Semiconductor process parameter tuning method and system

Abstract

The application discloses a semiconductor process parameter optimization method and system. The method comprises the following steps: receiving a process target; in a closed-loop iterative optimization cycle, controlling a device to perform a process based on a current process recipe and obtaining an actual process result; if the result is not within a tolerance range, a large language model is combined with the current recipe, the actual result and historical data and expert rules in background data to generate a set of adjusted process parameters as the current recipe for the next iteration; the cycle continues until the actual process result meets the requirements, and the final recipe is applied. By constructing an automatic closed-loop optimization framework driven by a large language model, the application realizes multi-dimensional collaborative intelligent optimization of process parameters, improves the efficiency and success rate of process development, reduces the dependence on the experience of senior engineers, and has self-adaptive learning ability in different field environments.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and more specifically to a method and system for optimizing semiconductor process parameters. Background Technology

[0002] Among the many processes in semiconductor manufacturing, key steps include etching, photoresist coating, development, and resist removal. The results of these processes, such as the thickness of the film on the substrate, the etching depth, and uniformity, have a decisive impact on the final performance and yield of the finished integrated circuit chip.

[0003] The above process parameters include not only the temperature, time, concentration ratio of the chemical solutions used, and flow rate set in the equipment formula, but also the motion trajectory and path of the moving parts in the equipment, such as the motion trajectory and motion parameters of the swing arm supplying the spin coater or etching solution, or the control curve of the glue pump in the spin coater. In addition, the influence between these parameters is generally considered to be mutually coupled.

[0004] Currently, the debugging and optimization of semiconductor processes still largely rely on the prior knowledge and historical experience of process engineers. In short, it primarily depends on manual debugging and verification by process engineers from the equipment supplier or the host equipment supplier. The process engineer first needs to define the process objectives, and then set a set of initial process parameters based on these objectives and their own past experience. Based on these initial parameters, they use equipment to process one or more test substrates. After the process is completed, they obtain the process results through measurement equipment and analyze the deviation between these results and the preset objectives. Based on their understanding, they optimize the parameters and continue experimenting. Through this iterative debugging process, the final process result is obtained.

[0005] This manual debugging method has a very long debugging cycle and is limited by the knowledge level of process engineers, making it impossible to guarantee the convergence of the iterative process. A complete cycle of experimentation and measurement analysis often consumes a significant amount of time; engineers may need to spend weeks or even months to obtain a feasible process formulation, which places a considerable burden on the manpower of equipment companies and FAB plants. Moreover, the final debugging effect heavily depends on the individual experience level of the engineers, making it difficult to effectively replicate and pass on process development capabilities, and thus lengthening the training cycle for new engineers.

[0006] More importantly, due to the interdependent nature of process parameters, engineers tend to use single-variable control during manual debugging to manage variables, which may only lead to the discovery of a locally optimal process formulation. Currently, while some solutions use machine learning models for training and parameter recommendation, these models are typically trained on data from specific environments. When equipment is deployed to a customer's site where environmental conditions, material batches, or upstream processes have subtle changes, the model's performance significantly degrades, exhibiting poor adaptability and robustness, often requiring further time for matching and retraining. Summary of the Invention

[0007] To address the problems of existing semiconductor process parameter tuning methods, which heavily rely on manual experience, are inefficient, and struggle to achieve global optimization of multiple parameters, this invention provides a method and system for semiconductor process parameter tuning.

[0008] The present invention provides a method for optimizing semiconductor process parameters, comprising the following steps:

[0009] S1. Obtain process objectives;

[0010] S2. In the iterative optimization loop, repeat the following steps:

[0011] S21. Based on the current process formula, control the semiconductor process equipment to perform the corresponding process on the substrate, wherein the current process formula is defined by the values ​​of preset process parameters;

[0012] S22. Obtain the actual process results formed on the substrate after the process is executed;

[0013] S23. If the actual process result does not fall within the process target, the large language model generates adjusted process parameters based on process data including the current process formula and the actual process result, combined with background data, and uses the formula containing the adjusted process parameters as the current process formula for the next iteration.

[0014] S3. The iterative optimization loop is terminated and the current process formula is applied until the actual process result falls within the process objective.

[0015] Preferably, the background data includes:

[0016] An experience database is used to store historical process data, which includes at least historical process formulas and corresponding historical process results; and

[0017] A rules database is used to store pre-defined expert knowledge, physical laws, or process constraints.

[0018] Preferably, before the iterative optimization loop begins, the following steps are also included:

[0019] The large language model retrieves relevant historical process data from the experience database based on the process objective, and combines it with data obtained from the rule database to generate an initial process formula, which is then used as the current process formula in the first round of the iterative optimization loop.

[0020] Preferably, after terminating the iterative optimization loop in step S3, the method further includes:

[0021] The data, including the current process formula and the corresponding actual process results, are stored as success cases in the experience database to achieve dynamic growth of the experience database.

[0022] Preferably, it further includes:

[0023] The current process formula and the actual process result that fail to fall within the process objective during the iterative optimization loop are stored as failure cases in the experience database.

[0024] Preferably, the historical process data in the experience database also includes equipment environmental parameters when the historical process formula was executed.

[0025] Preferably, in step S23, the step of generating adjusted process parameters using the large language model specifically includes:

[0026] The large language model analyzes the deviation between the actual process results and the process target, and learns the potential patterns in the experience database based on the adjustment strategies or causal relationships related to the deviation in the rule database, and infers and generates the adjusted process parameters.

[0027] Preferably, if the actual process result still does not fall within the process target after the iterative optimization loop reaches a preset maximum number of iterations, then at least one of the following shall be executed:

[0028] Terminate the iterative optimization loop;

[0029] The iterative optimization loop is terminated, and the entire iterative process is analyzed by the large language model to generate adjustment suggestions, which are used to guide the modification of the process objectives or initial constraints.

[0030] Preferably, in step S1, the constraint range of the process parameters set by the user is also received;

[0031] When generating the adjusted process parameters, the large language model ensures that the adjusted process parameters are within the constraints; or,

[0032] The adjusted process parameters generated by the large language model are in a structured data format, allowing the semiconductor process equipment to directly parse and execute them; or,

[0033] The process parameters include at least one of the following: temperature, time, concentration ratio of chemical solutions, flow rate, discrete point set of the speed or position of moving parts, or a curve description of the speed or position of moving parts in the process formulation.

[0034] Preferably, the process is etching, and the actual process result is etching rate and / or etching rate uniformity; or,

[0035] The process is uniform coating, and the actual process result is film thickness and / or film thickness uniformity.

[0036] Preferably, in step S3, before generating the adjusted process parameters from the large language model, the method further includes:

[0037] Structured prompts are generated, which include the role assigned to the large language model, the current state of the process, relevant data extracted from the background data, and a clear task description, to guide the large language model to perform constrained reasoning.

[0038] This application also provides a semiconductor process parameter tuning system, including:

[0039] Semiconductor process equipment;

[0040] Background data;

[0041] Large language model service;

[0042] A control device, communicating with the semiconductor process equipment, the background data, and the large language model service, is configured to perform the following operations:

[0043] S1. Receiving process objectives;

[0044] S2. In the iterative optimization loop, repeat the following operations:

[0045] S21. Control the semiconductor process equipment to execute the process based on the current process recipe.

[0046] S22. Obtain the actual process results formed on the substrate after the process is executed;

[0047] S23. If the actual process result does not fall within the process target, the current process formula and the actual process result are sent to the large language model service to trigger the large language model service to generate an adjusted process formula, and the adjusted process formula is used as the current process formula for the next iteration.

[0048] S3. The iterative optimization loop is terminated and the current process formula is applied until the actual process result falls within the process objective.

[0049] Preferably, the background data includes:

[0050] An experience database is used to store historical process data, which includes at least historical process formulas and corresponding historical process results; and

[0051] A rules database is used to store expert knowledge, physical laws, or process constraints predefined by equipment manufacturers.

[0052] Preferably, before the iterative optimization loop begins, the control device is further configured to:

[0053] The process objective is sent to the large language model service to trigger the large language model service to generate an initial process recipe based on relevant historical data in the experience database and data in the rule database, and the initial process recipe is used as the current process recipe in the first round of the iterative optimization loop.

