JPH06224126A - Film quality estimating device for semiconductor manufacturing device - Google Patents

Abstract

(57) [Summary] [Purpose] A small number of learning film formation process data is prepared for the second device by mixing film formation process data of two semiconductor manufacturing devices with the same specifications and learning them in the neural network. The film quality can be predicted only by [Structure] For the film formation data 4A, a large number of learning film formation parameter data 23 and film quality data 24 are prepared for the first (old) semiconductor manufacturing apparatus, but learning is performed for the second (new) apparatus. The number of preparations of the work data 23 and 24 is small.
In the present invention, the neural network 31 is made to learn the film formation data 4A in which the old and new identification data 21 of the apparatus are added to the learning data 23 and 24 of the old and new apparatus through the normalizing means 46. And the learned neural network 31 '
The new equipment identification data 21 and film formation parameter data 23
The predicted film quality data 47 obtained under the film forming condition 23 is output from the neural network 31 ′ by the new device.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for predicting film quality data obtained by a semiconductor manufacturing apparatus for forming a film on a semiconductor wafer from a film forming parameter as a film forming condition using a neural network. In the drawings below, the same reference numerals indicate the same or corresponding parts.

[0002]

2. Description of the Related Art As a method for predicting film quality data obtained by a semiconductor manufacturing apparatus of this type in the related art, Japanese Patent Application No. 3-299326 and Japanese Patent Application No. 4-80101, which are prior applications of the present applicant, are known.
As shown in the above issue, the neural network is made to learn the film forming parameters and the film quality data of the film forming result in the film forming performed in the past in the semiconductor manufacturing apparatus in advance, so that a new generation is made from the learned neural network. There is a method of predicting the film quality obtained by the semiconductor manufacturing apparatus in question based on the film parameter.

[0003]

However, even in the semiconductor manufacturing apparatus having the same specifications, the film quality data may be different even under the same film forming conditions due to the design change of details or the assembly variation or error of the structural parts due to cost reduction or the like. Many. Therefore, in the conventional technique, in order to predict the film quality from the film forming conditions, it is necessary to obtain a large number of film forming process data to be learned by the neural network for each semiconductor manufacturing apparatus, which is a problem. Therefore, an object of the present invention is to provide a film quality predicting apparatus for a semiconductor manufacturing apparatus that can solve this problem.

[0004]

In order to solve the above-mentioned problems, the film quality predicting apparatus according to claim 1 defines condition items such as the type and flow rate of the film forming gas, the film forming temperature, and the pressure of the film forming parameter ( 23) and the like, and the film formation result is evaluated by film quality data (24 etc.) including evaluation items such as film refractive index, residual stress and etching rate. Manufacturing equipment (7,
7 ', etc.) is used as the input data together with the device identification data (21, etc.) for identifying the semiconductor manufacturing device in which the film formation has been performed, respectively.
The film formation for making one or a plurality of neural networks (31 or the like) perform learning using the film quality data corresponding to the input data as teacher data, and for the learned neural networks (31 ′ or the like) to form a film By inputting the parameters and the apparatus identification data of the semiconductor manufacturing apparatus for performing this film formation, the film quality data of the film formation result corresponding to this input condition is predicted and output from this neural network.

[0005]

When a new semiconductor manufacturing apparatus (referred to as a new apparatus) is manufactured, the film formation process data obtained by the existing apparatus (referred to as an existing apparatus) having the same specifications as the new apparatus and the new apparatus are obtained. By allowing the neural network to learn by mixing with film formation process data, it becomes possible to predict the film quality data of the film formation results with the new device by preparing a small number of film formation process data for the new device. .

FIG. 1 is a functional block diagram of the present invention. That is, the film formation parameter data 23 and the device identification data 21 are input, and the film quality data 24 is to be output to a neural network (film quality prediction neural network) 31 in which film formation process data (completion of film formation process data obtained in advance from the existing device and the new device The learning is performed using the film parameter data 23 and the film quality data 24) and the device identification data 21 as learning data, and the learned neural network (film quality prediction neural network) 31 'is used to form the device identification data 21 of the semiconductor manufacturing device. By inputting the film formation parameter data 23 for film formation, the film quality data 24 for the film formation conditions 23 in the semiconductor manufacturing apparatus is predicted and output from this neural network 31 '.

