Process recipe search device, etching recipe search method, and semiconductor device manufacturing system

Abstract

This invention relates to a process recipe search apparatus, an etching recipe search method, and a semiconductor device manufacturing system. In process recipe development utilizing machine learning, to facilitate the evaluation of predicted processing shapes, a process recipe search apparatus (1310) that searches for etching recipes as parameters of a plasma processing apparatus (100) set to etch a sample into a desired shape displays the difference between the predicted processing shape and the target shape whenever the predicted processing shape of a sample based on a candidate etching recipe, predicted using a machine learning model, is displayed on a display device (1322).

Description

Technical Field

[0001] This invention relates to a process recipe search device, an etching recipe search method, and a semiconductor device manufacturing system. Background Technology

[0002] In semiconductor devices, the demands for lower power consumption and increased storage capacity have driven miniaturization and the increasing three-dimensionality (complexity) of device structures. In the fabrication of miniaturized devices, the goal is not simply to reduce pattern size, but also to achieve more complex fabricated shapes compared to existing semiconductor devices. To achieve a target fabricated shape (called the target shape) through dry etching, numerous parameters of the etching apparatus's gas, power, and high-frequency systems need to be set and controlled. To achieve complex fabricated shapes, processing must be performed in multiple steps while adjusting numerous parameters over time measured in seconds; therefore, setting a vast amount of parameters is required for the processing of a single pattern. Thus, even for skilled operators, setting the parameters to enable such an etching apparatus to perform the process requires a significant amount of time. The etching conditions, or the parameters set to perform the etching, are called the etching recipe.

[0003] Machine learning is known as a method for obtaining large amounts of parameters in etching processes with high precision and speed. This involves observing the structure of a semiconductor device processed under multiple processing conditions using an electron microscope, measuring the resulting images, and quantifying the processing results. A machine learning model is then created based on etching recipes used in multiple processes and the quantified processing results. The machine learning model can virtually predict the post-etching shape by inputting etching recipes that seem to be candidate recipes. By searching a large number of prediction results for etching recipes that approximate the target shape, complex etching conditions can be efficiently identified. Patent Document 1 is an example of this.

[0004] Existing technical documents

[0005] Patent documents

[0006] Patent Document 1: U.S. Patent Application Publication No. 2018 / 0082873

[0007] Several topics exist in the search for etching recipes that utilize machine learning models.

[0008] The first challenge is selecting the appropriate candidate etch pattern. Due to the complexity of machine learning models, multiple candidate etch patterns are typically obtained. It is rare to actually process using all candidate etch patterns; it is necessary to narrow down the range of candidate etch patterns to be processed from multiple options. Therefore, users utilize predicted processing shapes based on candidate etch patterns. For example, taking typical trench processing in semiconductor manufacturing as an example, for a trench shape (target shape) with a depth of 100 nm and a width of 30 nm, the desired shape accuracy is sometimes 0.5 nm for both depth and width. Furthermore, the target shape is generally determined by the dimensions of multiple locations throughout the trench structure.

[0009] When the anticipated processing shapes of multiple candidate etch formulations for such a target shape are visualized and imaged, if the resolution is set to 0.1 nm / pixel, the processing accuracy for a trench shape with a depth of 1000 pixels and a width of 300 pixels is only about 5 pixels different. If the predicted processing shapes overlap within a range of about 5 pixels, it becomes difficult to grasp the characteristics of each predicted processing shape to determine the quality of the candidate etch formulation and thus narrow down the processing range.

[0010] The second challenge lies in the fact that, when narrowing down the range of candidate etch formulations based on the predicted processing shape, selecting a candidate etch formulation simply because it most closely approximates the target shape may not be the optimal solution. For example, candidate etch formulations with excessively high sensitivity to processing conditions are undesirable. Sensitivity, in this context, refers to the degree to which a slight change in one of the processing conditions (e.g., gas flow rate, voltage applied to the electrode, etc.) affects the predicted processing shape. From the perspective of reproducibility in etching processes, candidate etch formulations that significantly alter the predicted processing shape due to minute changes in, for example, gas flow rate, voltage applied to the electrode, or current, are undesirable. However, even skilled process engineers find it difficult to infer sensitivity from the predicted processing shape derived from the processing conditions of the candidate etch formulation and a machine learning model.

[0011] Furthermore, sometimes during the etching recipe search, situations arise where candidate etching recipes contain processing conditions that cannot be set in the etching apparatus used for processing, and selection is still based on predicting a processing shape that approximates the target shape. This can lead to the continued searching of unexecutable etching recipes. The third challenge is to avoid such meaningless searches. One approach is to add constraints to the computer program that automatically searches for etching recipes to avoid processing conditions that cannot be set. However, the conditions of etching apparatuses, including those controlling gas systems and high-frequency systems, often change frequently. Moreover, the direction of the recipe search needs to be input by the user. Therefore, it is desirable for users to interactively set the search direction using a graphical user interface (GUI). Summary of the Invention

[0012] An embodiment of the present invention provides a process recipe search device that searches for etching recipes as parameters of a plasma processing apparatus set to etch a sample into a desired shape. The process recipe search device includes: a target shape determination unit that determines a target shape by defining the desired shape through multiple shape elements; a machine learning model generation unit that generates a machine learning model that predicts the processing shape of the sample by the plasma processing apparatus based on the parameters of the plasma processing apparatus; a recipe search unit that uses the machine learning model to search for candidate etching recipes that become alternative etching recipes; a processing recipe determination unit that displays the predicted processing shape of the sample based on the candidate etching recipes predicted by the machine learning model on a display device, and determines a candidate etching recipe selected from the displayed candidate etching recipes as the processing recipe set for the plasma processing apparatus to etch the sample; and a shape emphasis processing unit that emphasizes the difference between the predicted processing shape and the target shape on the display device.

