JPH0737821A - Thin film manufacturing device and manufacture of semiconductor device - Google Patents

Abstract

PURPOSE:To provide a thin film manufacturing device capable of cutting down the fundamental experiments based on the trials and errors by a method wherein nonlinear separation and prediction power of a neural network are used thereby enabling the film forming process operation control amount to be predicted by the whole or a part of film forming specifications. CONSTITUTION:The kinds of reaction gas inputted by a film forming specification inputting means 1 and the kinds of deposited substrate are analyzed by an electronic structural property analyzer neutral network 70 so as to predict chemical reaction formula, reaction rate formula, etc. Next, the film forming process operation control amount corresponding to the film forming specifications is predicted according to these data by a fluid reaction engineering analyzer. Through these procedures, the film forming process operation control amount of not only well known reaction gas group but also new reaction gas group can be predicted thereby enabling the waste of time and resources based on the trials and errors to be notably cut down.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a thin film manufacturing apparatus and a method for controlling the thin film manufacturing apparatus, which facilitates setting of manufacturing conditions for a known film forming system, and which makes it possible to predict the operation control amount of the film forming process. The present invention relates to a method for manufacturing a semiconductor device.

[0002]

2. Description of the Related Art In a conventional thin film manufacturing apparatus, a thin film satisfying desired film forming specifications (designated thin film deposition on a designated substrate by designated reaction gas species, growth rate, growth film thickness, selectivity, film quality) is obtained. Therefore, it was necessary to find the optimum conditions by trial and error by changing the reaction chamber experimental conditions such as the substrate temperature, the reaction gas introduction pressure, and the mixing ratio in a matrix manner. This is because it is difficult to find the optimum experimental condition by linear prediction because most of the film formation specifications are complicated nonlinear functions of the reaction chamber experimental condition.

Further, in the case of a film forming process whose operation history is not known at all, for example, when a new reaction gas or a new deposition substrate is used, the repetition of trial and error is remarkably increased as compared with the known film forming system, and the process is started up. However, there was a problem that it lasted for a long time.

Therefore, by numerical simulation,
An attempt to analyze the pressure distribution, temperature distribution, and reaction gas concentration in the reaction chamber by a fluid reaction engineering method, predict the film formation rate and film formation shape in advance, reduce the number of trials and errors, and learn the limit performance. Has come to be done. For example, Journal of Electrochemical Society [J. Ele
ctrochem. Soc., 138 , 1728 (1991)], a simulation of tungsten metal thin film formation by monosilane reduction has been reported. In these film formation simulations, the reaction rate, diffusion, and heat transfer with the boundary conditions such as the partial pressure of the reaction gas at the gas inlet and the temperature on the deposition substrate are solved to predict the growth rate on the deposition surface and It is possible to predict the film-formation profile distribution from the position dependence.

Further, the details of the surface chemical reaction are analyzed by an electronic state analysis method such as a molecular orbital method, and an attempt is made to obtain a chemical reaction formula and a clue of a reaction rate by finding a growth reaction elementary process. Thin film formation simulation is aimed at. For example, a reaction elementary process analysis by a molecular orbital method of silicon epitaxial growth using dichlorosilane as a source gas is reported in Applied Physics, Vol. 57, page 1022. In these reaction elementary process analysis, prediction of reactive species that can be generated from the source gas, reaction between generated reactive species and surface attachment process are analyzed by the molecular orbital method, and attempts are made to predict dissociation energy and adsorption energy.

[0006]

However, in such a thin film formation simulation, reaction formulas of chemical reactions involved in film formation and growth reaction rate formulas are assumed or experimentally determined. It still has a problem that basic experiments based on many trials and errors are necessary. Therefore, when a new reaction gas or a deposition substrate is used, it is necessary to repeat the basic experiment.

Further, in a computer simulation for obtaining a clue of a reaction equation or a reaction rate equation relating to a chemical reaction by an electronic state analysis method such as a molecular orbital method, the number of atoms such as silylene is extremely limited due to the limitation of the calculation time and the computer memory capacity. There is a problem that only limited reaction species can be analyzed, and it is not possible to analyze the reaction between reaction intermediates that are considered to be generated from the reaction gas used in the general thin film forming process and the adhesion process including solid surface. .

An object of the present invention is to provide a thin film manufacturing apparatus capable of predicting a film formation process operation control amount from all or a part of a film formation specification and reducing a basic experiment work amount based on trial and error. Is.

Another object of the present invention is to provide a control method of a thin film manufacturing apparatus capable of reducing the calculation time and the computer memory capacity required for searching a chemical reaction route for obtaining a clue of a reaction equation or a reaction rate equation relating to a chemical reaction. To provide.

Still another object of the present invention is to provide a method of manufacturing a semiconductor device having a thin film forming step capable of easily forming a thin film having high uniformity and reliability.

[0011]

[Means for Solving the Problems] The above-mentioned object is tried by using a neural network for a prediction procedure in which all or a part of a film formation specification is a primary input and a film formation process operation control amount is a final output. This can be achieved by a thin-film manufacturing system that can reduce the number of basic experiments based on errors and can predict new reaction gases.

Further, the above object is to provide the neural network with a database access function and a learning function using the database information as a teacher pattern, and have a film formation process operation control amount prediction function by the neural network associative function. This can be achieved more effectively by the thin film manufacturing apparatus.

The above object is to input the molecular structure and crystal structure information of the reactive gas species and / or the deposition substrate species, which are part of the film formation specifications, and to input the electron orbital structure of the molecule and the reactivity index. , Bond order, film formation reaction order, film formation activation energy, surface superlattice structure, electronic structure consisting of crystal surface band structure, and an electronic structure physical property analyzer for outputting all or part of physical property information are provided. It is equipped with a neural network that inputs all or part of the output information from the structural property analyzer and part of other film formation specifications, and outputs the film formation process operation control amount as input. This can be achieved more effectively by the thin film manufacturing apparatus.

The above object is to input the molecular structure and crystal structure information of the reactive gas species and / or the deposition substrate species, which are part of the film formation specifications, and to input the electron orbital structure of the molecule and the reactivity index. A bond order, a surface superlattice structure, an electronic structure consisting of a crystal surface band structure, and an electronic structure physical property analyzer that outputs all or a part of the physical property information are provided, and all the output information from the electronic structure physical property analyzer or A neural network is provided, which takes as an input a part of the chemical reaction formula for thin film growth, a reaction formula for thin film growth reaction, and / or a thin film growth activation energy, and outputs the neural network, reaction gas introduction pressure, and chamber temperature. ,
Input all or any combination of substrate temperatures, and predict the film formation process operation control amount by fluid reaction engineering means that analyzes all or any combination of pressure distribution in chamber, temperature distribution, and thin film growth rate on substrate surface. Can be more effectively achieved by the thin film manufacturing apparatus characterized by the above.

Another object is to limit the degree of freedom of atomic arrangement of the analysis system when deriving a chemical reaction formula or reaction rate formula required for the film formation process operation control amount prediction procedure using a neural network. This can be achieved by a method for controlling a thin film manufacturing apparatus including a chemical reaction route searching step that enables reduction in calculation time and computer memory capacity.

Still another object is to use a predicted film forming process operation control amount obtained by using the neural network in an electronic structure physical property analyzing apparatus including a molecular orbital method having the chemical reaction path searching method. This can be achieved by a method for manufacturing a semiconductor device including a step of forming a thin film by using

[0017]

According to the present invention, when a neural network is used as the predicting means in predicting the film formation process operation control amount that satisfies the film forming specifications, the nonlinear response function of each neuron forming the neural network and the connection weight between neurons are used. Can be adjusted to reconstruct an arbitrary nonlinear response, so that the film formation process operation control amount can be predicted even if the complex nonlinear relation between the film formation process operation control amount and the film formation specification is not known analytically.

