DE10208044B4 - Method and device for monitoring a manufacturing process - Google Patents

Abstract

Method for computer-aided monitoring of a production process,
In which a large number of measured variables is detected during the production process,
In which a first time range is selected, which is before the occurrence of a predetermined event,
In which a second time range is selected, which is after the occurrence of the predetermined event,
In which at least one linear combination is formed using the plurality of measured variables, the weights of the linear combination being selected such that the curve shape of the linear combination has a predetermined first waveform in the first time domain and a predetermined second waveform in the second time domain,
In which at least one characteristic characterizing the production process is determined from the linear combination, and
- In which is monitored using the characteristic of the manufacturing process.

Description

The The invention relates to a method and an arrangement for monitoring a manufacturing process.

in the The scope of industrial production is the degree of automation always increased. Especially in the production of semiconductors, generally in the Semiconductor industry, is a high degree of automation of considerable Importance to ensure international competitiveness.

at the manufacture and processing of semiconductor substrates for formation of semiconductor integrated circuits is a variety of manufacturing processes and machining processes, often from a few 100 manufacturing processes and machining processes used.

Examples for such Manufacturing processes or processing processes are processes for Applying material to the semiconductor substrate, for example Method for epitaxially growing material on a substrate material, Separation processes such as deposition processes from the gas phase such as Chemical Vapor Deposition (CVD) or Physical Vapor Deposition (PVD), or etching processes, general processes for structuring material layers. These Processes are very complex and can occur unnoticed disorders or poorly adapted process conditions easily too lead incorrectly manufactured semiconductor circuits. For this reason, the Manufacturing processes monitored very closely become.

Of the Manufacturing process becomes common through the analysis of during of the manufacturing process, which determines the manufacturing process characterize, performed. With sufficiently accurate evaluation of the determined measured quantities or the correspondingly accurate analysis of the measured quantities determined it is thus possible possibly based on the analyzed data in the manufacturing process or to intervene in the editing process.

Especially as part of an etching process a crucial parameter is the etching time, that is the time, while the structuring of the respective material layer takes place.

Just with very accurate and correct adherence to the etching time, it is possible to produce high quality semiconductor product.

is the etching time namely too low, then the one to be etched Layer only incomplete away.

Becomes the etching but done too long, this means in other words, is the etching time too long, can For example, active areas of one transistor or another be etched active component or it can become an undercutting come from semiconductor structures.

It is therefore crucial for the quality of a manufactured semiconductor product to determine the etching time exactly. In other words it is of considerable importance, the right end time, the is called the time at which the etching should be ended, if possible to be able to determine exactly.

Out US 5,877,032 A For example, there is known a method of determining the endpoint of a plasma etch process using the optical emission of the plasma to determine the endpoint. In this method, the fact is exploited that in the etching process, a layer located on a substrate etched through, thereby exposing the underlying substrate or the underlying layer. This interaction of the etching gas with the exposed substrate or the exposed layer can be detected spectroscopically as a change in the emission spectrum of the plasma. The change is compared according to [1] with a multiplicity of predetermined reference curves and from the comparison it is concluded that the end point of the plasma etching process.

In US 5,739,051 A Another method for determining the end point of a plasma etching process is described. In this known method, the optical emission of the plasma is also used to determine the end point, wherein the emission lines, which are characteristic of the interaction of the etching gas with the exposed substrate, are used for the evaluation.

Usually optical filters are used in industrial plants depending on the chemical reaction taking place in the plasma in order to determine the temporal Ver in fixed wavelength ranges to be able to track the intensities of the signals of the respective wavelengths. However, this procedure is impractical in process chambers where alternating products with different chemical compositions are produced.

The Detection of the endpoint in different optical fields requires a high mechanical and temporal effort to change the Filter. Furthermore, the selection of the wavelength ranges is the responsibility of the process engineer. The selection of unfavorable wavelength ranges for endpoint detection is not excluded for this reason.

Furthermore, it is off US 5,658,423 A a method for monitoring a manufacturing process is known, which is based on the so-called principal component analysis. In this method, the time evolution of the entire emission spectrum from 240 nm to 600 nm is used to determine the end point. By means of the main component analysis, the resulting data volume is reduced to a few so-called basic patterns and their temporal development is used for the detection of the end point. The detection of the end point of the plasma etching process is therefore no longer based on the evaluation of a single emission wavelength but on the change in the total available emission spectrum. In principle, however, it is also necessary in the method according to [3] to provide reference values in order to be able to conclude from the comparison of the currently measured variables with the reference values to the correct end point. Moreover, in the method according to [3], the end point signal is determined from the difference of two spectra matrices, the relevance of this change for the determination of the end point having to be ensured anew each time.