[0013] The effects of the invention

[0014] In the development of process formulations using machine learning, to facilitate the evaluation of predicted processing shapes, it is essential to easily select process formulations suitable for diverse customer requirements. Other issues and new features will become clear from the descriptions and figures in this specification. Attached Figure Description

[0015] Figure 1A This is an example of predicting the shape of the processed part.

[0016] Figure 1B This is an example that emphasizes the predicted processing shape.

[0017] Figure 2 This is a schematic diagram of the plasma processing device.

[0018] Figure 3 This is a flowchart for searching etching formulas.

[0019] Figure 4A This is a diagram used to illustrate the method of highlighting.

[0020] Figure 4B This is a diagram used to illustrate the method of highlighting.

[0021] Figure 5 This diagram is used to illustrate the method for setting the magnification ratio.

[0022] Figure 6 It is a flowchart for highlighting specific elements.

[0023] Figure 7This is an example of a GUI for selecting a processing formula from a pool of candidate etching formulas.

[0024] Figure 8A These are other examples of emphasis.

[0025] Figure 8B These are other examples of emphasis.

[0026] Figure 8C These are other examples of emphasis.

[0027] Figure 9 This is an example of a GUI that displays the sensitivity of candidate etching recipes.

[0028] Figure 10 This is a flowchart for searching etching formulas.

[0029] Figure 11 This is an example of a GUI used to redefine target shapes.

[0030] Figure 12 This is an example of a GUI that specifies the likelihood of the target shape and the range of recipe searches.

[0031] Figure 13 This is an example of the system structure of a semiconductor device manufacturing system.

[0032] Figure 14 This is an example of the system structure of a semiconductor device manufacturing system.

[0033] Explanation of reference numerals in the attached figures

[0034] 8: Mask; 9: Substrate; 20, 30, 40: Trench width; 50: Trench depth; 60: Target shape; 100: Plasma processing device; 101: Container; 102: Cluster plate; 103: Dielectric window; 104: Vacuum processing chamber; 105: Gas supply mechanism; 106: Gas piping; 107: Space; 108: Fine aperture; 112: Variable conduction valve; 113: Turbomolecular pump; 114: Coarse pump; 115: Pressure gauge; 116: Microwave power supply; 117: Automatic matching device; 118: Square waveguide; 119: Square-circular waveguide converter; 120: Circular waveguide; 121: Cavity resonator; 122, 123, 124: Solenoid coil; 125: Electrode; 126: 127: Bias Voltage Generator; 128: Temperature Control Device; 135: Electrostatic Adsorption Unit; 136: Plasma; 137: Grounding; 139: Electrostatic Adsorption Power Supply; 140: Coil Power Supply; 150: Control Unit; 408, 409: Trench Profile; 501: Display Window; 504: Profile Ruler; 505: Emphasis Display Ruler; 701, 901, 1101, 1201: GUI; 707: Selection Area; 708: Selection Checkbox; 709: Processing Condition Display Table; 710: Predicted Length Value Table; 711: Display Window; 802, 805: Trench Sidewall; 803, 806: Trench Bottom; 808: Vector; 902, 903, 904: Predicted Processing Shape; 905: Arrow 908: Display Table; 909: Select Area; 910: Specify Area; 911: Selection Drop-down Box; 912: Change Setting Scroll Bar; 1102: Select Area; 1103: Shape Display Area; 1104: Target Shape; 1107: Existing Measurement Value Table; 1108: New Measurement Value Table; 1206: Target Shape Table; 1207, 1208: Selection Drop-down Box; 1209: Shape Reflection Button; 1211: Display Image Change Button; 1212: Formula Search Range Table; 1213: Condition Specifying Box; 1214: Search Range Display Chart; 1216: Search Range; 1217: Search Range Determination Button; 1220: Redefine Target Shape Before; 1221: Redefine Target Shape After. 1300: Semiconductor device manufacturing system; 1310: Process recipe search device; 1311: Central processing unit; 1312: Database; 1313: Target shape determination unit; 1314: Machine learning model generation unit; 1315: Recipe search unit; 1316: Process recipe determination unit; 1317: Device control unit; 1318: Display shape emphasis processing unit; 1321: Input device; 1322: Display device; 1325: Target shape data; 1326: Recipe search range data; 1330: Evaluation device; 1331: Image / camera condition data; 1340: Dimension measurement device; 1341: Input / output device; 1342: Length measurement data; 1400: Platform; 1401: Database.1402: OS, 1403: Middleware, 1405: Process Recipe Search Application, 1406: Dimensional Measurement Application, 1410: Terminal. Detailed Implementation

[0035] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0036] Figure 2 This is a schematic diagram showing a structural example of a plasma processing apparatus. The plasma processing apparatus 100 is a microwave ECR plasma etching apparatus. The plasma processing apparatus 100 includes a vacuum processing chamber 104, inside which electrodes 125 are disposed, serving as the wafer 126 to be processed. The wafer 126 is disposed on the electrodes 125 via an electrostatic adsorption unit 135. Gas is supplied to the vacuum processing chamber 104 from a gas supply mechanism 105 through a gas pipe 106, a space 107, and a shower plate 102. Plasma 136 is generated by the interaction of the electric field and magnetic field generated by an electric field and magnetic field generating device disposed outside the vacuum processing chamber 104. The plasma 136 contains ions and atomic clusters, and plasma etching is performed by their interaction with the surface of the wafer 126.

[0037] A variable conduction valve 112 is disposed at the lower part of the vacuum processing chamber 104, and a turbomolecular pump 113 connected to the variable conduction valve 112 is used to exhaust the gas in the vacuum processing chamber 104. The variable conduction valve 112, the turbomolecular pump 113 and the roughing pump 114 are respectively connected to the control unit 150, and the operation of each device is controlled by the signal from the control unit 150.

[0038] In plasma etching, controlling the vacuum level of the vacuum chamber 104 is crucial. A pressure gauge 115 is installed to measure the internal pressure of the vacuum chamber 104, and the control unit 150 controls the variable transmission valve 112 based on the value of the pressure gauge 115 to control the pressure of the vacuum chamber 104.

