System and method for optimizing semiconductor process parameters using multi-agents

Abstract

A semiconductor process parameter optimization system and method using multi-agents are disclosed. A semiconductor process parameter optimization system using multi-agents according to one embodiment of the present invention comprises: one or more processors; a memory storing instructions that cause the system to perform a predetermined operation when executed by the one or more processors; and an orchestrator implemented by the one or more processors, comprising: a first agent that learns a surrogate model based on experimental result data, calculates a quality indicator representing the quality of the surrogate model, and determines the confidence level of the surrogate model based on the quality indicator; a second agent that determines an optimization strategy between search and utilization based on the quality indicator and determines candidate parameters for the next experiment using an acquisition function; a third agent that controls the execution of a semiconductor process experiment corresponding to the candidate parameters and acquires the results of the semiconductor process experiment and adds them to accumulated data; and an orchestrator that determines whether to terminate a loop based on a termination condition, controls the operation of the first agent, the second agent, and the third agent to repeat if the termination condition is not satisfied, and outputs an optimal parameter if the termination condition is satisfied.

Description

Technology Field

[0001] The present invention relates to semiconductor process parameter optimization technology. More specifically, the present invention relates to a multi-agent-based semiconductor process parameter optimization system and method in which a plurality of artificial intelligence agents collaborate to autonomously perform semiconductor process parameter optimization using a surrogate model, and provide information regarding the decision-making process and basis of each agent to a semiconductor process engineer, thereby improving the verifiability of the optimization process. Background Technology

[0002] In semiconductor device manufacturing processes, particularly etching, deposition, chemical mechanical planarization (CMP), and ion implantation, numerous process parameters such as gas flow rate, chamber pressure, RF power, substrate temperature, and process time interact in a complex manner with the process results. Finding the optimal combination of these parameters is a critical task directly related to semiconductor production yield and quality.

[0003] Conventional semiconductor process optimization has relied on the experience of process engineers or primarily used grid search methods based on Design of Experiments (DOE). Since these methods require iterative experimentation with multiple combinations of process parameters, deriving optimal conditions requires a large number of experiments and a significant amount of time, which also leads to increased wafer consumption costs.

[0004] Furthermore, when machine learning-based optimization techniques are applied to semiconductor processes, they are often implemented as "black boxes" where the rationale for the selection of prediction models and the recommendation of subsequent experimental conditions is not sufficiently presented. This makes it difficult for process engineers to trust the recommendation results and limits their application in actual manufacturing environments.

[0005] Therefore, a new semiconductor process parameter optimization technology is required that allows process engineers to directly verify the basis for judgment at each stage of the optimization process, while simultaneously reducing the number of experiments and costs. Prior art literature

[0007] Patent Publication No. 10-2025-0060356 (May 7, 2025) The problem to be solved

[0008] The present invention provides a semiconductor process parameter autonomous optimization system and method that reduces dependence on machine learning expertise and enables process engineers at a semiconductor manufacturing site (fab) to directly operate parameter optimization by autonomously performing the core roles of machine learning experts—namely, surrogate model quality evaluation, exploration / utilization strategy switching, and convergence determination—without human intervention through a plurality of specialized AI agents (surrogate agents, optimizer agents, experimental agents, and orchestrators).

[0009] Furthermore, by implementing a function that converts each agent's decision-making process into reasoning information in natural language understandable to process engineers in real time, the aim is to provide a reliable system that enables engineers to verify the basis of the AI's judgments at each stage of optimization and intervene when necessary.

[0010] In addition, the invention provides a semiconductor process optimization method that significantly reduces the number of experiments required in conventional methods, ranging from tens to hundreds, through a structured event-based architecture and real-time streaming using Server-Sent Events (SSE).

[0011] The problems to be solved by the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by a person skilled in the art from the description below. means of solving the problem

[0013] A semiconductor process parameter optimization system using a multi-agent according to an embodiment of the present invention comprises: one or more processors; a memory storing instructions that cause the system to perform a predetermined operation when executed by the one or more processors; and a first agent implemented by the one or more processors, which learns a surrogate model based on experimental result data, calculates a quality indicator representing the quality of the surrogate model, and determines the confidence level of the surrogate model based on the quality indicator; a second agent that determines an optimization strategy between search and utilization based on the quality indicator and determines candidate parameters for the next experiment using an acquisition function; a third agent that controls the execution of a semiconductor process experiment corresponding to the candidate parameters, acquires the results of the semiconductor process experiment, and adds them to accumulated data; and an orchestrator that determines whether to terminate a loop based on a termination condition, controls the operation of the first agent, the second agent, and the third agent to repeat if the termination condition is not satisfied, and outputs optimal parameters if the termination condition is satisfied.

[0014] Additionally, it may further include an AI inference engine that receives information representing the decision-making processes and grounds of the first to third agents, converts it into a form that allows a semiconductor process engineer to verify the judgment of each step, and provides the converted information to a user interface in real time.

[0015] In addition, the proxy model learned by the first agent is a Gaussian process model, and the quality indicator may be the coefficient of determination (R²).

[0016] In addition, the second agent can dynamically switch between an optimization strategy for search and utilization by applying a search weighted acquisition function when R² is below a predetermined threshold and applying a utilization weighted acquisition function when R² is above the threshold.

