Method, System, and Semiconductor Process Apparatus for Obtaining a Semiconductor Process Recipe

Abstract

The present disclosure provides a semiconductor processing apparatus, a tray and a cassette, and relates to the field of semiconductor equipment. The semiconductor processing apparatus includes a process chamber, a transport chamber and a load chamber, the load chamber is used to accommodate a cassette, the cassette is used to place a tray, a transport device for transporting the tray between the process chamber and the load chamber is provided in the transport chamber, an image recognition device is provided at the boundary between the load chamber and the transport chamber to recognize a predetermined characteristic image on the tray in the process of the transport device moving the tray and determine characteristic information of the tray, and the semiconductor processing apparatus further includes a controller for acquiring the characteristic information, searching for transport parameters corresponding to the characteristic information from a predetermined tray information library, and controlling the transport device to transport the tray according to the transport parameters. The tray is applied to the semiconductor processing apparatus. The cassette is used to place the tray. The present application can solve problems such as low transport accuracy and laborious work.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor manufacturing, and more specifically, to a method, a system, and a semiconductor process apparatus for obtaining a semiconductor process recipe.

Background Art

[0002] In the field of microelectronics manufacturing, plasma etching is an important process in the device processing procedure. For example, by etching trench gates on discrete transistor devices, the short-channel effect of the devices can be suppressed, and problems such as the reduction in mobility caused by electron scattering in the channel can be solved. Therefore, discrete transistor devices have undergone a transformation from conventional planar gates to trench gates and are applied to silicon-based and silicon carbide-based metal-oxide-semiconductor field-effect transistor (MOSFET) devices. Also, for example, by etching superjunctions on discrete transistor devices, the on-resistance can be effectively reduced, and the switching speed of the devices can be improved. Since the drive current is small when adopting a laterally structured superjunction, the mainstream products in the market adopt a vertically structured superjunction, and the process technology has developed from a multi-epitaxial process to a deep trench process. Therefore, it is necessary to adopt silicon deep etching. Furthermore, for example, in gallium nitride high electron mobility transistor (HEMT) devices, as one of the methods to realize normally-off devices, a gate recess is etched on the gate, and the normally-off of the device is realized by changing the surface energy level of the gallium nitride of the gate through the recess. Discrete devices that originally required trench etching are also developing towards high aspect ratios and high verticality. For example, in conventional silicon capacitor devices, the deeper the trench, the more effectively the space of the device can be utilized, the capacitance value can be increased, and the higher the verticality, the more the actual device can match the theoretical model and better meet the design requirements. Additionally, for example, in advanced packaging, by etching through-silicon vias (TSV) structures, the performance of the devices can be improved, the power consumption can be reduced, and the devices can be miniaturized.

[0003] Therefore, due to the development in fields such as semiconductor discrete devices and advanced packaging, the requirements for silicon deep etching are becoming increasingly high. Since the silicon deep-etched microstructure has a large aspect ratio and high perpendicularity, it is difficult to form with conventional wet etching and needs to be formed by a dry etching method. To obtain a deep and perpendicular silicon microstructure, the "Bosch" process, which is a dry etching process divided by time, is mainly used. This process provides sidewall passivation protection and a chemical reaction of fluorine-based plasma by inducing a fluorocarbon polymer with plasma to etch silicon downward, and the sidewall passivation protection (deposition step) and the fluorine-based plasma chemical reaction (etching step) are performed alternately. Since the "Bosch" process performs deposition and etching alternately, a scallop structure is inevitably formed, and the sidewall roughness increases.

[0004] Generally, the scallop at the upper part of the silicon deep-etched structure has the largest size, the roughest sidewall, and obvious scallops and lateral wrinkles can be seen. Therefore, it is necessary to find a method to reduce the roughness of the upper part of the silicon deep-etched structure. In addition to considering the sidewall roughness, constant attention should be paid to the etching pattern (sidewall angle). The reason is that after the formation of the silicon deep-etched structure, it is difficult for the plasma to penetrate into the silicon deep-etched structure, and the etching product is also difficult to be extracted from the silicon deep-etched structure. In many cases, it is easy to form a silicon deep-etched structure with a large upper part and a small lower part. At this time, it is necessary to improve the mean free path of the plasma by adjusting the process parameters to assist the movement of the plasma to the wafer. Also, the etching depth, etching rate, mask selection ratio, etc. are also indicators for evaluating the quality of the silicon deep etching effect. To obtain a good silicon deep etching effect, it is necessary to optimize the process parameters, that is, adjust the process recipe. Currently, mainly the process personnel manually try various recipes one by one to determine a recipe that can meet the process requirements. This process is very complicated and inefficient, and often cannot achieve automation.

Summary of the Invention

Problems to be Solved by the Invention

[0005] An object of the present invention is to provide a method, a system, and a semiconductor process apparatus for automatically obtaining a corresponding process recipe according to predetermined process requirements and improving the acquisition efficiency of the process recipe.