As described above, in the present invention, the film forming process data obtained from the existing apparatus and the new apparatus are used for learning the neural network. In the learning of the neural network, the film formation parameter data 23 and the device identification data 21 indicating the difference between the devices are input to each neuron of the input layer, and the film quality data 24 is input to each neuron of the output layer. Or later,
A neuron that inputs the device identification data 21 is described as a device identification input neuron. The device identification data 21 is 1 for the new device and 0 for the existing device. Since the adjustment amount of the coupling strength between the device identification input neuron and the neurons in the middle layer during learning is proportional to the device identification data 21, the difference between the new device and the existing device is learned. Further, since the transmission path of the film formation parameter data 23 is the same in the new device and the existing device,
If there are many film forming process data of the existing apparatus, the relationship between each film forming parameter data 23 and each film quality data 24 can be learned even if the film forming process data of the new apparatus is small. Therefore, if there are many film quality process data of existing equipment,
By preparing a small number of film formation process data of the new device and learning both by mixing them, it is possible to create a neural network that predicts the film quality data 24 as a result of film formation by the new device.

[0008]

Embodiments of the present invention will be described below with reference to FIGS. FIG. 2 is a configuration diagram of a film quality prediction system of an ECR plasma CVD apparatus using a neural network as an embodiment of the present invention. ECR is an abbreviation for electron cyclotron resonance, and CVD is an abbreviation for chemical vapor deposition. In FIG. 2, a computer 1 is a computer that executes a film quality prediction system, and a keyboard 2, a CRT 3, an auxiliary storage device 4, a semiconductor manufacturing apparatus 7, and a semiconductor manufacturing apparatus 7'are connected to the computer 1.

The semiconductor manufacturing apparatus 7'is

It corresponds to the existing apparatus shown in "Function", and it is assumed that a large number of film forming experiments have already been carried out in this apparatus and there are many film forming process data.

Further, the semiconductor manufacturing apparatus 7 is

This is a semiconductor manufacturing apparatus that corresponds to the new apparatus shown in (1) and is newly manufactured with the same specifications as the semiconductor manufacturing apparatus 7 ', and is a semiconductor manufacturing apparatus to be predicted by the film quality prediction system. The auxiliary storage device 4 stores the film forming data 4A of the semiconductor manufacturing apparatus 7 and the semiconductor manufacturing apparatus 7 ', and network data 4B in which various data of the neural network in which the film forming data 4A is learned are recorded.

Further, a neural network board 5 as hardware for performing learning and recognition of the neural network at high speed is incorporated in the computer 1. The film quality predicting means 1A in the computer 1 is a main functional part of the computer 1 in the present invention. The film quality predicting means 1A predicts and outputs film quality data as a result of film formation by the semiconductor manufacturing apparatus 7 with respect to film formation parameter data input by an operator, using a neural network learned from the film formation data 4A. .

The semiconductor manufacturing apparatus 7 and the semiconductor manufacturing apparatus 7'are controlled by the semiconductor manufacturing apparatus control means 6 in the computer 1. FIG. 3 shows film formation data 4 of the silicon oxide film (SiO film) in the ECR plasma CVD apparatus.
The structure of A is shown. In the figure, the film formation data 4A is formed by arranging data consisting of a combination of the apparatus identification data 21, the substrate number 22, the film formation parameter data 23, and the film quality data 24 for the amount of the film formed in the past.

Each film formation data 4A is distinguished by a unique substrate number 22, and the film formation data can be uniquely searched by the substrate number 22. In addition, the device identification data 21
Indicates which of the semiconductor manufacturing apparatus 7 and the semiconductor manufacturing apparatus 7 ′ is the film formation data obtained by
1, if it is the film formation data obtained from the semiconductor manufacturing apparatus 7,
0 for film formation data obtained from the semiconductor manufacturing apparatus 7 '
I am trying.

The film forming parameter data 23 is E in this example.
Controllable of CR plasma CVD equipment ・ Oxygen (O ₂ ) gas flow rate 23A ・ Silane (SiH ₄ ) gas flow rate 23B ・ Reaction chamber pressure 23C ・ Microwave power 23D ・ Solenoid coil current 23E ・ Substrate applied AC Power 23F, film-forming time 23G, and seven items.

The film quality data 24 are the results of various film quality evaluations of the silicon oxide film formed under the conditions of the film formation parameter data 23. In this example, the refractive index is 24A, the residual stress is 24B, and the etching rate is 24C. , Film forming speed 24D, in-plane film thickness distribution 24E, film hardness 24F, and 6 items.

