Wafer manufacturing scheme generation method based on diffusion evolution guided by fuzzy knowledge

Abstract

The invention discloses a wafer manufacturing scheme generation method based on diffusion evolution guided by fuzzy knowledge, which comprises the following steps of: 1, expanding the number of initial schemes by utilizing a certain number of existing wafer schemes based on existing experts and related literature knowledge in combination with a large model; 2, more high-quality scheme possibilities are explored through diffusion evolution by utilizing the expanded initial scheme, and candidate schemes are provided; according to the invention, through professional knowledge expansion and diffusion evolution based on the existing few-sample wafer manufacturing scheme, more wafer manufacturing candidate schemes can be provided, and the stability and high efficiency of the wafer manufacturing process are ensured.

Description

Technical Field

[0001] The present invention belongs to the field of wafer manufacturing, and in particular is a method for generating a wafer manufacturing plan based on fuzzy knowledge guided diffusion evolution. Background Art

[0002] Wafer manufacturing is a core component of the semiconductor industry. The process is complex and sophisticated, involving multiple steps and the coordinated control of numerous process parameters. During the development of new processes or the optimization of existing ones, validating the rationality of process parameters often requires expensive tapeout experiments. However, the high cost and long lead time of a single tapeout experiment limit the amount of effective experimental data available during the R&D phase, posing a significant challenge to rapidly exploring and determining the optimal manufacturing solution.

[0003] Furthermore, the wafer manufacturing system itself is a dynamic and complex system, affected by factors such as equipment aging, differences in raw material batches, and fluctuations in ambient temperature and humidity, resulting in a time-varying optimal process window. Traditional manufacturing solution generation or parameter optimization methods, such as those based on static statistical models, often struggle to effectively capture and adapt to these dynamic changes. Existing methods typically rely on fixed models or parameter update mechanisms, and when faced with dynamic disturbances in the production line, they are unable to quickly and accurately identify and adjust the optimal parameter combinations under non-steady conditions. As a result, the generated manufacturing solutions often lack robustness, making it difficult to consistently guarantee stable product yield and performance. Summary of the Invention

[0004] In order to address the shortcomings of the above-mentioned existing technologies, the present invention proposes a wafer manufacturing plan generation method based on fuzzy knowledge guided diffusion evolution, in order to generate high-quality, robust and interpretable wafer manufacturing plans, thereby ensuring the stability and efficiency of the wafer manufacturing process.

[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0006] The wafer manufacturing plan generation method based on fuzzy knowledge guided diffusion evolution of the present invention is characterized in that it is performed according to the following steps:

[0007] Step 1: Obtain the initial wafer manufacturing plan set for diffusion evolution ,in, represents the qth initial wafer manufacturing plan; Q is the total number of initial wafer manufacturing plans;

[0008] Step 2: Perform diffusion evolution to obtain a set of wafer manufacturing candidate solutions .

[0009] The method for generating a wafer manufacturing plan based on fuzzy knowledge-guided diffusion evolution according to the present invention is also characterized in that step 1 comprises:

[0010] Step 1.1: Obtain a set of n wafer manufacturing solutions ,in, represents the i-th wafer manufacturing plan, and , express The jth process parameter in Indicates the total number of process parameters;

[0011] Step 1.2: Calculate the cosine similarity between the n wafer manufacturing plans and the target wafer manufacturing plan, and select the K wafer manufacturing plans with the highest cosine similarity as the historical wafer manufacturing plan set. ;in, express The kth historical wafer manufacturing plan in , express The jth process parameter in ;

[0012] Step 1.3: Input the knowledge documents in the wafer manufacturing field into the domain large model (LLM) for processing to extract the implicit knowledge set of wafer manufacturing. ;

[0013] Step 1.4, Utilization right Perform data enhancement to generate the i-th enhanced wafer manufacturing plan , resulting in an enhanced set of wafer manufacturing solutions ;

[0014] Step 1.5: Construct and train the explainable similarity learning model EMSL to obtain the trained explainable similarity learning model and The advantages and disadvantages of each enhanced wafer manufacturing solution are ranked, and the top M wafer manufacturing solutions with higher rankings are selected as the wafer domain knowledge enhanced manufacturing solution set , represents the mth wafer domain knowledge enhanced manufacturing solution, represents the total number of wafer domain knowledge-augmented manufacturing solutions, ;