It is still out US 5,737,496 A It is known to determine the end point of a plasma etch process using an artificial neural network. The neural network is trained on the basis of a large number of determined measured variables, so that it can be used to detect the end point after completion of the training. It has been found that artificial neural networks often learn false signals and patterns, so that there may be a misinterpretation of the signals applied to the artificial neural network in the application phase. An erroneous training of the artificial neural network arises for example by changing the emission spectra due to aging phenomena of the sensors or chamber contamination occurring.

In US 5,864,773 A. a so-called virtual sensor system is described in which these changes in the aging of the sensors or in the event of chamber contamination are taken into account before the actual evaluation of the measured variables. In this way, a virtual sensor is to be created, which is free of chamber-specific or process-specific error effects. However, since [5] requires the experience of the operator, unexpected errors and changes can not be automatically taken into account.

It is still out WO 01/56072 A1 known to form a linear combination of a plurality of measured variables which characterize the manufacturing process, wherein the weights of the linear combination are selected so that a predetermined waveform is approximated. From the linear combination, the end point of the plasma etching process is determined.

Of the Invention is based on the problem, a manufacturing process to monitor using simplified measures in a simplified manner the essential information characterizing the process, this means important parameters formed become.

The Problem is posed by the method and arrangement for monitoring a manufacturing process with the features solved according to the independent claims.

at a method for computer-aided monitoring A manufacturing process will produce a variety of metrics during the Manufacturing process recorded. A first and a second time range are selected. The first time range is selected to be before entry a given event, for example, before the actual event End point of a plasma etch process, lies. The second time range is selected to be safe upon entry the given event. Using the variety of measured variables formed at least one linear combination, wherein the weights of the linear combination chosen like that are that the waveform of the linear combination in the first time range a predetermined first waveform and in the second time range a having predetermined second waveform. From the linear combination At least one characteristic characterizing the manufacturing process is determined and using the characteristic supervised the manufacturing process.

A Arrangement for monitoring a manufacturing process has at least one sensor for detecting of measurands during a Manufacturing process on. Further, in the arrangement is a processor provided, which is arranged such that using the at least one of the measured variables determined by the at least one sensor a linear combination is formed, wherein the weights of the linear combination chosen like that are that the waveform of the linear combination in a selected first Time range that precedes the occurrence of a given event, a predetermined first waveform and in a previously selected second Time range, which after the occurrence of the given event has a predetermined second waveform. From the at least A linear combination will at least one the manufacturing process characterizing characteristic determined. Furthermore, a monitoring device provided, which monitors the production process using the characteristic.

clear the invention can be seen in that by very simple Specification of a pre-endpoint range, ie the first time range, and a post-endpoint area, this means the second time range, which is very simple and fault tolerant using the expertise, that is, using experience after trial runs of the respective process can be determined in the process chamber, the Linear combination can be formed without being an explicit mathematical endpoint model with its necessary complex and complicated parameterization by a user is required.

Out This is the reason for monitoring a manufacturing process, in particular a semiconductor manufacturing process, now much less computationally and can be very simple Manner performed without the accuracy of the endpoint determination of each Manufacturing process, in particular a plasma etching process, significantly reduced becomes.

In order to becomes the application of endpoint pattern generation under process conditions with usual staff Enabled computers.

Especially is also due to the ease of use of the software monitoring tools to point out by a process engineer as the process engineer no longer understand, maintain and monitor a very complex endpoint model must, but semi-automatically two fault-tolerant time ranges under Using the personal computer defines what time ranges as a basis for The supervision be used. The process engineer can thus due to his Expertise and his experience in the respective manufacturing process using known endpoint waveforms which a rough endpoint determination, d. H. the determination of an endpoint range, enable, very simply specify the two time ranges, which are the basis for the others Represent calculations.

One additional Advantage of the invention is the fact that the natural history the endpoint line in the transition area the endpoint and not, as in the case of the procedure from [6] compulsorily to a course of a mathematical predetermined Endpoint model is adjusted.

Consequently is descriptive in particular in comparison to the method according to [6] the explicit model specification for the course of an end point signal replaced by the selection of a Pre-endpoint range and post-endpoint range.

in the The scope of this invention is the term linear combination not just the process of linearly combining measures, but understood the signal obtained in this way.

One additional Advantage of the invention is the fact that in a change of the manufacturing process the required adjustment for extraction the characteristic only by Adaptation of the corresponding evaluation software for the implementation of the inventive method can be made and no manual steps, such as For example, the replacement of optical filters are required.

On This way, in particular, the maintenance of the monitoring system or the arrangement for monitoring the manufacturing process considerably cheaper feasible.

preferred Further developments of the invention will become apparent from the dependent claims.

