Semiconductor process recipe automatic acquisition method, system and semiconductor process equipment

Abstract

The present invention discloses a method, system and semiconductor process equipment for automatically acquiring semiconductor process recipes, the method comprising: randomly inputting a set of process parameters into a constructed deep neural network model; judging whether the difference between each process result evaluation index output by the deep neural network model and the corresponding set process requirements meets the set requirements; if so, using a set of process parameters as the process recipe for the actual process; otherwise, using a gradient algorithm to adjust the input deep neural network model process parameters, and based on a self-consistent iteration method, obtaining a set of process parameters that can make all process result evaluation indexes output by the deep neural network model meet the corresponding set process requirements as the process recipe for the actual process. The corresponding process recipe is automatically acquired according to the given process requirements, and the efficiency of acquiring the process recipe is improved.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and more specifically, to a method and system for automatically acquiring a semiconductor process recipe and semiconductor process equipment. Background Art

[0002] In the field of microelectronics manufacturing, plasma etching is an important process in the device processing. For example, etching a trench gate in a discrete transistor device can suppress the short channel effect of the device and solve the problems of reduced mobility caused by channel electron scattering. Therefore, discrete transistor devices are undergoing a transformation from traditional planar gate to trench gate, and are also applicable to silicon-based and silicon carbide-based MOSFET devices; for example, etching a super junction in a discrete transistor device can effectively reduce the on-resistance and increase the switching rate of the device. Since the driving current of the lateral structure super junction is small, the mainstream products in the market use vertical super junctions, and the process technology has developed from multiple epitaxial processes to deep trench processes, so deep silicon etching is required; for example, one method to achieve a normally-off device in a gallium nitride HEMT device is to etch a groove (gate recess) in the gate, and change the surface state energy level of the gate gallium nitride through the groove to achieve the device being normally off. For discrete devices that originally require trench etching, they are also developing in the direction of high aspect ratio and high verticality. For example, in traditional silicon capacitor devices, the deeper the trench, the more effectively the device space can be used and the capacitance value can be improved; the higher the verticality, the more consistent the actual device is with the theoretical model, and the better it can meet the design needs; for example, in advanced packaging, etching through silicon vias (TSV) structures can improve device performance, reduce power consumption and reduce device size.

[0003] Therefore, the development of fields such as semiconductor discrete devices and advanced packaging has put forward higher and higher requirements for deep silicon etching. Since the deep silicon microstructure has a large aspect ratio and high verticality, traditional wet etching is difficult to complete and must be obtained by dry etching. In order to obtain silicon microstructures with a deep depth and a vertical angle, a time-separated dry etching process is mainly used, namely the "Bosch" process developed by Robert Bosch. This process uses plasma-induced fluorocarbon polymers to provide sidewall passivation protection and fluorine-based plasma chemical reactions to etch silicon downward, in which sidewall passivation protection (deposition step) and fluorine-based plasma chemical reactions (etching step) are performed alternately. Since the "Bosch" process uses deposition and etching alternately, a scallop structure will inevitably be formed, resulting in large sidewall roughness.

[0004] Generally speaking, the scallop size is the largest / the sidewall is the roughest at the top of the deep silicon structure, and obvious scallops and lateral wrinkles can be seen. Therefore, it is necessary to find a way to reduce the roughness of the top of the deep silicon structure. In addition to the sidewall roughness, the etching morphology (sidewall angle) is often concerned. This is because after the deep silicon structure is formed, it is difficult for the plasma to drill into the deep silicon structure, and the etched product is also difficult to be extracted from the deep silicon structure. It is often easy to form a deep silicon structure that is large on the top and small on the bottom. At this time, it is necessary to adjust the process parameters to increase the mean free path of the plasma and help the plasma move toward the wafer. In addition, etching depth, etching rate, mask selectivity, etc. are also indicators for evaluating the quality of deep silicon etching. To obtain a good deep silicon etching effect, it is necessary to optimize the process parameters, that is, adjust the process recipe. At present, it is mainly the process personnel who manually try various recipes one by one to determine the recipe that can meet the process requirements. This process is often very complicated and inefficient, and cannot be automated. Summary of the invention

[0005] The purpose of the present invention is to propose a method, system and semiconductor process equipment for automatically acquiring semiconductor process recipes, so as to automatically acquire corresponding process recipes according to given process requirements and improve the efficiency of acquiring process recipes.