Further, in searching for a chemical reaction path, the calculation time and the computer memory capacity can be reduced by limiting the degree of freedom of atomic arrangement of the analysis system.

Further, by incorporating this chemical reaction path search method into a part of the film formation process operation control amount prediction using the above neural network, the chemical reaction formula and the reaction can be applied to the new reaction gas and the deposition substrate surface. The speed formula can be predicted at high speed, and the operation control amount prediction can be made highly efficient.

[0020]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is an overall configuration diagram of an embodiment of a thin film manufacturing apparatus according to the present invention. The thin film manufacturing apparatus includes a reaction chamber / control unit 1000 and a control amount prediction unit 2 that are directly involved in film formation.
It is divided into two parts of 000. First, the reaction chamber
The control unit 1000 includes a reaction chamber 10 and a reaction chamber 1.
0 in a high vacuum state, a vacuum evacuation unit 40, a gas introduction unit 20 that introduces a thin film forming reaction gas into the reaction chamber 10, a substrate introduction holding unit 30 that introduces and holds a thin film deposition substrate in the reaction chamber 10, Analytical means 50 for qualitatively or quantitatively analyzing temperature distribution, gas pressure distribution, reaction product gas species, deposition substrate surface temperature, surface adsorption species, etc. in the reaction chamber 10, and temperature and total pressure in the reaction chamber 10. , Control means 60 for controlling the temperature of the deposition substrate, the mixing ratio of the introduced reaction gas and the like. The output of the analysis means 50 is input to the database 80 and stored as experimental data information.
Next, the controlled variable predicting unit 2000 causes the neural network 70, the type of deposited substrate, the type of reaction gas used, the type of target deposited thin film, the thin film deposition rate, the thin film mechanical properties such as residual stress, to the neural network 70, A film formation specification input means 1 for inputting a thin film formation specification composed of thin film electrical properties such as electric conductivity and insulation to the neural network 70, a keyboard 2 used for inputting and displaying the film formation specification, a CRT (Cathode) ray tube) 3, the neural network 70 accesses the database 80, the database access means 90 required when the database information is used, and the film formation process operation control amount output by the neural network 70. The process operation control amount display unit 300 is provided. The film formation specification input means 1 used here has a function of directly inputting the film formation specification input from the keyboard 2 into the neural network 70 and a function of converting the input film formation specification into other data and inputting it into the neural network 70. is doing. For example, a polysilicon thin film to which phosphorus element is added is mixed with disilane and phosphine mixed gas to form C
When it is formed by the VD method, it has a function of converting the electrical conductivity of the film-forming specifications into the amount of phosphorus added in the film, a partial pressure ratio of the flow rate ratio, and a molar ratio. In the present embodiment, the keyboard 2 is used to input data, but the input means is not limited to the keyboard.

Now, the database 80 in this embodiment.
Maintains operational history and thin film performance of past thin film formation processes. The data structure is such that the setting input of the film formation specification, the temperature distribution in the reaction chamber 10 during the film formation process measured by the analysis means 50, the gas pressure distribution, the reaction product gas species, the deposition substrate surface temperature, and the surface adsorption species. Etc. qualitative and quantitative analysis result, temperature in reaction chamber 10, total pressure, deposition substrate temperature, operation history of control means 60 with setting and mixing ratio of introduced reaction gas, and thin film machine analyzed after film formation. It is made up of analysis results of electrical properties, film forming speed, film forming shape, etc., and is stored and stored in a relational database format, and the database access means 90 enables access from the neural network 70.

Next, the operation of the thin film manufacturing apparatus according to this embodiment will be specifically described. The experimenter is required to form the target thin film, specifically, the type of deposition substrate, the type of reaction gas used, the type of target deposition thin film, the thin film deposition rate, the mechanical properties of the thin film such as residual stress, the electrical conductivity, A thin film electrical property such as insulation is input from the keyboard 2. These input specifications are converted into inputs of the neural network 70 by the film formation specification input means 1. The neural network 70 outputs a film formation process operation control amount satisfying the film formation specification from this input. The film formation process operation control amount is displayed on the film formation process operation control amount display unit 300 and is also input to the control means 60 to start the operation of the film formation process.