Means for Solving the Problems

[0006] To achieve the above object, the present invention randomly inputs a group of process parameters required for a process into a constructed deep neural network model to obtain a process result evaluation index corresponding to the group of process parameters, where the group of process parameters includes a plurality of process parameters, the deep neural network model is a trained multi-layer deep neural network, 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 the group of process parameters, and the output layer is used to output the process result evaluation index corresponding to the group of process parameters; judging whether the difference degree between each process result evaluation index and a set process requirement corresponding thereto satisfies a set requirement; if YES, using the group of process parameters as a process recipe for performing an actual process; If it is NO, according to the process requirements, along the direction from the output layer to the input layer, until all process result evaluation indicators output from the deep neural network model in the input layer satisfy the process requirements based on the gradient algorithm and the self-collision-free iteration method, optimize the eigenvalues in each intermediate layer and the numerical values of each process parameter in the input layer layer by layer, and use the process parameter group that satisfies the process requirements as a process recipe for performing the actual process. A method for obtaining a semiconductor process recipe is provided, including the step of

[0007] Optionally, the method for constructing the deep neural network model is Determining the number of layers and structure of the deep neural network model according to the number of process parameters included in the process parameter group and the number of process result evaluation indicators corresponding to the process parameter group, where 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 model includes a plurality of neurons, along the direction from the input layer to the output layer, one neuron sequentially decreases between adjacent deep neural networks, each neuron is connected to all neurons in the previous layer of the deep neural network, each process parameter is used as the input feature of one 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 deep neural network, the output feature of each neuron in the output layer is the process result evaluation indicator, and the number of process parameters included in the process parameter group is larger than the number of process result evaluation indicators corresponding to the process parameter group. According to the number of layers of the deep neural network model, providing an activation function executed by the deep neural network model.

[0008] Optionally, the method for constructing the deep neural network model is using the process database as training data and training the deep neural network model using a deep learning method, where the process database includes historical process parameter data and historical process result data of a specific process, and further including determining weight terms and bias terms of deep neural networks in each layer in the deep neural network model according to the training result of the activation function.

[0009] Optionally, the step of optimizing the eigenvalues in each intermediate layer and the numerical values of each process parameter in the input layer layer by layer based on a gradient algorithm and a self-collision-free iteration method is step S1 of finding, from the output layer, one process result evaluation index with the largest difference from the process requirement, step S2 of finding, from the intermediate layer adjacent to the output layer, one process parameter with the largest numerical value of the weight term related to the process result evaluation index with the largest difference, fixing the numerical values of the weight term and the bias term of one process parameter with the largest numerical value of the weight term, and using a gradient algorithm to optimize the eigenvalue of one process parameter with the largest numerical value of the weight term until the difference between the process result evaluation index with the largest difference in the output layer and the process requirement meets the set requirement, and adjusting step S3, At this time, if there are other process result evaluation indexes that do not meet the set requirement of the difference from the process requirement among other process result evaluation indexes, repeating steps S1 to S3 until all process result evaluation indexes in the output layer meet the process requirement, and at the same time, step S4 of determining the eigenvalues of each process parameter in the intermediate layer, Based on the eigenvalues of each process parameter in the intermediate layer obtained in step S4, along the direction from the output layer to the input layer, continue to optimize the eigenvalues of each process parameter in other intermediate layers and the numerical values of the process parameters in the input layer layer by layer using a gradient algorithm and a self-collision-free iteration method until a group of process parameters that enables all process result evaluation indicators output from the deep neural network model in the input layer to meet the process requirements is obtained, step S5.

[0010] Optionally, after step S3, determine whether the difference degree between the process result evaluation indicator correspondingly adjusted in step S3 and the process requirement meets the set requirement. If it meets, execute step S4. If it does not meet, fix one process parameter with the largest numerical value of the weight term obtained by adjustment and optimization in step S3. Step S6 of finding one process parameter with the second largest numerical value of the weight term related to the process result evaluation indicator correspondingly adjusted in step S3 from the intermediate layer adjacent to the output layer. Fix the numerical value of the weight term of one process parameter with the second largest numerical value of the weight term, and optimize and adjust the eigenvalue of one process parameter with the second largest numerical value of the weight term using a gradient algorithm until the difference degree between the process result evaluation indicator correspondingly adjusted in step S3 in the output layer and the process requirement meets the set requirement, step S7.

[0011] Optionally, after step S5, After optimizing and adjusting all process parameters related to the process result evaluation indicator, if there is still a process result evaluation indicator in the output layer that does not meet the set requirement of the difference degree from the process requirement, Select a process parameter that enables the sum of the squares of the differences between each process result evaluation index and the process requirement to be minimized, and form a new process parameter group. Further include step S8 of using the new process parameter group as a process recipe for performing an actual process.

[0012] Optionally, the calculation formula of the gradient algorithm is as follows.

