WO1999063134A1

WO1999063134A1 - Device and method for growing crystals

Info

Publication number: WO1999063134A1
Application number: PCT/EP1999/003788
Authority: WO
Inventors: Joachim Aufreiter
Original assignee: Leybold Systems Gmbh
Priority date: 1998-06-04
Filing date: 1999-06-01
Publication date: 1999-12-09
Also published as: DE19824838A1; JP2002517367A; KR20010022606A; EP1007769A1

Abstract

The invention relates to a device and method for growing crystals. According to the invention, signals which come from a neuronal network are superimposed on the control signals provided for a PID controller. Through this, single-crystals and polycrystals having a very high quality with regard to the outer shape and inner constitution are obtained.

Description

DEVICE AND METHOD FOR GROWING CRYSTALS

description

The invention relates to a device and a method for growing crystals according to the preambles of claims 1, 6 and 16.

In semiconductor, optoelectronic, microwave and other technologies, polycrystalline or amorphous substances often do not meet the requirements. In such a case, efforts are therefore made to use single crystals, the production of which, however, is complex. Since single crystals are practically non-existent in nature, they have to be manufactured artificially.

Known processes for producing single crystals are the Czochralski process, the Kyropoulos process, the vertical Bridgman process or the crucible-free process of the vertically changing zone.

In the Czochralski process, which is of great importance for production, the raw material, e.g. As silicon, germanium or gallium arsenide, placed in a crucible, which is located in an evacuated or filled with inert gas. After the raw material has melted, a seed element is immersed in the melt and pulled up while rotating. The desired single crystal then grows on this seed element.

The other procedures are more limited to laboratory operation. In the Kyropoulos process, similar to the Czochralski process, the raw material is first melted in a crucible. The seed crystal is held on a water-cooled rod to remove the heat of crystallization and into the Melt introduced. The temperature of the melt is then lowered so that a single crystal grows into the melt on the seed crystal.

In the Bridgman process, the raw material is melted in a crucible, the lower end of which tapers and the bottom end has a seed crystal.

By exposing this area to a temperature gradient, a single crystal grows upward from the seed crystal into the melt.

With the process of the vertically changing zone, single crystals can be grown without a crucible. There are two drawing rods, each one

Has bracket. A seed crystal is connected to the lower holder, while a polycrystalline rod is connected to the upper holder. The lower end of the polycrystalline rod is then heated to the melting point by means of a high-frequency induction coil. The seed crystal is then gradually introduced into the melting area from below. The upper pull rod and the seed crystal can now be moved downwards, whereby they are connected to each other via a melting zone within the induction coil and rotate in the opposite direction.

Regulations that automatically carry out the individual processes are used in all of these single-crystal production processes. The following geometric parameters are regulated: crystal neck length, crystal shoulder shape, crystal diameter and - if possible - internal solid size such as crystal orientation, oxygen content, resistance gradient and doping. These variables are therefore the controlled variables. These control variables are influenced by the physical variables temperature, gas flow and gas pressure in the housing that surrounds the crucible, and possibly by the speed of rotation of a rod on which the crystal is hanging, or by the speed of rotation of a crucible in which the melt is located. These physical variables therefore represent the disturbance variables of the control loop.

For the production of a perfect single crystal, it is important that the melt pool from or in which the crystal grows has a certain temperature profile. Such temperature profiles can be recorded, for example, by thermal cameras (cf. DE 39 19 920 AI). It is then by means of special control devices possible to generate certain temperature gradients that are required for crystal formation.

So-called feedback loops are used in closed control loops. H. the output signal of an amplifier is fed back to the input of this amplifier. If this feedback is proportional, one speaks of a P controller, if it is differential, one speaks of an I controller. The combination of P and I feedback is called "yielding". The corresponding controller is then called the PI controller. Analogously, a PD controller is called a delayed controller. In crystal growth processes, however, PID

Controllers used, i.e. controllers with yielding and delayed feedback. If a controller is only to be operated stably, a PI controller is sufficient. It delivers very good steady-state accuracy and, with the correct setting of its parameters, good stabilization at moderate speed. However, a PI controller is no longer sufficient if, with good stabilization, high demands are made on the speed of the control. In this case, a PID controller is used.

With PID controllers it is possible to control temperature profiles quickly and precisely in order to obtain a uniform crystal. The PID controller does not take into account whether and to what extent the quality of the crystal changes. This means that a single crystal produced by means of a PID controller has a very uniform outer body, but its inner values, e.g. B. oxygen content and ohmic resistance, may be distributed unevenly.

So-called fuzzy controllers can also be used in complex systems, which include the manufacturing processes of single or polycrystals. Such controllers are used in fuzzy technology (cf. K. Goser, fuzzy technology: fuzzy controllers for complex systems, Phys. Bl. 50, 1994, No. 11, pp. 1064-1067; Preuss: fuzzy control -heuristic control using fuzzy logic ", atp- automation engineering practice, 4/92, pages 176 to 184 and 5/92, pages 239

- 244).

In contrast to the random fuzziness that can be treated with the tools of probability theory, the fuzzy theory is about the linguistic I see fuzzy descriptions of properties such as fast, large, small, etc.

The term "fuzzy logic" (fuzzy = blurry) is used as a collective term for fuzzy set theory and fuzzy logic. While the rule applies to classic control methods: If the pressure> 3 bar, then

When cooling is switched on, the "blurred" rule applies to fuzzy control: If the temperature is high, then cool it very strongly. According to fuzzy logic, the premise does not assume either the value "0" or "1", but truth values between "0" and "l".