Next, the structure and operation of the neural network 70 will be described. The neural network 70 uses the neuron model shown in FIG.
It is composed of a multi-layered neural network arranged hierarchically as shown in FIG. The neuron model, which is the basic unit, outputs a single output {Y} from multiple inputs {xj}. Its output {Y} is input {xj} to each weight {w}.
The product sum X (X = Σxj · wj) multiplied by j} is converted by the non-linear function f. Normally, the range of the input {xj} and the output {Y} is limited to the [0,1] range. The non-linear operation of the function f determines the non-linearity of the entire neural network. In the present embodiment, the sigmoid function (f = 1 / (1 + ex
p (-X)) is used. As the non-linear response function, even if a Macarock-Pitts model, a Boltzmann machine model, a chaos model, or the like is used in addition to the sigmoid function model, there is no problem in the basic operation. In the hierarchical neural network shown in FIG. 3, continuous two layers such as an input layer neuron and an intermediate layer neuron and an intermediate layer neuron and an output layer neuron are connected by a weight {wj} between neurons. The output of each stage is not fed back to the input side. It is well known that a hierarchical neural network configured as shown in FIG. 3 using a neuron model having such an input / output function has learning and associative functions. This point is described in detail in, for example, documents such as Kaoru Nakano's supervision, "Neurocomputer", Technical Review Co., Ltd., published in September 1989. In the present invention, this learning and associative function is positively used for predicting the film formation process operation control amount. Further, in the present invention, the neural network including three layers of the input layer, the intermediate layer and the output layer is used, but even if the number of layers of the intermediate layer is further increased, there is no problem in the learning and the association function. However, in order to express non-linearity, it is necessary to reproduce both positive output and negative output responses to the input, so the number of neurons included in the intermediate layer is the number of neurons included in the input layer. You need at least twice the number. Next, the learning and association functions will be described in some detail in accordance with the present invention. First, enter the manufacturing thin film specifications related to the past thin film forming process {x
j}, the neural network 70 is trained to output the operation history of the thin film manufacturing process.
This input / output data is stored in the database 80 and can be used as the input / output data of the neural network 70 by the database access means 90. Further, as described above, since the input / output value range of the neuron is usually limited to the [0,1] range, the film forming rate, residual stress and other thin film mechanical properties, electrical conductivity, insulation, etc. It is necessary to standardize the thin film manufacturing process operation history data such as the manufacturing thin film specification and temperature, total pressure, deposition substrate temperature, and introduced reaction gas mixture ratio within the above range in advance. This is solved by giving the film formation specification input means 1 and the analysis means 50 a standardization function. In particular,
The maximum values of these various quantities are set in advance, and each input specification is divided by these maximum values. Now, the learning function is to input the manufacturing thin film specifications {xj} (Xa, X in FIG. 3).
The output {Yi} of the neural network 70 (corresponding to b, Xc) (corresponding to Y ₁ , Y ₂ , Y ₃ in FIG. 3) is the thin film manufacturing process operation history data {Ti} (FIG. 3).
Is achieved by changing the coupling coefficient {wj} connecting the inter-layer neurons until they match (corresponding to T ₁ , T ₂ , and T ₃ ). {Ti} is a film forming specification relating to the target thin film manufacturing process, and can be selected and set from the database 80 via the database access means 90. In FIG. 3, the driving history data {Ti} that is the target value is described as a teacher signal. Specific coupling coefficient {wj}
With regard to the initial setting method and the changing method of, in this embodiment, the optimal solution of the coupling coefficient {wj} is obtained by using the backpropagation method by Ramelhart et al. The basis is {T
i} and {Yi} squared error sum is defined as an evaluation function,
The coupling coefficient {wj} is modified so that the evaluation function decreases. About this, Ramelhart et al., Nature, Volume 323 (Rumelhart. DE, et al, Nature, vo
l.323). In this embodiment, 1% of the teacher pattern is used as an error limit for learning. The associative function of the learned neural network is output calculation for new input, and its meaning is clear. With the learning function and the association function of the neural network described above, the operation control amount of the thin film manufacturing process can be predicted from the thin film specification input. For thin film manufacturing processes with operating history, without using such learning and associative functions of neural networks,
It is easy to think that the operation control amount of the thin film manufacturing process can be easily predicted from the past operation history based on the input of thin film specifications.However, the thin film deposition process includes various factors such as temperature, total pressure, deposition substrate temperature, and introduced reaction gas mixture ratio. It becomes a complicated non-linear function of the reaction control amount, and its analytical expression is unknown, and it is almost impossible in some cases to perform simple estimation by a proportional calculation formula or the like.
Therefore, the nonlinear prediction by the neural network can be effectively used. Next, an example in which the operating amount of the tungsten thin film growth process using the (monosilane-WF ₆ ) system is controlled using the thin film manufacturing apparatus shown in FIG. 1 will be described. As the reaction chamber 10, a cold wall single wafer type reaction vessel was used. This thin film growth system is a non-linear system in which two types of growth reaction modes exist depending on the ratio of the monosilane flow rate to the WF ₆ flow rate. When the monosilane / WF ₆ partial pressure ratio is 1.5 or less, the gas phase reaction hardly occurs, the W thin film growth selectively occurs only on the W film, and H ₂ and SiHF ₃ are generated as a product gas with the growth. . On the other hand, when the monosilane / WF ₆ partial pressure ratio is 1.5 or more, a vapor phase reaction also occurs, W thin film growth occurs not only on the W film, and H ₂ and SiF ₄ are generated as a product gas along with the growth. The physical properties of the grown thin film change significantly depending on the growth reaction mode. For example, the crystallinity of the grown film also changes to alpha phase, beta phase, mixed crystal of both, and WSi ₂ depending on the growth conditions. For this reason, a skilled person having experience and intuition based on a large number of preliminary experiments is required to accurately set the thin film growth conditions that satisfy the film forming specifications. Therefore, the neural network 7 is configured so that the electrical resistance value, the step coverage, the residual stress value, and the Si content in the film are input, and the growth temperature, the reaction gas flow rate ratio, and the carrier gas flow rate are output.
Configured 0. Past film forming experiment data (item data that becomes input and output of the neural network 70) is stored in the database 80 as a relational database, and the relationship between the film forming specifications and the operation control amount is stored in the neural network 70 by the back propagation method. I learned it. As a result, for example, to obtain an electric resistance value of 200 micro ohm centimeters, a Si content of the film of 2.0 atomic%, and a residual stress of 2 × ^{10 10} dyn / cm ² , monosilane, WF ₆ , and N as a carrier gas are used. ₂ , Ar partial gas pressure of 80, 40, 400, 150 mTorr,
It was predicted by the neural network 70 that the film should be formed at the film forming temperature of 236 degrees. After these condition values are displayed on the film formation process operation control amount display unit 300,
The gas was transferred to the control means 60 and the gas partial pressure and temperature were controlled.
When the W thin film obtained as a result was analyzed by the analysis means 50, it was found that the specifications were almost completely satisfied, and even a non-skilled person could easily set the film formation conditions that satisfied the film formation specifications. It was Furthermore, since there is a critical film forming specification that makes linear prediction such as interpolation and extrapolation impossible for a non-linear reaction system whose reaction mode changes depending on experimental conditions, the thin film manufacturing apparatus can be effectively used. From the above, it was found that the workability of the thin film manufacturing process can be greatly improved because the ability to set the film forming conditions is extremely high and it is effective for a wide range of thin film growth systems.

Next, FIG. 4 shows an overall configuration diagram of another embodiment of the thin film manufacturing apparatus according to the present invention. This has a configuration in which an electronic structure property analyzer 100 is newly added as compared with the previous embodiment, and it is possible to predict the film formation process operation control amount even for a new gas and a new deposition substrate. Therefore, the operations of the film formation specification input means 1, the neural network 70, the database 80, etc. are also different from those of the embodiment shown in FIG. Hereinafter, the operation of the thin film manufacturing apparatus for a new gas and a new deposition substrate will be sequentially described. First, the film formation specification input means 1 inputs both the reactive gas species, the deposition substrate, and / or the molecular structure and crystal structure of either one of them. For example, to input a monosilane molecule, the chemical formula SiH ₄ may be input from the keyboard 2. For the crystal structure, enter the crystal name and crystal plane. For example, enter W-bcc crystal, (110) plane. Although the example of the monosilane molecule and the W crystal has been described here, it goes without saying that the same can be applied to other molecules and crystal structures.
The electronic structure physical property analyzer 100 is composed of a molecular force field method, a molecular orbital method, an electron density functional method, a molecular dynamics method, etc.,
Regarding material gas, electron orbital structure of molecule, reactivity index,
Information such as bond order and stable structure, information such as surface superlattice structure and crystal surface band structure for the deposited crystal substrate, and the order of the film formation reaction or the activation energy for the film formation reaction itself Is output as an input to the neural network 70. For example, in the case of the monosilane molecule shown above as an input example, the occupied electron orbital energy levels (-68.7, -6.1, -4.2,-)
0.73: the unit is an atomic unit) and its degeneracy, Mulliken charge (silicon 13.43, hydrogen 1.14: the unit is an elementary charge), etc. are output. In addition, W- input as a deposition substrate
For the input of the bcc crystal (110) plane, the charge density distribution on the atom of the W ₉ cluster modeling this (−0.37 to
+0.19: unit is electric charge), state density distribution (DOS) from -25 eV to vacuum state, Fermi energy,
It outputs various quantities such as (1x2) and (2x2) superstructured lattices. As for the film-forming reaction, W generated by SiH ₄ and WF ₆ in on reaction heat 1.08eV and W ₉ side of monosilane dissociative adsorption reaction on W ₉ surface _{(SiH 4 + W 9 → SiH} 3 + H + W 9) The reaction intermediate formation energy (equivalent to activation energy) of 1.4 eV or the like is output. Then, in the database 80, regarding the thin film growth process accumulated and collected from the thin film manufacturing process experiments and literatures so far, both the reactive gas species and / or the deposition substrate, or the electronic structure and physical property information and the process operation control amount. The physical properties of the grown thin film are recorded and stored in the form of a relational database including the names of reactive gas species and the names of deposited substrate species. The neural network 70 learns, through the database access means 90, an electronic structure of the reaction gas and the deposition substrate, and a process operation control amount output function for inputting physical property information, and immediately receives an input from the electronic structure physical property analyzer 100 to form a film. The process operation control amount can be output, and the film formation process starts operation. As the neural network used in this example, a hierarchical neural network was used as in the case of the thin film manufacturing apparatus shown in FIG. 1, and learning was also performed by the back propagation method. In the currently available electronic structural property analyzers, it is often difficult to directly calculate the order of the film formation reaction and the activation energy with respect to the film formation reaction itself from the viewpoint of calculation time and computer memory capacity. Many days on leek are not preserved. However, since the film forming process itself is determined by the combination of the electronic structure, the physical properties, the temperature, the pressure, etc. of each of the reaction gas and the deposition substrate, by arranging these in a matrix form at the input of the neural network 70, It is possible to predict the operating process control amount without any problem.