[0015] Step 1.6: Use formula (1) to obtain the jth process parameter after disturbance ,like When remains unchanged, otherwise, Assign to , thus obtaining the wafer manufacturing plan after the kth disturbance , and then get the perturbed wafer manufacturing plan set ,in, represents the lower bound value of the jth process parameter, represents the upper bound of the jth process parameter;

[0016] (1)

[0017] In formula (1), represents the jth random perturbation; represents the jth disturbance intensity, represents a normal distribution;

[0018] Step 1.7, by 、 and Together they form the initial wafer manufacturing solution set ;Q=2K+M.

[0019] Furthermore, the interpretable similarity learning model EMSL in step 1.5 is trained according to the following steps:

[0020] Step 1.5.1 Embed and get the i-th initial embedding representation ;

[0021] Step 1.5.2 According to Build sample pairs and use the fuzzy membership function Set the quality label of each sample pair. If the Ath wafer manufacturing plan The initial embedding representation of The corresponding fuzzy membership function Greater than B-th wafer manufacturing solution The initial embedding representation of The corresponding fuzzy membership function , then let the sample pair Pros and cons labels , on the contrary, let , ;

[0022] Step 1.5.3 Use formula (4) to construct the loss function of the interpretable similarity learning model EMSL :

[0023] (4)

[0024] In formula (4), represents the sigmoid function, represents the fuzzy membership function;

[0025] Step 1.5.4 uses gradient descent optimization to train the interpretable similarity learning model EMSL and calculate the loss function To update the model parameters until Convergence, get the trained interpretable contrast learning model.

[0026] Furthermore, the step 2 includes:

[0027] Step 2.1, define and initialize the variable q=1;

[0028] Step 2.2: Define the current time step as t and initialize t=T, where T is the total time step.

[0029] Step 2.3, let The qth initial wafer fabrication plan in The qth wafer manufacturing plan at the current time step t is recorded as ;

[0030] Step 2.4: Use formula (2) to get the diffusion scheduling parameters of the current time step t :

[0031] (2)

[0032] Step 2.5, calculation and The similarity of each initial wafer manufacturing plan except itself is calculated, and the top J initial wafer manufacturing plans with the highest similarity are selected as A collection of similar wafer manufacturing solutions ;

[0033] Step 2.6, calculation Any h-th similar wafer manufacturing solution Adaptability evaluation metrics ;

[0034] Step 2.7: Use formula (3) to obtain the qth wafer manufacturing plan at time step t-1 :

[0035] (3)

[0036] In formula (3), for The normalization constant of ; is the weight of the current time step t, Prior knowledge reference for the large model LLM, is the mapping function;

[0037] Step 2.8: If t>0, assign t-1 to t and return to step 2.4 to execute sequentially. Otherwise, As the qth wafer manufacturing candidate ;

[0038] Step 2.9: After assigning q+1 to q, return to step 2.2 and execute sequentially until q=Q, thereby obtaining a preliminary set of wafer manufacturing candidate solutions. .

[0039] The electronic device of the present invention includes a memory and a processor, and is characterized in that the memory is used to store a program that supports the processor to execute the wafer manufacturing plan generation method, and the processor is configured to execute the program stored in the memory.

[0040] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the wafer manufacturing plan generating method when the computer program is run by a processor.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. The present invention effectively overcomes the limitation of insufficient historical data by adopting a fuzzy knowledge-guided diffusion evolution method. It can generate high-quality wafer manufacturing solutions by introducing domain knowledge when data samples are limited, thereby reducing dependence on wafer tape-out experiments.

[0043] 2. By simulating the diffusion evolution process, the present invention reduces the need for expensive tape-out experiments, significantly reduces R&D costs and cycles, and thus accelerates the optimization of wafer manufacturing processes and the introduction of new solutions.

[0044] 3. This invention improves the interpretability of manufacturing solutions by combining the domain large model (LLM) with the explainable similarity learning model (EMSL), making it easier for engineers to understand the solution generation process and providing clear guidance for subsequent process adjustments.