The Embodiments described below relate both to the method as well as the arrangement for monitoring a manufacturing process.

According to one Embodiment of the invention, only measured variables are used, which detects at times which are in the first time range and / or in the second Time range are.

By This embodiment of the invention is for monitoring the respective manufacturing process required computing time required further reduced.

at a method for computer-aided monitoring A manufacturing process will produce a variety of metrics during the Manufacturing process recorded.

According to one additional Embodiment of the invention, it is provided that the predetermined first waveform of a predetermined constant first function value is formed, for example, the function value "0".

The predetermined second waveform can be of a predetermined constant second function value are formed, preferably of the function value "1".

According to one Another embodiment of the invention are as measured variables the intensities of different wavelengths used in an emission spectrum. Emission spectra include usually very detailed information about the radiation emission underlying process, such as a chemical reaction. When recording such spectra, however, very large amounts of data are lost from which come with conventional Methods the crucial information only with great difficulty can be extracted. It is according to the invention but now possible, to process the data of an emission spectrum in such a way that the decisive Information, that is in other words the characteristics, on very simple way without using an overall model and thus determined in real time or in real time from the data can be. Accordingly, you can Emission spectra with the inventive method and the inventive arrangement for online monitoring of manufacturing processes or machining processes be used.

In In this context, it should be noted that under this Description under a manufacturing process also a machining process to understand, since this for the production of an intermediate product used in an overall manufacturing process.

It is particularly preferred when the intensities of different wavelengths in a spectral range between 200 nm and 950 nm can be used.

According to one Another embodiment of the invention, the emission spectrum of a Plasma process, in particular a plasma etching process used.

plasma etching for structuring material layers play in semiconductor manufacturing a significant role. The dimensions of the structures created hang decisive of the quality the etching process used from. The emission spectrum of the plasma contains a lot of information, to control the plasma etch process necessary. Sophisticated sensor systems such as the so-called "Optical Emission Spectroscopy (OES) "or the so-called "Residual Gas Analysis (RGA) "can be used for Recording the spectra can be used.

It It is particularly preferred that the parameter used is the end point of the plasma process. that is the one Time at which the plasma process or the plasma etching process is to be terminated is determined.

According to another preferred embodiment of the invention, to determine the weights of the linear combination, the measured variables determined at the predetermined times are arranged in a matrix and a principal component analysis (PCA) is carried out on the basis of the matrix. The principal component analysis is a method of expressing a matrix A by its main vectors, also called eigenvectors, or principal components, also called eigenvalues. A number of different methods can be used to calculate the main components. For example, the Singular Value Decomposition (SVD) of the matrix A can be calculated. As an alternative, the eigenvalues as well as eigenvectors of the correlation matrix ( A ^T A ) of the matrix A can also be calculated. In addition, other methods for decomposing a matrix, such as the so-called "non-linear partial least squares" of a matrix A can be used.

Preferably, the principal component analysis is realized using a singular value decomposition, such that the matrix A ^{mxn is decomposed} into two orthogonal matrices U ^mxm and V ^nxn and one diagonally ^mapped matrix Σ with the singular values σ _i , where:

The so-called "left" eigenvectors u _i form the matrix U ^mxm . The so-called "right" eigenvectors v _i form the matrix V ^nxn . Usually, the singular values (eigenvalues) σ _{i are} ordered in magnitude and represent the proportion that the associated dyadic product u _i · v _i ^{T has} at the formation of the total matrix A. The dyadic products u _i · v _i ^T , which have only small singular values σ _i , usually contain no information or only very little information about the process to be monitored. Accordingly, real measured matrices of measured variables are generally sufficiently well represented by the dyadic products u _i · v _i ^T with the 3 to 15 largest singular values on σ _i .

According to one another embodiment The invention is used to determine the weights of the linear combination a parameter-linear estimation problem using the given first waveform and the given one second waveform and the linear combination minimized or maximized.

It is particularly preferred if by minimizing or Maximization of the parameter-linear estimation problem, generally by its solution, the essential main components of the measured variable matrix are determined. By minimizing or maximizing the parameter-linear estimation problem can be very easily those main components determine which are considered Need to become, around the given waveforms, that is, the predetermined first waveform and the predetermined second waveform, with sufficient accuracy to be able to approximate.

It It should be pointed out in this connection that the parameter linear to be solved according to the invention estimation problem in a much simpler way than the complex non-linear one Optimization problem for determining the weights of the linear combination to minimize the quality measure according to [6].

Further It should be noted that the parameter linear estimation problem has a unique and explicitly specifiable solution, whereas the at solution of the optimization problem according to [6] is possible.

By This embodiment of the invention thus becomes the mathematical calculation, this means the release the estimation problem considerably simplified and accelerated.