[0017] In addition, the above-mentioned acquisition function includes a combination of Expected Improvement (EI) and Upper Confidence Bound (UCB), and the second agent can determine the candidate parameters using Bayesian Optimization (BO).

[0018] In addition, the above termination condition may be satisfied if the optimal value is not improved during a predetermined number of consecutive iterations.

[0019] In addition, each of the first to third agents issues a structured event when a decision is completed, the orchestrator receives the structured event and automatically triggers the next agent, and the AI ​​inference engine receives the structured event and converts it into a form that can be verified by the semiconductor process engineer.

[0020] Additionally, each of the first to third agents is configured to transition between states of waiting, analysis, execution, and completion, and the AI ​​inference engine can transform and provide the decision basis of the corresponding agent in response to the transition to the completion state.

[0021] In addition, the surrogate model learned by the first agent may be a Gaussian process model having a composite kernel including an ARD (Automatic Relevance Determination) kernel and a polynomial kernel.

[0022] In addition, the semiconductor process may include at least one of etching, deposition, chemical mechanical polishing, and ion implantation.

[0023] A method for optimizing semiconductor process parameters using a multi-agent system according to an embodiment of the present invention may include: a step in which a first agent learns a surrogate model based on experimental result data, calculates a quality indicator representing the quality of the surrogate model, determines the confidence level of the surrogate model based on the quality indicator, and outputs the result of the confidence level determination; a step in which a second agent receives the quality indicator, determines an optimization strategy between search and utilization based on the quality indicator, and determines candidate parameters for the next experiment using an acquisition function; a step in which a third agent controls the execution of a semiconductor process experiment corresponding to the candidate parameters, acquires the results of the semiconductor process experiment, and adds them to accumulated data; and a step in which an orchestrator determines whether to terminate the repetition based on a termination condition, repeats the operation by the first agent, the second agent, and the third agent if the termination condition is not satisfied, and outputs the optimal parameters if the termination condition is satisfied.

[0024] Additionally, the method may further include the step of an AI inference engine receiving information representing the decision-making processes and grounds of the first to third agents, converting it into a form that allows a semiconductor process engineer to verify the judgment of each step, and providing the converted information to a user interface in real time.

[0025] In addition, the above surrogate model is a Gaussian process model, and the above quality indicator may be the coefficient of determination (R²).

[0026] In addition, the second agent can dynamically switch between the optimization strategy between search and utilization by applying a search weighted acquisition function when R² is less than a predetermined threshold and applying a utilization weighted acquisition function when R² is greater than or equal to the threshold.

[0027] Additionally, the above-mentioned acquisition function includes a combination of expected improvement (EI) and confidence upper bound (UCB), and the second agent can determine the candidate parameters using Bayesian optimization (BO).

[0028] In addition, the above termination condition may be satisfied if the optimal value is not improved during a predetermined number of consecutive iterations.

[0029] Additionally, the method may further include the step of issuing a structured event when each of the first to third agents has completed a decision; the step of the orchestrator receiving the structured event and automatically triggering the next agent; and the step of the AI ​​inference engine receiving the structured event and converting it into a form that can be verified by the semiconductor process engineer.

[0030] Additionally, each of the first to third agents transitions between states of waiting, analysis, execution, and completion, and the AI ​​inference engine can transform and provide the decision basis of the corresponding agent in response to the transition to the completion state.

[0031] In addition, the above surrogate model may be a Gaussian process model having a composite kernel including an ARD kernel and a polynomial kernel.

[0032] In addition, the semiconductor process may include at least one of etching, deposition, chemical mechanical polishing, and ion implantation. Effects of the invention

[0034] A semiconductor process parameter optimization system and method using multi-agents according to one embodiment of the present invention can reduce the number of experiments compared to conventional methods through strategic search based on a surrogate model, and accordingly, can reduce wafer consumption costs and optimization time.

[0035] In addition, by having multiple agents share the tasks of model quality evaluation, strategy determination, and termination judgment, reliance on machine learning experts can be reduced and operational convenience for process engineers can be improved.

[0036] In addition, since the basis for the agent's judgment can be provided in real time at each optimization step, process engineers can more easily verify the AI's recommendations and intervene when necessary, thereby improving the reliability of on-site application.

[0037] In addition, strategic search performance can be improved through automatic search / utilization strategy switching based on the coefficient of determination (R²).

[0038] In addition, the present invention has excellent scalability as it can be applied to various semiconductor processes such as etching, deposition, chemical mechanical polishing (CMP), and ion implantation.

[0039] The effects of the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below. Brief explanation of the drawing

[0041] FIG. 1 is an overall configuration block diagram of a semiconductor process parameter optimization system using a multi-agent according to one embodiment of the present invention. FIG. 2 is a drawing showing a change in process engineer work according to an embodiment of the present invention, comparing the conventional method with the method after the application of the present invention. FIG. 3 is an overall operation flowchart of a semiconductor process parameter optimization method using a multi-agent according to an embodiment of the present invention. FIG. 4 is a flowchart illustrating an automatic switching mechanism for the search / utilization strategy of a second agent according to an embodiment of the present invention. FIG. 5 is a block diagram showing an event-based agent communication architecture according to an embodiment of the present invention. FIG. 6 is a state diagram showing four state transitions of each agent according to one embodiment of the present invention. Figure 7 is a diagram comparing the parameter space search patterns of the conventional experimental design method and the strategic search method of the present invention. Figure 8 is a diagram showing the actual inference log in an etching process optimization scenario. FIG. 9 is a flowchart of a semiconductor process parameter optimization method using a multi-agent according to an embodiment of the present invention. Specific details for implementing the invention

[0042] Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement the invention. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the embodiments of the present invention in the drawings, parts unrelated to the explanation have been omitted.