[0006] To achieve the above object, the present invention proposes a method for automatically acquiring a semiconductor process recipe, comprising:

[0007] A set of process parameters required by the process is randomly input into the constructed deep neural network model to obtain a process result evaluation index corresponding to the process parameter group; wherein the deep neural network model is a multi-layer deep neural network that has completed training, and the deep neural network model includes an input layer, at least one intermediate layer and an output layer, the input layer is used to input a plurality of the process parameters, and the output layer is used to output the corresponding process result evaluation index;

[0008] Determine whether the difference between each of the process result evaluation indicators and the corresponding set process requirements meets the set requirements;

[0009] If yes, using the set of process parameters as a process recipe for performing the actual process;

[0010] Otherwise, according to the set process requirements, along the direction from the output layer to the input layer, the characteristic values ​​in each intermediate layer and the process parameter values ​​in the input layer are optimized layer by layer based on the gradient algorithm and the self-consistent iteration method, until a set of process parameters is obtained in the input layer so that all process result evaluation indicators output by the deep neural network model meet the corresponding set process requirements, as the process recipe of the actual process.

[0011] Optionally, the method for constructing the deep neural network model includes:

[0012] The number of layers and structure of the deep neural network model are determined according to the number of the process parameters and the number of the process result evaluation indicators; wherein the number of the intermediate layers is the difference between the number of the process parameters and the number of the process result evaluation indicators, each layer of the deep neural network in the deep neural network includes a plurality of neurons, and along the direction from the input layer to the output layer, one neuron is successively reduced between adjacent deep neural network layers, each neuron is connected to all neurons in the previous layer of the neural network, each process parameter is used as an input feature of a neuron in the input layer, each neuron in the intermediate layer is used to calculate the output features of all neurons in the previous layer of the neural network, and the output feature of each neuron in the output layer is the corresponding process result evaluation indicator, and the number of the process parameters is greater than the number of process result evaluation indicators;

[0013] An activation function for running the deep neural network model is given according to the number of layers of the deep neural network model.

[0014] Optionally, the method for constructing the deep neural network model further includes:

[0015] Using a process database as training data and using a deep learning method to train the deep neural network model, the process database includes historical process parameter data and historical process result data of a specific process;

[0016] The weight term and bias term of each layer of the deep neural network in the deep neural network model are determined according to the training result of the activation function.

[0017] Optionally, the optimizing the characteristic values ​​in each of the intermediate layers and the process parameter values ​​in the input layer layer by layer based on the gradient algorithm and the self-consistent iteration method includes:

[0018] Step S1: Finding a process result evaluation index that has the largest difference from the corresponding set process requirement from the output layer;

[0019] Step S2: Finding a process parameter with the largest weight value associated with the process result evaluation index with the largest difference from the intermediate layer adjacent to the output layer;

[0020] Step S3: The weight value of the process parameter with the largest weight value is fixed unchanged, and the characteristic value of the process parameter with the largest weight value is adjusted step by step using a gradient algorithm to optimize, until the difference between the process result evaluation index with the largest difference in the output layer and the corresponding set process requirement meets the set requirements;

[0021] Step S4: If there are other process result evaluation indicators whose difference degree with the corresponding set process requirements does not meet the set requirements among the other process result evaluation indicators, repeat steps S1-S3 until all process result evaluation indicators in the output layer meet the corresponding set process requirements, and determine the characteristic value of each process parameter in the intermediate layer;

[0022] Step S5: Based on the characteristic value of each process parameter in the intermediate layer obtained in step S4, continue to use the gradient algorithm and self-consistent iteration method to optimize the characteristic values ​​of other intermediate layers and the process parameter values ​​in the input layer layer by layer along the direction from the output layer to the input layer, until a set of process parameters is obtained in the input layer that can make all process result evaluation indicators output by the deep neural network model meet the corresponding set process requirements.

[0023] Optionally, before step S4, the method further includes: if the difference between the process result evaluation index adjusted in step S3 and the set process requirement still cannot meet the set requirement, the process parameters corresponding to the maximum weight value obtained after adjustment and optimization in step S3 are fixed, and the following steps are performed:

[0024] Step S6: Finding a process parameter with the second largest weight value related to the corresponding adjusted process result evaluation index from the intermediate layer adjacent to the output layer;

[0025] Step S7: fix the weight value of the process parameter with the second largest weight value unchanged, and use the gradient algorithm to optimize the step-by-step adjustment of the feature value of the process parameter with the second largest weight value until the difference between the corresponding adjusted process result evaluation index in the output layer and the corresponding set process requirement meets the set requirements.

[0026] Optionally, after step S5, the method further includes:

[0027] Step S8: If after optimizing and adjusting all process parameters related to the process result evaluation index, there are still process result evaluation indexes in the output layer whose difference from the corresponding set process requirements does not meet the set requirements, then:

[0028] A set of process parameters that can minimize the sum of the squares of the differences between each process result evaluation index and the corresponding set process requirements is selected, and the set of process parameters is used as a process recipe for the actual process.

[0029] Optionally, the calculation formula of the gradient algorithm is:

[0030]

[0031] Among them, x iis the adjusted process parameter in the middle layer, y is the adjusted process result evaluation index, and θ is the adjustment step size.