Further can for determining the characteristic a smoothing performed the linear combination become. Usually have the measured variables one Noise on, which is also in a linear combination of Measurements reflected. In order to avoid that due to the noise a misjudgment of the determined Measured variables made It is preferred, the noise component by smoothing the Suppress linear combination so far that the characteristic is correct can be determined. In this context, it is particular preferred when smoothing the linear combination using a zero-phase filter is achieved. By suppression of the noise component can qualitative criteria, such as the occurrence of a local maximum or the occurrence of a turning point used to determine the characteristic become. Qualitative criteria have opposite quantitative criteria, like overwriting certain Thresholds, the advantage that they are far less of the respective absolute size of the measured values depend and thus störunanfälliger about one larger value range can be used.

to increase the robustness of the surveillance It may be useful to compare the results of several trial runs respectively production runs in the determination of the weights, in other words the coefficients for the linear combination, flow in allow. Therefore, according to a another preferred embodiment the invention for determining the weights of the linear combination the measured variables of several Trial and / or production runs used. It is particularly preferred if the measured variables more Trial and / or production runs be arranged in a matrix (block matrix) and based on this matrix a principal component analysis is performed.

One embodiment The invention is illustrated in the figures and will be discussed below explained in more detail.

It demonstrate

1 a schematic representation of an arrangement according to an embodiment of the invention;

2a to 2c a representation of the first three left eigenvectors of a spectra matrix;

3a to 3c a representation of the corresponding first three right eigenvectors of a spectra matrix;

4 a representation of CN lines for nine contact hole etches in a wafer and the representation of selected time ranges before and after reaching the end point of the plasma etching process; and

5a and 5b 1 is a sketch of an end point pattern generated according to the invention in the spectral range ( 5a ) as well as a sketch of resulting endpoint signals in the time domain ( 5b ).

1 shows a schematic representation of an arrangement 100 for monitoring a plasma etching process according to an embodiment of the invention.

The arrangement has a process chamber 101 , that is to say a reaction chamber, in which, according to this exemplary embodiment, a plasma etching method, in particular a method of Reactive Ion Etching (RIE), can be carried out. About an inlet 102 the etching gas is in the reaction chamber 101 in which the etching gas is ionized and into a plasma 109 is transferred.

The plasma 109 is from a capacitively applied RF voltage coming from an RF source 103 is provided, generated and maintained. Through an outlet 104 become the reaction products of the etching from the reaction chamber 101 away. As the etching gas according to this embodiment of the invention, an etching gas mixture of CF ₄ , CHF ₃ and argon is used, wherein the plasma 109 is improved in its homogeneity by a magnetic field (not shown) of about 60 Gauss. The for the generation and maintenance of the plasma 109 used power is about 1200 watts. The pressure in the reaction chamber 101 is about 150 mTorr.

According to this embodiment, the plasma is used 109 to etch via holes in a silicon dioxide layer deposited on a silicon wafer 105 is arranged.

For this purpose, the silicon wafer 105 on a bracket 106 in the reaction chamber 101 filed and fastened there. In conventional semiconductor products, such as semiconductor memories or semiconductor processors, the contact holes, which usually serve to connect the diffusion regions (active regions) of transistors, have a very low aperture ratio. Accordingly, the requirements for detecting the correct end point of the etching, that is, the detection of the time at which the contact holes reach the silicon substrate, are correspondingly high. If the etching time is too short, the contact holes are not fully opened and the transistors are not connected. However, if the etching time is too long, the active regions of the transistors are strongly etched. Both can lead to the total failure of the integrated circuit, which has the respective transistor or the respective transistors.

To determine the correct end point, also referred to as the end time, of the plasma etching is at a window of the reaction chamber 101 an optical fiber 110 arranged, leading to a multichannel spectrometer 107 leads and is connected to this.

According to this embodiment, a multi-channel spectrometer 107 Hamamatsu ^{™ company} with the type designation C7460 for the detection of the optical emission of the plasma 109 used. The multichannel spectrometer 107 has 1024 channels that distribute evenly over a wavelength range of 200 nm to 950 nm. With the used multichannel spectrometer 107 For example, a maximum sampling rate of 50 samples per second is achievable. According to this embodiment, the emission spectrum of the plasma averaged over this period of time every two seconds 109 measured. The multichannel spectrometer 107 is with a data processing system 108 coupled, which continuously records the registered spectra net, evaluates and saves. The data processing system 108 has in particular a processor and a memory, which are coupled together via a computer bus.