[0043] The terms used in this specification are used merely to describe specific embodiments and are not intended to limit the invention. Singular expressions may include plural expressions unless the context clearly indicates otherwise.

[0044] In this specification, terms such as “comprising,” “having,” or “having” are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not excluding in advance the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0045] Furthermore, the components shown in the embodiments of the present invention are illustrated independently to represent different characteristic functions and do not imply that each component consists of separate hardware or a single software unit. That is, for convenience of explanation, each component is described as a separate component, and at least two of the components may be combined to form a single component, or a single component may be divided into multiple components to perform a function. Such integrated and separated embodiments of each component are also included within the scope of the present invention as long as they do not deviate from the essence of the invention.

[0046] In addition, the following embodiments are provided to explain more clearly to those with average knowledge in the industry, and the shapes and sizes of the elements in the drawings may be exaggerated for clearer explanation.

[0047] FIG. 1 is an overall configuration block diagram of a semiconductor process parameter optimization system using a multi-agent according to an embodiment of the present invention, and FIG. 2 is a diagram showing a change in process engineer tasks according to an embodiment of the present invention, comparing a conventional method with the method after the application of the present invention. FIG. 3 is an overall operation flowchart of a semiconductor process parameter optimization method using a multi-agent according to an embodiment of the present invention, and FIG. 4 is a flowchart showing an automatic switching mechanism for the search / utilization strategy of a second agent according to an embodiment of the present invention.

[0048] Figure 7 is a diagram comparing the parameter space search patterns of the conventional experimental design method and the strategic search method of the present invention.

[0049] With reference to FIG. 2, the difference between a conventional semiconductor process optimization method and a semiconductor process optimization method according to an embodiment of the present invention will be explained.

[0050] In the conventional method (BEFORE), process engineers directly perform most of the tasks of setting the experimental range, conducting experiments, analyzing results, and determining optimal conditions. Consequently, the number of experiments increases, and the objectivity of the analysis results may not be sufficient.

[0051] On the other hand, in the method (AFTER) according to one embodiment of the present invention, the system can recommend the next experimental conditions and provide the rationale for them, and the process engineer can review this to determine whether to perform the experiment and reflect the results. Accordingly, the engineer can focus on verification and experiment execution rather than repetitive search tasks. Specifically, after inputting existing experimental data into the system, the AI ​​can recommend the next experimental conditions, presenting conditions such as a gas flow rate of 235 sccm, a pressure of 10 mTorr, and an RF power of 165 W along with the rationale. The engineer can verify the rationale; if they are convinced, they proceed with the experiment, and if they determine that the recommended RF power exceeds the acceptable range, they can reject it and request a re-recommendation. Furthermore, if the engineer performs the experiment and inputs the results into the system, the system can relearn by reflecting these results. Additionally, if there is no continuous improvement for a predetermined number of times (e.g., 3 times), the system can automatically determine convergence and terminate. This method drastically reduces the number of experiments, significantly lowers costs, and enables the completion of optimization in a short period of time.

[0052] Figure 7 is a diagram comparing the parameter space search patterns of the conventional experimental design method and the strategic search method of the present invention.

[0053] Referring to Fig. 7, in conventional DOE or grid search methods, the parameter space is divided into a grid shape to evenly distribute possible combinations for search. For example, if there are two parameters (pressure and RF power), dividing each parameter into 8 and 7 levels generates 56 grid points, all of which must be searched. In this method, it may pass near the optimal point, but it relies on luck, and most experiments are performed in areas unrelated to the optimal point.

[0054] In contrast, the strategic search method of the present invention allows for the rapid narrowing of promising regions based on the results. For example, in the same two-parameter optimization, in iterations 1 to 5, regions that have not yet been explored are prioritized for exploration (broad search), in iterations 6 to 9, a surrogate model that has learned the search results determines a promising direction and intensively explores the area around that region (direction finding), and in iterations 10 to 12, the area around the optimal value is precisely explored (intensive search) to converge to the optimal point. According to this method, it is possible to strategically converge to the optimal region with only 12 experiments.

[0055] Referring to FIG. 1, a semiconductor process parameter optimization system (100) using a multi-agent according to one embodiment of the present invention includes a first agent (110, surrogate model agent), a second agent (120, optimizer agent), a third agent (130, experimental agent), an orchestrator (140), and an AI inference engine (150).

[0056] The first agent (110) can create a surrogate model by learning the relationship between input parameters and output results from past process experiment data.

[0057] In one embodiment, the first agent (110) may use a Gaussian Process (GP) model having a composite kernel composed of the sum of an Automatic Relevance Determination (ARD) kernel and a second-order polynomial kernel. Since the GP model can provide uncertainty information as well as predicted values, it can be utilized to determine a subsequent search strategy. Additionally, the first agent (110) may calculate a coefficient of determination (R²) as a quality indicator after the completion of training, and may determine the confidence level of the surrogate model based on the R² value. For example, as iterations proceed, the R² value may increase, and accordingly, the search strategy of the second agent (120) may also change.