[0054] Preferably, after terminating the iterative optimization loop, the control device is further configured to:

[0055] The process objective, the applied current process formulation, and the corresponding actual process results are sent as success cases to the experience database for storage; or,

[0056] The large language model service is configured to analyze the deviation between the actual process results and the process target, and based on the adjustment strategies or causal relationships related to the deviation in the rule database, infer and generate an adjusted process formula; or,

[0057] The control device is also configured to monitor the number of iterations of the iterative optimization loop. If the actual process result still does not fall within the process target after reaching the preset maximum number of iterations, the iterative optimization loop is terminated, and the large language model service is triggered to generate adjustment suggestions.

[0058] Preferably, the control device communicates with the semiconductor process equipment via an industrial bus protocol, and the control device communicates with the large language model service via an HTTP protocol.

[0059] Alternatively, after receiving the adjusted process formula generated by the large language model service, the control device performs a format verification on it. If the verification fails, it requests the large language model service to regenerate it.

[0060] Alternatively, the control device may also provide a user interface for receiving operator commands to monitor, manually stop, or intervene in the iterative optimization loop.

[0061] Alternatively, the semiconductor process equipment may further include an online measurement module, which is integrated into or connected to the semiconductor process equipment, for automatically acquiring the actual process results and transmitting them to the control device.

[0062] Preferably, it further includes:

[0063] The relay service layer acts as middleware between the control device and the large language model service.

[0064] The control device sends data including the current process formula and the actual process results to the relay service layer.

[0065] The relay service layer receives data from the control device, retrieves relevant historical data and rules from the background data, integrates all data into structured prompt words and sends them to the large language model service, receives the adjusted process formula generated by the large language model service, and forwards only the adjusted process formula to the control device.

[0066] Preferably, the background data is managed and maintained by the relay service layer to achieve physical isolation between the control device and the background data; or,

[0067] During the iterative optimization loop, the control device sends data including the current process formula and the actual process results to the transfer service layer, which processes the data and stores it in the background data.

[0068] The semiconductor process parameter optimization method and system disclosed in this invention significantly improves process development efficiency and success rate compared to existing technologies. Its automated closed-loop iterative optimization framework transforms the traditional reliance on manual engineer debugging cycles into an iterative process automatically executed by the system. This technical solution utilizes the powerful analysis and reasoning capabilities of large language models to comprehensively consider the complex coupling relationships between multiple process parameters based on current formulations, actual results, and massive historical data. It performs multi-dimensional collaborative adjustments to find a globally optimal process window, thereby obtaining a better-performing process formulation or improving debugging efficiency. Furthermore, it enables self-updating and adjustment of data, providing better field adaptability and robustness. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the basic architecture of the semiconductor process parameter tuning system described in an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of the core process of the semiconductor process parameter optimization method described in this embodiment of the invention;

[0071] Figure 3 This is a detailed schematic diagram illustrating the composition of the background data described in an embodiment of the present invention;

[0072] Figure 4 This is a schematic diagram of the architecture of the semiconductor process parameter optimization system using a relay service layer as described in an embodiment of the present invention;

[0073] Figure 5 This is a detailed flowchart illustrating the semiconductor process parameter optimization method including anomaly handling and knowledge integration as described in an embodiment of the present invention.

[0074] Figure 6 This is the process result change curve in Embodiment 4 of the present invention.

[0075] List of reference numerals

[0076] 10: Semiconductor process equipment; 11: Online measurement module; 20: Background data; 21: Experience database; 22: Rule database; 30: Control device; 31: Industrial control computer; 32: User interface; 40: Large language model service; 50: Transit service layer. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0078] In the description of this invention, it should be understood that the terms center, longitudinal, transverse, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0079] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "multiple" means two or more, unless otherwise explicitly specified.

[0080] Example 1

[0081] This embodiment provides a semiconductor process parameter tuning system, which aims to solve the technical problems of traditional semiconductor process parameter tuning, which relies heavily on manual experience, is inefficient, has difficulty in achieving global optimization of multiple parameters, and has poor adaptability to different customer sites.

[0082] Figure 1 This is a schematic diagram of the system architecture. The optimization system mainly includes semiconductor process equipment 10, background data 20, control device 30, and large language model service 40.

[0083] Semiconductor process equipment 10 is a physical unit that performs a specific process. For example, it can be a spin coater or developer in photolithography, an etching machine in etching, or a photoresist remover. In this embodiment, an etching device used for wet etching is used as an example. Generally, based on existing technology, this device can integrate process cavities such as etching chambers, as well as measurement modules for measuring results, such as film thickness gauges. After etching, without removing the substrate from the process equipment, parameters such as film thickness on the substrate surface can be automatically measured. Based on the film thickness data at various points on the substrate, we can also obtain etching rate data calculated based on the difference in film thickness before and after the process and time. The etching rate is typically a key process parameter in the etching process.

[0084] The control device 30 is the central hub of the entire system, typically an industrial control computer. The control system is not necessarily used independently to implement the control functions of this solution; it may, and in most cases, integrate multiple different functional requirements. The industrial control computer runs equipment control software and can be configured or connected to a user interface. The control device 30 communicates at high speed and reliably with the various motion control modules and sensors of the semiconductor process equipment 10 via agreed-upon industrial bus protocols, such as EtherCAT, to send process recipe instructions and collect and retrieve equipment status data. Simultaneously, the control device 30 communicates with the large language model service 40 via the standard HTTP protocol, a method that is highly versatile and easy to deploy and integrate.

[0085] The large language model service 40 is the core of the system's intelligent decision-making. It can be deployed on a specific cloud server to utilize the powerful computing resources of the cloud, such as cloud models provided by cloud service providers or large language model vendors. Alternatively, for data security and response speed, the large language model service can be deployed locally at the device end customer or equipment provider. For situations with sensitive data requirements or where the customer explicitly prohibits the transmission of device data to third parties, we recommend using a locally deployed large language model service. In this case, it is preferable for the equipment manufacturer to deploy a separate local large language model service, which is then invoked by the control terminals of the sold or used equipment. The service is deployed on a local server at the customer's factory. The large language model service 40 receives requests from the control device 30, performs complex analysis and reasoning, and generates adjusted process parameters. Generally, to accurately utilize the thinking and reasoning capabilities of the large language model, we may provide additional information, presented as background information. Finally, the large language model service 40 can also be a model fine-tuned based on a specific base model and combined with the process optimization scenario applied in this solution.

[0086] Background data 20 provides the background data and knowledge required for decision-making in the large language model service 40. Background data 20 can be singular or distributed, meaning that background data can be distributed in different locations. Background data can be appropriately functionally decomposed according to this principle to form background data modules for different purposes, as detailed later.

[0087] The core of this semiconductor process parameter tuning method is an iterative optimization process that involves a large language model forming a closed loop, such as... Figure 2 As shown. First, in step S1, the process engineer inputs the required process objectives, such as film thickness control objectives and etching rate data, through the user interface 32 of the control device 30. Taking etching rate as an example, the process objective is usually defined as the target value of the etching rate or with an acceptable tolerance range. For example, we can set the target etching rate to 500 angstroms per minute, requiring an error within 5% on one side. The user input can be either a single target value or an error range. In this case, the unset value is the default value, which is generally the initial value saved in the system or a historical value recorded in the previous round. In other words, regardless of how the user inputs, the final process objective generated by the system is still based on the target value and the error range.

[0088] Then, the system enters the iterative optimization loop during process debugging, which is step S2. In each iteration of step S2, in each iteration of the loop, step S21 is executed first. The control device 30 controls the semiconductor process equipment 10 to perform the process on the substrate based on the determined current process recipe. The current process recipe includes detailed process parameters, such as the segmented speed settings of the substrate spin coating speed curve during etching, the position curve of the robotic arm, the supply flow rate of the etching solution, etc.

[0089] In step S22, after the above etching process is completed, the online measurement module 11 inside the device 10 automatically measures the film thickness at various points on the substrate to obtain the actual process results. For the etching rate, the film thickness of the substrate before the process starts is known, or it can be obtained through the operation before the process starts. The calculation of the process time is not difficult. Therefore, the etching rate data at each measurement point can be calculated by measuring the film thickness data at various points on the substrate after etching.