FIG. 4 shows the structure of the film quality prediction neural network 31 used in the film quality prediction system. The film quality prediction neural network 31 has a three-layer structure, and the numbers of neurons in the input layer 32, the intermediate layer 33, and the output layer 34 are 7, 20, and 3, respectively. The ECR plasma CVD apparatus according to the present invention has seven controllable film formation parameters as described in FIG. In the film quality prediction neural network 31, 6 items (23A to 23F) other than the film forming time 23G of the film forming parameter data 23 and the device identification data 21 are input items, and the refractive index 24A of the film quality data 24 and the remaining amount are used. The output items are stress 24B and etching rate 24C.

FIG. 5 shows a flow of creating a learning pattern 41 to be learned by the film quality prediction neural network 31 from the film formation data 4A. Here, the learning pattern (which is a combination of the input pattern presented on the input side of the neural network and the teacher pattern presented on the output side) is limited to 0 or more and 1 or less, so the film formation parameter data 23 and the film quality data 24 Is normalized to 0 or more and 1 or less and the film quality prediction neural network 31 is made to learn. In FIG. 5, the learning pattern 41 of the film quality prediction neural network 31 uses the device identification data 21 and the normalized film formation parameter data 43 as input patterns, and the normalized film quality data 44 as teacher patterns.

Further, in the normalization method 46, each item (23A to 23F) of the film formation parameter data 23 which is an input item of the film quality prediction neural network 31 and each item (24A to 24C) of the film quality data 24 which is an output item are selected. On the other hand, the minimum value of each item can be set to 0 by different methods,
The maximum value is normalized to 1, and normalized film formation parameter data 43 and normalized film quality data 44 are created.

The post-normalization film formation parameter data 43 is composed of up to 6 items of normalized oxygen gas flow rate, silane gas flow rate, reaction chamber pressure, microwave power, solenoid coil current, and substrate-applied AC power. In addition, the regular post-film quality data 44 is composed of three items of normalized refractive index, residual stress, and etching rate. Learning pattern 4
1 is configured as a collection of a plurality (n) of sets of apparatus identification data 21, normalized film formation parameter data 43, and normalized film quality data 44. Here, in the learning pattern, if all seven input pattern items have the same value, only one of them is selected as a learning parameter. This is because it is not possible to learn a plurality of different teacher patterns for the same input pattern. Next, the created learning pattern 41 is learned by the film quality prediction neural network 31. Here, the number of pieces of data learned by the film quality prediction neural network 31 is: 110 data of the semiconductor manufacturing apparatus 7 ′, 20 data of the semiconductor manufacturing apparatus 7.

In the learning of the film quality prediction neural network 31, the oxygen gas flow rate input neuron 32A of the input layer 32 is the normalized oxygen gas flow rate, and the silane gas flow rate input neuron 32B is the normalized silane gas flow rate. -The normalized reaction chamber pressure is input to the reaction chamber pressure input neuron 32C.-The microwave power input neuron 32
D is the normalized microwave power, · The solenoid coil current input neuron 32E is the normalized solenoid coil current, · Substrate applied AC power input neuron 32
The normalized substrate-applied AC power is presented to F, the device identification data 21 is presented to the device identification input neuron 32G, and the normalized refractive index is shown to the refractive index output neuron 34A of the output layer 34, .Output neuron 34B for residual stress
The residual stress after normalization is presented to the output neuron for etching rate, and the etching rate after normalization is presented to the output neuron 34C for etching rate.

Also, the film quality prediction neural network 3
1 is the data of the semiconductor manufacturing equipment 7 and the semiconductor manufacturing equipment 7 '.
The data of is mixed and learned. The device identification data 21 input to the device identification input neuron 32G is 1 for learning the data of the semiconductor manufacturing device 7 and 0 for learning the data of the semiconductor manufacturing device 7 ′. When the device identification data 21 of the semiconductor manufacturing device for film formation and the normalized film formation parameter data 43 are input to the film quality prediction neural network 31 ′ learned by the method described below, the device identification data 21 is obtained from this neural network. Is 0, the semiconductor manufacturing apparatus 7 ′ outputs the predicted film quality data 47 of the semiconductor manufacturing apparatus 7 as the output pattern when the apparatus identification data 21 is 1. If the output pattern 47 is converted in the opposite manner to the normalization method 46, the absolute value of the predicted film quality can be obtained.