When a contact hole etch is performed, the intensities of different wavelengths (1024 different wavelengths in the present case) are recorded over a period of about 3 minutes. The resulting measured values, that is to say measured variables, are arranged in the form of a matrix Λ (λ, t). The time course of a specific wavelength corresponds exactly to one row of the matrix Λ (λ, t). Accordingly, the emission spectrum at any given time is given by exactly one column of the matrix Λ (λ, t). However, this arrangement is not mandatory, for example, the time courses of the wavelengths as the columns of the matrix Λ (λ, t) can be arranged. In the present embodiment, the matrix Λ (λ, t) has 1024 rows and 100 columns.

In order to determine the weights for the linear combination from the intensities of the measured wavelengths, a principal component analysis is carried out on the basis of the matrix Λ (λ, t). For this purpose, once a matrix Λ ₀ (λ, t) is created in a sample etching and determines the weights, which are then used for subsequent etching. The principal component analysis is performed using singular value decomposition, such that the matrix Λ ₀ (λ, t) is decomposed into two orthogonal matrices U ₀ and V _0, and a diagonally occupied matrix Σ _{0 is} decomposed with the singular values σ 0 / i.

Where:

The so-called "left" eigenvectors u 0 / i form the matrix U ₀ and the so-called "right" eigenvectors v 0 / i form the matrix V ₀ .

The 2a to 2c represent representations of the first three left eigenvectors u 0 / i of the matrix Λ ₀ (λ, t), that is, the left eigenvectors u 0 / i with the three largest singular values σ 0 / i.

The 3a to 3c representations of the corresponding first three right eigenvectors v 0 / i of the matrix Λ ₀ (λ, t).

The columns u 0 / i of the matrix U ₀ can be interpreted as a basic wavelength pattern, while the columns v 0 / i of the matrix V ₀ can be interpreted as basic time signals.

wobei mit dem Index i die wichtigsten linken Hauptkomponenten u 0 / i der Matrix Λ ₀(λ, t) indiziert werden. Wie viele und welche der wichtigsten Hauptkomponenten u 0 / i tatsächlich berücksichtigt werden, wird durch die im Weiteren beschriebene Minimierung eines parameterlinearen Schätzproblems bestimmt.The following approach is used to determine a suitable linear combination of the measured variables:

where the index i indicates the most important left main components u 0 / i of the matrix Λ ₀ (λ, t). How many and which of the main components u 0 / i are actually taken into account is determined by the minimization of a parameter-linear estimation problem described below.

The as yet unknown coefficients c _i are also defined in the context of the solution of the parameter-linear estimation problem explained below. If the most important main components u 0 / i are fixed and if the coefficients c _{i are} determined, then the linear combination of the measured quantities is determined as well. Due to the orthogonality of the left main components u 0 / i, the function y ₀ (t) can also be written in the following form:

The function y ₀ (t) is a function of time which can be adjusted to a given waveform using the coefficients c _i .

The predetermined waveform is selected according to the invention in only two predetermined Given time ranges, according to these embodiments in the form of a constant in each case in the time domain function value.

4 In this connection, in a plasma etching cycle for a first, a middle and a last wafer of a wafer lot in the wetclean cycle, the progressions of the CN lines are shown 401 along the time 402 , In other words, the intensities 403 the CN signals over time 402 applied. These CN lines 401 be in trial runs for the etching chamber 101 determined.

At the CN lines 401 It is very easy to see that in a range between 60 and 100 seconds during the etching process, ie in an endpoint range 404 the transition from a region in which the layer to be etched should still be etched into a region in which the continuation of the etching process would result in damage to the active regions of a transistor takes place. 4 It can also be seen that an exact determination of the endpoint in the endpoint area 404 based on the course of the CN lines 401 not possible.

It is just a rough definition of an area that is before reaching the endpoint is possible, this but for that in a very simple, vivid way. The same applies to the area which lies after reaching the end point.

According to the invention by means of the multi-channel spectrometer 107 recorded, ie usually measured, CN lines 401 from a user a first time range 405 selected according to this embodiment in a time range of 20 to 60 seconds after the start of the etching process. The first time range 405 is safe from reaching the end point of the plasma etching process.

There will also be a second time range 406 selected according to this embodiment in the range of 100 to 140 seconds after the start of the etching process. The second time range 406 surely lies after reaching the end point of the plasma etching process.

The user will thus be able to illustrate this manually, alternatively also automatically by a computer, after the evaluation of the CN lines 401 the first time range and the second time range are determined, whereby the endpoint range 404 , is set according to this embodiment in a time range between 60 and 100 seconds.