Next, FIG. 5 shows an overall view of another embodiment of the thin film manufacturing apparatus according to the present invention. Compared with the embodiment shown in FIG. 1, this has a configuration in which a fluid reaction engineering analyzer 200 is newly added, and the film formation process operation control amount prediction in consideration of the improvement of the uniformity of the film formation process in the reaction chamber 10 is also made. Is possible. Therefore, the film formation specification input means 1,
The operations of the neural network 70, the database 80, etc. are also different from those of the embodiment shown in FIGS. The operation of the thin film manufacturing apparatus will be described below. The film formation specification input means 1 inputs the molecular structure and / or crystal structure of the reactive gas species and / or the deposition substrate as one of inputs to the neural network 70. Regarding the input method, the film formation specification input means 1 shown in the previous embodiment also has a function of extracting the input molecular structure, molecular symmetry from the crystal structure, crystal symmetry, etc.
These can also be inputs to the neural network 70. The neural network 70 outputs a thin film growth chemical reaction formula, a thin film growth reaction formula, or both depending on the combination of the reaction gas and the deposition substrate. For example, monosilane on W crystal (110) plane, W by WF ₆
For thin film growth, the film formation chemical reaction formula, WF ₆ + 2SiH ₄ = W +
Output 2SiHF ₃ + 3H ₂ . Also, the reaction rate formula R∝C
Is output. Where C represents the monosilane concentration,
It can be seen that the film formation reaction proceeds in the first order. The thin film growth chemical reaction formula here is an overall chemical reaction formula,
It is not a reaction equation showing the individual elementary processes actually occurring on the deposition surface or in the reaction chamber. Similarly, the thin-film growth reaction rate equation mentioned here is a rate equation corresponding to the above-described thin-film growth chemical reaction equation, specifically, an equation showing concentration dependence. The database 80 associates the reaction gas species, the molecular structure of the deposition substrate, the crystal structure with the thin film growth chemical reaction formula and the thin film growth reaction rate formula with respect to the experiments and the thin film growth processes accumulated and collected from the literature. Recorded and saved. For example, regarding the polysilicon growth reaction by thermal decomposition of monosilane, the growth temperature is 620 ° C. and the pressure is 0.
3 Torr, gas flow rate range 70-100sccm, deposition rate 1
0 to 12 nm / min, thin film growth chemical reaction formula SiH ₄ =
Information such as Si + 2H ₂ and thin film growth reaction rate equation R∝C is stored. Here, C is the monosilane concentration. Since the growth of polysilicon is hardly affected by the underlying substrate, no information about the underlying substrate is given. The neural network 70 is a database access means 90.
Through the reaction gas and the molecular structure and the crystal structure of the deposition substrate, and has already learned to output the thin film growth chemical reaction equation and / or the thin film growth reaction rate equation. Reaction engineering analyzer 20
Output as 0 input. Neural network 70
The learning function of was carried out by the back propagation method as in the previous embodiment. Now, the fluid reaction engineering analysis device 20
0 is composed of a mass transport equation solving system in which a three-dimensional diffusion equation and a chemical reaction rate equation are combined. Since the chemical reaction formula or the chemical reaction rate formula is given, the reaction gas consumption amount and the film forming rate accompanying the reaction progress can be analyzed at each position in the reaction chamber 10. By inputting the shape, size, and symmetry of the reaction chamber 10 as a line drawing on the CRT screen using the keyboard 2 and the CRT 3, the boundary condition position of the above transport equation is determined, and the reaction gas supplied by the gas introduction means 20 is determined. The boundary value at the boundary condition position is determined by the introduction pressure and the chamber temperature set by the control means 10, and the thin film formation simulation on the substrate becomes possible. In this embodiment, the shape of the reaction chamber 10 is the keyboard 2 and the CRT.
Although the input was made by 3, it is needless to say that the input can be made by using a mouse, a light pen or the like. From this, the uniformity of the film forming process in the chamber can be evaluated.
Specifically, the uniformity of the film forming rate on the deposition substrate may be the film forming process uniformity. On the contrary, it is not a difficult problem to use this film formation process uniformity as an evaluation function and obtain the film formation process operation control amount such as the reaction gas introduction pressure, the substrate temperature, and the vacuum exhaust speed that minimizes the evaluation function. The optimum controlled variable can be set by the steepest descent method using the first derivative obtained by the sensitivity analysis for each controlled variable. In particular,
The results used for optimizing the nitrogen carrier gas flow rate of the gas head static single-wafer CVD apparatus will be described. The gas head static single-wafer CVD apparatus has a reaction chamber structure having a width of 15 cm, a depth of 18 cm, and a height of 15 cm, and a 5-inch wafer is loaded in the horizontal plane at the center of the chamber. Further, the structure is such that the reactive gas is blown from the vertical direction from the direction perpendicular to the wafer surface. Separately from this, a nitrogen gas supply head provided on the side wall of the reaction chamber introduces a nitrogen carrier gas in a horizontal direction parallel to the wafer surface. Further, the reaction chamber is exhausted from an exhaust port provided on the opposite side of the nitrogen gas supply head. In this example, a doped polysilicon film forming reaction was performed using monosilane, oxygen, and phosphine as supply gases. Here, a fluid reaction engineering analysis device 200 is used as a reaction chamber region having a height of 4 mm over the upper surface of the wafer.
Typed in. Further, the flow rates of the introduced reaction gas and the nitrogen carrier gas were input to the fluid reaction engineering analysis device 200 as boundary values. The boundary value is monosilane 0.52 l / min (20%
Nitrogen dilution), oxygen 2.45 l / min, phosphine 1.83
Enter the value of l / min (1% nitrogen dilution), enter the value of nitrogen gas flow rate in the range of 0 to 40 l / min, and use the fluid reaction engineering analyzer 200 to determine the reaction gas flow on the wafer surface. The film velocity distribution was analyzed and the carrier gas amount was optimized. The results analyzed by the fluid reaction engineering analyzer 200 are as follows. First, if the nitrogen gas flow rate is zero, 500 nm / min.
Although the film formation rate distribution was in the range of 0 nm / min, since the nitrogen carrier gas flow rate was zero, a large number of stagnation points were observed in the gas flow within the wafer surface. Nitrogen carrier gas flow rate is sequentially 40
When increasing to l / min, the deposition rate distribution becomes 10 in sequence.
Although it decreased to 0 nm / min, the stagnation related to the gas flow was improved, and at a flow rate of 40 l / min, it was improved to almost a laminar flow state. From the point of gas flow, the carrier gas flow rate is 40
It can be said that the l / min state is optimal, but due to the decrease in film formation rate,
Since the throughput was reduced, it was concluded that a carrier gas flow rate of about 10 l / min was optimal for optimizing the entire apparatus.