[0013] x i =x i -θ×(∂y / ∂x i ) In the formula, x i is the adjusted process parameter in the intermediate layer, y is the adjusted process result evaluation index, and θ is the adjustment step.

[0014] Optionally, the difference is calculated by the following formula, As the set requirement, the difference between the process result evaluation index and the process requirement set corresponding to it is 0 to 10%.

[0015] Δ=|(y - y 実際 ) / y 実際 | In the formula, Δ is the difference between the process result evaluation index and the set process requirement, y is the process result evaluation index output from the deep neural network model, and y 実際 is the given process requirement.

[0016] In the second aspect, the present invention is a trained deep neural network with a multi-layer structure, and is capable of outputting an accurate process result evaluation index corresponding to the process parameter group based on the input process parameter, and further provides a semiconductor process recipe automatic acquisition system including a calculation module for executing the semiconductor process recipe acquisition method according to the first aspect.

[0017] In a third aspect, the present invention further provides a semiconductor processing apparatus including the automatic acquisition system of the semiconductor process recipe described in the second aspect.

[0018] The beneficial effects of the present invention are as follows.

[0019] When process requirements are given, 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, all process result evaluation indicators output from the deep neural network model are enabled to meet the process requirements. The gradient algorithm and the self-collision-free iteration method are used to optimize and adjust the eigenvalues in each intermediate layer and the process parameter group of the input layer layer by layer until the process parameter group that can meet the process requirements is derived in the reverse direction as the process recipe of the actual process. Compared with the conventional method of repeatedly manually debugging the process recipe parameters, the present invention can automatically obtain the process recipe parameters that meet the process requirements by the deep neural network model according to the process requirements, improve the acquisition efficiency of the process recipe, and improve the degree of automation of the etching apparatus.

[0020] The system of the present invention has other characteristics and advantages, and these characteristics and advantages will become clear from the drawings incorporated herein and the subsequent specific embodiments, or will be described in detail in the drawings incorporated herein and the subsequent specific embodiments, and both these drawings and the specific embodiments are used to interpret the specific principles of the present invention.

Brief Description of the Drawings

[0021] By describing the exemplary embodiments of the present invention in more detail with reference to the drawings, the above and other objects, features and advantages of the present invention will become clearer. In the exemplary embodiments of the present invention, the same reference numerals generally denote the same members.

[0022]

Figure 1

Figure 2

Figure 3A

Figure 3B

Figure 3C

Figure 3D

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

Figure 18

Figure 19

Embodiments for Carrying Out the Invention

[0023] Prior art 1 discloses a deep neural network model applied to a photolithography apparatus, and the deep neural network model is shown in FIG. 1. Based on the deep neural network model that completes deep learning, the photolithography process result can be automatically output according to the photolithography process parameters input into the model.

[0024] The deep neural network model is prior art. Specifically, first, the activation function of the model needs to be defined, and a linear transformation needs to be performed on the parameters of the input layer. It is y = f(Σ i w i x i + b), where in the formula, x i is the input eigenvector (feature), w i is the weight of each eigenvalue, and b is the bias term. This is the case of the simplest neural network having only one layer. When there is a multi-layer neural network, the activation function of the two-layer model is y = g(Σ j w j f(Σ i w i x i + b1) + b2), and the activation function of the three-layer model is y = h(Σ k w k g(Σ j w j f(Σ i w i x i + b1) + b2) + b3), and so on.

[0025] Next, the input-output values (x i , y) obtained using the activation function of the model and the actual dataset (xi , y r ) and the sum of squared differences with respect to it, the objective function J(w i , b) = (Σ i w i x i + b - y r ) 2 is defined. If there are n actual data sets (x i,n , y r,n ), it is necessary to take the arithmetic mean of the objective function, and J(w i , b) = (1 / n)Σ n (Σ i w i x i,n + b - y r,n ) 2 .

[0026] In the neural network model, data training is an important part, mainly regarding parameter optimization. Currently, the mainstream optimizers mainly include three types: stochastic gradient descent, stochastic gradient descent with momentum, and Adam optimization method. Taking stochastic gradient descent as an example, its algorithm is w i = w i - η × [∂J(w i , b) / ∂w i , b = b - η × [∂J(w i , b) / ∂b].

[0027] That is, calculate the partial derivative function for the relevant parameters, move according to the product of the step η and the partial derivative function each time, and finally, the objective function obtains the minimum value, that is, the deviation between the data predicted by the model and the actual value stabilizes until it is the smallest. In the case of a multi-layer neural network, it is possible to execute in the reverse direction layer by layer to minimize the objective function of each layer, that is, the deviations between the data predicted by the model and the actual values are all the smallest.

[0028] The inventor found through research that the prior art 1 has the following problems.

[0029] The neural network model only provides forward learning and prediction, that is, it can only derive the process result in the forward direction from the process parameter recipe. When performing forward derivation, the variable in the gradient optimization algorithm is the weight value w i and cannot realize the reverse derivation of obtaining process parameters according to process requirements, and is not accompanied by machine automation.