An intelligent controller based on fuzzy logic has several advantages, including simple construction and simple design. The fact that fuzzy controllers are largely insensitive to inaccurately measured input parameters makes it possible to work with the simplest measurement sensors to obtain the input parameters.

The fuzzy technique is also known in connection with the pulling of single crystals from a melt (DE 43 01 072 AI = US Pat. No. 5,485,802). Here, the quantities of the crystal pulling process determined with the help of sensors or measuring devices, such as the melt bath height and the melt temperature, are continuously fed as input signals to a controller, to which a fuzzy processor is implemented that follows a control structure that takes into account empirically determined quantities Outputs control signal for a conveyor. Here, an empirically determined data field is used which contains at least two parameters from a first group and at least one parameter from a second group, the first

Group includes the melting temperature, the melting level and the crystal diameter, while the second group includes the material feed rate, the drawing rate and the heating power. The two parameters from the first group are then measured, and said at least one parameter from the second group is regulated by means of a fuzzy processor, so that the two at least two measured parameters from the first group and said at least one parameter the second group correspond to the empirically determined data field.

Although fuzzy logic has a number of advantages, it also has some disadvantages. As the complexity of the control system increases, it becomes very fast difficult to determine the correct set of rules - the so-called "expert" rules based on the experience of professionals - and member functions to accurately describe system behavior. In addition, an expert is not always able to express his knowledge explicitly.

If the complexity increases further, it can become extremely difficult and time-consuming to improve the linguistic variables and an existing rule set with regard to a given optimization goal. Sometimes the trial and error method for determining the fuzzy rules and the membership functions fails completely.

Furthermore, in particular in the case of a "feed forward system", no recurring information is included. In other words, conventional fuzzy logic rules do not hold back information about previous results or decisions. As a result, the ability to describe system behavior is limited.

Attempts have therefore been made to solve this problem by means of arrangements which are capable of learning. Such arrangements capable of learning are the artificial neural networks (KNN). They are used in particular when there is no algorithm between a sought and an existing size or only with a large one

Effort can be determined. In such cases, general algorithms are necessary that use a vector - e.g. B. from pressure and temperature - to another vector - z. B. the concentration of certain substances - map.

The neural network algorithm can be used to learn a relationship between certain quantities.

A neural network is the mathematical modeling of nerve cells and their connections, such as those found in the human brain. Neural networks are among the networks by the two properties "adaptation" and

Marked "Association". The adaptation consists in the ability to recognize and store knowledge. The association enables a network to compare information with stored knowledge so that what is known can be recognized. A neural network is a network whose nodes are called neurons and whose connections of the neurons carry threshold values, also called weights. These neurons can assume two states, namely the state of rest and the state of excitation. A neuron can also have multiple inputs, but only one output. The output of a neuron is connected to inputs of other neurons or to the outside world. A neuron changes to the excited state when a sufficient number of inputs are excited above the respective threshold value of the neuron. A well-known neural network is the feedforward network, a network whose neurons are organized in three layers. The input information flows from the input layer of the neurons via the connections to the hidden intermediate layer and from there to the output layer. The weights of the connections can be attached to the inputs of the neurons.

A neural network for regulating a CVD process for the production of semiconductor wafers is already known (US Pat. No. 5,649,063). This neural network interacts with a signal processor, in such a way that one

Feedback control is implemented, which reduces the process fluctuations.

However, neural networks are not always the most effective way to use an intelligent controller, since the implementation of a neural network is more expensive than the implementation of fuzzy logic. For example, the fuzzy logic is more effective for certain applications and - with appropriate programming - a conventional controller can be used to use the fuzzy logic. It is also possible to implement a neural network by programming a conventional inserted controller, but it is much slower. In addition, the use of hardware in fuzzy logic is easier than in neural logic.

Efforts are therefore made to combine fuzzy logic with neural networks (cf. B. Kosko: Neural Networks and Fuzzy Systems, Prentice-Hall, 1992;

Mahyon et al .: Neural Network and Fuzzy, Models For Real Time Control of a CVD Epitacral Reactor, Proc. SPIE-Int. Soc. Opt. Eng., 1992). From the fuzzy point of view, the weights of the network are selected as fuzzy sets of the linguistic terms that are in the rules of the fuzzy controller. A neuro-fuzzy controller is thus the implementation of a fuzzy controller on a neural network, so that the network work function is determined by a fuzzy inference scheme.

A fuzzy logic system based on a repeating neural network is already known (US Pat. No. 5,606,646). This system has neurons in a regular base layer, each neuron having a repeating architecture with an output-input feedback branch that contains a time delay element and a neural weight. Furthermore, this system has a new, network-based fuzzy logic machine in the finite state, the fuzzy logic system based on the neural network having a recurring architecture with an output-input feedback path which contains at least one delay element. Also included is a recurring fuzzy logic rule generator based on a neural network, where a neural network receives and fuzzifies the input data and provides data related to the fuzzy logic membership functions and the recurring fuzzy logic rules correspond.

Process optimization using a neural network is also known (US Pat. No. 5,671,335). This process is a complex process with multiple inputs, for example an injection molding process that is optimized so that it is integrated due to a neural network trained for the process

Target delivers. A test input is propagated forward over the neural network and the output of the network is compared to the target output. The difference is retransmitted over the network to determine an input error value in the network. This error value is used to correct the test entry. The correction process is repeated until the test input generates the target output within a predetermined degree of accuracy.

The object of the invention is to improve the quality of single and / or polycrystals by using special regulators.