Next, FIG. 6 shows an overall view of another embodiment of the thin film manufacturing apparatus according to the present invention. The neural network 70 receives the output of the electronic structure physical property analyzer 100 as an input, and inputs the output to the fluid reaction engineering analyzer 200. The final film-forming process operation control amount is set by the fluid reaction engineering analysis device 200. The electronic structure physical property analysis device 100 and the fluid reaction engineering analysis device 200 are respectively shown in FIG.
The same operation as that shown in FIG. 5 is performed. The operation of the entire thin film manufacturing apparatus will be described below. First, the film formation specification input means 1 inputs the molecular structure and / or crystal structure of both the reactive gas species and the deposition substrate species, which are part of the film formation specifications. The electronic structure physical property analyzer 100 is a molecular force field method,
It is composed of molecular orbital method, electron density functional method, molecular dynamics method, etc., and information about the electron orbital structure of molecules, reactivity index, bond order, etc. about the material gas, and surface superlattice for the deposited crystal substrate. Information such as structure and crystal surface band structure
It is output as an input to the neural network 70. The database 80 contains information on the molecular orbital structure of the reactive gas species, the reactivity index, the bond order, etc. regarding the thin film growth process accumulated and collected from the thin film manufacturing process experiments and literatures so far, and the surface superposition related to the deposited substrate. Information such as a lattice structure and a crystal surface band structure, a thin film growth chemical reaction formula, a thin film growth chemical reaction rate formula, and a thin film growth activation energy are associated and recorded and stored. The neural network 70 learns, through the database access means 90, output functions of these electronic structures relating to the reaction gas and the deposition substrate, chemical reaction formulas for inputting physical property information, chemical reaction rate formulas, activation energy, or any combination thereof. Therefore, upon receiving an input from the electronic structure physical property analyzer 100, these chemical reaction information can be output immediately.
The neural network used in this embodiment uses a hierarchical neural network as in the case of the thin film manufacturing apparatus shown in FIG. 1, and performs learning by the back propagation method. All or any combination of chemical reaction formulas, chemical reaction velocity formulas, and activation energies output by the neural network 70 is the fluid reaction engineering device 20.
It becomes 0 input. The fluid reaction engineering device 200 is composed of a mass transport equation solving system in which a three-dimensional diffusion equation and a chemical reaction rate equation are combined, as in the previous embodiment. Since the chemical reaction formula and the chemical reaction rate formula are given, the reaction gas consumption amount and the film forming rate accompanying the reaction progress can be analyzed at each position in the reaction chamber 10. Further, since the activation energy is applied, it is possible to predict even a local reaction rate change due to a temperature change at each position in the chamber 10. If the shape of the reaction chamber 10 is given, the value of the boundary condition is determined by the position of the boundary condition of the above transport equation, the reaction gas introduction pressure given by the gas introducing means 20 and the chamber temperature set by the control means 10. As described above, the thin film formation simulation on the substrate becomes possible, and the uniformity of the film forming process in the chamber can be evaluated, as in the case of the embodiment described above with reference to FIG. Further, the final setting of the film formation process operation control amount can be similarly performed. On the other hand, if the thin film manufacturing apparatus according to the present embodiment is used, the electron orbital structure, reactivity index, bond order, surface superlattice structure of the molecule which is analyzed and output by the electronic structure physical property analyzer 100 with respect to the new reaction gas and the deposited substrate. Since the chemical reaction formula and the chemical reaction formula can be predicted by the neural network 70 from the crystal surface band structure, the operating amount of the film forming process can be predicted more accurately than the thin film manufacturing apparatus shown in FIG.
The process start-up time can be shortened.

In the thin-film manufacturing apparatus described with reference to FIGS. 1, 4, 5, and 6, the reaction gas and the reaction gas are used to predict information relating to the chemical reaction path such as the chemical reaction formula, the chemical reaction rate formula, and the activation energy. The electronic structure of the deposited substrate, the learning of the neural network that inputs the physical property information, and the association function were used.
There is no problem even if the interaction analysis between the reaction gas and the deposition substrate and the reaction path analysis are performed directly. This is possible by using the electronic structure physical property analyzer 100. In principle, the structural degrees of freedom of the reaction gas and the deposition substrate are changed,
By monitoring the energy of the entire system, the chemical reaction path can be searched, and the chemical reaction formula, chemical reaction rate formula, activation energy, etc. can be predicted. FIG. 7 is a conceptual diagram when the chemical reaction path analysis is directly performed. Here, for simplification of the description, the reaction progress direction in the case where the system has two degrees of freedom, that is, the reaction structure starting from the original system to the transition structure to the production system is shown. As shown in FIG. 7, the chemical reaction proceeds in a portion corresponding to a stream where both walls are sandwiched by energy curved surfaces of the entire system. The position that becomes the saddle point during the progress is the transition structure of the reaction element process, and the difference between the energy at that position and the original system energy becomes the activation energy. When the degree of freedom of the system is small as shown in FIG. 7, the molecular orbital method and the electron density functional method included in the electronic structure physical property analyzer 100 are used while changing the coordinates corresponding to the degree of freedom. It is possible to calculate the total potential energy, create an energy curved surface, and search for a path. FIG. 8 shows a specific route search method. FIG. 8 shows a method of path analysis from the point A to the point B on the reaction path. First, the coordinates of three points are randomly updated with respect to the degree of freedom variable considered from the point A. The update amount depends on the accuracy of the route search, but there is no problem if it is within 0.1Å. For example, when this is set as the X point, the Y point, and the Z point, the respective energies are first calculated by the molecular orbital method or the electron density functional method. From these three points, the most stable point is obtained by the parabolic approximation. Next, if this operation is performed for all degrees of freedom, this becomes the update point B on the reaction path. Then, from point B, if the coordinates of the system are updated by the procedure described above, the entire chemical reaction path can be finally searched. The chemical reaction route search according to the above procedure can be used without problems when the degree of freedom of the system is low, such as a two-molecule reaction in the gas phase, but the system can be used freely like the chemical reaction on the solid surface. When the number of degrees increases, the calculation time and the computer memory for the molecular orbital method calculation and the electron density functional method calculation, which are performed each time with the coordinate update, become enormous and the route search becomes impossible. In the present invention,
In such a case, this problem is solved by updating the coordinates not in all the degrees of freedom of the target system but in some of them. Various methods can be considered as a method of limiting the degree of freedom. In the present invention, this is done by the following three methods. First, in the first method, of the standard vibration coordinates for the entire chemical reaction system,
The reference vibration coordinates with a large frequency are fixed, and the coordinates are updated for other degrees of freedom. This is a standard equivalent to fixing the vibration coordinate having a large curvature component on the potential curved surface as shown in FIG. 7, that is, avoiding a steep wall and moving in a gentle swath direction. A more precise basis for these two criteria is as follows. At the transition structure position in the middle of the reaction path, there is a reference vibration with an imaginary frequency in the reaction path traveling direction, and a reference vibration with a low frequency near the original system is connected to the imaginary reference vibration at this transition structure position. I will go. The other two criteria literally fix the atomic coordinates. One of them is a method of fixing the coordinates of atoms which are considered not to be directly involved in the reaction in the chemical reaction system. The other is to fix the bond distance between some atoms that are not considered to participate in the reaction. For example, when searching for a deposition reaction path on the crystal surface, it is not possible to fix the atomic coordinates of the reaction gas and its adjacent surface, but fixing other surface atomic coordinates does not directly affect atoms that do not participate in the reaction. Corresponding to not considering, shortening the calculation time,
The computer memory capacity can be reduced and its effect is great.

