Intelligent optimization method for silicon material purification process

Through the combination of deep learning algorithms and digital twin systems, the problem of low parameter optimization efficiency and disconnection of results in the silicon purification process is solved, and efficient and precise optimization of the silicon purification process is achieved, and production efficiency and product quality are improved.

CN118942564APending Publication Date: 2024-11-12HEFEI UNIV OF TECH
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410994701.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing silicon purification process parameter optimization method has problems such as low historical data utilization, lack of overall global process flow considerations, low parameter search and adjustment efficiency, and disconnection of optimization results from actual production, making it difficult to effectively improve the purity and yield of silicon materials. .

Method used

Deep learning algorithm is used to build a parameter optimization model for silicon purification process flow. Through classification and weighted design, iterative optimization is performed with digital twin systems to find the optimal parameter combination, and realize process parameters through visual assisted decision-making and dynamic feedback correction mechanisms. intelligent adjustment.

Benefits of technology

The production efficiency, product quality and profit level of the silicon purification process have been significantly improved, the time and cost of traditional trial and error methods have been reduced, and a closed-loop feedback mechanism has been formed, which has continuously improved the level of the silicon purification process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118942564A_ABST
    Figure CN118942564A_ABST
Patent Text Reader

Abstract

The invention discloses an intelligent optimization method for a silicon material purification process, which comprises the following steps of: 1, dividing the existing silicon material purification process flow into a plurality of categories to obtain process flow samples; 2, constructing a silicon purification technological process parameter optimization model, and training by adopting a technological process sample; 3, inputting process parameters of all silicon material purification process flows in the existing industrial production into the trained silicon purification process flow parameter optimization model in the step 2 so as to find an optimal parameter combination of each silicon material purification process flow; and step 4, inputting the optimal parameter combination of each silicon material purification process flow obtained in the step 3 into a digital twin system so as to control the actual production process to operate with the optimal parameter combination. The production efficiency, the product quality and the income level of the silicon purification process can be remarkably improved, and intelligent upgrading of the silicon purification industry is powerfully promoted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of silicon purification process optimization methods, and in particular to an intelligent optimization method for silicon material purification process. Background Art

[0002] Silicon materials are key raw materials for industries such as photovoltaics and semiconductors. Improving the purity and yield of silicon materials is an important goal of silicon purification process. Traditional process parameter optimization mainly relies on expert experience and "trial and error", which is inefficient and costly. With the development of artificial intelligence technology, the use of machine learning to optimize process parameters has become a new research hotspot. However, the existing silicon purification process parameter optimization methods generally have the following problems: 1. The utilization rate of historical process data is low, making it difficult to fully explore the intrinsic relationship between key parameters and optimization goals.

[0003] 2. Lack of overall consideration of the global process flow leads to limited effect of local parameter optimization.

[0004] 3. Parameter search and adjustment are inefficient, and it is difficult to quickly find the optimal parameter combination.

[0005] 4. The optimization results are out of touch with actual production, and lack visual decision-making assistance and dynamic feedback correction mechanisms. Summary of the invention

[0006] In response to the above technical problems, the present invention provides a method for intelligent optimization of silicon material purification process.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is: A method for intelligent optimization of silicon material purification process, comprising the following steps: Step 1: According to the grade of silicon material output by each silicon material purification process in existing industrial production and the final purification purity, the various silicon material purification process flows in existing industrial production are divided into several categories, and the silicon material purification process flows of each category are sorted and weighted according to their importance, thereby obtaining a process flow sample; Step 2: Use a deep learning algorithm to construct a silicon purification process parameter optimization model, and use the process parameters of each sub-process of each silicon material purification process in the process sample in step 1 to train the silicon purification process parameter optimization model, so that the silicon purification process parameter optimization model can learn to maximize the differences between different silicon material purification process flows and minimize the differences in process parameters within each sub-process in each silicon material purification process flow; Step 3: Input the process parameters of all silicon material purification process flows in existing industrial production into the silicon purification process flow parameter optimization model trained in step 2, and perform iterative optimization through the silicon purification process flow parameter optimization model to find the optimal parameter combination for each silicon material purification process flow; Step 4: Input the optimal parameter combination of each silicon material purification process obtained in step 3 into the digital twin system, and control the actual production process through the digital twin system to run with the optimal parameter combination.