[0045] 4. The present invention improves the optimization efficiency of the process plan by dynamically adjusting the manufacturing plan parameters. When faced with complex changes in the production environment, it can quickly respond and optimize the process plan, thereby improving overall production efficiency. DETAILED DESCRIPTION

[0046] In this embodiment, a method for generating a wafer manufacturing plan based on fuzzy knowledge-guided diffusion evolution is to simulate the diffusion evolution process and combine external domain knowledge to generate a high-quality, robust and interpretable wafer manufacturing plan even when the available historical sample data is limited. Specifically, the core of the present invention is to use fuzzy knowledge to guide the diffusion evolution of the wafer manufacturing plan, and through the fusion of historical data and domain knowledge, generate a manufacturing plan that can adapt to dynamic changes, thereby ensuring the stability and efficiency of the wafer manufacturing process. Specifically, the method is carried out in accordance with the following steps:

[0047] Step 1: Obtain the initial wafer manufacturing plan set for diffusion evolution ,in, represents the qth initial wafer manufacturing plan; Q is the total number of initial wafer manufacturing plans;

[0048] Step 1.1: Obtain a set of n wafer manufacturing solutions ,in, represents the i-th wafer manufacturing plan, and , express The jth process parameter in Indicates the total number of process parameters.

[0049] Step 1.2: Calculate the cosine similarity between the n wafer manufacturing plans and the target wafer manufacturing plan, and select the K wafer manufacturing plans with the highest cosine similarity as the historical wafer manufacturing plan set. ;in, express The kth historical wafer manufacturing plan in , express The jth process parameter in .

[0050] Cosine similarity is used to measure the similarity between two wafer manufacturing solutions in multidimensional parameter space. The calculation formula is: , the value range is [-1,1], the larger the value, the more similar the schemes are. Here each wafer manufacturing scheme can be regarded as a p-dimensional vector , the target wafer manufacturing plan is By calculating the similarity, we can select the k historical solutions that are closest to the target requirements as references. For example, when n=50, we can select the k=10 solutions with the highest similarity to enter the historical wafer manufacturing solution set to ensure that the reference solutions have a high degree of relevance.

[0051] Step 1.3: Input the knowledge documents in the wafer manufacturing field into the domain large model (LLM) for processing to extract the implicit knowledge set of wafer manufacturing. ;

[0052] Specifically, textual materials such as wafer production process specifications, expert technical manuals, and academic papers can be fed into an LLM model trained on wafer manufacturing domain knowledge. By analyzing this unstructured text, the LLM summarizes implicit knowledge that is beneficial to wafer manufacturing solutions, such as trends in the impact of certain process parameters on product yield and empirical correlations between parameters. This implicit knowledge represents domain experience that is typically difficult to acquire through explicit programming. The extracted implicit knowledge provides intelligent guidance for subsequent steps.

[0053] Step 1.4, Utilization right Perform data enhancement to generate the i-th enhanced wafer manufacturing plan , resulting in an enhanced set of wafer manufacturing solutions ;

[0054] Step 1.5: Construct and train the explainable similarity learning model EMSL to obtain the trained explainable similarity learning model and The advantages and disadvantages of each enhanced wafer manufacturing solution are ranked, and the top M wafer manufacturing solutions with higher rankings are selected as the wafer domain knowledge enhanced manufacturing solution set , represents the mth wafer domain knowledge enhanced manufacturing solution, represents the total number of wafer domain knowledge-augmented manufacturing solutions, ;

[0055] In this embodiment, the interpretable similarity learning model EMSL is trained according to the following steps:

[0056] Step 1.5.1 Embed and get the i-th initial embedding representation ;

[0057] Step 1.5.2 According to Build sample pairs and use the fuzzy membership function Set the quality label of each sample pair. If the Ath wafer manufacturing plan The initial embedding representation of The corresponding fuzzy membership function Greater than B-th wafer manufacturing solution The initial embedding representation of The corresponding fuzzy membership function , then let the sample pair Pros and cons labels , on the contrary, let , ;

[0058] Step 1.5.3 Use formula (4) to construct the loss function of the interpretable similarity learning model EMSL :

[0059] (4)

[0060] In formula (4), represents the sigmoid function, represents the fuzzy membership function;

[0061] Step 1.5.4 uses gradient descent optimization to train the interpretable similarity learning model EMSL and calculate the loss function To update the model parameters until Convergence, get the trained interpretable contrast learning model.