[0030] The photolithography apparatus has high accuracy requirements and its optimizer algorithm is complex (on the other hand, other semiconductor process apparatuses such as etching apparatuses have a certain tolerance range for accuracy requirements with respect to the calculation processing speed).

[0031] Prior art 2 discloses a self-collision-free iteration method, that is, when solving a transcendental equation, an analytical solution cannot be obtained and only a numerical solution can be obtained. When obtaining the numerical solution, a continuous iteration method is used and finally converges to the optimal solution to form self-collision-free.

[0032] When an operator of an equation is a function of an unknown, the self-collision-free iteration method can only be used. First, the operator is obtained for a set of assumed numerical values of the unknown, substituted into the original equation to obtain another set of numerical solutions of the unknown, and the iteration is repeated until the assumed numerical value of the unknown completely coincides with the finally calculated numerical value of the unknown, that is, until self-collision-free. This method is used when solving the Hartree-Fock equation in the field of quantum calculation.

[0033] This method is not currently applied to the neural network model and is not applied to the field of semiconductor process apparatuses either.

[0034] Prior art 3 discloses a neural network model for predicting the results of an etching process. As shown in FIG. 2, the inventors have studied and found that only a single-layer neural network model is used in this means, which belongs to an ordinary neural network rather than a deep neural network, and can only predict the results of the etching process, and cannot derive the process parameters in the reverse direction from the results of the etching process.

[0035] The present invention can automatically provide a process recipe during a related process (such as dry etching, etc.) according to actual process requirements by using a software algorithm. In a specific application, when a user needs to obtain a silicon deep trench structure with a specific etching depth, sidewall angle, sidewall roughness, and mask selection ratio, and expects to complete it at a specific etching rate, the means of the present invention can automatically give the corresponding process parameter recipe such as chamber pressure, upper and lower electrode power, intake flow rate, and etching time.

[0036] Hereinafter, the present invention will be described in more detail with reference to the drawings. Although the preferred embodiments of the present invention are shown in the drawings, it should be understood that the present invention is not limited to the embodiments described herein and can be realized in various forms. Rather, these embodiments are provided so that the present invention can be made more clear and complete, and can fully convey the scope of the present invention to those skilled in the art.

[0037] FIGS. 3A to 3D are some flowcharts of the steps of a method for obtaining a semiconductor process recipe according to an embodiment of the present invention. Referring to FIGS. 3A to 3D together, the method for obtaining a semiconductor process recipe includes the following steps S101 to S104.

[0038] Step S101: Randomly input a group of process parameters (including a plurality of process parameters) required for the process into a constructed deep neural network model, and obtain a process result evaluation index corresponding to the group of process parameters.

[0039] The deep neural network model is a trained multi-layer deep neural network. 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 group of process parameters, and the output layer is used to output a process result evaluation index corresponding to the group of process parameters. That is, each of the input layer, at least one intermediate layer, and the output layer is a layer of deep neural network.

[0040] Specifically, before the above step S101, first, a deep neural network model is constructed. As shown in FIG. 3B, the construction method of the deep neural network model includes the following steps S201 to S202.

[0041] Step S201: Determine the number of layers and the structure of the deep neural network model according to the number of process parameters included in the group of process parameters and the number of process result evaluation indexes corresponding to the group of process parameters.

[0042] 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 included in the group of process parameters and the number of process result evaluation indexes corresponding to the group of process parameters. Each layer of deep neural network in the deep neural network model includes a plurality of neurons. Along the direction from the input layer to the output layer, one neuron decreases sequentially between adjacent deep neural networks. Each neuron is connected to all neurons in the previous layer of deep neural network. Each process parameter is used as the input feature of one 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 deep neural network. The output feature of each neuron in the output layer is the corresponding process result evaluation index. The number of process parameters included in the group of process parameters is greater than the number of process result evaluation indexes corresponding to the group of process parameters.

[0043] Step S202: According to the number of layers of the deep neural network model, provide an activation function to be executed by the deep neural network model.

[0044] The deep neural network model of this embodiment is shown in FIG. 4. Subsequently, optionally, the method for constructing the deep neural network model further includes the following steps S203 to S204.

[0045] Step S203: Use the process database as training data and use a deep learning method to train the deep neural network model.

[0046] The process database includes historical process parameter data and historical process result data of a specific process.

[0047] Step S204: Determine the weight terms and bias terms of each layer of the deep neural network in the deep neural network model according to the training result of the activation function, and finally complete the construction of the deep neural network model.

[0048] In a specific implementation process, in the case of a processing machine in the semiconductor field such as an etching device, the number of process result evaluation indicators is generally less than the number of process parameters. Therefore, based on the method for constructing a neural network model disclosed in the prior art 1, the deep neural network model of the present invention can be constructed.