This object is achieved according to the features of claims 1, 6 and 16.

The advantage achieved with the invention is, in particular, that a very high quality of single and poly crystals with regard to the outer shape and the inner Quality is achieved because process errors are recognized and corrected immediately. In addition, the invention can easily be integrated into a computer-integrated manufacturing process.

Furthermore, it is possible to map the conditions in the manufacture of crystals by means of multi-stage artificial neural networks and thus to recognize early on during the manufacturing process whether the crystal has defects. The corresponding parameter correction can then be carried out in time using a fuzzy controller. In addition, operator training can be carried out on the cold model.

Furthermore, additional quality data of the aids, such as bubble profile of the quartz crucible, service life of the heater parts, degree of aging, number of dips per seed, batch and doping material, etc., can be used as soft sensors. It is also possible to use virtual analyzers, e.g. B. for melt pool viscosity, to correct the timing behavior of the control loop accordingly. The influence of sequences or time slices of different lengths such as stabilization, dip, neck, shoulder, body, endcone and shutdown on the subsequent step, such as B. Diameter overshoots in the transition from shoulder to body are corrected in the control strategy in accordance with the model analysis by superimposed fuzzy controllers and robustness filters. With corresponding process data, a prediction of measured values such as temperature, diameter, drawing speed etc. can be made. The forecast says e.g. B. off: If the process continues in this state, the product will receive a certain quality value.

The invention thus relates to a device and a method for growing crystals, in which the control signals for PID controllers are superimposed on signals that come from a neural network. In this way, a very high quality of single and poly crystals is achieved with regard to the outer shape and inner properties.

An embodiment of the invention is shown in the drawing and will be described in more detail below. Show it:

Fig. 1 is a schematic diagram of a device for pulling single crystals according to the Czochralski method with neuro-fuzzy control; 2 is a schematic diagram of the links between a neuro-fuzzy

Controller and a PID controller; 3 shows an artificial neural network; 4 shows a neuron with several inputs and one output;

5 shows a fuzzy controller; Fig. 6 shows a PID controller arrangement with superimposed neuro-fuzzy control.

1 shows a device 1 with which it is possible to pull a crystal 2, the underside 3 of which is connected to the surface 4 of a melt 5, from this melt 5. The crystal 2 is rotated in the direction of an arrow 6 or raised or lowered in the direction of an arrow 6a. The melt 5 is located in a crucible 7, which is driven by an electric motor 9 by means of a shaft 8. Shaft 8 and motor 9 are connected to one another via flanges 10, 11. The crucible 7 is located in a housing which consists of an upper part 12, a middle part 13 and a lower part 14. It can be rotated in the direction of an arrow 15 and raised and lowered in the direction of an arrow 6b. An electric heating device 16 is placed around the crucible 7 and is controlled by a programmable logic controller 17 (= PLC).

The rotation of the crystal 2 takes place by means of a rod or a rope 18, e.g. B. a threaded rod which is driven by an electric motor 19. This motor 19 is also controlled by the PLC 17.

The rod 18, the vertical axis of which is designated by 22, is surrounded by a tube 23 which is connected to the upper part 12 of the housing 12, 13, 14. This tube 23, which is connected to the part 12 via a flange 180, has a gas inlet opening 24. Gas outlet openings 25, 26 are provided in the bottom part 14 of the housing 12, 13, 14 and can be blocked off by valves 31, 32. The valves can be combined with suction pumps so that the gases in the housing 12, 13, 14 can not only be sealed off, but can also be extracted.

As can be seen from the arrows 6 and 6a, the rod 18 can not only be rotated by the electric motor 19, but also raised. The control of the elec- tromotors 19 takes place via the PLC 17, which is supplied with various operating status information. This operating state information is supplied by various sensors, for example a sensor 43 for detecting the level of the melt 5, a sensor 44 for detecting the temperature above the melt 5, a sensor 45 for detecting the gas pressure and a sensor 47, for example a camera to detect the shape of the crystal 2. Your data are fed to the PLC 17 and from there via lines 60 to 64 to a neuro-fuzzy logic 52 which is implemented as software in a computer 50.

The sensors 43 to 47 are only shown schematically in FIG. 1. It goes without saying that further sensors, e.g. B. for detecting the gas flow or for detecting the heat of the heating coil 16 can be provided.

The PLC 17 can also be used to control a valve 27 which is arranged between a gas reservoir 28 and the gas inlet opening 24. With this valve 27, the inflow of a gas into the tube 23 and thus into the container 12, 13, 14 can be controlled. The pumps or valves 31, 32 are also controlled via the PLC 17.

The PLC 17 works together with the computer 50, preferably a PC, which in turn is connected to a monitor 51. A data exchange takes place between PC 50 and PLC 17, which is indicated by data line 20. Likewise, data is exchanged between the neuro-fuzzy logic 52 and the PLC 17 via the data lines 60 to 64 or 65, 66.

A somewhat more detailed representation of the control processes important here is shown in FIG. 2. It can be seen in this illustration that the neuro-fuzzy logic 52 has a neuro part 53 and a fuzzy part 54 and that both parts 53, 54 are contained in the computer 50. In addition, it can be seen that the programmable controller 17 contains a PID controller 55 which is generated by the computer

50 is controlled. This PID controller can be converted into a P or D or ID or PD or PID controller by simply switching over. A more detailed illustration of the arrangement according to FIG. 2 can be found in FIG. 6, which will be described further below. PID controllers are mainly used in conventional systems. They are the most widely distributed. There are various reasons for this. For example, the use of PID controllers is not restricted to a special class of controlled system, but is possible for almost all process types. Practically all classic design processes are suitable for PID controllers. Thereby stand out

PID controllers are particularly characterized by their robustness. Structurally, a PID controller can be represented as a parallel connection made up of the P, I and D components, the P component ensuring generally favorable control behavior, the I component ensuring steady-state accuracy and the D component ensuring rapid regulation. A modern hardware PID controller works by means of an operational amplifier which has a series RC circuit in the feedback direction and on which there is an input of a parallel circuit comprising a resistor and a capacitor.