FIG. 6 is a neural network 3 for predicting film quality.
1 shows the flow of learning. The film quality prediction neural network 31 is made to learn a plurality (n) sets of learning patterns 41 (learning patterns 41 (1) to 41 (n)).
The learning is started after giving a random number to the threshold and weight of the film quality prediction neural network 31 to initialize the learning. Here, the threshold is a parameter that represents the input / output characteristics of each neuron that constitutes the neural network, and the weight is a parameter that represents the strength of connection between neurons. And first, the film quality prediction neural network 31
The device identification data 21 (1) and the normalized film formation parameter data 43 (1) are input to the output pattern 51 (1).
(Consists of three items: refractive index, residual stress, etching rate). Next, this output pattern 51
An error 52 (1) (consisting of three items of refractive index, residual stress, and etching rate) which is the difference between (1) and the normalized film quality data 44 (1) is obtained. And the error 52 (1)
The sum of squares of each item is calculated to obtain an error sum of squares 53 (1). In this way, the error 2 of the learning pattern 41 (1)
When the sum of squares 53 (1) is obtained, the error sum of squares 53 is also obtained from the learning pattern 41 (2) to the learning pattern 41 (n).
The error sum of squares 53 (n) is obtained from (2). In this way, all learning patterns 41 (learning pattern 41 (1)
To learning pattern 41 (n)) error sum of squares 53
When the sum of error squares 53 (1) to the sum of error squares 53 (n) is obtained, the sum of error square sums 54 is obtained. When the error sum of squares summation 54 is obtained in this way, the error sum of squares summation 54
So that the film quality prediction neural network 3
The weight 1 and the threshold adjustment amount are calculated and adjusted.

As described above, the learning from the presentation of all the learning patterns to the adjustment of the weights and the threshold value is one time, and the film quality prediction neural network 31 is made to learn 50,000 times. Learned film quality prediction neural network 3
1'input layer 32 oxygen gas flow input neuron 3
2A is the normalized oxygen gas flow rate, and silane gas flow input neuron 32B is the normalized silane gas flow rate.
The normalized reaction chamber pressure is input to the reaction chamber pressure input neuron 32C, the normalized microwave power is input to the microwave power input neuron 32D, and the normalized solenoid is input to the solenoid coil current input neuron 32E. The coil current, the normalized substrate-applied AC power to the input neuron 32F for the substrate-applied AC power, and the 1 to the device-identification input neuron 32G are input to the output layer 34.
Of the film quality data of the refractive index from the refractive index output neuron 34A, the residual stress from the residual stress output neuron 34B, and the etching rate from the etching rate output neuron 34C in the semiconductor manufacturing apparatus 7. Output the predicted value. However, the output neuron 34A
The values output from 34C to 34C are values normalized from 0 to 1. Therefore, output neurons 34A to 3
The absolute value of the predicted film quality can be obtained by performing the reverse conversion of the value output from 4C to the normalization method 46.

The thresholds and weights obtained by learning are
Each learned neural network is recorded in the network data 4B as a network file. FIG. 7 shows the structure of the network data 4B. In the figure, network data 4B includes a network structure 62 that defines the number of layers and the number of neurons in each layer, which is the structure of a neural network, a threshold value 63 obtained by learning, a weight 64, and a plurality of learning patterns used for learning. 41
It is composed of. The learning pattern 41 includes the substrate number 22, the apparatus identification data 21, the normalized film formation parameter data 4
3 and the normalized film quality data 44.

FIG. 8 shows six parameters of film formation parameter data (oxygen gas flow rate, silane gas flow rate, reaction chamber pressure, microwave power, solenoid coil current, substrate application AC) using the learned film quality prediction neural network 31 '. The flowchart for operation | movement description of the film quality prediction system which predicts film quality data (refractive index, residual stress, etching rate) of three items from electric power is shown. The film quality prediction system includes an input data receiving unit 71, an input data determination unit 72, an input data normalization unit 73, a film quality prediction unit 75, and an output pattern conversion unit 7.
6 and a predicted film quality display unit 78.