[0032] Optionally, the degree of difference is calculated by the following formula:

[0033] Δ=|(yy 实际 ) / y 实际 |

[0034] Among them, Δ is the difference between the process result evaluation index and the set process requirement, y is the process result evaluation index output by the deep neural network model, and y 实际 For a given process requirement;

[0035] The setting requirement is that the difference between the process result evaluation index and the corresponding set process requirement is between 0 and 10%.

[0036] In a second aspect, the present invention further provides a semiconductor process recipe automatic acquisition system, comprising:

[0037] A deep neural network model, wherein the deep neural network model is a deep neural network with a multi-layer structure that has been trained, and the deep neural network model can output accurate process result evaluation indicators corresponding to the process parameter group based on the input process parameters;

[0038] A calculation module is used to execute the method for automatically acquiring semiconductor process recipes described in the first aspect.

[0039] In a third aspect, the present invention further proposes a semiconductor process equipment, comprising the semiconductor process recipe automatic acquisition system described in the second aspect.

[0040] The beneficial effects of the present invention are:

[0041] The present invention is based on a deep neural network model that has been trained. Under given process requirements, based on the constructed deep neural network model, according to the set process requirements, along the direction from the output layer to the input layer, a gradient algorithm and a self-consistent iteration method are used to optimize and adjust the characteristic values ​​of each intermediate layer and the process parameters of the input layer layer by layer, until a set of process parameters that can make all process result evaluation indicators output by the deep neural network model meet the process requirements are inversely deduced as the process formula of the actual process. Compared with the existing method of manually repeatedly debugging the process formula parameters, the present invention can automatically provide process formula parameters that meet the process requirements through the deep neural network model according to the process requirements, thereby improving the efficiency of obtaining the process formula and enhancing the degree of automation of the etcher.

[0042] The system of the present invention has other characteristics and advantages, which will be apparent from the drawings incorporated herein and the following detailed description, or will be described in detail in the drawings incorporated herein and the following detailed description, which together serve to explain the specific principles of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, in which like reference numerals generally represent like components.

[0044] Figure 1 A schematic diagram of a deep neural network model applied to a lithography machine according to prior art 1 is shown.

[0045] Figure 2 A schematic diagram of a neural network model for predicting etching process results according to prior art three is shown.

[0046] Figure 3 A flow chart showing the steps of a method for automatically acquiring a semiconductor process recipe according to an embodiment of the present invention is shown.

[0047] Figure 4 A schematic diagram of a deep neural network model in a method for automatically acquiring a semiconductor process recipe according to an embodiment of the present invention is shown.

[0048] Figure 5 A schematic diagram of an etching process with a characteristic size of 3 microns in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0049] Figure 6 A schematic diagram showing the influence of chamber pressure on evaluation indicators of trench etching process results in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0050] Figure 7 A schematic diagram showing the influence of the final power of the lower electrode in the etching step on the evaluation index of the groove etching process result in a semiconductor process recipe automatic acquisition method according to a specific embodiment of the present invention is shown.

[0051] Figure 8 A schematic diagram showing the influence of the final time of the etching step on the evaluation index of the groove etching process result in a semiconductor process recipe automatic acquisition method according to a specific embodiment of the present invention is shown.

[0052] Fig. 9 A schematic diagram showing the influence of the final time of a deposition step on an evaluation index of a groove etching process result in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0053] Fig.10 A schematic diagram showing the influence of SF6 gas intake in an etching step on evaluation indicators of trench etching process results in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0054] Fig.11 A schematic diagram showing the influence of C4F8 gas intake in a deposition step on evaluation indicators of a groove etching process result in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0055] Fig.12 A SEM image showing the trench etching process requirements in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention.

[0056] Fig.13 A schematic diagram showing the influence of chamber pressure on evaluation indicators of hole etching process results in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0057] Fig.14 A schematic diagram showing the influence of the final power of the lower electrode in the etching step on the evaluation index of the hole etching process result in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0058] Fig.15 A schematic diagram showing the influence of the final time of an etching step on an evaluation index of etching process results in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0059] Fig.16 A schematic diagram showing the influence of the final time of a deposition step on an evaluation index of a hole etching process result in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0060] Fig.17 A schematic diagram showing the influence of SF6 gas intake in an etching step on evaluation indicators of hole etching process results in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0061] Fig.18 A schematic diagram showing the influence of C4F8 gas intake in a deposition step on evaluation indicators of hole etching process results in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention is shown.

[0062] Fig.19 A SEM image showing the hole etching process requirements in a method for automatically acquiring a semiconductor process recipe according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0063] Prior art 1 discloses a deep neural network model applied to a lithography machine. The deep neural network model is as follows: Figure 1As shown in the figure, the deep neural network model based on deep learning can automatically output the lithography process results according to the lithography process parameters of the input model.