[0090] In step S23, the control device 30 determines whether the actual process result falls within the preset tolerance range. In this embodiment, this means confirming whether the etching rate at each point meets the preset range of 500 ± 25 angstroms per minute. If the etching rate does not fall within the preset range, it means that the process formulation cannot meet the conditions and needs further optimization. The control device 30 then sends the process data, including the current process formulation and the actual process result, to the large language model service 40. The large language model service 40 generates a set of adjusted process parameters based on this information and the existing background data. After receiving this set of adjusted parameters, the control device 30 updates it to the current process formulation for the next iteration. In a more preferred embodiment, the controller can repeatedly execute the same process, such as etching, several times, for example, three times, to confirm whether the process operation is stable and controllable. If the results of several runs meet the conditions or the range after execution is reasonable, the current process formula is output. However, if the result of a single run is still unstable, it means that the current process formula that we obtained, which seems to meet the conditions, is incorrect. In this case, an alarm can be output and the process can wait for processing. Alternatively, it is preferable to choose to continue to return to the loop to iterate the current process formula.

[0091] Then, in the final S3 step, the system continuously repeats the loop of step S2 until the actual process results from a certain iteration, or more rigorously, several consecutive iterations, all successfully fall within the preset tolerance range. At this point, the iterative optimization loop terminates. In the advanced operation steps above, after obtaining a formula that meets the preset tolerance range, the iterative optimization loop can also terminate after repeatedly verifying the stability of the formula. The control device 30 saves the last used current process formula as the final, verified process formula and prompts the engineer that this formula can be applied for subsequent mass production. This essentially frees engineers from tedious repetitive experiments, significantly improving the efficiency and success rate of process development. Moreover, similar solutions can make full use of historical data and other information, and their potential ceiling is expected to be higher than relying solely on the engineer's experience.

[0092] Example 2

[0093] like Figure 3 As shown, this embodiment is intended to explain in detail the composition of background data 20 based on embodiment one, and also to explain in detail the specific ways and methods of its application.

[0094] In addition to process data such as the current process formula, actual process results, and process objectives, the information sent by the control device 30 to the large language model service 40 also requires additional background information to effectively utilize the reasoning capabilities of the large language model to generate highly usable adjustment suggestions. Providing additional background information is almost essential for the efficient and accurate operation of the entire process. If only the inputs and outputs of the current single iteration or the current process are provided, the large language model lacks the necessary process knowledge and equipment characteristic constraints. The adjustment suggestions it returns may lack directionality and fail to take into account the specific conditions of the equipment, or even deviate from natural laws. This can lead to slow convergence or even complete failure of the optimization process.

[0095] Therefore, we need to approximate the process optimization problem as closely as possible to a constrained reasoning task with complete information, which requires providing it with sufficient effective information—that is, the role of background data 20. In fact, in our improved design, the control device 30 sends more structured input to the large language model service 40. This request not only includes the current state but, more importantly, relevant knowledge extracted from the background data 20. For example, our request is generally organized into the following logical structure: The current goal is to achieve a specific etching rate and uniformity range; we adopted process formula A and obtained actual result B, which did not meet the goal; please refer to the following historical data and process rules, analyze the causal relationship between the current formula A and result B, and provide an adjustment suggestion, explaining how to modify specific parameters in formula A, such as the spindle speed curve or etching fluid flow rate, so that the result of the next execution is closer to the goal.

[0096] Background data 20 generally refers to the historical data and / or process rules indicated in the input prompts. Historical data primarily records data sets with process formulas and operational results. These data sets are typically filtered to ensure they are essentially identical to the operating conditions of the equipment for which the process formula is to be determined. The historical data and process rules indicated in the input above are usually what background data 20 provides. This background data 20 stores a large amount of process formulas and corresponding process result data collected and accumulated during actual equipment operation. We don't actually care about the specific composition of this historical data; we generally don't filter the usage results. In other words, it contains successful cases that achieve or approach the process target, and it also allows for failures or intermediate process cases that do not reach the target. By providing this data as background information to the large language model, the model can analyze the potential patterns and trends in the impact of specific process parameter adjustments on process results, thus providing biased, data-driven adjustment suggestions rather than random guesses. This helps the large language model discover effective adjustment patterns.

[0097] Process data may come from an experience database 21, which can be static or dynamic. Each record stored in the database constitutes a complete process case, which includes at least a historical process recipe, the corresponding historical process result, and the equipment environmental parameters when executing that historical process recipe. Equipment environmental parameters may include, for example, the temperature, humidity, and air pressure within the chamber. This data is crucial for accurately reproducing and predicting process results, as even small fluctuations in the environment can have a significant impact on the process. This data is continuously accumulated through actual operation at the customer's site.

[0098] In addition, background data may include predefined process rules and other rule databases, which are equivalent to explicit constraints. The sources of process rules are mainly twofold. First, there are pre-defined constraints on the process formulation based on equipment and process conditions. For example, rules might limit the maximum swing range, maximum speed, and maximum acceleration of the etching nozzle's swing arm based on its design, or limit the spindle speed range to 100-5000 RPM. The other part of the process rules consists of expert knowledge summarized by process engineers. For example, it might record information such as the etching rate range of 100-600 angstroms per minute for a specific photoresist for a particular type of etching solution, such as TMAH solution. Or, for example, the pre-collected rule data regarding the temperature-dependent etching rate of a specific photoresist for TMAH solution. Process rules can be predefined by equipment manufacturers, have a wide range of applications, and usually do not require frequent changes. Furthermore, the process rules included in these rules need to undergo an internal review process to ensure their guidance and accuracy, and to avoid transmitting erroneous information to the large language model. Process rules typically originate from a rules database 22, which is maintained and updated uniformly by the equipment manufacturer. For end users of FAB, it is usually read-only or invisible, which ensures the authority of the knowledge and protects the manufacturer's core technical secrets.

[0099] This is equivalent to providing guidance and constraints to the large language model from both top-level design and actual process dynamics when the large model replaces human-generated process recipes. Therefore, it can relatively constrain the directionality and effectiveness of the output of the large language model, ensure the reliability of the output scheme under unpredictable conditions, and guarantee that the process iteration can not only converge but also consume less time.

[0100] By utilizing this dual-database structure, the intelligence level of our optimization method can be further improved. In this case, the initial process recipe can be generated directly by the large language model. However, if the large language model has limited adjustment capabilities, directly using it to generate the initial process recipe as the starting point for refining the process recipe often results in a significant discrepancy between the initial recipe and the final result, making it difficult to converge to the optimal solution. Even if convergence occurs, the time consumption is often unacceptable. Therefore, while not entirely unacceptable, we usually have the process engineer determine the initial process recipe first. This can be using the equipment's existing process recipe or a modified version by the process engineer. If the large language model is used to generate the process recipe, an initial process recipe generation step can be added before the iterative optimization loop begins. After the engineer inputs the process objective, the control device 30 first sends the process objective to the large language model service 40. Based on this process objective, the large language model service 40 searches the background process data for the most relevant historical process data. Then, it combines the general physical laws and process constraints obtained from the process rules to perform comprehensive analysis and reasoning on the retrieved historical data, thereby generating an initial process recipe. This initial process recipe, as the first round of the iterative optimization loop, is usually closer to the optimal solution than a starting point set by engineers based on experience, thus significantly reducing the number of subsequent iterations.

[0101] Furthermore, to achieve the system's self-learning and personalized evolution capabilities, after terminating the iterative optimization loop in step 3, the method also includes a knowledge accumulation step. The control device 30 stores the complete record of this successful optimization, including the initial process objective, the current process formula of the application, and the corresponding actual process result, as a new success case in the experience database 21, enabling dynamic growth of the experience database. Thus, the content of the experience database 21 will dynamically grow with each successful application at the customer site. Further, to allow the system to learn from failures, all process formulas that failed to bring the actual process result within the preset tolerance range during the iterative optimization loop, the corresponding actual process results, and the adjustment process performed by the large language model can also be stored as failure cases in the experience database 21. These failure cases are extremely valuable for the large language model to understand the boundaries of the process window and avoid repeating mistakes in future decisions, significantly accelerating the convergence speed of subsequent optimization tasks. In this way, we combine dynamic customer experience and static expert rules, enabling the system not only to solve current problems but, more importantly, to continuously learn and self-optimize in specific environments. This can optimize and solve the problem of traditional models being unsuitable for different customer sites.

[0102] Example 3

[0103] like Figure 4 As shown, in fact, to address data security and intellectual property protection issues that may arise in actual deployments, we have also attempted to provide a more optimized system architecture. In some application scenarios with extremely high data security requirements, such as when customers explicitly prohibit any process-related data from leaving the factory, or when equipment manufacturers wish to protect their core process rules and accumulated process data from direct customer access, we can introduce a transit service layer 50. This transit service layer 50 serves as a secure isolation layer between the control device 30 and the large language model service 40, and also as a background data processing and interaction layer.