The PID algorithm is increasingly being implemented in software, for example on a memory-programmable controller (PLC) or a digital computer.

This form of implementation is characterized in particular by its flexibility.

FIG. 3 shows that the data given by the sensors 43 to 47 via the lines 33 to 37 to the PLC 17 go via lines 60 to 64 to the neuro part 53 of the computer 50. This neuro part 53 has a network 70 with neurons 72 to 76, a first hidden intermediate layer 77 (hidden layer) with six neurons 78 to 83, a second hidden intermediate layer 84 with five neurons 85, 86, 87, 88, 89 and one Output layer 90 with two neurons 91, 92. The neurons 91, 92 are connected to the output lines 65, 66.

3 shows a so-called feed-forward network. Neural networks are standardized structures and can therefore only process values between 0 and 1 or between -1 and +1. Need ^{to, J} are -gtewandelt real numbers.

As shown in FIG. 4, the weights wi, w ₂ , w _{3 of} the connections can be attached to the inputs xi, x ₂ , X3 of the neurons, respectively. Knowledge is stored in the weights of the inputs of a neuron and the threshold value of the neuron itself by adaptation. The association of input information with the stored information takes place by means of a comparison algorithm in the neuron. This comparison algorithm together with the weights defines one Network function. Is z. B. w _j = 0, then there is no connection between the previous and the following neuron. If wi = 0, the neuron inhibits the successor neuron, and if w> 0, the neuron activates the successor neuron. The possible types of learning using artificial neurons are: development of new connections, deletion of existing connections, modification of the strength w of

Connections, modification of the threshold value of neurons, modification of the activation and output function, insertion of new neurons and deletion of neurons.

The model reflection of a natural neuron is called Hebb's

Perceptron called. The inputs x _j , x, X3, which can be outputs from upstream data sources such as sensors, are multiplied within the neuron by the so-called weights w _j , w ₂ , W3 (e.g. 2; 0.1; 0.5) . The resulting weighted inputs are then processed via a so-called transfer function and result in the total output y. Since artificial neurons are simulated in the computer in the form of mathematical systems of equations, it is necessary to express the complete functions of the neuron as equations, the inputs of which are converted into an output.

If several perceptrons are connected in series and side by side, as shown in FIG. 3, with appropriate setting of the weights w _j , w, W3 and corresponding complexity of the network, any mathematical function and any formal connection can theoretically be represented. The entire artificial neural network then presents itself as a large equation with up to hundreds of coefficients (weights) and can - within certain value ranges - also map any other equation.

An exponential transfer function, which only allows the individual neuron to become active from a certain total value, ensures that certain neuron clusters each become active at a certain threshold value and thus rigid nonlinear relationships can also be represented by a network according to FIG. 3.

The weights of the neural network 70 can be trained - ie adapted - in one learning process. There are a number of learning methods, one of the best known of which is the back propagation method. Through the learning process, the Weights are set on the connections that connect the neurons of the successive layers to one another in accordance with the input signals in the process and the output signals from the process. When properly trained, the neuro network works like an inverse model of the crystal pulling process. In response to input signals, which represent process output calculators, the artificial neuro network generates several signals that determine the crystallization process. When these signals are added to the process, they cause a group of output signals to be generated that is closer to the desired group of output signals.

In multi-layer back propagation methods, the mathematical algorithm aims for zero errors by changing the weight.

Each neuron first calculates the weighted sum of the outputs of the neurons in front of it. Then it calculates the baseline by taking the weighted

Converted total using a transfer function. This output value is then passed on to the next neuron or to the output.

In addition to the back propagation method developed by Rummelhard, there are also other methods that use neural networks, e.g. B. the so-called

Kohonen networks developed by T. Kohonen around 1988. Here, all neurons are linked to the inputs and all other neurons, neighboring neurons become stronger, distant neurons inhibit each other and each input pattern correlates with an active center. Compared to that of David Rummelhard u. a. For example, the generalized backpropagation algorithm with delta rule proposed around 1985 for multilayered artificial neural networks, these alternatives have remained of little importance in practice.

With the back propagation algorithm, a training data set is first searched for, on which the neural network is trained. Such a training data record generally consists of a finite number of mutually assigned input and output vectors (series of numbers). Each input vector consists e.g. B. from three numbers (three inputs!) And each output vector from one number (one output). An error backpropation algorithm is then determined. Here the average error of all outputs of the last layer of the network is calculated. The weights of the individual neurons are then set in such a way that the mentioned error becomes iteratively zero. The Rummelhard artificial neural network is thus a type of iteration machine which, through small, targeted changes in the weights, approaches an optimum, the error minimum.

Computer programs that perform the above-described actions are already known and are offered by various companies, e.g. B. from the Californian companies Neuro DynamX (DynaMind) or CSS (Brain Maker).