[0049] 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). One intermediate feature variable decreases between the deep neural networks of each layer. As shown in FIG. 4, when taking an etching apparatus as an example, the input process parameter group includes six process parameters, namely, chamber pressure, upper electrode power, lower electrode power, SF6 flow rate, O2 flow rate, and etching time, that is, six variables. Finally obtained, there are three evaluation indicators, including etching depth, sidewall angle, and etching rate, in the process result corresponding to the process parameter group. A deep neural network model (a three-layer deep neural network) with two layers of intermediate feature variables is constructed.

[0050] Use the process database of process parameters and process results obtained during normal times for the deep learning training of the model, and use the same optimization algorithm as in the prior art 1 to obtain the optimal model function y = h(Σ k w k g(Σ j w j f(Σ i w i x i +b1)+b2)+b3), that is, the model relationship between the process parameters and the process results is obtained.

[0051] In an actual problem, when the process result / process requirement is known, it is necessary to derive the preferred process recipe in the reverse direction. As can be seen from the deep neural network model of the present invention, since the number of indicators of the process result is less than the number of process parameters in the process recipe, this problem is equivalent to solving an indeterminate equation, and theoretically there are infinite solutions. Therefore, in order to quickly find the optimal solution, the present invention uses a self-collision-free iterative method to find the optimal solution.

[0052] Step S102: Determine whether the difference degree between each process result evaluation indicator and the set process requirement meets the set requirement.

[0053] If it is YES in step S103, use the input process parameter group as a process recipe for performing the actual process.

[0054] Specifically, in steps S101 to S103, first, one process parameter group (including a plurality of process parameters x i ) is randomly given as the input feature of each neuron in the input layer of the deep neural network model, and the process result evaluation index y output from the output layer of the deep neural network model is compared with the process requirement y 実際 . If the set requirement is satisfied, the actual process is performed using the process parameter group.

[0055] The degree of difference is calculated by the following formula.

[0056] Δ = |(y - y 実際 ) / y 実際 | In the formula, Δ is the degree of difference between the process result evaluation index and the set process requirement, y is the process result evaluation index output from the deep neural network model, y 実際 is the set process requirement, Optionally, as the set requirement, the degree of difference between the process result evaluation index and the corresponding set process requirement is 0 to 10%, preferably 5%.

[0057] If it is NO in step S104, according to the set process requirement, along the direction from the output layer to the input layer, based on the gradient algorithm and the self-collision-free iteration method, optimize the eigenvalues in each intermediate layer and the numerical values of each process parameter in the input layer layer by layer until a process parameter group that enables all the process result evaluation indices output from the deep neural network model in the input layer to satisfy the corresponding set process requirement is obtained, and use the process parameter group that satisfies the process requirement as a process recipe for performing the actual process.

[0058] Specifically, as shown in FIG. 3C, the step of optimizing the eigenvalue in each of the intermediate layers and the process parameter value in the input layer based on the gradient algorithm and the self-collision-free iteration method layer by layer includes the following steps S301 to S 303.

[0059] Step S301: Find, from the output layer, one process result evaluation index with the largest difference from the correspondingly set process requirement.

[0060] Step S302: Find, from the intermediate layer adjacent to the output layer, one process parameter with the largest numerical value of the weight term related to the process result evaluation index with the largest difference.

[0061] Step S303: Fix the numerical value of the weight term and the numerical value of the bias term of the one process parameter with the largest numerical value of the weight term, and use the gradient algorithm to optimize the eigenvalue of the one process parameter with the largest numerical value of the weight term until the difference between the process result evaluation index with the largest difference in the output layer and the correspondingly set process requirement meets the set requirement, and then adjust the step.

[0062] After step S303, it further includes the following steps S304 to S307.

[0063] Judge whether the difference between the process result evaluation index adjusted correspondingly in step S303 and the set process requirement meets the set requirement. If it meets, execute step S304; if it does not meet, execute step S306.

[0064] Step S304: At this time, if there are other process result evaluation indexes whose differences from the correspondingly set process requirements do not meet the set requirements among other process result evaluation indexes, repeat steps S301 to S303 until all process result evaluation indexes in the output layer meet the correspondingly set process requirements, and at the same time, determine the eigenvalues of each process parameter in the intermediate layer.

[0065] Step S305: Based on the eigenvalues of each process parameter in the intermediate layer obtained in Step S304, along the direction from the output layer to the input layer, continue to optimize the eigenvalues of each process parameter in other intermediate layers and the numerical values of the process parameters in the input layer layer by layer using the gradient algorithm and the self-collision-free iteration method until a group of process parameters that enables all process result evaluation indicators output from the deep neural network model in the input layer to meet the set process requirements corresponding to them is obtained.

[0066] Step S306: Fix one process parameter corresponding to the weight term with the largest numerical value obtained by adjustment and optimization in Step S303, and find one process parameter with the second-largest numerical value among the weight terms related to the process result evaluation indicator adjusted correspondingly in Step S303 from the intermediate layer adjacent to the output layer.