[0104] In this architecture, the control device 30 deployed at the customer site does not communicate directly with the large language model service 40 as in the above embodiment, nor does it need to directly access the complete background data 20. Instead, the control device 30 sends a relatively concise request to the relay service layer 50. This request mainly includes contextual data of the currently ongoing tuning task, such as the process formula used in the current round, the actual process results obtained by measurement, and, preferably, all historical iteration data since the start of this tuning task, i.e., the formula and result data of each round, in order to help the model better understand the optimization process. It also includes data such as the specific operating conditions of the equipment.

[0105] The relay service layer 50 is typically deployed and maintained by the equipment manufacturer. It can be deployed on the same local server as the large language model service 40, or it can be deployed in the manufacturer's own cloud. The core responsibility of the relay service layer is to receive requests from the control device 30, and then refine them into complete and structured prompts required by the large language model before submitting them to the large language model service 40. Specifically, the relay service layer 50 generally manages complete background data 20, including a rule database 22 maintained by the manufacturer and a dynamically growing experience database 21. Upon receiving a request, the relay service layer 50, based on the process objectives and current status in the request, forms a complete input request to be sent to the large language model service 40, building upon the original request provided by the control device 30. In fact, to fully utilize the capabilities of the large language model and reduce noise information, the relay service layer 50 preferably retrieves the most relevant historical cases and process rules from the two databases and integrates this information with the real-time iterative data sent by the control device 30 to form the final complete input sent to the large language model.

[0106] This retrieval process can include at least two preferred screening criteria. For example, the first criterion is screening based on the similarity of process objectives. The relay service layer 50 analyzes the process objectives of the current tuning task, such as the target etching rate and target uniformity range, and uses this as a benchmark to search for a set of historical cases with similar or close process objectives in the experience database 21, or determines the selection dataset based on a given threshold range. That is, the relay service layer 50 attempts to provide the large language model with a direct reference for the current specific optimization direction by screening out historical data that meets the process objectives and sending it to the large language model. This helps the large language model quickly locate potentially effective parameter adjustment ranges, thereby focusing the large language model's attention on solving the limited problem and within the limited scope, rather than analyzing all historical data in a general way, in order to improve the speed of convergence to the target solution.

[0107] The second criterion is based on the similarity of the process environment and equipment conditions. The relay service layer 50 obtains the real-time status parameters of the current equipment from the data sent by the control device 30, such as cavity temperature, specific parameters of the processed wafer, and other environmental conditions that may affect the process results. Then, it matches historical cases executed under similar or identical conditions in the experience database 21. This filtering method provides historical data to the large language model that is usually highly referential and reproducible with the current equipment conditions. However, because semiconductor processes are extremely sensitive to environmental conditions, similar inputs may produce drastically different outputs in different environments. Therefore, by constraining this variable, the model can avoid erroneous reasoning due to differences in equipment or environment, making its suggestions more consistent with the current physical reality.

[0108] In practice, especially after the experience database has accumulated a sufficient amount of data, the two criteria mentioned above are usually used in combination. For example, a preliminary screening can be conducted based on the process environment and equipment conditions to identify a valid historical dataset that matches the current state. Then, within this dataset, the dataset can be sorted or further screened according to the similarity of the process objectives, and finally, the most relevant cases can be selected.

[0109] After processing the complete input, the large language model service 40 returns its generated adjustment suggestions to the relay service layer 50. Upon receiving the suggestions, the relay service layer 50 can first perform format validation or content parsing, and then forward only the core information—the adjusted process formula parameters—to the control device 30. After receiving the new formula parameters, the control device 30 continues to execute the next round of process iteration. The main purpose of introducing the relay service layer 50 in the optimization scheme is to achieve physical isolation between the client and the core knowledge base. The client's control device 30 can only access the inputs and outputs directly related to the current task, and cannot access the expert knowledge in the rule database 22 or the valuable data accumulated by other clients in the experience database 21, thus effectively protecting the intellectual property rights of the equipment manufacturer. At the same time, the update process of the dynamically growing experience database 21 is also more secure and centralized. After an optimization task is completed, whether successful or unsuccessful, the control device 30 will send the complete task record, including all iteration processes, to the relay service layer 50, which is responsible for archiving and storing this data in the experience database 21. Because the data is centrally managed, we can effectively ensure data consistency and security. Furthermore, we can leverage the quantity of local equipment from equipment manufacturers to obtain and dynamically update relevant process data. The system collects experience from all deployed equipment, thus forming a global and continuously evolving knowledge system, enabling the overall intelligence level of the system to continuously improve.

[0110] We also do not rule out another possible solution: each control device 30 maintains an independent database locally and sends relevant data to the relay service layer upon request. In this solution, the relay layer 50 can also maintain a basic rule database 22 and an experience database 21. In this case, the databases maintained by the control device 30 and the experience database 21 maintained by the relay layer 50 typically have separate responsibilities: the relay layer is responsible for updating the overall basic data, while the control device 30 provides updates related to the device itself. However, we still recommend the model where the relay service layer centrally manages all background data. This is not only because it reduces the burden of network transmission and simplifies the complexity of data synchronization and maintenance, but more importantly, because our equipment manufacturers have addressed the most critical data security and intellectual property protection issues in commercial solutions.

[0111] To illustrate the above embodiments, we will use a practical example. Taking the currently sold XX model single-wafer wet etching machine as an example, it supports 8-inch (200mm diameter) wafers. This etching process involves removing a silicon nitride (SiN) thin film from the wafer surface, therefore a high-temperature phosphoric acid etching solution is used. Our desired process targets are an average etching rate of 80 Å / min and a uniformity of less than 1.5%.

[0112] However, the current equipment process achieves an average etching rate of 82 Å / min, but the uniformity is severely out of control, reaching 3.5%, exhibiting a typical pattern of excessively fast etching in the central region and excessively slow etching in the edge regions. The system has undergone two rounds of automatic optimization, but the uniformity problem has not been significantly improved, and this is the third optimization request.

[0113] The following is the structured Markdown prompt content actually generated by the control system to address the above issues.

[0114] #Role You are a semiconductor wet etching process expert with over 20 years of experience, proficient in fluid dynamics, chemical reaction kinetics, and semiconductor equipment control principles. Your task is to analyze process data, diagnose process problems, and provide precise, actionable process formulation adjustments. #Current Situation We are currently working on an optimization task for a silicon nitride thin film phosphoric acid etching process on 200mm wafers. The current goal is to achieve an average etching rate of 80 Å / min and a uniformity of less than 1.5%. In the third round of optimization iterations, we encountered a bottleneck. Current process results show an average etching rate of 82 Å / min, which basically meets the requirements, but the uniformity is 3.5%, far exceeding the target range. Measurement data shows that the etching rate in the wafer center region is significantly higher than that in the edge region. #Background To help you make decisions, we provide relevant process rules and historical data. ##Process Rules (from Rule Database 22) ***Equipment Constraints**: *Phosphoric acid temperature controllable range: 140℃-165℃. *Spindle speed controllable range: 10-300RPM. *Etching solution flow rate controllable range: 100-800 ml / min. ***Expert Knowledge**: *Phosphoric acid temperature is the most critical parameter affecting the etching rate. For every 1°C increase in temperature, the etching rate of silicon nitride increases by approximately 1.5-2.0 Å / min. *Increasing the wafer spindle speed can enhance fluid disturbance on the wafer surface, helping to improve the etchant exchange efficiency between the center and edges, thereby improving uniformity. *When the spindle speed is too high (e.g., exceeding 200 RPM), excessive centrifugal force may cause a "spin-drying" effect at the wafer edges, which will reduce the edge etching rate and worsen uniformity. *At a constant spindle speed, increasing the etchant flow rate helps replenish the reactants consumed in the central region, which may have a positive effect on improving center-edge differences.##Historical Data (Source: Empirical Database 21) | Case Number | Phosphoric Acid Temperature (°C) | Spindle Speed ​​(RPM) | Average Etching Rate (Å / min) | Uniformity (%) | Remarks ||---|---|---|---|---|---|---||CASE-034|160|30|80.5|4.2|Center Fast||CASE-035|160|80|81.2|2.1|Uniformity Improved, But Still Exceeds Standard||CASE-036|161|80|83.0|2.0|Rate Slightly High||CASE-071|160| 180|81.5|3.8|Significant edge spin-drying effect|#Current Data***Current Process Target**:*Average Etching Rate: 80 Å / min (Tolerance ±2 Å / min)*Uniformity: <1.5%***Current Process Formulation (Round 3)**:*Phosphoric Acid Temperature: 160.5℃*Spindle Speed: 100 RPM*Etching Fluid Flow Rate: 500 ml / min*Etching Time: 120 s***Current Process Results (Round 3)**:*Average Etching Rate: 82 Å / min*Uniformity: 3.5%*Film Thickness Distribution: Thinnest remaining film thickness at the center, thickest at the edges. #Task 1.**Analysis**: Based on all the above information, please analyze the core reasons why the current formulation failed to achieve the uniformity target. 2.**Recommendation**: Generate a set of adjusted new process formulation parameters for the next round (Round 4) iteration. 3. **Explanation:** In your suggestion, please clearly indicate the parameters that need adjustment, the direction of adjustment, and the recommended values. Based on your understanding of the process, briefly explain the logical basis for this adjustment. #Constraints* Adjustment suggestions must strictly adhere to the equipment capability range defined in the "Process Rules". *Given that the current average etching rate is already close to the target, please prioritize adjusting parameters primarily aimed at improving uniformity (such as spindle speed), and then fine-tune the main parameters affecting the rate (such as phosphoric acid temperature) for compensation. *Please use JSON output format, containing two keys: `adjusted_parameters` (an object containing the specific parameters and their new values) and `reasoning` (an explanatory text).