[0039] A microwave power supply 116, serving as a first high-frequency power source, is installed at the top of the plasma processing apparatus 100. The frequency of this microwave power supply 116 is, for example, 2.45 GHz. Microwaves generated by the microwave power supply 116 propagate to the cavity resonator 121 via an automatic matching unit 117, a square waveguide 118, a square-to-circular waveguide converter 119, and a circular waveguide 120. The automatic matching unit 117 automatically suppresses reflected waves. The cavity resonator 121 adjusts the microwave electric field distribution to a distribution suitable for plasma processing. The microwave power supply 116 is controlled by a control unit 150.

[0040] Solenoid coils 122, 123, and 124, constituting an electromagnet, are arranged around the vacuum processing chamber 104 and the cavity resonator 121. A current flows through the solenoid coils 122, 123, and 124 via the coil power supply 140 controlled by the control unit 150, generating a magnetic field inside the vacuum processing chamber 104.

[0041] If a high-frequency power source and a magnetic field are generated inside the vacuum processing chamber 104, plasma based on electron cyclotron resonance (ECR) will be formed in the region where the strengths of the electric and magnetic fields have a specific relationship. Electrons existing inside the vacuum processing chamber 104 move while rotating along the magnetic field lines generated by the magnetic field produced by the solenoid coils 122, 123, and 124 due to the Lorentz force. If the frequency of the microwaves propagating from the microwave power source 116 matches the frequency of the electrons' rotation, the electrons are resonantly accelerated, effectively generating plasma.

[0042] The region where the ECR is generated (ECR surface) can be controlled by the magnetic field created by solenoid coils 122, 123, and 124. The magnetic field created by solenoid coils 122, 123, and 124 can be controlled by the control unit 150 through the current flowing to the coils. The diffusion of plasma can be controlled by controlling the magnetic field distribution. Through these effects, the distribution of plasma above the wafer 126 is controlled, improving the uniformity of plasma processing.

[0043] The electrode 125, which mounts the wafer 126, is located below the ECR surface and is fixed in the vacuum processing chamber 104. The electrode 125, like the wafer 126, is circular. The plasma processing apparatus 100 has a transport device, such as a robotic arm (not shown), which uses this transport device to place the wafer 126 onto the electrode 125 via an electrostatic adsorption unit 135 located on the electrode 125. An electrostatic voltage is supplied to the electrostatic adsorption unit 135 from an electrostatic adsorption power supply 139.

[0044] A bias voltage generating unit 127 is connected to electrode 125, and a bias voltage is applied to wafer 126 via the bias voltage generating unit 127. Since the degree to which ions within plasma 136 are drawn into wafer 126 depends on the bias voltage, the processing shape of wafer 126 can be controlled by controlling the bias voltage generating unit 127 via control unit 150. Furthermore, electrode 125 is also connected to a temperature control device 128, and the processing shape of wafer 126 is controlled by adjusting the chemical reaction process during processing at the surface of wafer 126 through temperature control. Temperature control device 128 is also controlled from control unit 150.

[0045] The control unit 150 is a computer that controls the timing and amount of the etching recipe, which consists of multiple steps required for the processing of the wafer 126, to operate in a suitable sequence. The etching recipe is based on a pre-set recipe.

[0046] exist Figure 13 An example of the system structure of a semiconductor device manufacturing system for searching etching recipes is shown. The semiconductor device manufacturing system 1300 includes a process recipe search device 1310, an input device 1321, a display device 1322, a plasma processing device 100, an evaluation device 1330, a dimensional measurement device 1340, and an input / output device 1341. The user inputs information such as target shape data 1325 and recipe search range data 1326 from the input device 1321, and interactively performs an etching recipe search based on information from candidate etching recipes prompted by the GUI displayed on the display device 1322.

[0047] The process formula search device 1310 receives target shape data 1325 from the input device 1321 and searches for an etching formula that the plasma processing device 100 can optimally obtain to the target shape. When formula search range data 1326 is input, a search for etching formulas in the parameter space corresponding to the search conditions of the formula search range data 1326 is performed.

[0048] Input device 1321 receives input data from the user or length measurement data 1342 from input / output device 1341 and inputs it to process recipe search device 1310. Examples include keyboards, pointing devices, touch panels, storage media reading devices, etc.

[0049] Display device 1322 is a display that shows the user information related to the etching recipe search from process recipe search device 1310. Other output devices may include a printer, a storage medium write-out device, etc.

[0050] The plasma processing device 100 is in Figure 2 The following is an example of a processing apparatus. The plasma processing apparatus 100 processes a semiconductor or semiconductor device based on an etching recipe input from the process recipe search device 1310, and then transfers the processed semiconductor or semiconductor device to the evaluation device 1330.

[0051] The evaluation device 1330 captures a cross-section of the semiconductor or semiconductor device processed in the plasma processing apparatus 100, obtaining a cross-sectional image as a processing result. The evaluation device 1330 includes charged particle beam measurement devices such as SEM (Scanning Electron Microscope) and TEM (Transmission Electron Microscope). Alternatively, a portion of the semiconductor or semiconductor device processed in the plasma processing apparatus 100 can be removed as a fragment and transported to the evaluation device 1330 for measurement. The acquired cross-sectional image, along with imaging conditions such as magnification, is transferred to the input / output device 1341 as image / imaging condition data 1331.

[0052] The dimension measuring device 1340 receives the definition of the target shape from the process recipe search device 1310 via the input / output device 1341, and image / camera condition data 1331 from the evaluation device 1330. It measures a given dimension based on the definition of the target shape from the cross-sectional image and outputs it as length measurement data 1342 to the input / output device 1341. The input / output device 1341 outputs the length measurement data 1342 to the input device 1321.

[0053] The process formula search device 1310 includes a central processing unit 1311, a database 1312, a target shape determination unit 1313, a machine learning model generation unit 1314, a formula search unit 1315, a processing formula determination unit 1316, a device control unit 1317, and a display shape emphasis processing unit 1318.