[0067] Step S307: Fix the numerical value of the weight term of one process parameter with the second-largest numerical value, and use the gradient algorithm to optimize the eigenvalue of one process parameter with the second-largest numerical value of the weight term and perform step adjustment until the difference degree between the process result evaluation indicator adjusted correspondingly in Step S303 in the output layer and the set process requirements meets the set requirements.

[0068] In some other selectable embodiments, after completing Step S303, without determining whether the difference degree between the process result evaluation indicator adjusted correspondingly in Step S303 and the set process requirements meets the set requirements, directly execute Step S304, and Steps S306 and S307 may also be omitted.

[0069] Preferably, after Step S305, Step S308 may further be included.

[0070] Step S308: After optimizing and adjusting all process parameters related to the process result evaluation index, if there is still a process result evaluation index that does not meet the set requirements with the difference degree from the corresponding process requirements set in the output layer,

[0071] Select the process parameters that enable the sum of squares of the differences between each process result evaluation index and the corresponding set process requirements to be the smallest, form a new group of process parameters, and use this new group of process parameters as the process recipe for performing the actual process.

[0072] In this embodiment, the calculation formula of the gradient algorithm is the following formula.

[0073] x i =x i -θ×(∂y / ∂x i ) In the formula, x i is the adjusted process parameter in the intermediate layer, y is the adjusted process result evaluation index, and θ is the adjustment step.

[0074] In the specific implementation process of the above steps S 301 to step S 308, first, compare the process result evaluation index y output from the output layer of the deep neural network model with the process requirement y 実際 . If it does not meet the set requirements, according to the difference degree Δ = |(y - y 実際 ) / y 実際 | between the output process result evaluation index y and the process requirement y 実際 , find one process result evaluation index with the largest difference degree from the set process requirements, and according to the magnitude of the numerical value w of the weight term of the process parameter group corresponding to this process result evaluation index, use the gradient algorithm x i =x i -θ×(∂y / ∂x i ) (θ is the step) to find the specific process parameter x i of the process parameter group corresponding to the numerical value of the maximum weight term.Adjust it until the result becomes self-collision-free, that is, until the process result obtained by the deep neural network model according to the given process parameters is substantially consistent with the process requirements (the output process result evaluation index only needs to reach 95% of the set process requirements, and the value can be between 90% and 100%), and repeat this process.

[0075] If the process requirements cannot be met even after completely traversing the process parameter with the largest numerical value of the weight term according to the gradient algorithm, then step S Fix one process parameter with the largest numerical value of the weight term obtained by adjustment and optimization in step 303, perform optimization on the eigenvalue of the process parameter with the second largest numerical value of the weight term using the gradient algorithm, and so on. If the process requirements cannot be met even after completely traversing all process parameters, then find the process parameter with the smallest sum of squares for all process indicators according to the difference Δ between the process result evaluation index y and the process requirement y 実際 and perform the actual process, that is, select a set of process parameters that can minimize the sum of squares of the differences between each process result evaluation index and the corresponding set process requirement, form a new group of process parameters, and use the new group of process parameters as the process recipe for performing the actual process.

[0076] The following further describes the process of obtaining the optimal solution using the above self-collision-free iteration method, taking an etching device as an example.

[0077] First, randomly use a group of process parameters in the process recipe, input it into the deep neural network model, and obtain the process result at the output layer. Next, find one process result evaluation index with the largest difference from the set process requirements from the process results. For example, assuming that the difference in the sidewall angle is the largest, at this time, search for one maximum weight term related to the sidewall angle in the previous layer of the deep neural network, and then fix the value of this weight term. Adjust the value of this weight term step by step according to the gradient algorithm until the sidewall angle meets the requirements (for example, meets 95% of the set value). At this time, other process result evaluation indicators output from the model may also change. If other process result evaluation indicators also exceed the set process requirements, optimize the process result evaluation indicators in the same way as the above method until all process result evaluation indicators meet the set process requirements. At the same time, obtain the eigenvalue of the process parameters in a set of intermediate layers adjacent to the value of the process result evaluation indicator. Similarly, obtain the data of the above process parameter group in the process recipe of the leftmost input layer in FIG. 4, and optimize the eigenvalue of the process parameters in this intermediate layer in the same way until the data of the above process parameter group in the process recipe of the leftmost input layer in FIG. 4 is obtained.

[0078] The embodiment of the present invention is a trained deep neural network with a multi-layer structure, and a deep neural network model capable of outputting an accurate process result evaluation index corresponding to the input process parameter group based on the input process parameter group, and a calculation module for executing the method for obtaining a semiconductor process recipe according to the above embodiment, and further provides a semiconductor process recipe automatic acquisition system including the same.

[0079] The present invention further provides a semiconductor process apparatus including the semiconductor process recipe acquisition system of the above embodiment.

[0080] In the means of the present invention, according to process requirements, a corresponding process recipe can be automatically generated by software, improving the degree of automation of the etching apparatus. That is, originally, process engineers input process parameters according to experience and repeat debugging. Now, it has evolved to directly input process requirements and further generate a corresponding process recipe by software.