The operator uses the film formation parameter data 23.
Enter. The film forming parameter data 23 is composed of six items of oxygen gas flow rate, silane gas flow rate, reaction chamber pressure, microwave power, solenoid coil current, and substrate applied AC power. The input data receiving unit 71 receives the input of the film formation parameter data 23 from the operator.
The input data determination unit 72 determines whether or not each item of the input film formation parameter data 23 is within the effective range of the present system. That is, if each item of the input film formation parameter data 23 is within the valid range of the present system, the process proceeds to the input data normalization section 73, and if even one item is outside the valid range, it is considered unpredictable. To finish. The input data normalization unit 73 normalizes the input film formation parameter data 23 by the same method as the normalization method 46, creates an input pattern 74, and proceeds to the film quality prediction unit 75.

The input pattern 74 is composed of six items of normalized oxygen gas flow rate, silane gas flow rate, reaction chamber pressure, microwave power, solenoid coil current, and substrate applied AC power. In the film quality prediction unit 75, the input pattern 7
4 and the device identification data 21 are input to the film quality prediction neural network 31 'to obtain an output pattern 51. At this time, the device identification data 21 is 1. Output pattern 51
Is composed of three items: refractive index, residual stress, and etching rate.

The output pattern conversion unit 76 outputs the output pattern 4
A pair of 1s is subjected to conversion reverse to that of the normalization method 46 to create predicted film quality data 77 that is the absolute value of the predicted film quality, and the process proceeds to the predicted film quality display unit 78. The predicted film quality data 77 is composed of three items: refractive index, residual stress, and etching rate. The predicted film quality display unit 78 displays the predicted film quality data 77 and ends the present system.

[0029]

According to the present invention, film forming process data obtained from a newly created semiconductor manufacturing apparatus (referred to as a new apparatus), an existing semiconductor manufacturing apparatus having the same specifications as the new apparatus (existing apparatus) Film forming process data obtained from
Since both of the two are mixed and the method of learning to the neural network is used, if a small number of film forming process data of the new device is prepared, the film quality data of the film formed by the new device can be predicted, The number of times of film formation and film evaluation for predicting film quality for each device can be reduced. Therefore, according to the present invention, it is possible to create a system that predicts film quality data as a result of film formation by a new apparatus under arbitrary film formation conditions at low cost and in a short time.

[Brief description of drawings]

FIG. 1 is a functional configuration diagram of the present invention.

FIG. 2 is a system configuration diagram as an embodiment of the present invention.

[Fig. 3] Similarly, a configuration diagram of film formation data

FIG. 4 is a diagram showing a structure of a film quality prediction neural network.

FIG. 5 is a diagram showing a flow of creating a learning pattern for making a neural network learn from film formation data.

FIG. 6 is a diagram showing a learning flow of the film quality prediction neural network.

FIG. 7 is a diagram showing the structure of a network file.

FIG. 8 is a flowchart for explaining the operation of the film quality prediction system.

[Explanation of symbols]

 1 Computer 1A Film Quality Predicting Means 2 Keyboard 3 CRT 4 Auxiliary Storage Device 4A Film Forming Data 4B Network Data 5 Neural Network Board 6 Semiconductor Manufacturing Equipment Controlling Means 7 Semiconductor Manufacturing Equipment 7'Semiconductor Manufacturing Equipment 21 Device Identification Data 23 Film Forming Parameter Data 24 Film quality data 31 Film quality prediction neural network 31 'Film quality prediction neural network 43 Normalized film forming parameter data 44 Normalized film quality data 46 Film forming process data normalizing means 47 Predicted film quality data (output pattern)

Claims (1)

[Claims] 1. A film is formed by using condition items such as a kind and flow rate of a film forming gas, a film forming temperature, a pressure, etc., and a film forming result is evaluated such as a film refractive index, a residual stress and an etching rate. In a semiconductor manufacturing apparatus evaluated by film quality data consisting of items, an apparatus for identifying the semiconductor manufacturing apparatus in which the film forming parameters of the past film forming results of two similar semiconductor manufacturing apparatuses are respectively formed. The identification data is used as input data, the film quality data corresponding to the input data is used as teacher data to cause one or a plurality of neural networks to perform learning, and the learned neural networks are provided with the film forming parameters for forming films from now on. By inputting the device identification data of the semiconductor manufacturing device that causes this film formation, Quality prediction apparatus for a semiconductor manufacturing device, characterized in that to predict output quality data for deposition results corresponding to the force conditions.