[0058] The above ARD kernel can automatically determine how much each parameter affects the result. The detailed configuration of the above GP model may be as follows.

[0059] Kernel: Sum of ARD kernel + second-order polynomial kernel

[0060] Accuracy Indicator: Coefficient of Determination (R²) ― Ranges from 0 to 1; higher values ​​indicate greater accuracy.

[0061] Learning method: Retraining with the entire accumulated data in every iteration

[0062] Strategy Switching: Automatically switches between exploration and exploitation based on R² value

[0064] The GP model adopted in one embodiment can achieve optimal performance in semiconductor process optimization situations where initial data is scarce (one embodiment: 5 to 20). Since the GP model provides not only predicted values ​​but also uncertainty, it can be directly utilized in determining the search strategy (which region is not yet well known). The ARD kernel can provide a function to automatically determine how much each parameter affects the result. For example, in an etching process, information such as "RF power has the greatest influence on cd_bias, and temperature has almost no influence" can be automatically extracted. This information can be utilized in the search strategy of the second agent (120), which is the optimization agent, and can serve as the basis for natural language inference information in the AI ​​inference engine (150), such as "pressure-RF power interaction is the main uncertainty factor."

[0065] To explain the operation of the first agent (110) with a specific etching process example, a prediction formula is constructed by learning a pattern from past process experiment data (example), and conditions such as a gas flow rate of 235 sccm, a pressure of 10 mTorr, and an RF power of 165 W can be recommended, for example. Additionally, as more data is added, the accuracy of the first agent (110) can be improved (R²: 45% → 72% → 95%).

[0066] After the surrogate model training is complete, the first agent (110) can autonomously determine the confidence level by calculating the coefficient of determination (R²). The determination of the confidence level can be made according to various criteria, and Table 1 shows an example. In one embodiment, the determination result can be published as a MODEL_TRAINED event (see FIG. 5) and transmitted to the second agent (120) and the AI ​​inference engine (150).

[0068] R² value Confidence level meaning 0.45 (Repeat 1) lowness Insufficient data, need to maintain a wide search 0.72 (Repeat 5) commonly Under improvement but still insufficient, maintaining exploration 0.95 (Repeat 9) height Fully reliable, switching to intensive search 0.97 (Repeat 12) Very high Final convergence level, optimal value can be determined

[0070] The second agent (120) can receive the MODEL_TRAINED event from the first agent (110) and use the prediction formula of the surrogate model to find the next experiment candidate parameter closest to the desired result.

[0071] Referring to FIG. 4, the second agent (120) can select at least one of a search-weighted strategy or a utilization-weighted strategy based on the quality indicator (R²) transmitted from the first agent (110).

[0072] (a) Exploration-weighted mode (R² < threshold θ, one example: 0.9): For example, when the quality metric is below the threshold, an Upper Confidence Bound (UCB)-based exploration-weighted strategy may be applied to prioritize the exploration of high-uncertainty regions. This prioritizes the exploration of regions that the model does not yet know well, namely parameter combinations with high uncertainty. The purpose of the exploration mode is to collect information on parameter spaces that the model has not yet encountered to improve the prediction accuracy of the surrogate model.

[0074] (b) Utilization-weighted mode (when R² >= threshold θ): On the other hand, when the quality metric is greater than or equal to the threshold, an EI (Expected Improvement)-based utilization-weighted strategy may be applied to prioritize the exploration of areas with high potential for improvement around the current optimal value. This finds a better value by precisely exploring the area around the currently known optimal value. The purpose of the utilization mode is to rapidly converge to the global optimal point using a sufficiently accurate surrogate model.

[0075] This autonomous transition of exploration / utilization strategies indicates that AI autonomously performs decisions regarding model reliability and exploration scope, which were previously judged manually by machine learning experts.

[0076] After determining the strategy, the second agent (120) can derive candidate parameters that maximize the gain function using Bayesian optimization and an optimization technique that considers boundary constraints. If necessary, the second agent (120) can additionally issue a STRATEGY_SWITCH event indicating that a transition of the search strategy has occurred. In one embodiment, the second agent (120) can determine candidate parameters that maximize the gain function in a short time using the L-BFGS-B (Limited-memory Broyden-Fletcher-Goldfarb-Shanno Bounded; Limited-memory BFGS supporting boundary constraints) optimization algorithm. This is a calculation that involves substituting numbers into a formula hundreds of times, and can be performed without actual experimentation or cost. The determined candidate parameters can be issued as a CANDIDATE_FOUND event and transmitted to the third agent (130) and the AI ​​inference engine (150).

[0077] The third agent (130) receives the CANDIDATE_FOUND event of the second agent (120), controls the execution of a semiconductor process experiment corresponding to the candidate parameters transmitted from the second agent (120), and can acquire the results and reflect them in the accumulated data. In one embodiment, the method of acquiring the experiment results may include at least one of a semi-autonomous method in which a process engineer directly inputs the results, an automatic method that directly interacts with manufacturing equipment or a simulator, or a method of querying results of similar conditions from an existing database.

[0079] (a) Manual experiment method (semi-autonomous)

[0080] When the system presents recommended parameters and their rationale, the process engineer directly performs the experiment on the equipment and inputs the results into the system. In this case, the third agent (130) waits for the input of results, receives the engineer's input, and transmits it to the first agent (110). The engineer checks the rationale for the recommendation, and if they are convinced, proceeds with the experiment; if an anomaly is detected, they may reject it and request a re-recommendation.