[0115] Example 4

[0116] This embodiment details how the large language model service 40 generates a set of adjusted process parameters in step 2 of the iterative optimization loop. Taking an etching process as an example, its process objective is to achieve a specific etching depth and less than 1% etching non-uniformity. The process parameters are very complex, including not only scalar parameters such as temperature, time, chemical concentration ratio, and flow rate in the equipment formulation, but also curves representing the speed or position of moving parts of the equipment, or the speed or position of moving parts defined by a series of discrete point sets. For example, the oscillation speed curve of the etching spray arm is a key process parameter. This curve is usually defined by a series of discrete point sets, where each point represents a position and a corresponding speed. These points together determine the distribution of the chemical solution on the substrate surface and have a decisive influence on etching uniformity.

[0117] Suppose that after one iteration, the actual process results show that the etching depth at the center of the substrate meets the standard, but the etching depth in the edge area is insufficient, causing the overall uniformity to exceed the preset tolerance range. The control device 30 sends data containing the current etching formula and the non-uniformity result to the large language model service 40.

[0118] After receiving the data, the large language model service 40 first analyzes the deviation between the actual process results and the process target, i.e., the center meets the target, but the edges are insufficient. Then, the large language model service 40 infers based on the adjustment strategies or causal relationships related to this deviation in the rule database 22. The rule database 22 may store the following rules: increasing the residence time of the spray arm in the substrate edge region can improve the etching rate of the edge region; increasing the total flow rate of the chemical solution can improve the overall etching rate, but may have a greater impact on the center region; adjusting the terminal acceleration of the spray arm swing speed curve can improve the uniformity of the edge transition zone.

[0119] Unlike process engineers who only adjust one parameter at a time, large language models can understand the complex coupling relationships between these parameters. They comprehensively consider the aforementioned rules and perform multi-objective optimization reasoning. It may astutely determine that simply extending the edge dwell time, while solving the problem of insufficient edge etching, might lead to an excessively long total etching time, resulting in over-etching in the central region. Therefore, a better approach is to adjust multiple parameters simultaneously.

[0120] Finally, the large language model service 40 will generate a set of adjusted process parameters. These parameters may include: slightly reducing the total flow rate of the liquid by 2%, while modifying the motion speed curve of the spray arm, specifically by reducing its oscillation speed in the edge region and fine-tuning the acceleration parameters of the curve.

[0121] To ensure direct execution by the equipment, the adjusted process parameters generated by the large language model service 40 are in a strictly structured data format, such as JSON or XML. Upon receiving this structured data, the control device 30 first performs format validation. If validation fails, for example, due to corrupted data format or missing necessary fields, the control device 30 requests the large language model service 40 to regenerate it. This validation mechanism is crucial for ensuring communication robustness and stable system operation, preventing erroneous instructions from being sent to the underlying equipment, thereby preventing potential equipment damage or process accidents. After successful validation, the control device 30 can directly parse the data and convert it into control instructions for the underlying hardware of the semiconductor process equipment 10.

[0122] Furthermore, during the initial task setup, engineers can set specific process parameter constraints on the user interface 32, i.e., receive user-defined process parameter constraints, such as the liquid temperature must be between 20 and 25 degrees Celsius. The large language model service 40 strictly adheres to these user-defined constraints when generating all adjusted process parameters, ensuring the safety and controllability of the entire optimization process and guaranteeing that all generated parameters are within the stated constraints.

[0123] Taking photoresist etching and development as an example, and using our single-wafer wet etching equipment supporting 6-12 inch wafers as an example, this etching process, because it involves controllable etching after photoresist development, requires precise control of critical dimensions. We use 2.38% TMAH developer / etcher, with a target etching rate of 500 Å / min and a uniformity target of ≤3.5% on a 12-inch wafer (measured at 49 points). Currently, after the third iteration of optimization, the average etching rate is 565 Å / min, deviating from the target value; more seriously, the uniformity is 5.8%, exhibiting a significant phenomenon of faster etching at the center and slower at the edges, far exceeding the specification requirements. Furthermore, the scanning path of the swing arm is defined from 40 degrees (one edge of the wafer) to 57 degrees (the other edge of the wafer), with 48.5 degrees corresponding to the wafer center. Its scanning speed can be programmed in segments to form a speed curve. The intermediate service layer needs to combine the current iteration data, historical success cases, and equipment process rules to request the optimized formula for the fourth iteration from the large language model.

