and change.

Claims (10)

1. the method for the error checking in the manufacturing equipment, wherein said equipment has transducer, and transducer has the output of the current state of indicating equipment, and this method comprises the following steps:

(a) set up fault fingerprint, this fault fingerprint comprises the difference of sensing data sensor values specified with it, and sensing data is represented equipment state under the error condition, wherein, fault fingerprint comes down to constant having the same nominal standard and moving between the different manufacturing equipments of same nominal process;

(b) fault fingerprint is stored in the fault fingerprint storehouse;

(c) use transducer to determine the current state fingerprint of equipment;

(d), detect mistake based on the comparison between the fault fingerprint in current state fingerprint and the fault fingerprint storehouse.

2. the method for claim 1, wherein in step (d), the vector set of the expression sensing data of fault fingerprint and the deviation between nominal value, and the respective vectors collection of the expression sensing data of current state fingerprint and the deviation between the nominal value between compare.

3. method as claimed in claim 2, the nominal value that wherein is used to calculate the vector set of current state fingerprint are the nominal values from the sensing data of the transducer of described manufacturing equipment.

4. method as claimed in claim 2, the nominal value that wherein is used to calculate the vector set of current state fingerprint is the nominal value from the sensing data of the transducer of different manufacturing equipments, and described different manufacturing equipments have identical nominal standard with the equipment of mentioning first and move identical nominal and handle.

5. method as claimed in claim 2 is wherein by measuring relevant the comparing between fault fingerprint collection and the current state fingerprint.

6. method as claimed in claim 2 wherein compares by the Euclidean distance that calculates between fault fingerprint collection and current state fingerprint.

7. the method for claim 1, wherein after described detection step, also comprise the influence of prediction error to particular process output.

8. the method for claim 1, wherein after described detection step, comprise that also control appliance input offsets mistake.

9. the method for claim 1, wherein fault fingerprint comes from and comprises the tool profile of one group of equipment input to the curve of transducer response.

10. the method for claim 1, wherein manufacturing equipment comprises plasma-reaction-chamber.

Images