[0081] Hereinafter, the present invention will be further interpreted and described with specific examples.

[0082] Specific Examples The present invention is applied to an etching apparatus, and the specific process was to realize an etching process with a critical dimension of 3 microns. The etching process with a critical dimension of 3 microns is shown in FIG. 5.

[0083] Since the Bosch process needs to be used for the etching pattern, a plurality of process parameters x among the process parameter group i are the chamber pressure in the deposition step, the central power of the upper electrode in the deposition step, the edge power of the upper electrode in the deposition step, the central C4F8 flow rate in the deposition step, the edge C4F8 flow rate in the deposition step, the start time of the deposition step, the deposition step Final time, the chamber pressure in the etching step, the central power of the upper electrode in the etching step, the edge power of the upper electrode in the etching step, the start power of the lower electrode in the etching step, the final power of the lower electrode in the etching step, the central C4F8 flow rate in the etching step, the C4F8 flow rate at the edge of the etching step, the start time of the etching step, and the final time of the etching step. The process Result evaluation index y includes etching depth, upper opening dimension, lower opening dimension, selectivity, upper sidewall roughness (scallop dimension), lower sidewall roughness (scallop dimension), etc.

[0084] Taking trench etching as an example, according to the process database shown in FIGS. 6 to 11, the deep neural network model of the above embodiment is trained.

[0085] From the process databases shown in FIGS. 6 to 11, the numerical values of the weight terms of the model function having the following tendencies can be trained.

[0086] The chamber pressure shows a high correlation with the etching pattern, and the numerical value of the weight term is large. The lower electrode final (bias) power also shows a high correlation with the pattern, and the numerical value of the weight term is large.

[0087] In addition to constructing the process database by itself in this embodiment, the process tendencies in the above process database are also described in various documents. Therefore, it is also possible to construct the corresponding process database by referring to the documents.

[0088] After obtaining the trained model, using the recursive iteration method, according to the process results shown in FIG. 12, the process recipe that meets the process requirements is derived in the reverse direction. For example, when deriving the etching angle in the reverse direction, it has a large relationship (weight) with the lower electrode final power of the etching and the final single-step time of the etching. According to the means of the present invention, these two process parameters are simultaneously optimized by the gradient algorithm, and the finally obtained process recipe is shown in Table 1.

[0089] [Table 1]

[0090] * The lower electrode power increases from the start value to the final value. ** The single-step time increases from the start value to the final value.

[0091] In the case of the hole etching process, similarly, the process databases shown in FIGS. 13 to 18 are constructed to train the above deep neural network model. After obtaining the trained model, using the recursive iteration method, according to the process results shown in FIG. 19, the process recipe is derived in the reverse direction and is shown in Table 2.

[0092]

Table 2

[0093] As described above, when the means of the present invention is used, in an etching apparatus, a process recipe can be automatically output according to the process result, and the degree of automation of the machine can be improved. In addition, the present invention can be applied not only to etching apparatuses but also to other processes in the semiconductor field such as PVD, CVD, furnace tubes, and cleaning machines.

[0094] As described above, each embodiment of the present invention has been described. However, the above description is not exhaustive but illustrative, and is not limited to each disclosed embodiment. 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 obtaining a semiconductor process recipe, comprising: Randomly inputting a set of process parameters required for a process into a constructed deep neural network model, and obtaining a process result evaluation index corresponding to the set of process parameters, wherein the set of process parameters includes a plurality of process parameters, the deep neural network model is a trained multi-layer deep neural network, 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 the set of process parameters, and the output layer is used to output the process result evaluation index corresponding to the set of process parameters; Determining whether the difference between each of the process result evaluation indices and a set process requirement set corresponding thereto meets a set requirement; If YES, using the set of process parameters as a process recipe for performing an actual process; If NO, according to the process requirement, along the direction from the output layer to the input layer, based on the gradient algorithm and the self-collision-free iteration method, optimizing the eigenvalues in each intermediate layer and the numerical values of each process parameter in the input layer layer by layer until all the process result evaluation indices output from the deep neural network model in the input layer can meet the process requirement, and using the set of process parameters that meet the process requirement as a process recipe for performing an actual process. A method for obtaining a semiconductor process recipe, characterized by comprising the above steps.

2. The method for constructing the deep neural network model is as follows: A step of determining the number of layers and the structure of the deep neural network model according to the number of process parameters included in the process parameter group and the number of the process result evaluation indexes corresponding to the process parameter group, 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 indexes, each layer of the deep neural network in the deep neural network model includes a plurality of neurons, along the direction from the input layer to the output layer, one neuron sequentially decreases between adjacent deep neural networks, each neuron is connected to all neurons in the deep neural network of the previous layer, each process parameter is used as an input feature of one neuron in the input layer, each neuron in the intermediate layer is used to calculate the output features of all neurons in the deep neural network of the previous layer, the output feature of each neuron in the output layer is the process result evaluation index, and the number of the process parameters included in the process parameter group is larger than the number of the process result evaluation indexes corresponding to the process parameter group step, A step of providing an activation function executed by the deep neural network model according to the number of layers of the deep neural network model, characterized in that the method for obtaining a semiconductor process recipe according to claim 1 includes the above steps.