[0062] After the EMSL model is trained, it is applied to the enhanced set of wafer manufacturing solutions. Every two solutions are compared. After all comparisons are completed, the top M solutions are selected as the final knowledge-enhanced solution set to ensure that the selected solutions integrate domain knowledge and are of high quality.

[0063] Step 1.6: Use formula (1) to obtain the jth process parameter after disturbance ,like When remains unchanged, otherwise, Assign to , thus obtaining the wafer manufacturing plan after the kth disturbance , and then get the perturbed wafer manufacturing plan set ,in, represents the lower bound value of the jth process parameter, represents the upper bound of the jth process parameter;

[0064] (1)

[0065] In formula (1), represents the jth random perturbation; represents the jth disturbance intensity, Represents a normal distribution.

[0066] Step 1.7, by 、 and Together they form the initial wafer manufacturing solution set ;Q=2K+M.

[0067] Step 2: Perform diffusion evolution to obtain a set of wafer manufacturing candidate solutions ;

[0068] Step 2.1, define and initialize the variable q=1;

[0069] Step 2.2: Define the current time step as t and initialize t=T, where T is the total time step.

[0070] Step 2.3, let The qth initial wafer fabrication plan in The qth wafer manufacturing plan at the current time step t is recorded as .

[0071] Step 2.4: Use formula (2) to get the diffusion scheduling parameters of the current time step t :

[0072] (2)

[0073] Diffusion scheduling parameters It is used to adjust the weight of each influencing factor in the evolution step. Its value can decrease or increase with time step to control the importance of the current solution and the new information introduced when the solution is updated. In the actual process, it can also be replaced by other forms. For example, a linear decreasing scheduling strategy can be adopted: set the initial weight , the final weight , then each step is Calculate, so that as t decreases from T to 0, It also decreases linearly from 1 to 0. Similarly, strategies such as cosine function scheduling can be used to smoothly adjust In actual implementation, the appropriate scheduling function form can be selected based on experience or experiments.

[0074] Step 2.5, calculation and The similarity of each initial wafer manufacturing plan except itself is calculated, and the top J initial wafer manufacturing plans with the highest similarity are selected as A collection of similar wafer manufacturing solutions ;

[0075] Step 2.6, calculation Any h-th similar wafer manufacturing solution Adaptability evaluation metrics ;

[0076] Step 2.7: Use formula (3) to obtain the qth wafer manufacturing plan at time step t-1 :

[0077] (3)

[0078] In formula (3), for The normalization constant of ; is the weight of the current time step t, Prior knowledge reference for the large model LLM, is the mapping function.

[0079] Step 2.8: If t>0, assign t-1 to t and return to step 2.4 to execute sequentially. Otherwise, As the qth wafer manufacturing candidate ;

[0080] Step 2.9: After assigning q+1 to q, return to step 2.2 and execute sequentially until q=Q, thereby obtaining a preliminary set of wafer manufacturing candidate solutions. , it has achieved the expansion and diffusion evolution of existing small-sample wafer manufacturing solutions through professional knowledge to provide more wafer manufacturing candidate solutions.

[0081] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0082] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are executed.

Claims

1. A wafer manufacturing plan generation method based on fuzzy knowledge guided diffusion evolution, characterized in that: Follow these steps: Step 1: Obtain the initial wafer manufacturing plan set for diffusion evolution ,in, represents the qth initial wafer manufacturing plan; Q is the total number of initial wafer manufacturing plans; Step 2: Perform diffusion evolution to obtain a set of wafer manufacturing candidate solutions .