[0054] Figure 3This is a flowchart illustrating how the process recipe search device 1310 uses a machine learning model to search for etching recipes for the plasma processing apparatus 100. The central processing unit 1311 handles the input and output of data to the process recipe search device 1310 and controls the overall process. First, the process engineer determines the target shape 202 based on the desired shape after processing. The central processing unit 1311 stores the target shape data 1325 received from the input device 1321 into the database 1312. The definition of the target shape is sometimes updated during the search process; target shape management is performed by the target shape determination unit 1313. Next, a processing recipe determination 203 is made based on information such as the physical dimensions of the pattern disposed on the wafer and the material. In the initial stage of the search, due to the lack of information about the processing formula, the user inputs an initial etching formula into the input device 1321. The central processing unit 1311 forwards the initial etching formula received from the input device 1321 to the processing formula determination unit 1316. The processing formula determination unit 1316 then uses the device control unit 1317 to perform etching 204 and measurement 205 based on the initial etching formula.

[0055] After etching 204 in the plasma processing apparatus 100, the evaluation device 1330 visualizes the processing result, and the dimensional measurement device 1340 measures the processing result 205. The measured length data 1342 is input to the process formula search device 1310 via the input / output device 1341. The central processing unit 1311 stores the measured length data 1342 and the etching formula (processing conditions) together in the database 1312, and performs a processing size determination 206 to compare the measured value with the target shape. If a difference is observed between the measured result and the target size, the processing formula is further discussed 203.

[0056] Generally, when creating a machine learning model, measurement results from multiple processing recipes are required. Therefore, the central processing unit 1311 determines the number of measurement results (209). If the number of results is insufficient, a manual recipe discussion (208) is conducted to determine the processing recipe (203) and processing size (206). On the other hand, if the data determination (209) determines that there is sufficient measurement data for creating the machine learning model, the machine learning model creation unit 1314 is activated. The combination of measurement values ​​stored in the database 1312 and processing conditions (parameters of the plasma processing apparatus) is used as learning data to create the machine learning model (211). The recipe search unit 1315 then uses the created machine learning model to perform an automatic recipe search (210). In the processing recipe determination unit 1316, multiple candidate etching recipes extracted by the recipe search unit 1315 based on the machine learning model are displayed on the display device 1322, prompting the user to determine multiple processing recipes that are considered to achieve the target shape (203). After multiple processing recipes are determined (203), etching (204) and processing size (206) are performed. When the measured value and the target shape are consistent in the processing dimension determination 206, the development of the etching formula is completed 207.

[0057] for Figure 13 This refers to a semiconductor device manufacturing system that incorporates etching recipe searching and dimensional measurement as applications within a platform. Figure 14 An installation example is shown. Platform 1400 is built on the cloud and runs applications that perform processing on OS 1402 and middleware 1403. Platform 1400 includes: a process recipe search application 1405 that performs processing equivalent to the process recipe search device 1310, and a dimensional measurement application 1406 that performs processing equivalent to the dimensional measurement device 1340. Users can access the platform from terminal 1410 via a network to utilize the functions of the applications built on platform 1400. Platform 1400 has a database 1401 that stores the data required for application execution. Furthermore, plasma processing device 100 and evaluation device 1330 are connected to enable data exchange with platform 1400 via a network.

[0058]

Example 1

[0059] In the early stages of etching recipe development, measured values ​​often deviate significantly from the target shape. As multiple etching recipes are tested, the processed shape gradually approaches the target shape. As the actual measurement results approach the target shape, the predicted processed shape of the etching recipe proposed by the machine learning model through automatic recipe search also becomes closer to the target shape. However, it is difficult to visually understand the actual shape differences based solely on numerical data. Furthermore, as explained in the first problem description, even when displayed graphically, the differences between the predicted and target shapes are difficult for the user to visually discern, given the similarity between the predicted and target shapes. The process recipe search device 1310 includes a shape emphasis processing unit 1318, which can emphasize the differences between the predicted and target shapes and display them on the display device 1322.

[0060] exist Figure 1A The illustration schematically depicts one of the predicted processing shapes based on an etching recipe (candidate etching recipe) proposed from a machine learning model. Figure 1A The illustrated predicted processing shape is an example of the cross-sectional shape of the trench structure (hereinafter referred to as the trench shape). The width of the trench shape is determined by the mask 8, and the trench shape is formed on the substrate 9 by etching from the opening of the mask 8. Figure 1A The target shape 60 is also shown by thick lines. The widths 20, 30, and 40 of the trench shape are defined as different widths within the measured trench depth. Process engineers define the target shape, for example, using the widths 20, 30, and 40 of the trench shape and the depth 50 of the trench shape. Figure 1A In the case of predicted processing shape, the width of the groove shape (20, 30) is larger than the width of the target shape, and the width of the groove shape (40) is smaller than the width of the target shape. Furthermore, the depth of the groove shape (50) is deeper than the depth of the target shape. As the deviation between the predicted processing shape and the target shape decreases, it becomes increasingly difficult to visually grasp the problem of such predicted processing shape.

[0061] In this embodiment, by Figure 1B The method of depiction, even with Figure 1A The same predicted processing shape can also highlight the difference between the predicted processing shape and the target shape.

[0062] use Figure 4A and Figure 4B This embodiment describes an emphasis display method for highlighting errors in the target shape. Figure 4A This is the trench shape without emphasis. When the predicted depth pD, predicted width pW, target depth D, and target width W of the target shape are set as trench shape 408, the difference dD between the predicted depth pD and the target depth D, and the difference dW between the predicted width pW and the target width W are characterized as follows.

[0063] dD = D - pD

[0064] dW=(W-pW) / 2

[0065] Here, for depth and width, when the magnification (called emphasis ratio) used in the emphasis display is set to a and b, the errors with the target shape are set to dD1 and dW1 respectively, as follows.

[0066] dD1=a×dD

[0067] dW1=b×dW

[0068] It is emphasized that the magnification ratios a and b can be automatically obtained based on the difference between the predicted processing shape and the target shape, or they can be specified by the user on the GUI.