The reduction effect when the reaction path search method described above is applied is shown below. The reduction effect when the path analysis of the dissociative adsorption reaction of monosilane molecules on the tungsten surface cluster W ₂₀ is performed by the molecular orbital method will be described. FIG. 9 shows the initial setting structure of the reaction. In dissociative adsorption of monosilane, the tungsten surface atoms do not move and only monosilane is considered to decompose. Therefore, all surface tungsten atoms are fixed, and for the monosilane molecule, only the triple degenerate vibration of the minimum frequency 914 Kaiser is left, and other The standard vibration was fixed. Three of the four hydrogen atoms vibrate away from the surface, and silicon atoms are heading toward the surface.
Since the reaction is considered to be initiated by the surface approach of silicon atoms and the upward movement of hydrogen atoms, the initial structure shown in FIG. 9 is considered to be a suitable arrangement for the search of reaction paths. Figure 8
According to the procedure shown in, the reaction path search from the original system to the production system can be performed in 50 steps by a large-scale computer (100MI).
Calculation performance of PS). As a result, the CPU
Dissociative adsorption reaction of monosilane (SiH ₄ → S in about 10 hours)
iH ₃ + H) could be traced and a reaction heat of −3.6 eV could be obtained. On the other hand, as a result of carrying out a route search using the same computer without fixing the degrees of freedom of all atoms, C
It took about 5 times as much PU time. Table 1 shows the results together with the comparison of computer memory capacity.

[0030]

[Table 1]

In this route search, since the LEAF molecular orbital method described in Irie et al., Seoreticashi Mika Actor, 70, 239, is used, calculation and storage of overlap integral between electrons including inner-shell electrons and exchange integral are unnecessary. Therefore, it can be said that the effect of reducing the computer memory capacity is small, but the effect of reducing the CPU time is large. Unlike other molecular orbital calculation methods, such as the LEAF molecular orbital method, when using an ab initio molecular orbital that requires calculation of overlap integral and exchange integral across all electrons, CPU time and computer memory capacity Further reduction effect can be expected.

Next, an example in which a tungsten thin film is grown by hydrogen reduction of WF ₆ using the thin film manufacturing apparatus according to the present invention shown in FIG. 6 will be described. The growth of the tungsten thin film by hydrogen reduction of WF _{6 has} a good step coverage of nearly 100% under normal film forming conditions, and the trench structure can be completely filled. In this example, in order to evaluate the performance of the present apparatus, the film formation process operation control amount and the film formation were performed to obtain a step coverage of 80%.
With a step coverage of 80%, for example, a trench having a width of 4 μm and a depth of 10 μm cannot be completely filled with tungsten, resulting in a structure in which soot enters the center of the trench. The procedure for calculating the operation control amount will be described below. The electronic structure physical property analyzer 100 uses the LEAF molecular orbital method used in the above embodiment, the neural network 70 has a three-layer hierarchical network, and the fluid reaction engineering analyzer 200 has three layers.
We used a mass transport equation solving system that combines the dimensional diffusion equation and the chemical reaction rate equation. Electronic structure information (maximum occupied orbital energy, lowest unoccupied orbital energy, hydrogen molecule, WF ₆ molecule, W ₂₀ cluster by LEAF molecular orbital method,
State density, bond order, charge density on atom, electrostatic multipole moment) are input to neural network 70 as W.
The reaction order of the hydrogen reduction reaction of F ₆ was predicted and used as a part of the input of the fluid reaction engineering analyzer 200. Further, the path analysis of the WF ₆ hydrogen reduction reaction on the W ₂₀ cluster was performed by the electronic structure physical property analyzer 100 to obtain the chemical reaction formula and the activation energy. Neural network 7 for activation energy
As an input of 0, the reaction rate was predicted and input to the fluid reaction engineering analysis device 200. From the analysis of the reaction path between the hydrogen molecule and the WF ₆ molecule on the W ₂₀ cluster, it was found that the chemical reaction formula can be applied by the following formula (1).

[0033]

## EQU1 ## W ₂₀ + 3H ₂ + WF ₆ → W ₂₁ + 6HF (1) The activation energy of this reaction could be analyzed as 80 kJ / mol. Furthermore, as a result of predicting the reaction order from the neural network 70, the dependency of the WF ₆ molecule is low,
Since it was predicted that the dependence of hydrogen molecules was high, there was no dependence of WF _{6 and} the dependence of hydrogen molecules was a first-order dependence.
It took about 6 hours to complete the above information prediction. This is mainly the electronic structure physical property analyzer 100.
The time required for the reaction path search by the neural network 70 and the prediction by the neural network 70 takes almost no time. Next, based on these results, the analysis by the fluid reaction engineering analyzer 200 proceeds, and the film formation process operation control amount with the step coverage of 80% is displayed on the film formation process operation control amount display unit 300 as follows. The operation was started.

The time required for this is 10 minutes.

Deposition substrate temperature: 700 K WF ₆ introduction partial pressure: 2 Pa H ₂ introduction partial pressure: 20 Pa Predicted growth rate: 50 nm / min Growth operation time: 80 min After the operation was completed, the step coverage was inspected to find a step difference of 84%. It was found that the desired embedding characteristics were obtained, which consisted of the coverage. According to the thin film manufacturing apparatus of the present invention, a desired thin film can be obtained within about 8 hours from the input of the film forming specification, and it is possible to avoid waste of time and resources such as many times of sample exchange and vacuum exhaustion due to trial and error. You can

Next, using the thin film manufacturing apparatus according to the present invention, C
An example applied to the film composition control by reducing the hysteresis phenomenon in the VD thin film manufacturing process will be described. The CVD apparatus having the structure shown in FIG. 1 was used as the thin film manufacturing apparatus, and the composition of the ternary ferroelectric thin film PZT [Pb (Zr, Ti) O ₃ mixed crystal] was controlled.
Pb (C ₁₁ H ₁₉ O ₂ ) ₂ , Zr (Ot-Bu) ₄ and Ti (Oi-Pr) ₄ were used as raw materials, and a reaction gas was introduced into the apparatus by an N ₂ carrier gas bubble. Further, O ₂ gas was separately introduced into the reaction chamber 10 to adjust the oxygen composition of the PZT thin film. A film was formed on a Pt substrate at a N ₂ and O ₂ mixed carrier gas flow rate of 25 cc / min, a substrate temperature of 600 degrees, and a total pressure of 5 to 10 torr. However, the raw material temperature was set to 40 degrees from the saturated vapor pressure. As a result, 5
Target composition Pb: Zr: Ti = 1: 0.36: 0.58 P at the film formation rate of nm / sec
A ZT film was obtained, but each time the film forming experiment was repeated, it became a PZT film having a composition ratio different from that shown above, and the ferroelectric property was not exhibited. Compositionally, it appeared as an increase in Pb content. Upon intensive investigation, it was found that the saturated vapor pressure of the raw material Pb (C ₁₁ H ₁₉ O ₂ ) ₂ was low, and precipitation and solidification occurred at the equipment introduction part and the inner wall of the equipment, and the composition ratio changed with time. did. In the PZT film, the ferroelectric characteristics are caused by the ratio of four elements including Pb, Zr, Ti, and O, and simply Pb (C ₁₁ H
It is not possible to eliminate the deterioration of the ferroelectric properties with the lapse of time simply by reducing the amount of ₁₉ O ₂ ) ₂ introduced. Therefore, using the operation history data {Ti}, which is recorded and stored in the database 80, regarding the PZT thin film manufacturing process, including the film formation gas flow rate, temperature, accumulated operation elapsed time, and composition after film formation, A neural network was constructed. That is, the composition ratio of Pb, Zr, Ti, and O, the oxygen gas flow rate, the nitrogen gas flow rate, the temperature of the reaction chamber 10, the pressure, the deposition substrate temperature, and the accumulated operation elapsed time are input as inputs of the thin film neural network 70. Pb (C ₁₁ H ₁₉ O
₂ ) ₂ , Zr (Ot-Bu) ₄ , Ti (Oi-Pr) ₄ Experimental results up to now by the backpropagation method shown in the previous example by constructing a neural network that outputs the bubble gas flow rate for each raw material Learned. As a result, by using the bubble gas flow rate for the raw material predicted by the neural network 70, the PZT thin film could be formed with a stable composition ratio over the cumulative operating time of 500 hours. Thus, the thin film manufacturing apparatus according to the present invention has an effect of forming a complex multi-element oxide thin film with a stable composition for a long time.