[0064] The deep neural network model is an existing technology. Specifically, first, it is necessary to define the activation function of the model and perform a linear transformation on the parameters of the input layer: y = f(Σ i w i x i +b), where x i is the input feature vector, w i is the weight of each feature value, and b is the bias. This is the simplest case with only one layer of neural network. If there are multiple layers of neural network, the activation function of the two-layer model is transformed into y=g(Σ j w j f(Σ i w i x i +b1)+b2), the activation function of the three-layer model is transformed into y=h(Σ k w k g(Σ j w j f(Σ i w i x i +b1)+b2)+b3), and so on.

[0065] Then, according to the input and output values ​​(x i ,y) and the actual data set (x i ,y r ) defines the objective function: J(w i ,b)=(Σ i w i x i +by r ) 2 , if there are n actual data sets (x i,n ,y r,n ), then we need to perform arithmetic averaging on the objective function: J(w i ,b)=(1 / n)Σ n (Σ i w i x i,n +by r,n ) 2 .

[0066] In the neural network model, data training is an important step, which mainly involves parameter optimization. The current mainstream optimizers are mainly divided into the following three categories: stochastic gradient descent, momentum stochastic gradient descent and Adam optimization method. Taking stochastic gradient descent as an example, its algorithm is:

[0067] That is, the partial derivatives of the relevant parameters are calculated, and each time they are moved according to the product of the step size η and the partial derivative, and finally stabilized until the objective function obtains the minimum value, that is, the deviation between the data predicted by the model and the actual value is minimized. For the case of a multi-layer neural network, it can be reversed layer by layer, so that the objective function of each layer takes the minimum value, that is, the deviation between the data predicted by the entire model and the actual value is minimized.

[0068] The prior art 1 has the following problems:

[0069] The neural network model only provides forward learning and prediction, that is, the process results are only forward derived from the process parameter formula. When performing forward deduction, the variable in the gradient optimization algorithm is the weight value w i , it is impossible to achieve reverse deduction of process parameters according to process requirements; and it does not involve machine automation.

[0070] Photolithography machines have high requirements for precision and their optimizer algorithms are relatively complex (while other semiconductor process equipment such as etchers have a certain tolerance for precision requirements relative to computing processing speed).

[0071] Prior art 2 discloses a self-consistent iterative method, that is, when solving transcendental equations, no analytical solution can be obtained but only numerical solution can be obtained. When obtaining the numerical solution, a continuous iterative method is used, and finally converges to the optimal solution to form a self-consistent solution.

[0072] If an operator in the equation is a function of unknowns, the only method available is the self-consistent iterative method. First, a certain set of unknown values ​​is given, the operator is solved, and then the solution is obtained by substituting it into the original equation to obtain another set of numerical solutions for the unknowns. This process is repeated until the values ​​of the assumed unknowns are exactly the same as the values ​​of the finally calculated unknowns, i.e. they are self-consistent. This method is used in solving the Hartree-Fock equation in the field of quantitative computing.

[0073] This method has not been applied to neural network models or in the field of semiconductor process equipment.

[0074] Prior art 3 discloses a neural network model for predicting etching process results, such as Figure 2 As shown, this solution only uses one layer of neural network model, which is an ordinary neural network rather than a deep neural network, and can only predict the etching process results, but cannot reversely infer the process parameters through the etching process results.

[0075] The present invention can use software algorithms to automatically provide process recipes for related processes (such as dry etching, etc.) according to actual process requirements. Specifically for practical applications, if the user needs to obtain a deep silicon structure with a certain etching depth, sidewall angle, sidewall roughness, and mask selectivity, and expects to complete it at a certain etching rate, the corresponding process parameter recipes can be automatically given through the solution of the present invention, such as chamber pressure, upper and lower electrode power, air flow rate, and etching time.

[0076] The present invention will be described in more detail below with reference to the accompanying drawings. Although preferred embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0077] Figure 3 A flow chart showing the steps of a method for automatically acquiring a semiconductor process recipe according to an embodiment of the present invention is shown.

[0078] like Figure 3 As shown, a method for automatically acquiring a semiconductor process recipe comprises:

[0079] Step S101: randomly input a set of process parameters required by the process into the constructed deep neural network model to obtain a process result evaluation index corresponding to the process parameter group; wherein the deep neural network model is a multi-layer deep neural network that has completed training, and the deep neural network model includes an input layer, at least one intermediate layer and an output layer, the input layer is used to input multiple process parameters, and the output layer is used to output the corresponding process result evaluation index;

[0080] Specifically, before step S101, a deep neural network model is first constructed. The method for constructing the deep neural network model includes:

[0081] The number of layers and structure of the deep neural network model are determined according to the number of process parameters and the number of process result evaluation indicators; wherein the deep neural network model includes an input layer, at least one intermediate layer and an output layer, the number of intermediate layers is the difference between the number of process parameters and the number of process result evaluation indicators, each layer of the deep neural network in the deep neural network includes multiple neurons, and along the direction from the input layer to the output layer, one neuron is successively reduced between adjacent deep neural network layers, and each neuron is connected to all neurons in the previous layer of the neural network, each process parameter is used as an input feature of a neuron in the input layer, and each neuron in the intermediate layer is used to calculate the output features of all neurons in the previous layer of the neural network, and the output feature of each neuron in the output layer is the corresponding process result evaluation indicator, and the number of process parameters is greater than the number of process result evaluation indicators; the activation function for the operation of the deep neural network model is given according to the number of layers of the deep neural network model. The deep neural network model of this embodiment is as follows Figure 4 shown.

[0082] Afterwards, a process database is used as training data, and a deep learning method is used to train a deep neural network model. The process database includes historical process parameter data and historical process result data of a specific process.

[0083] According to the training results of the activation function, the weight items and bias items of each layer of the deep neural network model are determined, and finally the construction of the deep neural network model is completed.

[0084] In the specific implementation process, for semiconductor processing machines such as etching machines, the number of process result evaluation indicators is generally less than the number of process parameters. Therefore, the deep neural network model of the present invention is constructed on the basis of the neural network model construction method disclosed in the prior art:

[0085] The number of layers of the neural network model is determined by the difference between the number of input variables (the number of process parameters) and the number of output variables (the number of process result evaluation indicators). An intermediate feature variable is reduced between each layer of the neural network, such as Figure 3 As shown, taking the etcher as an example, if the input process parameters have 6 variables, including: chamber pressure, upper electrode power, lower electrode power, SF6 flow rate, O2 flow rate and etching time, and the final process results have 3 evaluation indicators, including etching depth, side wall angle and etching rate, then a deep neural network model (three-layer neural network) with two layers of intermediate feature variables is constructed.

[0086] The process database of process parameters and process results obtained in normal times is used for deep learning training of the model, and an optimization algorithm similar to the prior art can be used to obtain the best model function y=h(Σ k w k g(Σj w j f(Σ i w i x i +b1)+b2)+b3), that is, the model relationship between process parameters and process results is obtained.

[0087] For practical problems, when the process results / process requirements are known, it is necessary to infer the optimal process formula. According to the deep neural network model of the present invention, the number of indicators of the process results is less than the number of parameters in the process formula. Therefore, this problem is equivalent to solving an indeterminate equation. Theoretically, there are infinite sets of solutions. Therefore, in order to quickly provide the optimal solution, the present invention adopts a self-consistent iterative method to find the optimal solution.

[0088] Step S102: determining whether the difference between each process result evaluation index and the corresponding set process requirement meets the set requirements;

[0089] Step S103: If yes, the input set of process parameters is used as a process recipe for the actual process.

[0090] Specifically, in steps S101-S103, a set of process parameters x is first randomly given. i As the input features of each neuron in the input layer of the deep neural network model, the process result evaluation index y output by the output layer of the deep neural network model is compared with the process requirement y 实际 A comparison is performed, and if it meets the set requirements, the actual process is carried out using this set of process parameters.

[0091] The difference is calculated by the following formula:

[0092] Δ=|(yy 实际 ) / y 实际 |

[0093] Among them, Δ is the difference between the process result evaluation index and the set process requirement, y is the process result evaluation index output by the deep neural network model, and y 实际 To set the process requirements;

[0094] The setting requirement is: the difference between the process result evaluation index and the corresponding set process requirement is between 0 and 10%, preferably 5%.

[0095] Step S104: Otherwise, according to the set process requirements, along the direction from the output layer to the input layer, the characteristic values ​​in each of the intermediate layers and the process parameter values ​​in the input layer are optimized layer by layer based on the gradient algorithm and the self-consistent iteration method, until a set of process parameters is obtained in the input layer so that all process result evaluation indicators output by the deep neural network model meet the corresponding set process requirements, as the process recipe of the actual process.

[0096] Specifically, optimizing the characteristic values ​​in each intermediate layer and the process parameter values ​​in the input layer layer by layer based on the gradient algorithm and the self-consistent iteration method includes:

[0097] Step S201: Finding a process result evaluation index that has the largest difference from the corresponding set process requirement from the output layer;

[0098] Step S202: Finding a process parameter with the largest weight value associated with the process result evaluation index with the largest difference from the intermediate layer adjacent to the output layer;

[0099] Step S203: The weight value and the bias term of the process parameter with the largest weight value are fixed unchanged, and the value of the process parameter with the largest weight value is adjusted step by step using a gradient algorithm to optimize, until the difference between the process result evaluation index with the largest difference in the output layer and the corresponding set process requirement meets the set requirements;