[0082] (b) Equipment interlocking method (fully autonomous)

[0083] A third agent directly calls the API of semiconductor manufacturing process equipment (etching, deposition, etc.) or a process simulator to automatically execute experiments and collect results. In this case, the entire process of recommendation, execution, result collection, and retraining can be performed automatically. An engineer can monitor in real time through the AI ​​inference engine (150) and intervene when necessary.

[0085] (c) PoC (Proof of Concept) method

[0086] The results of conditions most similar to the recommended parameters can be automatically retrieved from an existing experimental data set. This method can be used to verify the strategic search capability of the system. In one embodiment, verification was performed by starting with 10 out of 50 total experimental data and automatically retrieving the data most similar to the recommended conditions from the remaining 40, and it was confirmed that strategic search achieved 25.8% better results compared to random and reached a global optimality in 12 to 15 iterations.

[0088] The acquired experimental results can be added to the accumulated data and published as an EXPERIMENT_COMPLETE event to be delivered to the orchestrator (140) and the AI ​​inference engine (150).

[0089] The orchestrator (140) controls the execution order from the first agent (110) to the third agent (130), receives a step completion event to automatically trigger the next agent, and autonomously determines whether to terminate the iteration loop based on a predetermined convergence condition.

[0090] In this embodiment, the convergence condition may be when the optimal value is not improved in a predetermined number of consecutive iterations (e.g., 3 times). When the convergence condition is satisfied, the orchestrator (140) issues a CONVERGENCE event and outputs the optimal parameter as a final output, and when the convergence condition is not satisfied, the above steps can be automatically repeated without human intervention.

[0091] By referring to Table 2 below, the operating principle of the AI ​​inference engine (150) according to one embodiment of the present invention and the difference from black box AI can be clearly understood.

[0092] In black-box AI, input data is processed, and only the final recommended value is output. Process engineers cannot understand "why this value was chosen" during this process, making it difficult to trust the AI ​​and conduct experiments that consume expensive wafers. If results are poor, it is even impossible to trace at which stage the problem occurred. This leads to a lack of trust among process engineers and, consequently, can hinder the field adoption of AI-based optimization.

[0093] In contrast, an AI inference engine (150) according to one embodiment of the present invention receives structured events generated by each agent (110, 120, 130) and converts the judgment basis included in the event into natural language inference information, summary information, or visualization information that a process engineer can understand and verify. Accordingly, the process engineer can verify the optimization process by checking the judgment basis of each step in real time, rather than simply checking the final recommendation result. In one embodiment, the AI ​​inference engine (150) receives structured events (e.g., JSON format) issued by each agent (110, 120, 130) and converts them into natural language inference information.

[0095] Agent Event Type Input data Inference information in the form of generated natural language Agent 1 MODEL_TRAINED Accuracy 45% (R²), wide search mode Model accuracy 45% (R²). Insufficient data. Exploring a wide area. 2nd Agent STRATEGY_SWITCH Accuracy 95% (R²), Search → Focus Switching Achieved 95% accuracy. Switched to intensive search to precisely explore the area around the optimal value. 2nd Agent CANDIDATE_FOUND Candidate parameters, predicted values, improvement rate Candidate found: RF power 165W, pressure 10 mTorr. Expected 8.3% improvement over existing. Third agent EXPERIMENT_COMPLETE Experimental results compared with the previous optimal Experimental result cd_bias=0.44. 8.3% improvement compared to the previous optimal value of 0.48. New optimal value updated. Orchestrator CONVERGENCE Number of consecutive unimproved cycles, optimal value No improvement for 3 consecutive cycles. Converging to optimal value 0.440. Terminating the optimization process.

[0098] In addition, the AI ​​inference engine (150) according to one embodiment of the present invention can clearly provide a point in time when an engineer can intervene. For example, depending on the result of the AI ​​inference engine (150), the engineer may stop the process if the model accuracy is low but it switches to a utilization mode, or reject it if the recommended parameters exceed the equipment limits. Additionally, the engineer may raise doubts if the interaction analysis is inconsistent with domain knowledge, or request an extension if convergence is too fast. In contrast, in a black-box approach, it is impossible to know at which stage the problem lies, so one has no choice but to either accept everything or reject everything.

[0100] In one embodiment, the inference information in the form of natural language from the AI ​​inference engine (150) can be streamed in real time to a client web interface via Server-Sent Events (SSE). This allows the engineer to check the optimization progress in real time and decide whether to intervene immediately by receiving the same event from the event bus and immediately displaying the inference information in the form of natural language converted by the AI ​​inference engine (150) on the engineer's screen.

[0101] FIG. 5 is a block diagram showing an event-based agent communication architecture according to an embodiment of the present invention.

[0102] Referring to FIG. 5, inter-agent communication of the present invention can be implemented as a structured event-based architecture. Each agent (110, 120, 130, 140) can publish the corresponding event to the event bus whenever it completes a decision. There are two entities receiving the event bus: an orchestrator (140) receives the event and automatically triggers the next agent, and an AI inference engine (150) receives the same event, converts it into inference information in the form of natural language, and delivers it to the engineer UI via SSE streaming.

[0103] For example, event types can be as follows.

[0104] - MODEL_TRAINED: Issued when the first agent completes training the surrogate model. May include quality metrics, confidence levels, uncertainty information, etc.