Next, an embodiment of a semiconductor device manufacturing method using the thin film manufacturing apparatus according to the present invention will be described with reference to FIG. In this embodiment, a basic cell of a DRAM having a memory function was constructed using the manufacturing apparatus having the configuration shown in FIG. First, a channel stopper 403 was formed on a p-type substrate 401 by an ion implantation method, an element isolation oxide film 405 was formed, and then a gate oxide film 404 was formed. After that, a transfer gate 406 and a storage gate 407 are formed, a source / drain region 402 is formed, and then a wiring 408 and the like are formed to finally form a basic cell of a DRAM having a memory function. Of the series of manufacturing processes, the transfer gate 406,
The storage gate 407 was made of doped polysilicon using this thin film manufacturing apparatus. Upon formation, HOMO orbital energies of both disilane and phosphine (energy values of highest occupied electron orbitals), LUMO orbital energies (energy of lowest unoccupied electron orbitals), all electron orbital symmetry obtained by the electronic structure physical property analyzer 100, With the activation energy of polysilicon generation, dipole moment (degree of charge separation), molecular volume (approximate size of molecule) and substrate temperature as input, the sticking coefficient of both source gases (polysilicon among molecules that collide with the surface). The ratio involved in the formation chemical reaction) was output by the neural network 70. HOMO,
The reason for selecting the LUMO orbital energy as an input is that these amounts are indices showing the oxidizing power and reducing power of the molecule and are considered to determine the basis of the chemical reaction. Similarly, the dipole moment is an index of the attractive force on the surface,
The molecular volume is an index that indicates the matching between the thin film surface gap and the source gas, and the substrate temperature is an index that indicates the thermal motion of the surface and source gas molecules, and is considered to be the basic quantity that determines the sticking coefficient. The input of the neural network. Now, the neural network 70 is monosilane,
It has already been learned by the back-propagation method to output the sticking coefficient with respect to the input, using the database 80 having the chemical reaction thermodynamic data of the semiconductor manufacturing gas such as phosphine and arsine and the sticking coefficient measured value. Using the value of the obtained sticking coefficient, the flow of raw material gas in the reaction chamber 10 and the distribution of film formation rate were analyzed by the fluid reaction engineering analyzer 200. However, the reaction chamber 10 has a diameter of 2
It is a 0 cm cylindrical hot wall type batch reactor.
Boundary conditions such as shape and gas partial pressure are keyboard 2, CR
Entered using T3. As a result, at the substrate temperature of 500 degrees, the mixed gas pressure of 0.5 torr (mixing ratio of 100: 1 and disilane excess) indicates that the film formation rate distribution is optimum, which is output by the fluid reaction engineering analyzer 200 and the control means 60. The gas flow rate and the gas pressure in the chamber were adjusted, and the polysilicon film formation was started. It took only a few seconds to predict the adhesion coefficient output and the distribution in the chamber, and there was no time hindrance to the operation start. As a result, the distribution of the memory cell structure within the wafer is extremely uniform, and the use of this manufacturing apparatus makes it possible to predict the optimum process operation value within a short time, and the manufacturing distribution within the wafer becomes particularly important. It has been shown to be very effective in the manufacture of semiconductor devices.

Next, another embodiment of the semiconductor device manufacturing method using the thin film manufacturing apparatus according to the present invention will be described with reference to FIG. In this embodiment, a liquid crystal display driving T
An FT transistor was constructed. The substrate 500 has a distortion temperature of 5
It is a glass substrate of 80 degrees. With the substrate kept at 400 ° C., a silicon layer 501 containing both a crystalline component and an amorphous component is formed by plasma CVD using monosilane gas diluted with hydrogen gas as a raw material to a thickness of about 600 Å.

The ratio of the crystalline component to the amorphous component is an important control factor because it determines the carrier mobility and thus the operating performance of the TFT transistor. So the gas dilution,
A database consisting of a film formation process operation history including applied power for plasma generation, substrate temperature, and accumulated operation time, and a crystal component / amorphous component ratio was constructed, and the plasma CVD apparatus was operated and controlled with the configuration shown in FIG. As a result, when the substrate temperature is close to 400 degrees, the plasma generation power is reduced to 0.
It was found that a value of about 4 W / cm ² was set and a gas dilution degree of about 5% was suitable, and it became possible to operate with minimal change over time. Next, after island-photographing this silicon layer, a gate oxide film 504 is formed by atmospheric pressure CVD,
Next, a doped silicon layer for a gate electrode was formed by the plasma CVD method. Next, a gate electrode 505 was formed by photoetching, and a source region 502 and a drain region 503 were formed by an ion implantation method. Next, PS
After G and Al are vapor-deposited to form a TFT, a transparent electrode 509 is formed.
Deposited by sputtering method, TF for driving liquid crystal display
T. In addition, 506 is wiring, 507 is PSG film, 5
Reference numeral 08 is an oxide film.

As can be seen from the above two embodiments, the use of the thin film manufacturing apparatus according to the present invention has the effect that the semiconductor device can be manufactured efficiently and reproducibly regardless of the process history.

[0041]

As described above, according to the present invention, first, the learning function of the neural network is used for the known film forming reaction system, and the optimum operation process control is performed even for the reaction system which operates nonlinearly. There is an effect that the amount can be calculated and predicted instantly. In addition, even for new reaction gases and deposition substrates, it is possible to predict the film formation reaction formula, reaction order, etc., and to predict the operation control amount, so many sample exchanges due to trial and error, evacuation, etc. The waste of time and resources can be largely avoided.

Further, by applying it to the actual manufacturing of the semiconductor device, the manufacturing can be performed efficiently and with good reproducibility regardless of the process history.

[Brief description of drawings]

FIG. 1 is a schematic diagram showing an overall configuration diagram of an embodiment of a thin film manufacturing apparatus according to the present invention.

FIG. 2 is an explanatory diagram showing a response of a neuron model which is a basic unit of a neural network.

FIG. 3 is an explanatory diagram showing a configuration and a learning function of a hierarchical neural network.

FIG. 4 is a schematic view showing an overall configuration diagram of another embodiment of the thin film manufacturing apparatus according to the present invention.

FIG. 5 is a schematic diagram showing the overall configuration of another embodiment of the thin film manufacturing apparatus according to the present invention.

FIG. 6 is a schematic view showing an overall configuration diagram of still another embodiment of the thin film manufacturing apparatus according to the present invention.

FIG. 7 is an explanatory view showing the proceeding direction of a chemical reaction.

FIG. 8 is an explanatory diagram of a procedure of a chemical reaction route search method.

FIG. 9 is a view showing an initial structure of monosilane dissociation reaction on a tungsten cluster.

FIG. 10 is a diagram showing a cross-sectional structure of a semiconductor device manufactured using the thin film manufacturing apparatus according to the present invention.

FIG. 11 is a diagram showing a cross-sectional structure of another semiconductor device manufactured using the thin film manufacturing apparatus according to the present invention.