[0100] Step S204: if there are other process result evaluation indicators in the other process result evaluation indicators whose difference degree with the corresponding set process requirements does not meet the set requirements, repeat steps S201-S203 until all process result evaluation indicators in the output layer meet the corresponding set process requirements, and determine the characteristic value of each process parameter in the intermediate layer;

[0101] Step S205: Based on the characteristic value of each process parameter in the intermediate layer obtained in step S204, continue to use the gradient algorithm and self-consistent iteration method to optimize the characteristic values ​​of other intermediate layers and the process parameter values ​​in the input layer layer by layer along the direction from the output layer to the input layer, until a set of process parameters is obtained in the input layer that can make all process result evaluation indicators output by the deep neural network model meet the corresponding set process requirements.

[0102] Preferably, before step S204, the method further includes: if the difference between the process result evaluation index adjusted in step S203 and the set process requirement still cannot meet the set requirement, the process parameter corresponding to the maximum weight value obtained after adjustment and optimization in step S203 is fixed, and the following steps are performed:

[0103] Step S206: Find a process parameter with the second largest weight value related to the corresponding adjusted process result evaluation index from the intermediate layer adjacent to the output layer;

[0104] Step S207: fix the weight value of the process parameter with the second largest weight value unchanged, and use the gradient algorithm to optimize the step-by-step adjustment of the feature value of the process parameter with the second largest weight value until the difference between the corresponding adjusted process result evaluation index in the output layer and the corresponding set process requirement meets the set requirements.

[0105] Preferably, after step S205, the following steps may also be included:

[0106] After step S205, the method further includes:

[0107] Step S208: If after optimizing and adjusting all process parameters related to the process result evaluation index, there are still process result evaluation indexes in the output layer whose difference from the corresponding set process requirements does not meet the set requirements, then:

[0108] A set of process parameters that can minimize the sum of the squares of the differences between each process result evaluation index and the corresponding set process requirements is selected, and the set of process parameters is used as a process recipe for the actual process.

[0109] In this embodiment, the calculation formula of the gradient algorithm is:

[0110]

[0111] Among them, x i is the adjusted process parameter in the middle layer, y is the adjusted process result evaluation index, and θ is the adjustment step size.

[0112] In the specific implementation process of the above steps 201-208, firstly, the process result evaluation index y output by the output layer of the deep neural network model and the process requirement y 实际 After comparison, if it does not meet the set requirements, the output process result evaluation index y and process requirement y 实际 The difference between them is Δ=|(yy 实际 ) / y 实际 |, find out the process result evaluation index that is the most different from the set process requirement, and according to the weight value w of the process parameter corresponding to the process result evaluation index, according to the gradient algorithm Adjust a process parameter x corresponding to the maximum weight value i, where θ is the step length, and this cycle is repeated until the result is self-consistent, that is, the process result obtained according to the model for given process parameters is close to the process requirement (the output process result evaluation index can reach 95% of the set process requirement, and can be between 90% and 100%).

[0113] If the process requirement cannot be met by completely traversing the process parameter with the largest weight value according to the gradient algorithm, the corresponding process parameter should be fixed after taking the eigenvalue of the item with the largest weight value, and the gradient algorithm should be optimized for the process parameter with the second largest weight value. Similarly, if the process requirement cannot be met by completely traversing all process parameters, the process parameter with the smallest sum of squares of all process indicators according to Δ is used for the actual process, that is, a set of process parameters that can minimize the sum of squares of the differences between each process result evaluation indicator and the corresponding set process requirement is selected, and the set of process parameters is used as the process formula for the actual process.

[0114] The following uses an etching machine as an example to further explain the process of finding the optimal solution using the self-consistent iteration method:

[0115] First, a group of process recipes are randomly used, and after inputting into the deep neural network model, the process results are obtained in the output layer. Then, a process result evaluation index with the largest difference from the set process requirements is found in the process results. For example, assuming that the side wall angle has the largest difference, it is necessary to find an item with the largest weight related to the side wall angle in the previous layer of the neural network. Next, the weight value is fixed and the value of this item is stepped according to the gradient algorithm until the side wall angle meets the requirement (such as reaching 95% of the set value). At this time, other process result evaluation indicators output by the model may also have changed. If other process result evaluation indicators also exceed the set requirements, the process result evaluation indicators are still optimized according to the above-mentioned similar methods until all process result evaluation indicators meet the set requirements. At the same time, a group of intermediate layer feature values ​​adjacent to the group of process result evaluation indicators are also obtained. A similar method is used to optimize the values ​​of the intermediate layer features of this layer, and so on, until the process parameter data of the leftmost input layer, i.e., the process parameter recipe group, is obtained.