5 shows the logical structure of a fuzzy controller 54 with one output variable and several input variables. Viewed from the outside, this controller has no blurring, i. H. Both input and manipulated variables are sharp values. Rather, the blurring is due to the interior of the controller. Both the input and the manipulated variables are linguistic variables and by the membership functions, i. H. Fuzzy sets of the individual linguistic threshing floor, characterized. The sharp input variables on the lines 65, 66 are converted into fuzzy variables by the fuzzification units 100, 101, 102. In the second step, an interference machine 103 generates based on a predetermined one

Control value, a fuzzy manipulated variable from these fuzzified input variables. This is finally converted back into a sharp signal by a defuzzification unit 104.

When fuzzifying, the input variables of the fuzzy controller are assigned to the linguistic variables and linguistic tennes. A temperature is so z. B. with different degrees of membership assigned the linguistic terms "high temperature", "medium temperature" and "low temperature". So a temperature of 900 ° C can belong to 0.26 (26%) of the amount of the "medium temperature" and at the same time to 0.68 of the amount of the "high temperature". This uncertainty of knowledge corresponds to linguistic uncertainty.

Two rules, which are provided with the premises "If temperature is high" and "If temperature is medium", would both affect the process in this case, but would be weighted differently. This weighting is carried out in the second step of information processing, the so-called inference. The inference engine of a fuzzy controller uses the linguistic variables from the fuzzification, looks at all linguistic rules and creates clear statements from contradicting rules. Comes z. For example, if a rule with the plausibility 0.36 concludes that a valve must be closed and another rule closes with the plausibility 0.65 that a valve only has to be closed half, the inference machine in the fuzzy controller forms a compromise value according to various possible methods and closes the valve z. B. 60% total opening.

The defuzzification is used in the last step of the information processing of the fuzzy controller 54 in order to convert the linguistic statements back into physical (manipulated) variables and to now lead to the contradictory consequences of these rules to a clear result. It is possible to convert fuzzy statements such as "wide open", "slight opening" or "high flow rate" back into physical quantities. B. Lead setpoints from PID controllers.

6 shows a block diagram which shows the principle of the neuro-fuzzy control of the crystal pulling systems once again in principle.

The crystal pulling process is a typical multi-size problem, since different sizes such as crystal rotation, crucible rotation, crystal or crucible lifting or lowering speed, gas flow, gas chamber pressure, heating temperature, melting temperature, crystal diameter, melt pool height etc. have to be taken into account. In such processes, model-based methods are used in connection with the artificial neural network. This is essentially done in two steps: First the model is formed and then the controller is designed. When creating a model, there are two options: the creation of a control observer model or the development of a reaction-kinetic model. The formation of kinetic models for the purpose of regulation is, however, relatively complex, because all known material conversions, heat transfers, diffusion processes and balance equations must be recorded, quantified and formulated in the form of differential equations, which then represent the behavior of the modeled process as a whole. In contrast, when the control engineering models are created, the process considered is considered "Black Box" viewed, from which analysis of the reactions of the process to disturbances draws conclusions about the dynamic route behavior. The observed reactions are in turn formulated in higher order differential equations and form the model. Since most models do not consist of closed mathematical equations, but e.g. B. include partial differential equations, closed controller development is rarely possible.

Drawing a crystal using the Czochralski method consists of several steps. Ultimately, the aim of process control is to operate the process in such a way that certain quality parameters are kept at a level that guarantees constant product quality. These quality parameters include e.g. B. the crystal orientation, the oxygen content of the crystal, the electrical resistance gradient and the doping.

When pulling the crystal, the problems of excessive dead time or inaccurate measurement of sizes can arise.

Various attempts have been made to solve these problems with empirical control systems based on pure fuzzy controllers. The problem of a pure fuzzy controller - as well as a classic controller - consists essentially in the fact that this, like the plant operator, is subject to the problems described above due to the long dead times.

In order to correctly parameterize a fuzzy controller in such a system for every situation, a very large effort has to be made, so that such a system is ultimately burdened with too much maintenance and configuration effort.

According to the present invention, therefore, all relevant data of the crystal pulling system 1 are recorded cyclically and stored in a database. In addition to the temperature values of the heating zones 16, gas flow, drawing speed etc. are also stored here.

With the help of the neuromodel 202, a neuronal process model is then generated from this data, which predicts the properties of the crystal. By entering certain sizes in the neuromodel 202, various properties of the crystal to be drawn.

The neuromodel 202 completed in this way is used e.g. B. in a process computer, such as in the computer 50 of FIG. 1, implemented and connected to the process on line. A superimposed fuzzy controller 201 registers on the basis of the

Predictions of the neuromodel 202 constantly the expected quality of the crystal and thus counteracts quality drops even before they can be measured.

Before the fuzzy controller 201 writes changes to the setpoints of the subordinate process controllers, it tries changes z. B. from the temperature control for the heater 16 on the neuro model 202. There are so-called test setpoints for the neuromodel.

To a certain extent, the neuromodel 202 runs alongside the process and is corrected via a statistical model monitor 200 and a robustness filter 203. The filtering of the robustness filter 203, which is a low-pass filter, is as sluggish as necessary, but as quick as possible. The signal and noise components of the data are separated. System-theoretical methods of autocorrelation analysis can be used for this, since these methods can identify the noise component and the signal component of the continuously recorded data.

In the statistical model monitoring 200, to which the material and analysis data 94 are fed, the KNN is checked with test data which are not part of the

Training database were. Each recommendation of the fuzzy controller 201 therefore initially goes to the neuromodel 202, is evaluated again by the fuzzy controller and only then written to the process via the robustness filter 203. The neuromodel 202, to which material and analysis data from the crystal pulling system (1) are fed, provides so-called quality data (QA (t + 1); QB (t + 1)).