[0124] #Role You are an expert system specializing in semiconductor wet process optimization. Your task is to analyze the causes of current process formulation failure based on the multi-dimensional data provided by you, and to provide a specific, executable, and complete process formulation adjustment plan for the next iteration, paying particular attention to the synergistic optimization between multi-dimensional parameters. #Current Situation Currently executing a photoresist controllable etching process optimization task for 12-inch wafers. The process target is an etching rate of 500 Å / min and a uniformity of ≤3.5%. After 3 iterations, the current process results are seriously substandard. -**Current Round**: #3 -**Average Etching Rate**: 565 Å / min (Target: 500 Å / min) -**Etching Rate Uniformity**: 5.8% (Target: ≤3.5%) -**Main Problem**: The etching rate is too fast, and the etching rate in the center of the wafer is significantly higher than that in the edge areas, resulting in poor uniformity. Previous adjustment strategies have failed to effectively solve this problem. #Background - **Equipment Model**: xxxx - **Wafer Size**: 12 inches (300mm) - **Process Fluid**: TMAH, concentration 2.38% - **Process Temperature**: 23.0℃ - **Process Type**: Photoresist Etching #Data### 1. Iteration History of This Optimization Task | Round | Spindle Speed ​​Curve (RPM) | Etching Fluid Flow Rate (ml / min) | Swing Arm Scanning Speed ​​Curve (Angle: mm / s) | Average Etching Rate (Å / min) | Uniformity (%) | |:---|:---|:---|:---|:---|: ---|:---||#1(Initial)|[0,500,500,0]|800|[(40:150),(57:150)]|610|7.5||#2|[0,400,400,0]|800|[(40:150),(57:150)]|580|6.8||#3|[0,300,300,0]|800|[(40:150),(57:150)]|565|5.8|*Note 1: The spindle speed curve format is [acceleration segment target, constant speed segment target, deceleration segment target, stop]. **Note 2: The arm scanning speed curve is defined by key control points, and the format is [(angle:speed),...]. For example, `[(40:150),(57:150)]` means maintaining a constant speed of 150 mm / s from 40 degrees to 57 degrees.*###2. Detailed process results for the current round (#3) - **Etching rate distribution converted from multi-point film thickness measurement**: - Center point (R=0mm, approximately 48.5 degrees): 605 Å / min - Average of the middle region (R=75mm): 570 Å / min - Average of the edge region (R=145mm, approximately 40 / 57 degrees): 520 Å / min ###3. Relevant historical cases retrieved from the experience database - **Case A (Successful)** - **Process target**: Etching rate 450 Å / min, uniformity ≤1.5% (slightly slower at the center) - **Successful formulation**: Spindle speed [0,600,600,0] RPM, flow rate 750 ml / min, **Hand speed curve [(40:180),(48.5:220),(57:180)] mm / s**. -**Key Conclusion**: By increasing the scanning speed in the central region (48.5 degrees) of the wafer, the liquid interaction time was reduced, effectively compensating for the problem of faster central etching caused by the smaller centrifugal force, and ultimately achieving high uniformity. -**Case B (Failure)** -**Process Target**: Etching rate 600 Å / min, uniformity ≤3% -**Failed Recipe**: Spindle speed [0,200,200,0] RPM, flow rate 900 ml / min, **Operating arm speed curve [(40:200),(48.5:100),(57:200)] mm / s**. -**Failure Cause Analysis**: The spindle speed was too low, and the operating arm speed in the central region was too slow, which aggravated the accumulation of liquid in the center, leading to uncontrolled central etching rate and uniformity deterioration to over 10%. ###4. Relevant Process Rules Retrieved from the Rule Database -**R-01 (Equipment Constraint)**: Spindle speed range: 100-3000 RPM. -**R-02 (Equipment Constraints)**: Etching fluid flow rate range: 500-2000 ml / min. -**R-03 (Equipment Constraints)**: Swing arm angle range: 40-57 degrees; point speed range at any position: 50-300 mm / s. -**R-04 (Expert Knowledge)**: For TMAH etching PR process, the most direct way to reduce the overall etching rate is to reduce the etching fluid flow rate, but too low a flow rate (<600 ml / min) will increase the risk of defects. -**R-05 (Expert Knowledge)**: When the center is fast and the edges are slow, simply reducing the spindle speed will worsen the uniformity due to the weakened centrifugal force. Adjusting the swing arm speed curve should be considered as a priority to compensate. -**R-06 (Expert Knowledge)**: Increasing the spindle speed helps enhance the fluid transport capacity to the edges and is an auxiliary means to improve the slow edge rate, but it needs to be adjusted in conjunction with the swing arm speed curve.**R-07 (Expert Knowledge)**: To improve the phenomenon of fast etching rate at the center and slow etching rate at the edges, the scanning speed of the swivel arm in the wafer center region (around 48-49 degrees) should be increased, while its scanning speed in the wafer edge region (around 40 and 57 degrees) should be appropriately reduced to adjust the effective contact time of the etchant in different regions. **#Task 1. **Analysis:** Combining iteration history, detailed results, historical cases, and process rules, analyze the root cause of the excessively fast rate and poor uniformity in the third-round formulation, especially the limitations of the single-parameter adjustment strategy. 2. **Reasoning:** Based on the analysis, propose a complete and directly executable new process formulation for the fourth iteration. The new formulation must simultaneously aim to reduce the average etching rate and significantly improve uniformity; **the core is designing a reasonable swivel arm scanning speed curve**. 3. **Explanation:** Briefly explain the adjustment ideas and expected effects of your proposed new formulation, especially how the parameters work synergistically. #Constraints - A complete formulation containing three key parameters must be provided: `spindle speed curve`, `etchant flow rate`, and `swivel arm scanning speed curve`. - All parameter settings must strictly adhere to the device constraints defined in the rules database. - The adjustment logic must align with patterns observed in expert knowledge and historical cases, particularly the coordinated adjustment strategy for uniformity control. - The final output formula should be presented in a clear format for easy subsequent analysis. The swing arm scanning speed curve should be provided in the form of angle-speed control point pairs. - The output format should be JSON, containing `adjusted_parameters` (an object containing specific parameters and their new values; for the spindle speed curve and swing arm scanning speed curve, please use key-value pairs for data output) and `reasoning` (an explanatory text).

[0125] If we send the above content to the large language model service, we can get the following output. The control device can parse the data in the adjusted_parameters key of the JSON data in the following output to obtain the corresponding process parameters, which is the updated current process formula.

[0126] {"adjusted_parameters":{"spindle speed curve":[{"step":1,"description":"acceleration stage","type":"ramp","target_rpm":500,"duration_s":3.0},{"step":2,"description":"uniform etching stage","type":"hold","target_rpm":500,"duration_s":45.0},{"step":3,"description":"deceleration and stopping stage","type":"ra mp","target_rpm":0,"duration_s":3.0}],"etcher flow rate_ml_per_min":700,"swing arm scanning speed curve":[{"angle_deg":40.0,"speed_mms":120},{"angle_deg":44.0,"speed_mms":180},{"angle_deg":48.5,"speed_mms":240},{"angle_deg":53.0,"speed_mms":180},{"angle_deg":57.0,"speed_mms ":120}]},"reasoning":"###1. Cause Analysis\n\nAfter in-depth analysis of the data from the first three iterations, the limitations of the single parameter (spindle speed) adjustment strategy were confirmed. The main problems with the current formulation (#3) are: 1) the overall etching rate (565 Å / min) is too high; 2) the uniformity is poor (5.8%), exhibiting a typical distribution of fast at the center and slow at the edges. The root cause is that the constant swing arm scanning speed (150 mm / s) cannot compensate for the radial etching rate difference caused by wafer rotation and hydrodynamic effects, while simply reducing the rotation speed may weaken the transport capacity of the solution to the edges, failing to effectively improve uniformity. \n\n## #2. Refined Reasoning & Recipe Design\n\nTo achieve the goals of the 4th iteration, a refined strategy of multi-parameter collaborative optimization must be adopted. The core of this adjustment is to construct a nonlinear swing arm velocity profile and coordinate the adjustment of flow rate and rotation speed to achieve precise spatial modulation of the residence time of the liquid on the wafer surface.\n\n1. **Swing Arm Scanning Velocity Curve (Core Optimization)**: The issue of the overly sparse control points you pointed out is crucial. To achieve smooth and effective velocity changes, I increased the number of key control points from 3 to 5 to define a more precise parabolic velocity curve that is "fast in the middle and slow at both ends".This not only defines the speed at key locations but also implies the acceleration variation of the swing arm in different zones, which is crucial for controlling the uniformity of the edge transition zone. ***Central Region (48.5°)**: Speed ​​increased to **240 mm / s** to minimize the residence time of the etching solution in the center, directly suppressing over-etching in the center. ***Edge Region (40° & 57°)**: Speed ​​decreased to **120 mm / s**, significantly extending the contact time of the etching solution at the edge to improve the edge etching rate. ***Transition Region (44° & 53°)**: A transition point with a speed of **180 mm / s** is added to ensure smooth and controllable acceleration of the swing arm from the edge to the center and deceleration from the center to the edge, avoiding local flow instability due to sudden speed changes, thereby optimizing the uniformity of the entire wafer. 2. **Etching Solution Flow Rate Adjustment**: To address the issue of an overall high rate, according to rule R-04, the flow rate is reduced from 800 ml / min** to 700 ml / min**. This reduction (-12.5%) aims to bring the average rate back from 565 Å / min to near the target value of 500 Å / min, while maintaining a flow rate above 600 ml / min to mitigate potential defect risks. 3. **Spindle Speed ​​Curve (Time-Dimensional Definition)**: Adopting your suggestion, the spindle speed curve is defined as a precise step with timing control. The constant speed range speed is increased from 300 RPM** to 500 RPM**. According to rule R-06, higher speeds provide stronger centrifugal force, helping to deliver the solution more evenly to the wafer edge. This synergizes with the slower scanning of the swing arm in the edge region, jointly improving the edge rate and serving as an important auxiliary means to improve uniformity. *`Step 1 (Ramp)`: Accelerate from 0 to 500 RPM within 3 seconds to ensure a smooth process start-up. *`Step 2 (Hold)`: Stabilize at 500 RPM for 45 seconds as the main etching time window. *Step 3 (Ramp): Decelerate to 0 within 3 seconds to ensure stability at the end of the process. ###3. Expected Outcome This meticulously designed multi-parameter synergistic formulation, through nonlinear spatial modulation of the swing arm speed curve, overall rate control of the flow rate, and centrifugal force-assisted enhancement of the rotation speed, is expected to achieve two major goals simultaneously in the fourth iteration: ***The average etching rate** will return to near the target of 500 Å / min. ***Etching rate uniformity** will be decisively improved, with the rate difference between the center and the edge effectively compensated, and the uniformity index is expected to enter the target range of ≤3.5%.