[Explanation of symbols]

1 ... Film formation specification input means, 2 ... Keyboard, 3 ... CRT,
10 ... Reaction chamber, 20 ... Gas introduction means, 30 ... Substrate introduction holding means, 40 ... Vacuum exhaust means, 50 ... Analysis means,
60 ... Control means, 70 ... Neural network, 80
... database, 90 ... database access means, 1
00 ... Electronic structure physical property analyzer, 200 ... Fluid reaction engineering analyzer, 300 ... Film formation process operation control amount display unit, 40
1 ... Si substrate, 402 ... Source / drain region, 403
... Channel stopper, 404 ... Gate oxide film, 405 ...
Inter-element isolation film, 406 ... Transfer gate, 407 ... Storage gate, 408 ... Wiring, 500 ... Glass substrate, 501 ... Silicon layer, 502 ... Source region, 503 ... Drain region,
504 ... Gate oxide film, 505 ... Gate electrode, 506 ...
Wiring, 507 ... PSG, 508 ... Oxide film, 509 ... Transparent electrode.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification number Reference number within the agency FI Technical indication location H01L 21/31 (72) Inventor Jiro Ushio 1-280 Higashi-Kengikubo, Kokubunji-shi, Tokyo Hitachi Central (72) Inventor Ken Yamaguchi Ken-ji, Kokubunji, Tokyo 1-280, Higashi Koigokubo Central Research Laboratory, Hitachi, Ltd.

Claims (18)

[Claims] 1. A reaction chamber, means for introducing a reaction gas into the reaction chamber, means for holding a substrate provided in the reaction chamber, pressure adjusting means for adjusting the pressure in the reaction chamber, and reaction in the reaction chamber. In a thin film manufacturing apparatus comprising an analyzing means for analyzing the above, and a controlling means for controlling the reaction in the reaction chamber, the whole or a part of the film forming specification is used as a primary input, and the film forming process operation control amount is used as a final output. A thin film manufacturing device characterized by using a neural network for the prediction procedure. 2. The neural network comprises access means to a database provided separately from the thin film manufacturing apparatus, and has a learning function of using information in the database as an input teacher pattern. The thin film manufacturing apparatus described. 3. The thin-film manufacturing, wherein the information in the database is any one of a film formation specification, a film formation process operation control amount, a film formation process operation elapsed time, a film formation process operation accumulated elapsed time, etc., or any combination thereof. 3. The thin film manufacturing apparatus according to claim 2, which has apparatus operation history data. 4. The film formation specification has molecular structure and crystal structure information of both reactive gas species and / or deposition substrate species. An electronic structure physical property analyzer that outputs all or part of the electronic structure and physical property information consisting of the property index, bond order, film formation reaction order, film formation activation energy, surface superlattice structure, crystal surface band structure, A neural network is provided which inputs all or part of output information from the electronic structure physical property analyzer and part of other film forming specifications and outputs the film forming process operation control amount. The thin film manufacturing apparatus according to any one of claims 1 to 3, wherein 5. The electronic structure physical property analyzer is a molecular force field method,
5. The thin film manufacturing apparatus according to claim 4, wherein all or a combination of the molecular orbital method, the electron density functional method, the molecular dynamics method, and the energy band calculation method is used. 6. The control means receives a reaction gas species and / or a deposition substrate, or a molecular structure and / or a crystal structure of either one of them as an input, and uses both or one of a thin film growth chemical reaction formula and a thin film growth reaction rate formula. Neural network for output, the thin film growth chemical reaction formula, and / or the thin film growth reaction velocity formula, reaction gas introduction pressure, chamber temperature, substrate temperature are input, and pressure distribution in chamber, temperature distribution, substrate surface 4. The control according to any one of claims 1 to 3, wherein the control is performed based on the operation control amount of the film forming process predicted by a fluid reaction engineering analysis device that analyzes all or a combination of any of the upper thin film growth rates. Thin film manufacturing equipment. 7. The operation amount of the film forming process has all or any of a thin film growth chemical reaction formula, a thin film growth reaction rate formula, and a thin film growth activation energy, and an output from the neural network and a reaction gas introduction pressure. , Chamber temperature, substrate temperature, or all combinations of them,
The thin film production apparatus according to claim 4 or 5, wherein the film reaction control amount is predicted by a fluid reaction engineering analysis apparatus that analyzes all or any combination of the pressure distribution in the chamber, the temperature distribution, the thin film growth rate on the substrate surface. . 8. A step of inputting a molecular structure and / or a crystal structure of both reactive gas species and / or deposition substrate species, molecular electronic structure, reactivity index, bond order, chemical reaction path from the input information. A step of analyzing all or any combination of electronic structure consisting of film formation reaction order, film formation reaction activation energy, surface superlattice structure, crystal surface band structure and physical property information using an electronic structure physical property analysis method, the electronic structure Inputting all or part of the physical property information and part of other film forming specifications to the neural network and outputting the film forming process operation control amount, controlling the film forming process according to the output film forming process operation control amount In the method for controlling a thin film manufacturing apparatus, the degree of freedom of the atomic arrangement (positional coordinates) of the input molecule, crystal, etc. is limited, and the chemical reaction path (the atomic arrangement is defined). Potential and - a control method of a thin film manufacturing apparatus characterized by searching the minimum point and saddle) on the energy function. 9. By fixing the vibration coordinate having a large curvature component among the reference vibration coordinates of a certain minimum point on the potential-energy function whose domain is the atomic arrangement,
9. The method for controlling a thin film manufacturing apparatus according to claim 8, wherein the degree of freedom of the atomic arrangement (positional coordinates) is limited. 10. Freeness of the atomic arrangement (positional coordinates) by fixing the vibration coordinates having a large frequency among the reference vibration coordinates of a certain minimum point on the potential-energy function having the atomic arrangement as the definition range. 9. The method for controlling a thin film manufacturing apparatus according to claim 8, wherein the degree is limited. 11. The thin film manufacturing apparatus according to claim 8, wherein the degree of freedom of the atomic arrangement (positional coordinates) is limited by fixing the positional coordinates of some atoms of the atomic assembly. Method. 12. The thin film manufacturing apparatus according to claim 8, wherein the degree of freedom of the atomic arrangement (positional coordinates) is limited by fixing a bond distance between some atoms of the atomic assembly. Control method. 13. The method for analyzing the physical properties of electronic structures, wherein all or a combination of a molecular force field method, a molecular orbital method, an electron density functional method, a molecular dynamics method, an energy band calculation method is used. The method for controlling a thin film manufacturing apparatus according to claim 8, 14. A semiconductor device having a thin film of any one of an amorphous silicon film, a silicon oxide film, a nitride film, a metal film, a metal silicide film, and a composite oxide film, or a combination of any two or more thin films. In the method, a step of introducing and holding the semiconductor device substrate on a sample stage in a reaction chamber, a step of introducing the thin film raw material gas into the reaction chamber, all of the film forming specifications of the thin film or the combination thin film or It is characterized by comprising a step of inputting a part thereof to a neural network and predicting a film formation process operation control amount as the neural network output, and a step of controlling the operation of the reaction chamber according to the film formation process operation control amount. Manufacturing method of semiconductor device. 15. The film formation process operation control amount predicting step by the neural network comprises a database access step and a learning step in which information in the database is used as an input teacher pattern. Of manufacturing a semiconductor device of. 16. The film forming process comprises PVD and thermal CV.
16. The method of manufacturing a semiconductor device according to claim 14, wherein the method is one of D, plasma CVD, and photo CVD, or any combination thereof. 17. The semiconductor device according to claim 14, wherein the semiconductor device is one of a silicon bipolar semiconductor, a silicon field effect semiconductor, and a compound semiconductor, or any combination thereof. A method of manufacturing a semiconductor device according to item 1. 18. The method of manufacturing a semiconductor device according to claim 14, wherein the semiconductor device is a semiconductor memory device having a memory function such as RAM and ROM.