The conventional PID controllers 204 to 206, which are implemented in the PLC 17, remain. They are only corrected by the signals coming from the robustness filter 203. If only controllers 204 to 206 were used, there would be no optimal control because of the multi-size problems. The individual controllers 204 to 206 would have to be corrected manually by the system operator.

However, manual controllers are always associated with the disadvantage that, in addition to the reproducibility of the operations of the operating personnel being low, work is always delayed and therefore less precise.

The experience of the plant operator is recorded in the fuzzy controller 201.

The conventional controllers 204 to 206 are, however, already able to operate a system, although not optimally. Therefore, the fuzzy controller 201 becomes

Setpoint control or parameterization is superimposed on the PID controllers 204 to 206 as an optimization controller. If undefined states of a controller occur, this ensures that conventional control technology can keep the process at least within non-critical limits. 93 is a network and 94 is the data logging for material and analysis data. With the network

93 can be a QA network.

The method according to the invention can thus be used to produce crystals, in particular single crystals, process variables such as melting temperature or gas flow on the one hand and quality characteristics such as analysis data of raw materials on the other hand

- e.g. B. purity or doping of silicon, or dimensional tolerances in crucible walls or quantities of base and dopants can be detected and processed.

The control devices for heating, pumps and the like are controlled in the usual way via PID controllers.

The current process variables are recorded and stored in the database 200 as target and actual values. In contrast to conventional control technology, the setpoints are also variable.

The current quality characteristics are recorded once as actual values up to the "Charging" process step, ie until the end of the preparation for the automatic process control, and stored in the database with fixed, optimal target values. The process variables are then processed in the database 200 using SPC methods. SPC methods ensure quality in production based on statistical assumptions. They are based on the assumption of normal distributions in data series. Here, a sliding corridor is driven over the data sets for data analysis, and the mean value and empirical spread are determined therein.

When processing process variables with the S PC method, a table is compressed to typical representative data after statistical validation by a fuzzy cluster method and used for training the neural network.

SE rules, d. H. "Standard Electric Rules" contain guidelines for the automatic analysis of data records for process capability. According to this, z. B. Records if a point is outside + 1-3 sigma or if two of three consecutive points are outside + 1-2 sigma or if five consecutive points show an upward or downward trend etc.

A neural process model is generated in 202 from the process variables processed in the database 200 and a local error assessment is output for each statement of the neural network. The selection of the algorithm and the configuration of the neural network is carried out automatically in 202 system-controlled systems. It is determined which input variables are not relevant, after which they are removed. As a result, the neuromodel only maps significant quantities to one another. A sensitivity analysis is carried out on these variables, by means of which the strength is recognized with which each input variable of the model acts on each output variable, which predicts the target values of the process variables or quality variables. A table on the quality characteristics, the process variables and the quality variables is given below.

Färämeteπ ualitäts- ProcessSensorik quality features sizes sizes

Raw materials, control diameter, image evaluation, diameter cir- cuit (camera) or weight sensors

(Si, analysis average encoder RRG - Dopant) values, pulling radial

Quantity, type, speed Resistance manufacturer gradient heater a) ORG - radial oxygen gradient several s. Heater a) u. a. m. Heater b)

Semi-finished shoulder function growth diameter / time

(Quartz, dimension, approach pyrometer

Graphite glazing temperature crucible) depth,

Number of bubbles,

Manufacturer,

Type etc.

Vacuum / pressure measurement

Pressure cans

Gas flow flow meter

Functions magnetic field calculated bath height trianulation position laser or calculated

Melting resistance contact measurement

Crucible rotation encoder crystal encoder rotation

Temperature pyrometer (e.g.

Increment function) Feed the crucible encoder

Crystal encoder feed

Dislocation image ejection / tung or

Structurally visual with manual loss

input

In the fuzzy controller 201, the target values of the process variables predicted by the process model are changed into test target values by means of a genetic algorithm parameter optimization. The calculation for a quality function g (x, y,... Z), a vector [x, y. . . z], is possible for individual variable sets that designate a searched extreme point.

Genetic algorithms represent their information in the form of bit strings, which are also called "chromosomes". For example, the bit string can be 66 bits long. The genetic algorithm is given an initial population consisting of approximately one hundred "chromosomes", the value of which is determined using a random generator. The chromosomes are then converted into real numbers and individually transferred to the trained neuro model. This calculates the values of the quality quantities, e.g. B. the crystal diameter, which would result from the combination of different parameters. These quality quantities are converted into a quality value, e.g. B. "relative fitness" converted. The greater the quality value, the greater the fitness of the chromosome. The genetic algorithm performs a selection, whereby two randomly chosen chromosomes compete against each other. There is a high probability that the better chromosome wins and is transferred to the target population while the other is discarded. In the next phase, the selected individuals are crossed and mutated. Selection of the mutation and crossing probability as well as the position of the "crossing over" in the chromosome are determined in a directed random generator. The process is repeated until the average quality value of the population is greater than a predetermined one Limit is.

After the predicted target values of the process variables have been changed to test target values in the fuzzy controller by means of the genetic algorithm parameter optimization, a process model is again generated from the test target values in 200, which defines the target values of the process variables. The setpoint process variables determined by the neural process model are subjected to an estimation error correction in the database. It is a cumulative error correction in which the weights are corrected with cumulative weight changes after each epoch.

The process data target values changed by the database estimation error correction are now passed to the robustness filter 203, in which the changed process variable target values are low-pass filtered. The robustness filter 203 in turn acts upon the PID controllers 204 to 206 with respectively assigned setpoints of the process variables.