[0069] Figure 4B This refers to the trench shape for emphasis display. In trench shape 409, using errors dD1 and dW1 as defined above, the prediction depth eD and prediction width eW are set as follows:

[0070] eD=D+dD1

[0071] eW = W + 2 × dW1.

[0072] If with Figure 4A Compared to the usual display, it can not only visually depict the differences from the target shape more clearly, but also show the shape changes of different parts such as groove width and depth based on the target shape.

[0073] use Figure 5 This section explains how to set the magnification ratios a and b. In the GUI, the display window 501 showing the predicted processing shape has a length T and a width Y. To prevent the shape from exceeding the dimensions of the display window 501, the magnification ratio a for depth and the magnification ratio b for width must satisfy the following relationships:

[0074] a<(TD) / dD

[0075] b < (YW) / (2×dW).

[0076] Depending on how the window is displayed, the white space is sometimes considered as offset.

[0077] In addition, such as Figure 5As shown, when emphasizing the predicted machining shape, it is advisable to display three scales as needed. One is the outline scale 504 that displays the target shape, and the remaining one is the emphasis display scale 505 that indicates the predicted machining shape with emphasized errors. In this example, since the width and depth of the groove shape are emphasized with their respective emphasis levels, two orthogonal scales are shown.

[0078] This section explains a method for emphasizing the most basic processing shapes in etching: the width and depth of the trenches. In etching, besides these simple objectives, control is often required for tapered structures where trench sidewalls gradually narrow, for suppressing surface roughness by etching the trench sidewalls, and for controlling shapes such as booming sidewalls that expand midway. Among these, surface roughness is generally evaluated with smaller values, and a magnification of several times is insufficient to emphasize the shape's characteristics. Therefore, using the RA value, which represents the degree of surface roughness, as a benchmark, and setting the RA value at a magnification of several to several hundred times is effective.

[0079] In addition, Figure 3 , Figure 4B The example described above emphasizes the error between the predicted processed shape and the target shape. However, the object of emphasis in this embodiment is not limited to the target shape and the predicted processed shape; it can be applied to situations where differences between any shapes are emphasized. An example is given below. Let's assume that image B is emphasized relative to image A, where A is called the reference image and B is called the comparison image. For example, consider the following method: measuring the processing results based on two processing recipes according to the definition of their target shapes, and comparing the processed shapes obtained from the measurements of the two processing results with each other.

[0080] If the groove depth DA of the reference image, the groove width WA of the reference image, the groove depth DB of the comparison image, and the groove width WB of the comparison image are set as follows, the difference dD3 between the groove depth of the reference image A and the groove depth of the comparison image B, and the difference dW3 between the groove width of the reference image A and the groove width of the comparison image B are represented as follows:

[0081] dD3=DA-DB

[0082] dW3 = (WA - WR) / 2.

[0083] Here, for depth and width, when the emphasis ratio used in the emphasis display is set to c and d, the errors with the reference image are defined as dD4 and dW4 respectively, as follows.

[0084] dD4=c×dD3

[0085] dW4=d×dW3

[0086] The magnification values ​​c and d can be automatically determined based on the difference between the reference image A and the comparison image B, or they can be specified by the user in the GUI. Using the errors dD4 and dW4 defined above, the depth fD and width fW of the groove shape in the comparison image B to be highlighted are set as follows:

[0087] fD=D+dD4

[0088] fW = W + 2 × dW4.

[0089] Thus, not limited to the target shape, it is possible to visually and clearly depict the differences of any reference image, and by comparing the processed shapes based on good etching formulas with each other, the selection of a better etching formula becomes easy.

[0090] Figure 6 This is a flowchart depicting the process of highlighting a display using the shape emphasis processing unit 1318. To perform the highlighting, the display shape is first selected (602). For example, elements defining the target shape, such as the width and depth of the groove, are selected. Next, the emphasis ratio is specified in the display emphasis ratio specification / calculation (603). This can be automatically calculated based on the groove shape, etc. Next, the emphasis shape is calculated (604), and the display is performed on the display device 1322 (605). The user judges the quality of the displayed content based on the display results (606). If a re-evaluation of the emphasis ratio is needed, the emphasis ratio is adjusted through the emphasis ratio discussion (607), and the display emphasis ratio specification / calculation (603) and display result judgment (606) are repeated.

[0091] exist Figure 7 An example of a graphical user interface (GUI) is shown. To allow the engineer to select a processing recipe from candidate etching recipes derived from a machine learning model in processing recipe determination 203, the processing recipe determination unit 1316 causes the display device 1322 to display this GUI. Above the GUI 701 is a display window 711 that normally displays and emphasizes the trench shape. Below it is a selection area 707 from which the trench shape to be displayed in the display window 711 is selected. The selection area 707 displays candidate etching recipes (predicted recipes) selected by the machine learning model and the target shape, each configured with a selection checkbox 708 for selection to display / not display. Figure 7 In this case, the expected processing shape and target shape based on the predicted formulas No.1 and No.4 in the candidate etching formulas will be displayed in the display window 711. By changing the checkbox in the selection area 707, the expected processing shape of other predicted formulas can be displayed.

[0092] Below the selection area 707 is a processing condition display table 709 for candidate etching formulations. The processing condition display table 709, displayed in display window 711, shows the processing conditions of the predicted formulation for the desired processing shape. These processing conditions are recommended by a machine learning model. By comparing the processing conditions of the predicted formulation with the desired processing shape, the user can advance the selection of processing conditions from the perspective of each specific issue.

[0093] Below the processing condition display table 709 is a table 710 displaying predicted length values ​​that represent the dimensions of the elements of the predicted processing shape (in this example, width W1, W2, W3, and depth D1) shown in the display window 711. The predicted length values ​​are calculated by inputting the processing conditions of the predicted recipe to a machine learning model. While the predicted length value table 710 may not be particularly useful in recipe selection, it allows users to easily identify situations where there are obvious anomalies, such as incorrect target shape input or shapes recommended by the machine learning model.