[0127] Figure 6 The variation curves of the process objective during the process iteration in this embodiment are presented. In this experimental process debugging, our initial process formulation was designed manually, specifically as follows:

[0128] Spindle speed curve (RPM): 500, the constant speed range is 500 RPM.

[0129] Etching fluid flow rate (ml / min): 800

[0130] The arm scanning speed curve (mm / s): 150, the arm scans at a constant speed of 150mm / s throughout the entire process.

[0131] It's evident that there's a clear uniformity issue with faster speeds at the center and slower speeds at the edges, because the process doesn't compensate for the hydrodynamic differences at different wafer radii. The first three adjustments essentially continued the initial process formulation adjustments, controlling the process by simply reducing the spindle speed, resulting in slight improvements in both rate and uniformity. In the fourth adjustment, the swing arm speed was segmented, which better addressed uniformity while simultaneously reducing the rate. Subsequent operations were optimizations based on the fourth adjustment, which we won't elaborate on here. The process formulation was successfully debugged and finalized after the eighth adjustment. Subsequent tests primarily focused on repeatability. The results showed that the process remained stable within the target window in most cases, but two minor deviations occurred, indicating that our design may still have room for optimization. The final process formulation is as follows:

[0132] Spindle speed curve (RPM): 600

[0133] Etching fluid flow rate (ml / min): 660

[0134] The scanning speed curve of the swing arm (mm / s): [(40:130),(48.5:230),(57:130)], which represent the angle-speed respectively.

[0135] As can be seen, the main improvement is in the speed variation of the swing arm, which ensures uniformity by using a speed at the center point that is much higher than that at the edge point.

[0136] Example 5

[0137] In fact, the above solution is not limited to etching operations; it can also be applied to other fields such as photoresist coating. Taking a single-wafer rotary photoresist coater (model xxxx) for processing 12-inch (300mm) wafers as an example, this device 10 includes not only a vacuum chuck and spindle for rotating the substrate, but also a pumping system for precise photoresist delivery, a coating nozzle that can move on the substrate surface, an edge removal module driven by a swing arm to remove excess photoresist from the substrate edges, and a hot plate baking unit for curing the photoresist film after the process. To achieve closed-loop process control, the device 10 also integrates an online measurement module 11, specifically a multi-point optical film thickness gauge. After the substrate has undergone photoresist coating and subsequent baking processes, it can automatically measure the film thickness at multiple preset points on the substrate surface without manual transfer, thus obtaining actual process results including average film thickness and film thickness uniformity.

[0138] The control device 30 is the central hub of the entire system; in this embodiment, it is specifically an industrial control computer 31. This industrial control computer 31 runs equipment control software and a user interface 32 for process engineers. The control device 30 communicates at high speed and reliably with various motion control modules, sensors, and heating units of the semiconductor process equipment 10 via industrial bus protocols, such as EtherCAT, to send complex process recipe instructions containing multiple steps and to collect equipment status data in real time. Simultaneously, the control device 30 communicates with the large language model service 40 via a standard Hypertext Transfer Protocol, a method that is highly versatile and easy to deploy and integrate in different network environments.

[0139] The large language model service 40 is the core of the system's intelligent decision-making. It can be deployed on a cloud server to leverage the powerful computing resources of the cloud for complex reasoning; alternatively, to ensure data security and reduce network latency, it can be deployed on a local server within the customer's factory. The large language model service 40 receives requests from the control device 30 containing the current process status, performs in-depth analysis and reasoning, and generates a set of optimized process parameters for the next round of experiments. Furthermore, in more advanced scenarios, the large language model service 40 can also be a dedicated model formed by fine-tuning a specific base model combined with a large amount of spin coating process data accumulated in this solution, thereby making its recommendations more domain-specific.

[0140] The background data 20 provides the data and knowledge required for decision-making in the large language model service 40, and is the foundation for ensuring that it generates high-quality recommendations.

[0141] The system's operating method, a semiconductor process parameter tuning method, is based on a closed-loop iterative optimization process. First, in step S1, the process engineer inputs a process objective and a preset tolerance range through the user interface 32 of the control device 30. For example, the process objective is to form a 5000 angstrom thick photoresist film on a 300mm silicon wafer using a specific type of photoresist, with a preset tolerance range within 100 angstroms of the average film thickness, and the film thickness uniformity (i.e., the percentage of the film thickness standard deviation to the average film thickness) being less than 2%.

[0142] Then, the system enters the iterative optimization loop of step S2. In each iteration of the loop, step S21 is executed first, where the control device 30 controls the semiconductor process equipment 10 to perform a spin coating process on a substrate based on a current process recipe. This current process recipe is a multi-step procedure that includes a set of detailed process parameters, such as the movement trajectory and dispensing amount of the photoresist nozzle, the acceleration curve of the substrate rotation, the low-speed rotation speed and time for photoresist spreading, the high-speed rotation speed and time for photoresist spinning, the parameters of the edge stripping step, and the final baking temperature and time on the hot plate, etc.

[0143] Next, in step S22, after the spin coating and baking processes are completed, the online measurement module 11 inside the equipment 10 automatically measures the film thickness at forty-nine points on the substrate, obtains the actual process results, and transmits the complete measurement data, including the average value, standard deviation, maximum and minimum values, to the control device 30.

[0144] Subsequently, in step S23, the control device 30 determines whether the actual process result falls within the preset tolerance range. If it does not fall within this range, for example, if the measured actual average film thickness is 6000 angstroms and the uniformity is 4%, the control device 30 will send the current process formula containing the complete steps, the substandard actual process result, and the target of this optimization task to the large language model service 40. The large language model service 40 will generate a set of adjusted process parameters based on this information and the correlation knowledge obtained from the background data 20. After receiving this set of adjusted parameters, the control device 30 updates it to the current process formula for the next iteration.

[0145] Finally, in step S3, the system continuously repeats the loop of step S2 until the actual process result generated in a certain iteration successfully falls within the preset tolerance range. At this point, the iterative optimization loop terminates, and the control device 30 saves the last used current process formula as the final, verified process formula, and prompts the engineer that this formula can be applied for subsequent mass production. In this way, the entire process of optimizing the spin coating process parameters has been transformed from relying on a lot of trial and error and experience-based judgment by engineers to a highly efficient and intelligent closed-loop system, significantly improving the efficiency and success rate of process development, especially for the development of new adhesive types or new film thickness targets, where the acceleration effect is particularly obvious.

[0146] Example 6

[0147] like Figure 5 As shown, this embodiment mainly illustrates the anomaly handling mechanism and auxiliary functions of the system during the optimization process, so as to ensure its stability and practicality in the actual production environment.

[0148] The system incorporates a protection mechanism to limit the number of iterations. When initiating an iterative optimization loop, the control device 30 monitors the number of iterations. Process engineers can preset a maximum number of iterations, such as 10, through the graphical user interface, or the system may have a default value. If the actual process results still fail to fall within the preset tolerance range after the iterative optimization loop reaches this preset maximum number of iterations, the system will determine that the optimization has failed to converge. In this case, the control device 30 will proactively terminate the iterative optimization loop to avoid endless experiments that waste test substrates and chemicals.

[0149] After the loop terminates, the system does not end directly. Instead, preferably, the control device 30 triggers the large language model service 40 to perform a special analysis task, namely step S4. The large language model service 40 analyzes all attempts throughout the iteration process, including the parameters, results, and adjustment strategies of each round, and generates an adjustment suggestion. Its purpose is to help engineers diagnose problems from a higher perspective. For example, the adjustment suggestion can guide modifications to process objectives or initial constraints. The suggestion might indicate that target uniformity cannot be achieved under the current drug concentration and temperature constraints, suggesting checking the stability of the upstream film-forming process, or trying to relax the drug temperature constraint to 28 degrees Celsius and then retrying. Such advanced suggestions are invaluable for solving fundamental process problems. Then, the control device provides the debugging suggestions and the debugging log history to the process engineer through a graphical user interface or a reserved interface for reference when preparing and debugging process formulations.