In other words, the result is a predetermined overall waveform that is only in the first time range 405 and in the second time range 406 is defined as a first waveform having a first function value "0" in the first time range 405 and a second waveform having the constant function value "1" in the second time range 406 , In this overall waveform, which only in the two predetermined time ranges 405 . 406 is defined and not in the endpoint area 404 , the characteristic of the end point of the etching process according to the invention can be relatively easily extracted from the linear combination.

wobei mit

• • t₁ ein Startzeitpunkt des ersten Zeitbereichs 405,
• • t₂ ein Endzeitpunkt des ersten Zeitbereichs 405,
• • t₃ ein Startzeitpunkt zweiten Zeitbereichs 406, und
• • t₄ ein Endzeitpunkt des zweiten Zeitbereichs 406,

bezeichnet wird.This results in the following function for the overall curve shape:

being with

• T _{1 is} a start time of the first time range 405 .
• T _{2 is} an end time of the first time range 405 .
• • t ₃ a start time second time range 406 , and
• T _{4 is} an end time of the second time range 406 .

referred to as.

Clearly, a virtual model output, that is to say a desired end point signal, is defined during the defined pre-endpoint period, that is to say in the first time range 405 is equal to the value "0" and during the post-end point period, that is, in the second time range 406 , equal to the value "1".

There will be only intensity data from these two periods for the further process 405 . 406 that is, from the first time range 405 and from the second time range 406 , and using this spectral data, a linear estimation problem is set up, the content of which is the modeling of the virtual model output by the spectral data at the respective times.

When (regularized solution) this parameter-linear estimation problem receives you get the desired endpoint pattern.

Thus, the parameter estimation problem is constructed according to the following function:

The parameter-linear estimation problem is solved in the following way, whereby it is necessary to determine the weights of a superposition of the columns of X in order to approximate y _MO as well as possible, that is to say: X · C ≌ y NOT A WORD , (7)

die folgende Lösung: copt = (X T·X)–1·X T·yMO. (10) Using the quality criterion Q (according to this embodiment corresponds to the quadratic norm of the approximation error), which is formed according to the following rule: Q = ∥⁣ X · C - y NOT A WORD ∥⁣ 2 = ( X · C - y NOT A WORD ) T · ( X · C - y NOT A WORD ) = = c T · X T · X · C - 2 · c T · X T · y NOT A WORD + y T NOT A WORD · y NOT A WORD (8th) results from the following solution condition, which is necessary for minimizing the quality measure with regard to the individual parameters:

the following solution: c opt = ( X T · X ) -1 · X T · y NOT A WORD , (10)

For the specific application according to this embodiment, according to regulation (4): V · Σ · C = y 0 → x = v · Σ (11) and thus the solution due to the orthogonality of V formal: c opt = Σ -1 · V T · y NOT A WORD , (12)

These known solution leads at Presence of many parameters for over-parameterized solutions, which are characterized by strongly oscillating and noisy solutions.

wobei n so gewählt werden, um einerseits eine gute Annäherung des EP-Signale an die vorgegebene Sprungfunktion zu erreichen (hohe EP-Steilheilt). Anderseits sollten nicht zu viele Komponenten eingehen, um keine Störungen in das EP-Muster zu projizieren (rauhen Muster) und die EP-Robustheit damit zu gefährden. Die Wahl von n ermöglicht die Einstellung dieses Kompromisses.As an alternative, it is firstly possible to directly limit the number of components that go into the inversion of ( X ^T · X ) ^-1 and to calculate c _i for i = 1..n and ci = 0 for i > n to set. c REG-1 (n) = Σ -1 n · V T · y Not a word (13) With

where n are chosen so on the one hand a good approximation of the EP signals to the predetermined jump function (high EP steepness). On the other hand, not too many components should be included so as not to project disturbances into the EP pattern (rough pattern) and endanger the EP robustness with it. The choice of n allows the adjustment of this compromise.

regularisiert werden.Alternatively, the general solution may be appropriate

be regularized.

Specifically, then again with X = V · Σ (16) c REG-2 (ρ) = Σ -1 ρ · V T · y NOT A WORD , (17) With

The solution is now dependent on the regularization parameter ρ, whose variation of the Set a compromise between endpoint signal sharpness (approach to the desired 0-1-model function) and EP robustness (smoothness of the spectral pattern / long-term stability).

The Setting this parameter can be visually supported properly, by using this parameter with the aid of a slider shown function is varied and the resulting EP patterns and EP signals are immediately displayed on a screen to the user become.

Both Variants are mathematically equivalent, but differ in the visualization options for one support the optional setting of the two parameters component number or Regularisierungsparmeter.