Claims

claims

1. Device for growing crystals, in particular single crystals, in which the process data, for. B. melting temperature, gas pressure, melting level, gas flow and the like, detected and processed and adjusting devices, for. B. heaters, pumps and the like can be controlled via PID controllers, characterized by a device which overlays a PID controller (55; 204 to 206) with an artificial neural network (53, 202).

2. Device for growing crystals according to claim 1, characterized by a) a plurality of sensors (43 to 47), which the actual values of process data, for example the melting temperature, the gas pressure, the melting level, the gas flow and the like during the crystal growth process to capture; b) a programmable controller (17) to which the actual values detected by the sensors (43 to 47) are supplied; c) a neuro-fuzzy logic (52) which is connected to the programmable controller (17) via data lines (60 - 64 or 65, 66), d) a plurality of adjustable devices (9, 16, 19, 31, 32), e.g. B. heaters, pumps, drawing devices, rotary motors, valves and the like, which are controlled by PID controllers (55, 204 to 206), the setpoint control of the PID controller by the

Neuro-fuzzy logic takes place.

3. Device according to claim 1, characterized in that the PID controller (55, 204 to 206) is preceded by a robustness filter (203).

4. Device according to claim 1, characterized in that the PID controller is a P, PI, PD, etc. controller.

5. Device according to claim 2, characterized in that the neuro-fuzzy logic (52) is realized by the software of a computer (50).

6. A method for growing crystals, in particular single crystals, in which the process data, for. B. melting temperature, gas pressure, melting level, gas flow and the like, detected and processed and adjusting devices, for. B. heaters, pumps and the like can be controlled via PID controller, characterized by an input direction, the PID controller (55; 204 to 206) superimposed a control with an artificial neural network (53, 202).

7. The method according to claim 6, characterized in that the control of the PID controller (55, 204 to 206) is superimposed on a control of the fuzzy controller (54, 201).

8. The method according to claim 6, characterized in that the process data coming from a crystal pulling system (1) are fed via lines (33 to 37) to a neuromodel (202) which generates a neuronal model of a crystal type from these data.

9. The method according to claim 6, characterized in that the model formation is optimized by training.

10. The method according to claim 6, characterized in that the fuzzy controller

(201) constantly acts on the expected quality of the crystal on the basis of the predictions of the neuromodel (202).

11. The method according to claim 6, characterized in that a statistical model monitoring (202) is provided which checks the artificial neural network (202) with test data.

12. The method according to claim 11, characterized in that the material and analysis data (94) of the statistical model monitoring (202) are supplied.

13. The method according to claim 6, characterized in that the neuromodel

(202) provides quality data (QA (t + 1); QB (t + 1)) that are made available to the control process.

14. The method according to claim 6, characterized in that the influence of the sequences of different lengths such as stabilization, dip, neck, shoulder, body, end cone and shut down on the subsequent step, such as. B. Diameter overshoot during the transition from shoulder to body is corrected according to the model analysis by superimposed fuzzy controllers and robustness filters.

15. The method according to one or more of the preceding claims, characterized in that a correction of the model of the crystal type is initiated by the superimposed fuzzy control when a PLC (17) and / or a statistical model monitoring (200) via a robustness filter in the Intervene regulation.

16. A method for producing crystals, in particular single crystals, in which process variables, for. B. melting temperature or gas flow, as well as quality features, e.g. B. raw material purity, dimensions of a crucible wall, etc., recorded and processed and heaters, pumps and the like are controlled via PID controllers, characterized by the following steps: a) the process variables are recorded cyclically and as actual values and variable setpoints in a database ( 200) filed; b) the quality features are recorded once as actual values until the transition to automatic process control and are stored in the database (200) with fixed target values; c) the process variables are processed in the database (200); d) a neural process model is generated (in 202) from the process variables processed in the database (200); e) in a fuzzy controller (201), the target values of the process variables predicted by the process model are changed to test target values by means of a genetic algorithm parameter optimization; f) from the test setpoints of the process variables generated by the fuzzy controller (201), a process model is again generated (in 200) which defines the setpoints of the process variables; g) the setpoint process variables determined by the neural process model are subjected to an estimation error correction in the database (200); h) the process variables changed by the database estimation error architecture are placed on a robustness filter (203) in which the changed

Process variable setpoints are low-pass filtered; i) the robustness filter (203) acts upon PID controllers (204 to 206) with respectively assigned setpoints of the process variables.

17. The method according to claim 16, characterized in that the process variables in the database are processed using the SPC method, with a table after an algorithmic and statistical validation by a fuzzy cluster method compressed to typical representative data sets and used for training the neural network.

18. The method according to claim 16, characterized in that a local error assessment is carried out for each statement made by the neuromodel.

19. The method according to claim 16, characterized in that the selection of the algorithm and the configuration of the neural process model (in 202) is carried out automatically, controlled by an expert system.

20. The method according to claim 19, characterized in that it is determined which input variables are not relevant and that these input variables are removed, so that the neural process model only maps significant variables to one another.

21. The method according to claim 20, characterized in that a sensitivity analysis is carried out on the significant variables, whereby the strength is recognized with which each input variable of the model acts on each output variable that predicts the target values of the process variables or quality variables.

22. The method according to claim 16, characterized in that when changing the setpoints of the process variables in test setpoints, the calculation for a quality function g (x, y, z...) Is possible for individual sets of variables that designate a sought extreme point.