[0150] To facilitate process management by engineers, the user interface 32 of the control device 30 provides real-time monitoring of the entire iterative optimization cycle. Engineers can view the current iteration number, the process formula for each round, the actual process results, and the deviation curve from the target on the interface. Simultaneously, the interface also includes manual stop or intervention buttons. If, based on their experience, engineers determine that the current optimization direction has deviated significantly, or if a more urgent task requires the use of the equipment, they can manually stop the optimization process at any time, i.e., receive operator commands to monitor, manually stop, or intervene in the iterative optimization cycle.

[0151] The communication and control links of the entire system have also been carefully designed. As mentioned earlier, the control device 30, namely the industrial control computer 31, communicates with the semiconductor process equipment 10 via a real-time industrial bus protocol, ensuring the accuracy of process execution. The control device 30 communicates with the large language model service 40 via a flexible HTTP protocol, or vice versa, facilitating decoupling and integration with cloud-based or local AI services. This hardware and software combined architecture ensures the stable and efficient operation of the system.

[0152] In summary, the semiconductor process parameter tuning method and system based on a large language model provided by this invention achieves a high degree of automation and intelligence in process parameter tuning by constructing a closed-loop iterative optimization framework driven by a large language model and supported by dual databases. It has the potential to overcome the drawbacks of traditional manual tuning methods and can also optimize and aggregate datasets through continuous learning at the customer site. Due to the inherent flexibility of the large language model, its repeated deployment cost is almost zero, and it does not require customization for each machine like control code. We believe it has the potential to become an efficient, reliable, and adaptive process development paradigm.

[0153] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing semiconductor process parameters, characterized in that, Includes the following steps: S1. Obtain process objectives; S2. In the iterative optimization loop, repeat the following steps: S21. Based on the current process formula, control the semiconductor process equipment to perform the corresponding process on the substrate, wherein the current process formula is defined by the values ​​of preset process parameters. S22. Obtain the actual process results formed on the substrate after the process is executed; S23. If the actual process result does not fall within the process target, the large language model generates adjusted process parameters based on process data including the current process formula and the actual process result, combined with background data, and uses the formula containing the adjusted process parameters as the current process formula for the next iteration. S3. Terminate the iterative optimization loop until the actual process result falls within the process objective, and apply the current process formula.

2. The semiconductor process parameter optimization method as described in claim 1, characterized in that, The background data includes: An experience database is used to store historical process data, which includes at least historical process formulas and corresponding historical process results; and A rules database is used to store pre-defined expert knowledge, physical laws, or process constraints.

3. The semiconductor process parameter optimization method as described in claim 2, characterized in that, Before the iterative optimization loop begins, the following is also included: The large language model retrieves historical process data from the experience database based on the process objective, and combines it with data obtained from the rule database to generate an initial process recipe; the initial process recipe is used as the current process recipe in the first round of the iterative optimization loop.

4. The semiconductor process parameter optimization method as described in any one of claims 2 or 3, characterized in that, After terminating the iterative optimization loop in step S3, the method further includes: Data including the current process formula and the corresponding actual process results are stored as success cases in the experience database to achieve dynamic growth of the experience database.

5. The semiconductor process parameter optimization method as described in any one of claims 2 or 3, characterized in that, Also includes: The current process formula and the actual process result that fail to fall within the process objective during the iterative optimization loop are stored as failure cases in the experience database.

6. The semiconductor process parameter optimization method as described in claim 2, characterized in that, The historical process data in the experience database also includes equipment environmental parameters when the historical process formulas were executed.

7. The semiconductor process parameter optimization method as described in claim 2, characterized in that, In step S23, the step of generating adjusted process parameters using the large language model specifically includes: The large language model analyzes the deviation between the actual process results and the process target, and learns the potential patterns in the experience database based on the adjustment strategies or causal relationships related to the deviation in the rule database, and infers and generates the adjusted process parameters.

8. The semiconductor process parameter optimization method as described in claim 1, characterized in that, If the actual process result still does not fall within the process objective after the iterative optimization loop reaches the preset maximum number of iterations, then at least one of the following shall be executed: Terminate the iterative optimization loop; The iterative optimization loop is terminated, and the entire iterative process is analyzed by the large language model to generate adjustment suggestions, which are used to guide the modification of the process objectives or initial constraints.

9. The semiconductor process parameter optimization method as described in claim 1, characterized in that, In step S1, the user-defined constraint range of process parameters is also received; When generating the adjusted process parameters, the large language model ensures that the adjusted process parameters are within the constraints; or, The adjusted process parameters generated by the large language model are in a structured data format for direct parsing and execution by semiconductor process equipment; or, The process parameters include at least one of the following: temperature and time in the process formula, concentration ratio and flow rate of chemical solutions, discrete point set of the movement speed or position of moving parts, or curve description of the movement speed or position of moving parts.

10. The semiconductor process parameter optimization method as described in claim 1, characterized in that, The process is etching, and the actual process result is the etching rate and / or etching rate uniformity; or... The process is uniform coating, and the actual process result is film thickness and / or film thickness uniformity.

11. The semiconductor process parameter optimization method as described in claim 1, characterized in that, In step S3, before the adjusted process parameters are generated from the large language model, the following steps are also included: Structured prompts are generated, which include the role assigned to the large language model, the current state of the process, relevant data extracted from the background data, and a clear task description, to guide the large language model to perform constrained reasoning.

12. A semiconductor process parameter tuning system, characterized in that, include: Semiconductor process equipment; Background data; Large language model service; The control device communicates with the semiconductor process equipment, the background data, and the large language model service, and performs the following operations: S1, receiving process objectives; S2. In the iterative optimization loop, repeat the following operations: S21. Control the semiconductor process equipment to execute the process based on the current process recipe. S22. Obtain the actual process results formed on the substrate after the process is executed; S23. If the actual process result does not fall within the process target, the current process formula and the actual process result are sent to the big language model service to trigger the big language model service to generate an adjusted process formula, and the adjusted process formula is used as the current process formula for the next iteration. S3. The iterative optimization loop is terminated until the actual process result falls within the process target, and the current process formula is applied.

13. The semiconductor process parameter tuning system as described in claim 12, characterized in that, The background data includes: An experience database is used to store historical process data, which includes at least historical process formulas and corresponding historical process results; and A rules database is used to store predefined expert knowledge, physical laws, or process constraints.

14. The semiconductor process parameter tuning system as described in claim 12, characterized in that, Before the iterative optimization loop begins, the control device performs the following: The process objective is sent to the large language model service to trigger the large language model service to generate an initial process recipe based on relevant historical data in the experience database and data in the rule database, and the initial process recipe is used as the current process recipe in the first round of the iterative optimization loop.

15. The semiconductor process parameter tuning system as described in any one of claims 12 or 13, characterized in that, After terminating the iterative optimization loop, the control device executes: The process objective, the applied current process formula, and the corresponding actual process results are sent to the experience database for storage; or, The large language model service analyzes the deviation between the actual process results and the process target, and infers and generates the adjusted process formula based on the adjustment strategies or causal relationships related to the deviation in the rule database. or, Monitor the iteration count of the iterative optimization loop. If the actual process result still does not fall within the preset process target after reaching the preset maximum iteration count, terminate the iterative optimization loop, or terminate the iterative optimization loop and trigger the large language model service to generate adjustment suggestions.

16. The semiconductor process parameter tuning system as described in claim 12, characterized in that, The control device communicates with the semiconductor process equipment via an industrial bus protocol, and the control device communicates with the large language model service via an HTTP protocol. Alternatively, after receiving the adjusted process formula generated by the large language model service, the control device performs a format verification on it. If the verification fails, it requests the large language model service to regenerate it. Alternatively, the control device may also provide a user interface for receiving operator commands to monitor, manually stop, or intervene in the iterative optimization loop. Alternatively, the semiconductor process equipment may further include an online measurement module, which is integrated into or connected to the semiconductor process equipment, for automatically acquiring the actual process results and transmitting them to the control device.

17. The semiconductor process parameter tuning system as described in claim 12, characterized in that, Also includes: The relay service layer acts as middleware between the control device and the large language model service. The control device sends data including the current process formula and the actual process results to the relay service layer. The relay service layer receives data from the control device, retrieves relevant historical data and rules from the background data, integrates the information and sends it to the large language model service, receives the adjusted process formula generated by the large language model service, and forwards only the adjusted process formula to the control device.

18. The semiconductor process parameter tuning system as described in claim 17, characterized in that, The background data is managed and maintained by the relay service layer; or... During the iterative optimization loop, the control device sends data including the current process formula and the actual process results to the transfer service layer, which processes the data and stores it in the background data.

Images