After the determination of the coefficients c _i and the selection of the most important main components u 0 / i or u 0 / i, the sample etching is evaluated and the weights of the linear combination are determined from the measured quantities.

und den optimalen Koeffizienten c opt / i wird ein Endpunktmustersignal u_EP gebildet gemäß folgender Vorschrift:From the selected main components c opt = Σ -1 · V T · y NOT A WORD , (19) wherein one of the two regularization variants (rules (14) or (18)) described above is used for the inversion of Σ,

and the optimum coefficient c opt / i, an end point pattern signal u _{EP is} formed according to the following rule:

5a shows a determined endpoint pattern u _EP . The individual elements of the endpoint pattern u _EP are the sought-after weights for the linear combination of the intensities of the measured wavelengths of the emission spectrum.

In the present embodiment, the endpoint pattern u _{EP can be} a chemical interpretati be assigned on.

In 5a For example, you can see CN lines and H lines.

In the following, the end point of an etch, that is, a plasma etching process, during the manufacture of an integrated circuit is determined by, as described with reference to FIG 1 described, continuously the emission spectrum of the plasma etching is measured.

Using the weights u _EP, a linear combination y (t) -the endpoint signal-is formed from the measured variables. This can be done, for example, by creating a matrix Λ (λ, t) from the measured values and multiplying the endpoint pattern u _EP by the transposed matrix Λ (λ, t) after each measurement time according to the following rule: y (t) = Λ (λ, t) T · u EP , (22)

It follows thus, as in 5b shown, an endpoint signal with a distinct step shape, which allows a determination of the end point of the plasma etching very easy.

An evaluation of in 5b shown linear combination y (t) provides the desired end point of the plasma etching. One way of evaluating the linear combination y (t) is to specify a threshold value, for example, in the specification of the value "0.5." If the value of the linear combination y (t) exceeds the threshold value, then the etching process can be performed either immediately or after Expiration of a certain, permanently specified Nachätzzeit be terminated.

The Setting a threshold is very easy to implement and delivers for many processes produce satisfactory results.

The Method for determining the characteristic value usually depends on the absolute values of the Measured values. In the present embodiment, this depends Achieving the threshold, for example, the total intensity of the measured Radiation off. The total intensity the measured radiation is not known a priori, so that by specifying a threshold, the correct endpoint is not is exactly defined.

Out For this reason, a qualitative signal analysis is preferably performed. to The exact definition of the endpoint will therefore be the turning point in the following the linear combination y (t), that is, the maximum of the first derivative the linear combination y (t) after the time used. The turning point is essentially independent from the total intensity the measured signal. According to the preferred embodiment the invention, the linear combination y (t) is filtered first, before the first time derivative is formed. Through the filtering the noise is suppressed and the actual signal is stronger.

In order to in the filtering, the maximum of the first derivative of the linear combination y (t) does not move, the filtering is done with a so-called Zero-phase filter performed. For filtering the linear combination y (t), a number of methods or filters are used.

In the present embodiment, the discrete-time variant (sampling time T _a ) of a first-order Butterworth filter (filter time T _F ) is used.

The Butterworth filter has the following transfer function G:

For the concrete application, a filter time constant T _F of 10 seconds was selected at a sampling time T _a of two seconds.

The concrete filtering is carried out by first applying the filter G to the signal to be filtered, that is to say the linear combination y (t). The intermediate signal y * generated in this way is passed to compensate for the phase shift generated in the first filtering via a zero and Polstellenspiegelung filter created by the filter. Due to the instability of this filter, it is necessary to perform the implementation in the reverse time direction. However, this is equivalent to the renewed application of the original filter G to the time reversed signal y *. The result of this filtering is in turn reversed in its time direction to finally obtain the desired filtered signal y _F. This two-fold application of the first-order Butterworth filter yields a zero-phase filter that does not shift the inflection point of the function y (t). Since one works usually with a fixed Nachätzzeit, the so-called over-etch time, this method can be applied despite their temporally akausalen character, which is in the order of the filter time constant.

Out the temporal derivatives of the filtered signals become the maximum determines what corresponds to the turning points of the linear combination y (t).

To Expiration of a fixed over-etch time, that is, after-etching time, becomes the plasma etching process stopped with the presented methods end point controlled. Especially vias with a very low aperture ratio can on be produced reproducibly and with high quality.

According to the above described embodiment The invention has used only one trial etch to the right one Weights, that is the coefficients for determine the linear combination of the measured quantities. To increase the robustness of the method, however, it is alternatively provided the results of several test runs or production runs in the determination of weights for incorporate the linear combination allow. So can For example, the measured Spektrenmatrizen multiple etches, for example at the beginning, in the middle and at the end of a so-called wet-clean Cycle, be included in the determination of weights.