[0116] The embodiment of the present invention further provides a semiconductor process recipe automatic acquisition system, comprising:

[0117] A deep neural network model, wherein the deep neural network model is a deep neural network with a multi-layer structure that has been trained. The deep neural network model can output accurate process result evaluation indicators corresponding to the process parameter group based on the input process parameters;

[0118] The calculation module is used to execute the semiconductor process recipe automatic acquisition method of the above embodiment.

[0119] The present invention further provides a semiconductor process equipment, including the semiconductor process recipe automatic acquisition system of the above embodiment.

[0120] The solution of the present invention can automatically provide corresponding process recipes by software according to process requirements, thereby improving the automation level of the etching machine. That is, the original process engineer inputs process parameters based on experience and repeatedly debugs to directly input process requirements, and then the software provides corresponding process recipes.

[0121] The present invention is further explained below by means of specific embodiments. Specific embodiments

[0123] The present invention is applied to an etching machine, and the specific process is to realize an etching process with a critical dimension of 3 microns. Figure 5 shown.

[0124] The etching pattern needs to use the Bosch process, so the process parameter x i Including deposition step chamber pressure, deposition step upper electrode center power, deposition step upper electrode edge power, deposition step center C4F8 flow, deposition step edge C4F8 flow, deposition step start time, deposition step start time, etching step chamber pressure, etching step upper electrode center power, etching step upper electrode edge power, etching step lower electrode start power, etching step lower electrode final power, etching step center C4F8 flow, etching step edge C4F8 flow, etching step start time, etching step final time; process evaluation index y includes etching depth, upper opening size, lower opening size, selectivity, upper side wall roughness (scallop size), lower side wall roughness (scallop size), etc.

[0125] Taking groove etching as an example, Figure 6-Figure 11 The process database shown is used to train the deep neural network model of the above embodiment.

[0126] from Figure 6-Figure 11 The weight value trend of the model function that can be trained in the process database shown is as follows:

[0127] The chamber pressure has a strong correlation with the etching morphology, and the weight value is large. The final bias power also has a strong correlation with the morphology, and the weight value is large.

[0128] It should be noted that, in addition to the process database established by itself in this embodiment, the process trends in the above process database have also been reported in various documents, so a corresponding process database can also be established by consulting the documents.

[0129] After getting the trained model, we use the recursive iterative method to Fig.12The process results shown in the figure are used to infer the process formula that meets the process requirements. For example, when the etching angle is inferred, it has a large relationship (weight) with the final lower electrode power and the final single-step time of etching. According to the scheme of the present invention, the two process parameters will be optimized by gradient algorithm at the same time, and the final process formula is shown in Table 1:

[0130] Table 1 Etching process recipe

[0131]

[0132] * The lower electrode power increases from the starting value to the final value

[0133] ** Single step time increases from the starting value to the final value

[0134] For the hole etching process, the process is similar. Figure 13-Figure 18 The process database shown in the figure is used to train the above deep neural network model. After obtaining the trained model, the recursive iterative method is used according to Fig.19 The process results shown in Table 2 are used to infer the process formula:

[0135] Table 2 Etching process formula

[0136]

[0137] In summary, the solution of the present invention can realize the automatic provision of process recipes according to process results in the etching machine, thereby improving the automation rate of the machine. It should be noted that the present invention is not only applicable to etching machines, but also to other process processes in the semiconductor field, such as PVD, CVD, furnace tubes, cleaning machines, etc.

[0138] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for automatically acquiring a semiconductor process recipe, characterized in that: include: A set of process parameters required by the process is randomly input into the constructed deep neural network model to obtain a process result evaluation index corresponding to the process parameter group; wherein the deep neural network model is a multi-layer deep neural network that has completed training, and the deep neural network model includes an input layer, at least one intermediate layer and an output layer, the input layer is used to input a plurality of the process parameters, and the output layer is used to output the corresponding process result evaluation index; Determine whether the difference between each of the process result evaluation indicators and the corresponding set process requirements meets the set requirements; If yes, using the set of process parameters as a process recipe for performing the actual process; Otherwise, according to the set process requirements, along the direction from the output layer to the input layer, the characteristic values ​​in each of the intermediate layers and the process parameter values ​​in the input layer are optimized layer by layer based on the gradient algorithm and the self-consistent iteration method, until a set of process parameters is obtained in the input layer so that all process result evaluation indicators output by the deep neural network model meet the corresponding set process requirements, as the process recipe of the actual process; The step of optimizing the characteristic values ​​in each intermediate layer and the process parameter values ​​in the input layer layer by layer based on the gradient algorithm and the self-consistent iteration method includes: Step S1: Finding a process result evaluation index that has the largest difference from the corresponding set process requirement from the output layer; Step S2: Finding a process parameter with the largest weight value associated with the process result evaluation index with the largest difference from the intermediate layer adjacent to the output layer; Step S3: The weight value and bias term of the process parameter with the largest weight value are fixed unchanged, and the value of the process parameter with the largest weight value is adjusted step by step using a gradient algorithm to optimize, until the difference between the process result evaluation index with the largest difference in the output layer and the corresponding set process requirement meets the set requirements; Step S4: If there are other process result evaluation indicators whose difference degree with the corresponding set process requirements does not meet the set requirements among the other process result evaluation indicators, repeat steps S1-S3 until all process result evaluation indicators in the output layer meet the corresponding set process requirements, and determine the characteristic value of each process parameter in the intermediate layer; Step S5: Based on the characteristic value of each process parameter in the intermediate layer obtained in step S4, continue to use the gradient algorithm and self-consistent iteration method to optimize the characteristic values ​​of other intermediate layers and the process parameter values ​​in the input layer layer by layer along the direction from the output layer to the input layer, until a set of process parameters is obtained in the input layer that can make all process result evaluation indicators output by the deep neural network model meet the corresponding set process requirements.

