Data processing: generic control systems or specific application – Generic control system – apparatus or process – Optimization or adaptive control
Reexamination Certificate
2001-04-24
2003-09-09
Patel, Ramesh (Department: 2121)
Data processing: generic control systems or specific application
Generic control system, apparatus or process
Optimization or adaptive control
C700S014000, C700S055000, C700S121000, C700S159000, C700S110000, C700S143000
Reexamination Certificate
active
06618632
ABSTRACT:
BACKGROUND OF THE INVENTION
Field of the Invention
The present invention relates to a method of monitoring or controlling machining plants, in particular a method of monitoring or controlling machining plants for the production of integrated semiconductor products.
Modern industrial production is generally characterized by a high degree of automation. In the semiconductor industry in particular, a high degree of automation is an essential requirement for being internationally competitive.
During manufacture, the products pass through a series of machining steps, which are performed in an automated manner on the product to be machined by various machining plants. For example, in the production of an integrated semiconductor product, the product passes through up to 600 process steps, which for a large part can be carried out only with highly specialized machining plants in a clean-room environment. In this case, a number of identical machining plants are often combined to form a machining area (“bay”).
The costs for such automated production are influenced to a great extent by the question as to how well and efficiently the manufacturing process can be monitored or controlled, so that the ratio of defect-free products to the overall number of products manufactured (=yield) assumes as great a value as possible.
Unfortunately, the individual machining steps are subject to fluctuations and irregularities, which in the worst case may mean, for example, the defect of a number of chips or the entire wafer or the failure of a machining plant. Therefore, each individual machining step must be carried out as stably as possible in order to ensure an acceptable yield after the completed processing of a wafer.
In the prior art there are extensive methods available for localizing possibly present defects on a wafer and assigning these defects to particular machining steps. However, some of these methods use data which are only obtained from the complete machining of a wafer. Therefore, tracing back a defect is often only possible with very great effort.
For example, after the completion of a wafer, various electrical and functional parameters, such as current yields, turn-on voltages of transistors, stand-by currents, etc., are measured under different timing conditions in the wafer test area and subsequently evaluated by the production engineers. The defect evaluation is extremely complicated, since it has to be established retrospectively which machining steps were carried out improperly.
The time taken for a wafer to pass through a production line may be up to 8 weeks. If, for example, a serious defect occurs after a machining time of only 2 weeks, this defect may possibly be detected only after a further 6 weeks of processing. On the one hand, the production line is consequently unnecessarily occupied with reject products, on the other hand, due to the time delay, many further defective wafers are produced before the cause of the defect can be eliminated.
To alleviate this problem, there are, in addition, control instruments which analyze data acquired during or after a production step or a number of production steps.
For example, after certain machining steps, wafers are removed from the production process and visually analyzed with a scanning electron microscope (SEM). Unfortunately, these methods are very time-consuming and labor-intensive and therefore take place only on the basis of random sampling.
In the method of “Univariate Statistical Process Control (SPC)”, the process results (for example line widths or layer thicknesses on 2 wafers of a batch) are checked after a logically related sequence of process steps (for example coating, exposure to light and subsequent etching). If deviations from the desired value occur measures are taken to eliminate the causes of the defects. However, this gives rise to the difficulty of establishing at which of the machining steps coming into question the cause of the deviations is to be found.
In the method of “Univariate Statistical Equipment Control (SEC)”, measurement data, such as the process temperature, voltage, power, gas flow and pressure for example, are acquired in every process step via the so-called equipment link and entered on control cards. These control cards are checked at several machining steps to ascertain whether the parameter acquired is within the range of an upper control line (UCL) and a lower control line (LCL). If the measured values are between the LCL and UCL, the current process is considered to be normal. Since, however, it is not possible to investigate manually all parameter curves for all plants, because of the great effort this involves, one restricts oneself to the most important parameters. However, even if one restricts oneself to the most important parameters, this primarily visual method of control can be applied at most to a 3-dimensional set of parameters. Even with a 2-dimensional set of parameters, the method becomes very unclear. For example, the combination of two normal parameters may be abnormal, i.e. possibly lead to a chip or plant defect. Such multidimensional combinational effects cannot be resolved with the conventional methods. Furthermore, the management and analysis of the control cards are very time-consuming.
SUMMARY OF THE INVENTION
The publication “Fault Detection and Isolation in Technical Processes with Neural Networks” by B. Köppen-Seliger et al., appearing in Proceedings of the 34th Conference on Decision & Control New Orleans, La. December 1995, describes a concept for the detection and isolation of defects in technical processes which is based on the use of an RCE network (“Restricted Coulomb Energy Neural Network”). Furthermore, the document U.S. Pat. No. 5,361,628 (Kenneth A. Marko et al.) discloses a method of evaluating an engine test with the aid of neural networks. However, the methods mentioned require suitably controlled preliminary tests in order to allow the neural networks used to be appropriately trained. Corresponding preliminary tests with “good” machining plants under stable conditions require a very great effort, however, which generally cannot be provided in a production environment.
It is therefore the object of the present invention to provide a method of monitoring and/or controlling machining plants which avoids or alleviates the disadvantages mentioned of the conventional methods.
Further advantageous embodiments, refinements and aspects of the present invention emerge from the subclaims, the description and the attached drawings.
According to the invention, a method of monitoring and/or controlling machining plants which have time-dependent machining parameters is provided. The method according to the invention comprises the following steps:
a) desired time-dependent machining parameters are measured as a measured curve,
b) time-independent numerical values are generated from the measured machining parameters, and
c) the time-independent numerical values are entered into a classifier, which distinguishes between normal states of the machining plant and abnormal states of the machining plant.
The method according to the invention is characterized in that the classifier is trained with training vectors of which the components are time-independent numerical values, and, to prepare for the training of the classifier, those training vectors which are regarded as “abnormal” are filtered out from the training vectors available, the distance of each training vector from every other training vector being determined by means of a suitably selected measure of distance for the filtering of the abnormal training vectors.
The method according to the invention has the advantage that the analysis of the data takes place fully automatically. A time-intensive, and consequently cost-intensive, “manual” monitoring of machining plants can be avoided. All the desired parameters can be processed simultaneously and in parallel, immediately after the ending of a machining step. A time delay between the occurrence of a defect and its detection, as is customary with conventio
Federl Peter
Pöppel Gerhard
Wachtmeister Frank
Greenberg Laurence A.
Infineon - Technologies AG
Mayback Gregory L.
Pham Thomas
Stemer Werner H.
LandOfFree
Process for monitoring processing plants does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Process for monitoring processing plants, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Process for monitoring processing plants will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3056804