3. The method for constructing the deep neural network model is A step of training the deep neural network model using a deep learning method with a process database as training data, wherein the process database includes historical process parameter data and historical process result data of a specific process step, A step of determining the weight term and the bias term of each layer of the deep neural network in the deep neural network model according to the training result of the activation function, characterized in that the method for obtaining a semiconductor process recipe according to claim 2 further includes the above steps.

4. The step of optimizing the eigenvalue in each intermediate layer and the numerical value of each process parameter in the input layer layer by layer based on the gradient algorithm and the self-collision-free iteration method is Step S1 of finding, from the output layer, one process result evaluation index with the greatest difference from the process requirements; Step S2 of finding, from the intermediate layer adjacent to the output layer, one process parameter with the greatest numerical value of the weight term related to the process result evaluation index with the greatest difference; Step S3 of fixing the numerical value of the weight term of one process parameter with the greatest numerical value of the weight term and the numerical value of the bias term, and optimizing and adjusting the eigenvalue of one process parameter with the greatest numerical value of the weight term using a gradient algorithm until the difference between the process result evaluation index with the greatest difference in the output layer and the process requirements meets the set requirements; At this time, if there are other process result evaluation indexes that do not meet the set requirements for the difference from the process requirements among other process result evaluation indexes, steps S1 to S3 are repeated until all process result evaluation indexes in the output layer meet the process requirements, and at the same time, step S4 of determining the eigenvalues of each process parameter in the intermediate layer is performed; Based on the eigenvalues of each process parameter in the intermediate layer obtained in step S4, along the direction from the output layer to the input layer, continue to optimize the eigenvalues of each process parameter in other intermediate layers and the numerical values of the process parameters in the input layer layer by layer using a gradient algorithm and a self-collision-free iterative method until a group of process parameters that enables all process result evaluation indexes output from the deep neural network model in the input layer to meet the process requirements is obtained. Step S5, which is a method for obtaining a semiconductor process recipe according to claim 3, characterized by including the above.

5. After step S3, determine whether the difference between the process result evaluation index correspondingly adjusted in step S3 and the process requirements meets the set requirements. If it meets, perform step S4. If it does not meet, fix one process parameter with the greatest numerical value of the weight term obtained by adjustment and optimization in step S3. Step S6 of finding, from the intermediate layer adjacent to the output layer, one process parameter with the second greatest numerical value of the weight term related to the process result evaluation index correspondingly adjusted in step S3. Fix the numerical value of the weight term of one process parameter with the second-largest numerical value among the weight terms, and use the gradient algorithm until the difference between the process result evaluation index adjusted correspondingly in step S3 in the output layer and the process requirement meets the set requirement to optimize the eigenvalue of one process parameter with the second-largest numerical value among the weight terms and perform step adjustment in step S7, and a step of executing the step, further comprising the method for obtaining a semiconductor process recipe according to claim 4.

6. After step S5, After optimizing and adjusting all process parameters related to the process result evaluation index, if there is still a process result evaluation index in the output layer that does not meet the requirement with the set difference from the process requirement, Select process parameters that minimize the sum of squares of the differences between each process result evaluation index and the process requirement to form a new process parameter group, and further comprising step S8 of using the new process parameter group as a process recipe for performing an actual process, characterized in that the method for obtaining a semiconductor process recipe according to claim 5.

7. The calculation formula of the gradient algorithm is the following formula, characterized in that the method for obtaining a semiconductor process recipe according to claim 1. x ｉ = x ｉ - θ × (∂y / ∂x ｉ ) (where x ｉ is the adjusted process parameter in the intermediate layer, y is the adjusted process result evaluation index, and θ is the adjustment step.)

8. The difference is calculated by the following formula, As the set requirement, the difference between the process result evaluation index and the correspondingly set process requirement is 0 to 10%, characterized in that the method for obtaining a semiconductor process recipe according to claim 1. Δ = |(y - y 実際 ) / y 実際 | (where Δ is the difference between the process result evaluation index and the set process requirements, y is the process result evaluation index output from the deep neural network model, and y 実際 is the given process requirement.)

9. An automatic acquisition system for a semiconductor process recipe, A trained deep neural network with a multi-layer structure, which can output an accurate process result evaluation index corresponding to the process parameter group based on the input process parameter, and a deep neural network model, A calculation module for executing the method for obtaining a semiconductor process recipe according to any one of claims 1 to 8, characterized in that the automatic acquisition system for a semiconductor process recipe.

10. A semiconductor process apparatus, characterized in that it includes the automatic acquisition system for a semiconductor process recipe according to claim 9.

Images