2. The method for automatically acquiring semiconductor process recipes according to claim 1, characterized in that: The method for constructing the deep neural network model includes: The number of layers and structure of the deep neural network model are determined according to the number of the process parameters and the number of the process result evaluation indicators; wherein the number of the intermediate layers is the difference between the number of the process parameters and the number of the process result evaluation indicators, each layer of the deep neural network in the deep neural network includes a plurality of neurons, and along the direction from the input layer to the output layer, one neuron is successively reduced between adjacent deep neural network layers, each neuron is connected to all neurons in the previous layer of the neural network, each process parameter is used as an input feature of a neuron in the input layer, each neuron in the intermediate layer is used to calculate the output features of all neurons in the previous layer of the neural network, and the output feature of each neuron in the output layer is the corresponding process result evaluation indicator, and the number of the process parameters is greater than the number of process result evaluation indicators; An activation function for running the deep neural network model is given according to the number of layers of the deep neural network model.

3. The method for automatically acquiring semiconductor process recipes according to claim 2, characterized in that: The method for constructing the deep neural network model also includes: Using a process database as training data and using a deep learning method to train the deep neural network model, the process database includes historical process parameter data and historical process result data of a specific process; The weight term and bias term of each layer of the deep neural network in the deep neural network model are determined according to the training result of the activation function.

4. The method for automatically acquiring semiconductor process recipes according to claim 1, characterized in that: Before step S4, the method further includes: if the difference between the process result evaluation index adjusted in step S3 and the set process requirement still cannot meet the set requirement, the process parameter corresponding to the maximum weight value obtained after adjustment and optimization in step S3 is fixed, and the following steps are performed: Step S6: Finding a process parameter with the second largest weight value related to the corresponding adjusted process result evaluation index from the intermediate layer adjacent to the output layer; Step S7: fix the weight value of the process parameter with the second largest weight value unchanged, and use the gradient algorithm to optimize the step-by-step adjustment of the feature value of the process parameter with the second largest weight value until the difference between the corresponding adjusted process result evaluation index in the output layer and the corresponding set process requirement meets the set requirements.

5. The method for automatically acquiring semiconductor process recipes according to claim 4, characterized in that: After step S5, the method further includes: Step S8: If after optimizing and adjusting all process parameters related to the process result evaluation index, there are still process result evaluation indexes in the output layer whose difference from the corresponding set process requirements does not meet the set requirements, then: A set of process parameters that can minimize the sum of the squares of the differences between each process result evaluation index and the corresponding set process requirements is selected, and the set of process parameters is used as a process recipe for the actual process.

6. The method for automatically acquiring semiconductor process recipes according to any one of claims 1, 4 and 5, characterized in that: The calculation formula of the gradient algorithm is: Among them, x i is the adjusted process parameter in the middle layer, y is the adjusted process result evaluation index, and θ is the adjustment step size.

7. The method for automatically acquiring semiconductor process recipes according to any one of claims 1, 4 and 5, characterized in that: The degree of difference is calculated by the following formula: Δ=|(yy 实际 ) / and 实际 | Among them, Δ is the difference between the process result evaluation index and the set process requirement, y is the process result evaluation index output by the deep neural network model, and y 实际 For a given process requirement; The setting requirement is that the difference between the process result evaluation index and the corresponding set process requirement is between 0 and 10%.

8. A semiconductor process recipe automatic acquisition system, characterized in that: include: A deep neural network model, wherein the deep neural network model is a deep neural network with a multi-layer structure that has been trained, and the deep neural network model can output accurate process result evaluation indicators corresponding to the process parameter group based on the input process parameters; A computing module, used to execute the method for automatically acquiring a semiconductor process recipe according to any one of claims 1 to 7.

9. A semiconductor process equipment, characterized in that: Includes the semiconductor process recipe automatic acquisition system as described in claim 8.

Images