2. The wafer manufacturing plan generation method based on fuzzy knowledge guided diffusion evolution according to claim 1 is characterized in that: The step 1 comprises: Step 1.1: Obtain a set of n wafer manufacturing solutions ,in, represents the i-th wafer manufacturing plan, and , express The jth process parameter in Indicates the total number of process parameters; Step 1.2: Calculate the cosine similarity between the n wafer manufacturing plans and the target wafer manufacturing plan, and select the K wafer manufacturing plans with the highest cosine similarity as the historical wafer manufacturing plan set. ;in, express The kth historical wafer manufacturing plan in , express The jth process parameter in ; Step 1.3: Input the knowledge documents in the wafer manufacturing field into the domain large model (LLM) for processing to extract the implicit knowledge set of wafer manufacturing. ; Step 1.4, Utilization right Perform data enhancement to generate the i-th enhanced wafer manufacturing plan , resulting in an enhanced set of wafer manufacturing solutions ; Step 1.5: Construct and train the explainable similarity learning model EMSL to obtain the trained explainable similarity learning model and The advantages and disadvantages of each enhanced wafer manufacturing solution are ranked, and the top M wafer manufacturing solutions with higher rankings are selected as the wafer domain knowledge enhanced manufacturing solution set , represents the mth wafer domain knowledge enhanced manufacturing solution, represents the total number of wafer domain knowledge-augmented manufacturing solutions, ; Step 1.6: Use formula (1) to obtain the jth process parameter after disturbance ,like When remains unchanged, otherwise, Assign to , thus obtaining the wafer manufacturing plan after the kth disturbance , and then get the perturbed wafer manufacturing plan set ,in, represents the lower bound of the j-th process parameter, represents the upper bound of the j-th process parameter; (1) In formula (1), represents the jth random perturbation; represents the jth disturbance intensity, represents a normal distribution; Step 1.7, by 、 and Together they form the initial wafer manufacturing solution set ;Q=2K+M.

3. The wafer manufacturing plan generation method based on fuzzy knowledge guided diffusion evolution according to claim 2 is characterized in that: The interpretable similarity learning model EMSL in step 1.5 is trained according to the following steps: Step 1.5.1 Embed and get the i-th initial embedding representation ; Step 1.5.2 According to Build sample pairs and use the fuzzy membership function Set the quality label of each sample pair. If the Ath wafer manufacturing plan The initial embedding representation of The corresponding fuzzy membership function Greater than B-th wafer manufacturing solution The initial embedding representation of The corresponding fuzzy membership function , then let the sample pair Pros and cons labels , on the contrary, let , ; Step 1.5.3 Use formula (4) to construct the loss function of the interpretable similarity learning model EMSL : (4) In formula (4), represents the sigmoid function, represents the fuzzy membership function; Step 1.5.4 uses gradient descent optimization to train the interpretable similarity learning model EMSL and calculate the loss function To update the model parameters until Convergence, get the trained interpretable contrast learning model.

4. The wafer manufacturing plan generation method based on fuzzy knowledge guided diffusion evolution according to claim 3 is characterized in that: The step 2 includes: Step 2.1, define and initialize the variable q=1; Step 2.2: Define the current time step as t and initialize t=T, where T is the total time step. Step 2.3, let The qth initial wafer fabrication plan in The qth wafer manufacturing plan at the current time step t is recorded as ; Step 2.4: Use formula (2) to get the diffusion scheduling parameters of the current time step t : (2) Step 2.5, calculation and The similarity of each initial wafer manufacturing plan except itself is calculated, and the top J initial wafer manufacturing plans with the highest similarity are selected as A collection of similar wafer manufacturing solutions ; Step 2.6, calculation Any h-th similar wafer manufacturing solution Adaptability evaluation metrics ; Step 2.7: Use formula (3) to obtain the qth wafer manufacturing plan at time step t-1 : (3) In formula (3), for The normalization constant of ; is the weight of the current time step t, For the prior knowledge reference of the large model LLM, is the mapping function; Step 2.8: If t>0, assign t-1 to t and return to step 2.4 to execute sequentially. Otherwise, As the qth wafer manufacturing candidate ; Step 2.9: After assigning q+1 to q, return to step 2.2 and execute sequentially until q=Q, thereby obtaining a preliminary set of wafer manufacturing candidate solutions. .

5. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the wafer manufacturing plan generation method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the wafer manufacturing plan generation method according to any one of claims 1 to 4 are executed.

Images