[0094] To enable users to identify the predicted processing shape by emphasizing the differences from the target shape, display methods other than the emphasis display described so far are also considered. Figures 8A to 8C This will illustrate this display method. Figure 8A This method uses color to indicate the difference from the target shape. For example, a large difference is displayed in red, a small difference in green, and vice versa. Figure 8A For example, the trench bottom 803 is shown in red, and the trench sidewalls 802 are shown in green. Thus, the user can visually select shape elements (in this case, depth, etc.) that have high priority through color. Figure 8B The example shown uses line thickness to represent the difference from the target shape. In this case, the difference between the thicker line and the target shape is greater, so the trench bottom 806 is shown as a line thicker than the trench sidewall 805. Figure 8C This method displays the deviation as a vector, represented by arrows. In this case, it's easy to understand the degree to which the corresponding measurement point deviates from the target shape in two dimensions. Figure 8C In this model, since the shape elements are width and depth, vector 808 becomes a vector oriented laterally and vertically. Conversely, when defining the tapered sidewalls and the slope of the bottom surface as shape elements, the errors of the tapered shape and the slope are expressed as vectors oriented towards the normal directions of the sidewalls and bottom surfaces, respectively. By performing such two-dimensional vector annotation, predicting the understanding of the machining shape becomes easier.

[0095]

Example 2

[0096] As Example 2, a method and GUI for evaluating the sensitivity of candidate etching recipes using a machine learning model, which is described as the second issue, will be explained. During the processing recipe determination 203 based on the candidate etching recipes searched using the machine learning model, the processing recipe determination unit 1316 displays the GUI described below on the display device 1322.

[0097] exist Figure 9 An example of a GUI 901 is shown, displaying the sensitivity of the candidate etching formula obtained by the processing formula determination unit 1316 based on a machine learning model. Above the GUI 901 is a shape display window showing the magnitude of the predicted shape change. In addition to displaying the predicted processing shape 902 obtained from the candidate etching formula as a solid line in the shape display window, the window also displays the predicted processing shape 903 of a sensitivity evaluation etching formula with a +10% change in parameter A (one of the processing conditions of the etching formula) and the predicted processing shape 904 of a sensitivity evaluation etching formula with a -10% change in parameter A as dotted lines. Arrows 905 indicating the amount of change in the predicted processing shape represent the amount of change in the predicted processing shape caused by changing parameter A; the smaller arrow is displayed as the smaller change in the predicted processing shape. Figure 9 As shown in the case of predicted formula No. 1, it can be seen that, regarding the trench depth, changing parameter A by +10% results in a small change, while changing it by -10% results in a large change. Conversely, regarding the trench width, changing parameter A by -10% results in a small change, while changing it by +10% results in a large change.

[0098] exist Figure 9 The example illustrates the shape change when only parameter A of the etching formula is changed. GUI 901 includes a selection area 909 for parameters used in sensitivity evaluation, and multiple parameters can be changed simultaneously via a drop-down box 911. The amount of change of the parameters used in the evaluation is displayed in a designated area 910, and the amount of change can be set for each parameter using a scroll bar 912.

[0099] In the middle layer of GUI901, there is a selection area 707 where candidate etching recipes for sensitivity evaluation can be selected. During the evaluation of these candidate etching recipes, the machine learning model can be fed with the sensitivity evaluation etching recipe that causes parameter variations each time to obtain the predicted processing shape. Alternatively, the predicted processing conditions of the etching recipe surrounding the predicted processing shape can be pre-calculated, and the predicted processing shape corresponding to the specified parameter variation amount can be retrieved from the computer's storage device.

[0100] Below GUI901 is an etching recipe display table 908. Display table 908 shows the etching recipe set as the sensitivity evaluation target (here, No. 1, "Original" column), the sensitivity evaluation etching recipe when parameter A is changed by ±10% for sensitivity evaluation, and various predicted processing shapes. Target shapes can also be noted in display table 908.

[0101]

Example 3

[0102] Figure 10 This is a flowchart of the process recipe search device 1310 using a machine learning model to develop etching recipes for the plasma processing device 100. In the search for processing recipes utilizing a machine learning model, the recipe search process 1001, which involves adding and changing the target shape, is shown as a dotted line. Figure 3 The flowcharts shown illustrate the differences. Generally, during the development of an etching formulation, process engineers identify challenges in the formulation development process, add new shape elements to address these challenges, and redefine the target shape. By emphasizing the added shape elements as illustrated in Example 1, the differences from the target shape can be highlighted, making it easier to identify new challenges.

[0103] Regarding Figure 3 The same process is used in the flowchart, with the same reference numerals and repeated descriptions omitted. In the processing dimension determination 206, if there is a difference between the target shape and the measurement result, the target shape determination 1002 is performed. If the target shape is good, the processing shape obtained in the previous measurement 205 and the etching formula of etching 204 are added as new learning data, and the process proceeds to the creation of the machine learning model 211. On the other hand, if a problem is found in the target shape determination 1002, the target shape determination unit 1313 performs target shape correction 1003, and creates a new dataset by performing the measurement 1004 of the corrected target shape, and proceeds to the creation of the machine learning model 211. Thus, whenever a new shape element is added to the target shape, the target shape determination unit 1313, through the device control unit 1317, performs length measurement of the new shape element on the size measuring device 1340 based on the large amount of experimentally completed image / camera condition data 1331, recreates the dataset, and updates the machine learning model.

[0104] exist Figure 11An example of a GUI is shown, displayed on a display device 1322 by a target shape determination unit 1313, used to add shape elements to redefine the target shape and thus recreate the dataset. Above the GUI 1101, a selection area 1102 is configured for implementing etching 204 and selecting the etching formula from which the processing results are obtained in measurement 205. Checkboxes are present in the selection area 1102, and the processing results of the selected prediction model are displayed in a middle-layer shape display area 1103. In the display to the shape display area 1103, in addition to the target shape, an electron microscope image can also be overlaid. By overlaying the electron microscope image, process engineers can easily detect any new shape elements that were missed.