[0105] - CANDIDATE_FOUND: Issued when the second agent determines candidate parameters. May include candidate parameter values, prediction results, expected improvement information, etc.

[0106] - EXPERIMENT_COMPLETE: Issued when a third-party agent acquires experiment results. May include experiment result values, information reflecting accumulated data, etc.

[0107] - CONVERGENCE: Issued by the orchestrator when termination conditions are met. May include consecutive unimproved cycles, the optimal value, the total number of iterations, etc.

[0108] - STRATEGY_SWITCH: An event optionally issued by the second agent when a transition occurs between a search-weighted strategy and a leverage-weighted strategy. It may include the reason for the strategy transition, baseline quality metrics, and information about the transitioned strategy.

[0110] The advantage of this event-based structure is that the orchestrator (140) and the AI ​​inference engine (150) can independently process the same event stream, thereby clearly separating the functions of each agent and increasing the scalability and maintainability of the system.

[0112] FIG. 6 is a state diagram showing four state transitions of each agent according to one embodiment of the present invention.

[0113] Referring to FIG. 6, each agent (110, 120, 130, 140) according to one embodiment of the present invention can perform transitions between four states.

[0115] The idle state may be a state of waiting for a trigger from the orchestrator (140).

[0116] The THINKING state may be a state in which decisions are made, such as model training, quality metric calculation, strategy determination, or candidate parameter calculation.

[0117] The RUNNING state may be a state in which experiment execution control or result collection is performed.

[0118] The Complete state may be a state where a task is terminated and an event containing the result and the basis for judgment is issued.

[0120] After the COMPLETE state transition, if the convergence condition of the orchestrator (140) is not satisfied, the first agent (110), which is the first agent of the next iteration, transitions from a waiting state to an analysis state, and a new iteration can start.

[0122] Figure 8 is a diagram showing the actual inference log in an etching process optimization scenario.

[0123] Referring to FIG. 8, an embodiment for optimizing the Critical Dimension Bias (cd_bias) in an etching process is illustrated. In one embodiment, input parameters may include gas flow rate, chamber pressure, RF power, substrate temperature, and etching time, and the output target value may be cd_bias. For example, when optimization was started based on 8 initial experimental data points, significant improvement was confirmed through approximately 12 iterations. In addition, superior results were confirmed compared to a random search method in a PoC environment. As an example of the finally derived parameter combination, a cd_bias of 0.440 can be achieved at a gas flow rate of 235 sccm, a pressure of 10 mTorr, and an RF power of 165 W.

[0124] FIG. 9 is a flowchart of a semiconductor process parameter optimization method using a multi-agent according to an embodiment of the present invention.

[0125] Referring to FIG. 9, a semiconductor process parameter optimization method (S100) using a multi-agent according to an embodiment of the present invention may include steps S110 to S140. Additionally, the method may further include step S150 as needed. Each step may be performed by one or more processors of a computer system, and data generated at each step may be stored in memory or transferred to the next step.

[0126] First, in step S110, the first agent (110) can train a surrogate model based on experimental result data, calculate a quality indicator representing the quality of the surrogate model, determine the confidence level of the surrogate model based on the quality indicator, and output the result of the confidence level determination. Here, the experimental result data may include a plurality of process parameters and process result values ​​corresponding to each combination of process parameters. For example, the process parameters may include at least one of gas flow rate, chamber pressure, RF power, substrate temperature, and process time, and the process result values ​​may include at least one of critical dimension deviation (cd_bias), film thickness uniformity, etching depth, surface roughness, resistance value, or yield-related value. Additionally, the first agent (110) can retrain the surrogate model in every iteration using the accumulated total experimental data.

[0127] In one embodiment, the surrogate model in step S110 may be a Gaussian process model, and the quality indicator may be a coefficient of determination (R²). Additionally, the Gaussian process model may have a composite kernel including an Automatic Relevance Determination (ARD) kernel and a polynomial kernel. Accordingly, the influence of each process parameter and the interaction between parameters may be reflected together. Furthermore, the first agent (110) may determine the confidence level of the surrogate model as low, medium, high, or very high based on the calculated R² value.

[0128] In step S120, the second agent (120) receives the quality indicator, determines an optimization strategy between search and utilization based on the quality indicator, and can determine candidate parameters for the next experiment using an acquisition function. Specifically, the second agent (120) may apply a search-weighted strategy when the R² is below a predetermined threshold, and apply a utilization-weighted strategy when the R² is above the threshold. The search-weighted strategy is intended to prioritize the investigation of areas that have not yet been sufficiently evaluated or areas of high uncertainty, and the utilization-weighted strategy may be intended to precisely search for areas where further improvement is possible around the optimal value known to date.

[0129] Additionally, the gain function used in step S120 may include a combination of Expected Improvement (EI) and Upper Confidence Bound (UCB). For example, in a search-weighted strategy, the weight of UCB may be increased, and in a utilization-weighted strategy, the weight of EI may be increased. Furthermore, the second agent (120) may determine the candidate parameters using Bayesian optimization (BO), and the maximization of the gain function may be performed through a boundary-constrained optimization technique such as the L-BFGS-B algorithm. Accordingly, computationally promising candidate parameters can be quickly derived before conducting actual additional experiments.