CHANGED REQUIREMENTS

[Received at the International Office on November 16, 1999 (November 16, 1999); original claims 1-22 replaced by amended claims 1-21 (4 pages)]

1. Device for growing crystals, in particular single crystals, in which the process data, for. B. melting temperature, gas pressure, melting level, gas flow and the like, detected and processed and adjusting devices, for. B. heaters, pumps and the like can be controlled via PID controllers, characterized by neuro-fuzzy logic (52; 201, 202), through which the setpoint control of the PID controller (55; 204 to 206) takes place.

2. Device for growing crystals, in particular single crystals, in which the process data, for. B. melting temperature, gas pressure, melting level, gas flow and the like, detected and processed and adjusting devices, for. B. heaters, pumps and the like can be controlled via PID controllers, characterized by a) a plurality of sensors (43 to 47) which record the actual values of process data, for example the melting temperature, the gas pressure, the melting level, the gas flow and detect the like during the crystal growth process; b) a programmable logic controller (17) to which the actual values detected by the sensors (43 to 47) are supplied; c) a neuro-fuzzy logic (52; 201, 202) which is connected to the programmable controller (17) via data lines (60-64 or 65, 66), d) a plurality of adjustable devices (9, 16, 19, 31, 32), e.g. B. heaters, pumps, drawing devices, rotary motors, valves and the like, which are controlled via PID controllers (55, 204 to 206), the setpoint control of the PID controllers (55; 204 to 206) by the neuro-fuzzy logic (52; 201, 202).

3. Device according to claim 1 or claim 2, characterized in that a robustness filter (203) is connected upstream of the PID controllers (55, 204 to 206).

4. Device according to claim 1 or claim 2, characterized in that the PID controller (55; 204 to 206) are P, PI, PD etc. controllers.

5. Device according to claim 1 or claim 2, characterized in that the neuro-fuzzy logic (52; 201, 202) is realized by the software of a computer (50).

6. A method for growing crystals, in particular single crystals, in which the process data, for. B. melting temperature, gas pressure, melting level, gas flow and the like, detected and processed and adjusting devices, for. B. heaters, pumps and the like can be controlled via PID controllers, characterized by neuro-fuzzy logic (52; 201, 202), through which the setpoint control of the PID controller (55; 204 to 206) takes place.

7. The method according to claim 6, characterized in that the process data coming from a crystal pulling system (1) are fed via lines (33 to 37) to an artificial neural network (202) which generates a neuronal model of a crystal type from these data.

8. The method according to claim 7, characterized in that the generation of the neural model of the crystal type is optimized by training.

9. The method according to claim 7, characterized in that a fuzzy controller

10. The method according to claim 6, characterized in that a statistical model monitoring (200) is provided, which checks the artificial neural network (202) with test data.

11. The method according to claim 10, characterized in that the statistical model monitoring (200) material and analysis data (94) are supplied.

12. The method according to claim 7, characterized in that the neural network

(202) outputs quality data (QA (t + 1); QB (t + 1)), which are fed to a fuzzy controller (201).

13. The method according to claim 6, characterized in that the influence of the sequences of different lengths such as stabilization, dip, neck, shoulder, body, end cone and shut down on the subsequent step, such as. B. Diameter overshoot during the transition from shoulder to body is corrected according to the model analysis by superimposed fuzzy controllers (201) and robustness filters (203).

14. The method according to one or more of the preceding claims, characterized in that a correction of the model of the crystal type is initiated by the fuzzy controller (201) when a PLC (17) and / or a statistical model monitoring (200) via a Intervene in the robustness filter (203) in the control.

15. A method for producing crystals, in particular single crystals, in which process variables, for. B. melting temperature or gas flow, as well as quality features, e.g. B. raw material purity, dimensions of a crucible wall, etc., recorded and processed and heaters, pumps and the like are controlled via PID controllers, characterized by the following steps: a) the process variables are recorded cyclically and as actual values and variable setpoints in a database ( 200) filed; b) the quality features are recorded once as actual values until the transition to automatic process control and are stored in the database (200) with fixed target values; c) the process variables are processed in the database (200); d) a neural process model is generated (in 202) from the process variables processed in the database (200); e) in a fuzzy controller (201), the target values of the process variables predicted by the process model are changed to test target values by means of a genetic algorithm parameter optimization; f) from the test setpoints of the process variables generated by the fuzzy controller (201), a process model is again generated (in 200) which defines the setpoints of the process variables; g) the setpoint process variables determined by the neural process model are subjected to an estimation error correction in the database (200); h) the process variables changed by the database estimation error correction are placed on a robustness filter (203) in which the changed

16. The method according to claim 15, characterized in that the process variables in the database (200) are processed using the SPC method, with a table after an algorithmic and statistical validation by a fuzzy clustering method. drive compressed to typical representative data sets and used for training the neural network.

17. The method according to claim 15, characterized in that a local error assessment is carried out for each statement made by the neuromodel.

18. The method according to claim 15, characterized in that the selection of the algorithm and the configuration of the neural process model (in 202) is carried out automatically, controlled by an expert system.

19. The method according to claim 18, characterized in that it is determined which input variables are not relevant and that these input variables are removed, so that the neural process model only maps significant variables to one another.

20. The method according to claim 19, characterized in that a sensitivity analysis is carried out on the significant variables, whereby the strength is recognized with which each input variable of the model acts on each output variable that predicts the target values of the process variables or quality variables.

21. The method according to claim 15, characterized in that when changing the setpoints of the process variables in test setpoints, the calculation for a quality function g (x, y, z...) Is possible for individual variable sets, which is a sought extreme point describe.