[0105] exist Figure 11 The shape display area 1103 shows an example of the processed shape corresponding to the redefined target shape. Here, only the opening width W1 and depth D1 are defined in the target shape before redefinition. The existing measurement value table 1107 is displayed in the lower layer. As is clear from this, regardless of the processing result of any predicted formula, the opening width W1 and depth D1 are equal to the target shape. However, the actual groove shapes vary, which means that the definition of the target shape before redefinition is insufficient. Therefore, the process engineer specifies the roughness R and bottom width W2 as new shape elements for the target shape 1104 displayed in the shape display area 1103. The dimensions of the newly defined shape elements are measured from the electron microscope image, the display of the shape display area 1103 is updated, and the new measurement value table 1108 is displayed in the lower layer. This measurement can be performed manually by the user or automatically using automated measurement techniques based on image processing such as edge detection. By preparing such a GUI, the dataset can be prepared correctly and quickly from the electron microscope image, thereby advancing to the creation 211 of a machine learning model based on the corrected target shape.

[0106]

Example 4

[0107] As an example 4, a method and GUI are described as follows: by visualizing the constraints in the etching recipe search, users can easily specify complex target shapes or limit the recipe search range as described in the third subject. The GUI described below is intended to continue the display of the GUI 1101 for target shape correction described in Example 3, but the GUI described below can also be used to adjust the search range without modifying the target shape.

[0108] exist Figure 12An example of GUI1201 is shown, displayed on display device 1322 by recipe search unit 1315, specifying the likelihood of the target shape and the recipe search range. The target shape 1220 before redefinition and the target shape 1221 after redefinition are shown in the upper display area. Typically, the likelihood of the target shape managed by target shape determination unit 1313 is set for each shape element of the defined target shape; for example, the width of the trench is determined to be ±1 nm, and the bottom of the trench is determined to be ±10 nm. The magnitude of the likelihood can also be adjusted, and sometimes a slightly wider likelihood set in the early stages of etching recipe development is gradually narrowed. Likelihood information is essential in the selection of candidate etching recipes; however, if likelihoods are set for multiple shape elements of the target shape, target management becomes correspondingly more difficult for the user.

[0109] GUI1201 visualizes the likelihood of the target shape. If we compare the likelihood δD1 of the groove depth and the likelihood δW1 of the groove width in the target shape 1220 before redefining, the likelihood of the groove depth is greater. Thus, by labeling shape features with lower likelihood as finer and shape features with higher likelihood as coarser, likelihood can be identified simultaneously with the target shape.

[0110] On the other hand, in the redefined target shape 1221, the likelihood δW2 of the groove width is greater than the likelihood δW1 of the groove width in the original target shape 1220, while the likelihood δD2 of the groove depth is less than the likelihood δD1 of the groove depth in the original target shape 1220. By comparing and displaying such changes, the changes in the target shape, including the likelihoods, can be visualized, thereby reducing the burden on the user.

[0111] The target shape table 1206 displays the size (W, D) and likelihood (δW, δD) of the target shape. The likelihood is specified using the target shape's shape element selection dropdown 1207 and can be changed using the likelihood selection dropdown 1208. This change is determined by pressing the shape reflection button 1209 and is reflected in the target shape table 1206.

[0112] In the likelihood selection dropdown (1208), you can choose a numerical value or a qualitative relationship, similar to the importance options "high, medium, low." This is to account for situations where, in the early stages of formulation development, there are still undetermined numerical targets.

[0113] In addition, the target shape and likelihood are sometimes determined with reference to the best conditions during formula development. Therefore, users can also change the background image by clicking the image change button 1211 and input the target shape with just mouse action without being aware of the image ruler.

[0114] The lower layer of GUI1201 visualizes the etching recipe search range based on a machine learning model. As shown in Recipe Search Range Table 1212, the parameters of the etching recipe are constrained by factors such as gas flow rate limitations, applicable electrode voltages, and controllable electrode temperatures. Furthermore, given the relationships between various parameters arising from the device's control method and the directionality of the etching recipe sought by the user (e.g., searching for recipes that avoid using a certain gas), searching for the etching recipe within the range that satisfies these relationships and directions makes it easier to obtain the user's desired recipe more efficiently from a vast parameter space.

[0115] The relationship and directionality between such parameters are entered in the condition specification box 1213 as a constraint relationship. The constraint relationship can be in a simple form such as four arithmetic operations, or it can be represented by a complex elementary mathematical relationship implied by the simulator. For example, selecting "A > B" from the condition specification box 1213 will display the search range based on this relationship as the search range 1216 in the search range display chart 1214. By visualizing the search range 1216 based on the recipe search range table 1212, the user can confirm that the entered constraint relationship is correctly reflected. By pressing the search range confirmation button 1217, the central processing unit 1311 forwards the determined recipe search range data 1326 to the recipe search unit 1315, so that it is reflected in the recipe search.

Claims

1. A process recipe search device for searching etching recipes as parameters of a plasma processing apparatus set to etch a sample into a desired shape. The process formula search device is characterized by having: The target shape determination unit determines the target shape by defining the desired shape through multiple shape elements. The machine learning model generation unit generates a machine learning model that predicts the processing shape of the sample by the plasma processing device based on the parameters of the plasma processing device. The recipe search unit uses the machine learning model to search for candidate etching recipes that can become the etching recipe; A processing formula determination unit displays a predicted processing shape of the sample based on the candidate etching formulas, predicted using the machine learning model, on a display device; and determines a processing formula for setting the plasma processing apparatus to etch the sample from the displayed candidate etching formulas; and The display shape emphasis processing unit emphasizes the difference between the predicted processing shape and the target shape and displays it on the display device. The processing formula determination unit causes the display device to display the predicted processing shape of the sample, based on a sensitivity evaluation etching formula that varies at least one parameter contained in the candidate etching formula, predicted using the machine learning model. The display shape emphasis processing unit displays the predicted processing shape based on the candidate etching formula and the predicted processing shape based on the sensitivity evaluation etching formula on the display device, respectively, while emphasizing the difference from the target shape.