[0130] In step S130, the third agent (130) controls the execution of a semiconductor process experiment corresponding to the candidate parameter and can acquire the results of the semiconductor process experiment and add them to the accumulated data. In one embodiment, step S130 may be performed in a semi-autonomous manner in which a process engineer directly operates the experiment equipment to perform the experiment and inputs the results after confirming the recommended candidate parameter. In another embodiment, step S130 may be performed in a fully autonomous manner in which the third agent is directly linked with the semiconductor manufacturing equipment or process simulator to automatically perform the experiment, collect results, and store data. In yet another embodiment, step S130 may be performed by querying the results with conditions most similar to the candidate parameter from an existing experiment database.

[0131] In step S140, the orchestrator (140) determines whether to terminate the iteration based on a termination condition, and if the termination condition is not satisfied, repeats steps S110 through S130, and if the termination condition is satisfied, outputs the optimal parameter. The termination condition may include at least one of the following: if the objective function value is not improved by a reference value during a recent predetermined number of iterations, if the amount of change of the recommended candidate parameter is below an allowable range, or if the total number of iterations reaches the maximum number of iterations. If the termination condition is satisfied, the orchestrator (140) can output the final optimal parameter and the corresponding process result value.

[0132] Additionally, in step S150, the AI ​​inference engine (150) receives information indicating the decision-making process and the basis therefor of the first to third agents (110, 120, 130), converts it into a form that allows a semiconductor process engineer to verify the judgment of each step, and provides the converted information to a user interface in real time. Here, the information may include at least one of model accuracy, reason for strategy switching, basis for selecting candidate parameters, expected improvement rate, experimental result comparison information, and basis for convergence judgment. Accordingly, the process engineer can verify the optimization process by checking the basis of the AI's judgment at each step, rather than simply receiving only the final recommended value.

[0133] Additionally, each of the first to third agents (110, 120, 130) may issue a structured event when a decision is completed, and the orchestrator (140) may receive the structured event and automatically trigger the next agent, and the AI ​​inference engine (150) may receive the structured event and convert it into a form that can be verified by the semiconductor process engineer. The structured event may include an event type, a time of creation, an agent identifier, a result value, a basis for judgment, and metadata. For example, the first agent (110) may issue a MODEL_TRAINED event when the surrogate model training is completed, the second agent (120) may issue a CANDIDATE_FOUND event when candidate parameters are determined, and the third agent (130) may issue an EXPERIMENT_COMPLETE event when experimental results are obtained. Additionally, the orchestrator (140) may issue a CONVERGENCE event when a termination condition is met.

[0134] Additionally, each of the first to third agents (110, 120, 130) can transition between states of waiting, analysis, execution, and completion. For example, in the waiting state, they wait for a trigger signal; in the analysis state, they perform decisions such as model training, strategy determination, or candidate parameter calculation; in the execution state, they control the execution of experiments or collect results; and in the completion state, they can issue a structured event containing results and the basis for the decision. The AI ​​inference engine (150) can respond to the transition to the completion state by converting the agent's decision basis into a natural language sentence, summary information, visualization information, or warning message and providing it.

[0135] The method according to one embodiment of the present invention may be applied to at least one semiconductor process among etching, deposition, chemical mechanical polishing (CMP), and ion implantation. For example, in the etching process, the purpose may be to minimize cd_bias or control the etching profile within a target range; in the deposition process, the purpose may be to optimize film thickness uniformity or deposition rate; and in the CMP process, the purpose may be to improve planarization performance or residual film thickness variation. Additionally, in the ion implantation process, the purpose may be to optimize values ​​related to doping concentration distribution, implantation depth, or electrical characteristics. Accordingly, the method of the present invention can be commonly applied to various semiconductor manufacturing processes in which a number of process parameters interact.

[0137] The embodiments described in this specification may be implemented by hardware, software, firmware, or a combination thereof. For example, the embodiments may be implemented by one or more processors, controllers, microprocessors, Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), or combinations thereof.

[0138] Additionally, each module, unit, component, or agent mentioned in this specification may be implemented as individual hardware or may be implemented integrated within one or more hardware or software components.

[0139] Such hardware, etc., may be implemented within the same device or in individual devices to support the various operations and functions described in this specification. Additionally, components, units, modules, components, etc., described as "parts" in the invention may be implemented together or individually as separate but interoperable logic devices. Descriptions of different features for modules, units, etc., are intended to highlight different functional embodiments and do not necessarily imply that they must be realized by individual hardware components. Rather, functions associated with one or more modules or units may be performed by individual hardware components or integrated within common or individual hardware components.

[0140] The present invention has been described with reference to the embodiments illustrated in the drawings, but this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the present invention should be determined by the technical spirit of the appended claims. Explanation of the symbols

[0142] 100: Semiconductor Process Parameter Optimization System 110: 1st Agent 120: Second Agent 130: Third Agent 140: Orchestrator 150: AI Inference Engine