In order to take into account a number of trial runs or production runs, it is preferable to combine the measured variable matrices of the individual processes into a single, large block matrix. For the example of contact hole etching, this means that a large spectral matrix is formed from n time-aligned spectral matrices. This means that, for a fixed wavelength (that is to say one row of the matrix), the temporal intensity profiles of the different etching processes are entered one after the other into a row of the matrix. If, for example, three sample etchings are to be taken into account in accordance with the embodiment described above, a matrix Λ ₀ (λ, t) with 1024 lines (wavelengths) and 300 (3 × 100) columns is formed (3 etchings with 100 time steps each).

To determine the weights of the linear combination y (t), it is then again possible to carry out a singular value decomposition on the basis of the matrix Λ ₀ (λ, t) which comprises a plurality of block matrices Λ _n (λ, t) according to the following rule:

The matrix U again contains the linearly independent spectral base patterns u _i (λ). However, the vectors v _i (t) now embody the temporal signal components associated with the spectral base patterns over several etchings. Accordingly, the data from the pre- and post-end ranges of all etches used in the analysis are used for modeling. The endpoint pattern u _EP then generated by the regularized modeling process now additionally has the property of compensating for the differences between the individual etches (drifts, window absorptions,...) In the resulting endpoint signals and thus leads to end-point-stable end-point signals.

100

arrangement to monitor a manufacturing process at line

101

process chamber

102

inlet

103

RF source

104

outlet

105

Silicon wafer

106

bracket

107

Multichannel spectrometer

108

Data processing system

109

plasma

110

optical fiber

401

CN-line

402

Time

403

Intensity CN line

404

Endpoint range

405

first time range

406

second time range

501

endpoint signal time range

502

endpoint signal spectral

Claims (19)

Method for computer-aided monitoring of a production process, - in which a variety of metrics during the Production process is recorded, - in which a first time range selected which occurs before the occurrence of a given event, - in which a second time range selected which occurs after the occurrence of the given event, - in which using the plurality of measured variables at least one linear combination is formed, wherein the weights of the linear combination are chosen such that the waveform of the linear combination in the first time range a predetermined first waveform and in the second time range has a predetermined second waveform, - in which from the linear combination at least one the manufacturing process characterizing characteristic determined will, and - at monitored using the characteristic of the manufacturing process becomes. Method according to Claim 1, in which only measured variables are used which have been detected at times which are in the first time range and / or in the second time range. The method of claim 1 or 2, wherein the predetermined first waveform of a predetermined constant first function value is formed. The method of claim 3, wherein the first function value the value is zero. Method according to one of claims 1 to 4, wherein the predetermined second waveform of a predetermined constant second function value is formed. The method of claim 5, wherein the second function value the value is one. Method according to one of claims 1 to 6, in which as measured variables the intensities of different wavelengths be used in an emission spectrum. Method according to claim 7, wherein the intensities of different wavelengths be used in a spectral range between 200 nm and 950 nm. Method according to claim 7 or 8, wherein the emission spectrum a plasma process is used. The method of claim 9, wherein the emission spectrum a plasma etching process is used. Method according to Claim 9 or 10, in which the endpoint is the characteristic of the plasma process is determined. Method according to one of claims 1 to 11, - in which to determine the weight of the linear combination, the measured quantities to the predetermined times are arranged in a matrix, and - in which Based on the matrix, a principal component analysis is performed. The method of claim 12, wherein for performing the Principal Component Analysis a singular value decomposition is used. Method according to one of claims 1 to 13, wherein for the determination The weights of the linear combination a parameter-linear Schatzproblem using the first waveform of the linear combination in FIG the first time range and using the second waveform the linear combination in the second time range minimized or is maximized. The method of claim 14, wherein minimizing or maximizing the parameter-linear estimation problem the essential ones Main components of the matrix of measured quantities for approximation of the waveform the linear combination in the first time range and in the second Time range are determined. Method according to one of claims 1 to 15, wherein for the determination the weights of the linear combination used the measured variables of several test or production runs become. Method according to claim 16, - in which to determine the weights of the linear combination the measured quantities several Trial and / or production runs be arranged in a matrix, and - based on the matrix a principal component analysis is performed. Arrangement for monitoring a manufacturing process, - with at least one sensor for acquiring measured variables during a Manufacturing process, - With at least one processor arranged such that the following Procedural steps are feasible: - a first Time range is selected which is before the occurrence of a given event, - a second Time range is selected which is after the occurrence of the given event, - under Use of the plurality of measured variables is at least one linear combination formed, wherein the weights of the linear combination are chosen such that the waveform of the linear combination in the first time range a predetermined first waveform and in the second time range has a predetermined second waveform, - from the Linear combination becomes at least one characterizing the manufacturing process Characteristic determined, and - With a monitoring device to monitor the manufacturing process using the determined characteristic. Arrangement according to claim 18, with a control device, with which the manufacturing process is controlled using the characteristic becomes.