2. The process formula search device according to claim 1, characterized in that, The display shape emphasis processing unit amplifies the difference between the target shape and the predicted processing shape by an emphasis ratio determined for each shape element, thereby displaying the predicted processing shape on the display device.

3. The process formula search device according to claim 1, characterized in that, The display shape emphasis processing unit can visually and identifiably display the difference between the predicted processing shape and the target shape on the predicted processing shape on the display device. The difference between the predicted processing shape and the target shape can be displayed by the color or line width of the shape element in the predicted processing shape, or by overlaying a vector representing the deviation between the predicted processing shape and the target shape onto the shape element for display.

4. The process formula search device according to claim 1, characterized in that, The formula search unit displays the range of parameters that the candidate etching formula of the plasma processing apparatus can take on the display device, and searches for the candidate etching formula in the constrained parameter space when the constraints between the parameters of the plasma processing apparatus are specified.

5. An etching recipe search method, utilizing a process recipe search device, said process recipe search device searching for etching recipes as parameters of a plasma processing apparatus set to etch a sample into a desired shape. The etching recipe search method is characterized in that... The target shape is determined by defining the desired shape using multiple shape elements. A machine learning model is created to predict the processing shape of the sample by the plasma processing device based on the parameters of the plasma processing device. The machine learning model is used to search for candidate etching recipes to become the etching recipe. The predicted processing shape of the sample based on the candidate etching formula, predicted using the machine learning model, is displayed on the display device. The predicted processing shape is displayed on the display device, emphasizing the difference between the predicted processing shape and the target shape. The machine learning model is used to predict the processed shape of the sample based on the sensitivity evaluation of the etching formulation, which varies at least one parameter contained in the candidate etching formulation. The predicted processing shape based on the candidate etching formula and the predicted processing shape based on the sensitivity evaluation etching formula are respectively displayed overlaid on the display device, emphasizing their differences from the target shape.

6. The etching formula search method according to claim 5, characterized in that, The difference between the target shape and the predicted processing shape is magnified by an emphasis factor determined for each shape element, thereby displaying the predicted processing shape on the display device.

7. The etching formula search method according to claim 5, characterized in that, The range of parameters that the candidate etching formula of the plasma processing apparatus can take is displayed on the display device, and the candidate etching formula is searched in the constrained parameter space when the constraints between the parameters of the plasma processing apparatus are specified.

8. An etching recipe search method utilizing a process recipe search device, said process recipe search device searching for etching recipes as parameters of a plasma processing apparatus set to etch a sample into a desired shape. The etching recipe search method is characterized in that... The target shape is determined by defining the desired shape using multiple shape elements. A machine learning model is created to predict the processing shape of the sample by the plasma processing device based on the parameters of the plasma processing device. The machine learning model is used to search for candidate etching recipes to become the etching recipe. The predicted processing shape of the sample based on the candidate etching formula, predicted using the machine learning model, is displayed on the display device. The predicted processing shape is displayed on the display device, emphasizing the difference between the predicted processing shape and the target shape. The processing formula for etching the sample is determined by selecting from a plurality of candidate etching formulas displayed on the display device and setting the plasma processing device accordingly. When the plasma processing apparatus is configured with the determined processing formula, and the processed shape based on the first processing formula is used as a reference image and the processed shape based on the second processing formula is used as a comparison image, the difference between the comparison image and the reference image is magnified by an emphasis ratio determined for each shape element, and the comparison image is displayed on the display device. The determined processing formula is set for the plasma processing apparatus. If the processed shape obtained by etching the sample is inconsistent with the target shape, the target shape is corrected. Update the machine learning model. The updated machine learning model is used to search for candidate etching recipes to become the etching recipe.

9. A semiconductor device manufacturing system, comprising: terminal; Plasma processing equipment; and The platform, connected via a network to the terminal and the plasma processing apparatus, is equipped with a process recipe search application that searches for etching recipes based on parameters set by the plasma processing apparatus to etch samples into desired shapes. The semiconductor device manufacturing system is characterized in that... The process formula search application performs the following steps: Determine the target shape by defining the desired shape using multiple shape elements; A machine learning model is created to predict the processing shape of the sample by the plasma processing device based on the parameters of the plasma processing device. The machine learning model is used to search for candidate etching recipes to become the etching recipe; and The predicted processing shape of the sample based on the candidate etching formula, predicted using the machine learning model, is displayed on the terminal. For the predicted processing shape displayed on the terminal, the difference between the predicted processing shape and the target shape is emphasized. The process formula search application performs the following steps: The machine learning model is used to predict the processed shape of the sample based on the sensitivity evaluation of the etching formulation by changing at least one parameter contained in the candidate etching formulation; and The predicted processing shape based on the candidate etching formula and the predicted processing shape based on the sensitivity evaluation etching formula are respectively overlaid on the terminal in a state emphasizing the difference from the target shape.

10. The semiconductor device manufacturing system according to claim 9, characterized in that, The difference between the target shape and the predicted processing shape is magnified by an emphasis factor determined for each shape element to display the predicted processing shape on the terminal.

11. The semiconductor device manufacturing system according to claim 9, characterized in that, The process formula search application performs the following steps: The range of parameters that the candidate etching formulation of the plasma treatment apparatus can take is displayed on the terminal. When constraints are specified between parameters of the plasma processing apparatus from the terminal, the search for the candidate etching formula is performed in the constrained parameter space during the step of searching for the candidate etching formula.

12. The semiconductor device manufacturing system according to claim 9, characterized in that, The process formula search application performs the following steps: The candidate etching formula selected from the candidate etching formulas displayed on the terminal is determined as the processing formula for setting the plasma processing device to etch the sample; and The processing formula is set for the plasma processing device.

Images