Claims

Claim 1 A semiconductor process parameter optimization system using multiple agents, comprising: one or more processors; a memory storing instructions that cause the system to perform a predetermined operation when executed by the one or more processors; and a first agent implemented by the one or more processors, which learns a surrogate model based on experimental result data, calculates a quality indicator representing the quality of the surrogate model, and determines the confidence level of the surrogate model based on the quality indicator; a second agent that determines an optimization strategy between search and exploit based on the quality indicator and determines candidate parameters for the next experiment using an acquisition function; a third agent that controls the execution of a semiconductor process experiment corresponding to the candidate parameters and acquires the results of the semiconductor process experiment and adds them to accumulated data; and an orchestrator that determines whether to terminate a loop based on a termination condition, controls the operation of the first agent, the second agent, and the third agent to repeat if the termination condition is not satisfied, and outputs an optimal parameter if the termination condition is satisfied. Claim 2 A semiconductor process parameter optimization system according to claim 1, further comprising an AI inference engine that receives information indicating the decision-making process and basis of the first to third agents, converts it into a form that allows a semiconductor process engineer to verify the judgment of each step, and provides the converted information to a user interface in real time. Claim 3 A semiconductor process parameter optimization system according to claim 1, wherein the proxy model learned by the first agent is a Gaussian process model and the quality indicator is the coefficient of determination (R²). Claim 4 A semiconductor process parameter optimization system according to claim 3, wherein the second agent applies a search weighted acquisition function when R² is less than a predetermined threshold value and applies a utilization weighted acquisition function when R² is greater than or equal to the threshold value to dynamically switch between search and utilization optimization strategies. Claim 5 A semiconductor process parameter optimization system according to claim 4, wherein the acquisition function includes a combination of Expected Improvement (EI) and Upper Confidence Bound (UCB), and the second agent determines the candidate parameters using Bayesian Optimization (BO). Claim 6 A semiconductor process parameter optimization system according to claim 1, wherein the termination condition is satisfied when the optimal value is not improved during a predetermined number of consecutive iterations. Claim 7 A semiconductor process parameter optimization system according to paragraph 2, wherein each of the first to third agents issues a structured event when a decision is completed, the orchestrator receives the structured event and automatically triggers the next agent, and the AI ​​inference engine receives the structured event and converts it into a form that can be verified by the semiconductor process engineer. Claim 8 A semiconductor process parameter optimization system according to paragraph 2, wherein each of the first to third agents is configured to transition between states of waiting, analysis, execution, and completion, and the AI ​​inference engine converts and provides the decision basis of the corresponding agent in response to the transition to the completion state. Claim 9 A semiconductor process parameter optimization system according to claim 1, wherein the surrogate model learned by the first agent is a Gaussian process model having a composite kernel including an ARD (Automatic Relevance Determination) kernel and a polynomial kernel. Claim 10 A semiconductor process parameter optimization system according to claim 1, wherein the semiconductor process comprises at least one of etching, deposition, chemical mechanical polishing, and ion implantation. Claim 11 A method for optimizing semiconductor process parameters using a multi-agent system performed by a computer system, comprising: a step in which a first agent learns a surrogate model based on experimental result data, calculates a quality indicator representing the quality of the surrogate model, determines the confidence level of the surrogate model based on the quality indicator, and outputs the result of the confidence level determination; a step in which a second agent receives the quality indicator, determines an optimization strategy between search and utilization based on the quality indicator, and determines candidate parameters for the next experiment using an acquisition function; a step in which a third agent controls the execution of a semiconductor process experiment corresponding to the candidate parameters, acquires the results of the semiconductor process experiment, and adds them to accumulated data; and a step in which an orchestrator determines whether to terminate the repetition based on a termination condition, repeats the operation by the first agent, the second agent, and the third agent if the termination condition is not satisfied, and outputs the optimal parameter if the termination condition is satisfied. Claim 12 A semiconductor process parameter optimization method according to claim 11, further comprising the step of an AI inference engine receiving information representing the decision-making process and basis of the first to third agents, converting it into a form that allows a semiconductor process engineer to verify the judgment of each step, and providing the converted information to a user interface in real time. Claim 13 A semiconductor process parameter optimization method according to claim 11, wherein the surrogate model is a Gaussian process model and the quality indicator is the coefficient of determination (R²). Claim 14 A semiconductor process parameter optimization method according to claim 13, wherein the second agent applies a search weighted acquisition function when R² is less than a predetermined threshold value and applies a utilization weighted acquisition function when R² is greater than or equal to the threshold value to dynamically switch between an optimization strategy for search and utilization. Claim 15 A semiconductor process parameter optimization method according to claim 14, wherein the acquisition function includes a combination of expected improvement (EI) and confidence upper bound (UCB), and the second agent determines the candidate parameters using Bayesian optimization (BO). Claim 16 A semiconductor process parameter optimization method according to claim 11, wherein the above termination condition is satisfied when the optimal value is not improved during a predetermined number of consecutive iterations. Claim 17 A semiconductor process parameter optimization method according to claim 12, further comprising: a step of issuing a structured event when each of the first to third agents has completed a decision; a step in which the orchestrator receives the structured event and automatically triggers the next agent; and a step in which the AI ​​inference engine receives the structured event and converts it into a form that can be verified by the semiconductor process engineer. Claim 18 A semiconductor process parameter optimization method according to claim 12, wherein each of the first to third agents transitions between states of waiting, analysis, execution, and completion, and the AI ​​inference engine converts and provides the decision basis of the corresponding agent in response to the transition to the completion state. Claim 19 A semiconductor process parameter optimization method according to claim 11, wherein the surrogate model is a Gaussian process model having a composite kernel including an ARD kernel and a polynomial kernel. Claim 20 A method for optimizing semiconductor process parameters according to claim 11, wherein the semiconductor process comprises at least one of etching, deposition, chemical mechanical polishing, and ion implantation.

Images