[0008] Furthermore, in step 2, the silicon purification process parameter optimization model includes a first Transformer module, a global average pooling layer, a pooling operation layer, and a second Transformer module that introduces a relative position bias. The data processing process during the training of the silicon purification process parameter optimization model is as follows: Assume that each silicon material purification process has m sub-processes, and the number of process parameters of the i-th sub-process is Ni; First, the first Transformer module is used to extract features and fuse information of the process parameters of the i-th sub-process, and the output is a tensor tensor_i with a shape of (b, Ni, c), where b is the training batch size and c is the dimension of parameter encoding; then, the second dimension Ni of tensor tensor_i is globally averaged pooled through the global average pooling layer, and the feature representation of each process parameter is compressed into a vector of length c, thereby obtaining a tensor local_tensor_i with a shape of (b, 1, c); Therefore, for each m-channel sub-process in the silicon material purification process, the first Transformer module outputs tensors tensor_1, tensor_2, ..., tensor_m of a total of m-channel sub-processes; The first Transformer module outputs a total of m sub-process tensors tensor_1, tensor_2, ..., tensor_m, which are pooled through the pooling operation layer and then concatenated in the second dimension to obtain a tensor global_tensor with a shape of (b, m, c); then the second Transformer module with relative position bias is used to perform y times of feature extraction and information fusion on the tensor global_tensor to obtain a tensor tensor_y with a shape of (b, m, c).

[0009] Furthermore, in step 2, the loss function during training adopts the weighted sum of the global difference loss and the local consistency loss of each sub-process, where the calculation process of the local consistency loss of each sub-process is as follows: Calculate the mean square error (MSE) between the tensor local_tensor_i and its corresponding tensor tensor_i as the local consistency loss loss_i of the i-th sub-process; The calculation process of global difference loss is as follows: Perform global average pooling on tensor tensor_y to obtain the global average pooling processing result refined_global_tensor; then calculate the mean square error MSE between refined_global_tensor and tensor global_tensor as the global difference loss loss_global.

[0010] Furthermore, in step 2, a back propagation algorithm and an optimizer are used during training to update the parameters of the silicon purification process parameter optimization model.

[0011] Further, the process of step 3 is as follows: (3.1) Input the process parameters of all silicon material purification process flows in existing industrial production into the silicon purification process flow parameter optimization model trained in step 2 for inference. By comparing the inter-class mean square errors corresponding to different silicon material purification process flows, determine which silicon material purification process flow process parameter combinations need to be further optimized; for the silicon material purification process flow that needs to be optimized, use the intra-class mean square error to determine the specific process parameters that need to be optimized; (3.2) After determining the silicon material purification process and corresponding process parameters that need to be optimized, use the silicon purification process parameter optimization model to make small changes to the process parameters, and then re-input the modified process parameters into the silicon purification process parameter optimization model, and compare the mean square errors corresponding to the process parameters before and after the modification to determine the effect and direction of the adjustment; If the mean square error decreases, keep the current adjustment direction and continue to optimize; if the mean square error increases, first determine whether the mean square error decreases continuously for n times and then suddenly increases. If so, it is considered that the parameters of the previous step have reached the optimal point and stop adjusting. If not, reverse the adjustment direction and continue to search for the optimal parameter combination.

[0012] (3.3) Iterate the process of adjusting the process parameters and comparing the mean square error in step (3.2) until the iterative optimization termination condition is met.

[0013] Furthermore, in step (3.3), the iteration is stopped when any of the following two iterative optimization termination conditions is met: (1) If the mean square error decreases n times continuously and then suddenly increases, the parameters of the previous step are considered to be optimal; (2) If the mean square error still does not reach the preset convergence threshold after m iterations, the current optimal parameters are considered to be the global optimal point.

[0014] Compared with the prior art, the present invention has the following advantages: 1. Through classification and weighted design, make full use of limited historical data, improve sample utilization efficiency, and take into account both global optimization and local focus.

[0015] 2. Using deep learning and contrastive learning, we built an efficient and accurate process parameter optimization model, which maximizes the differences between processes and minimizes the differences in parameters within processes.

[0016] 3. Introduce a large model to intelligently adjust parameters, quickly search for the optimal parameter combination through iterative optimization, and reduce the time and cost of the traditional "trial and error method".

[0017] 4. Apply the optimized parameters to the digital twin system, and use visualization and virtual suggestions to assist actual production decisions. The combination of digital twins and intelligent optimization forms a closed-loop feedback to continuously improve the silicon purification process.

[0018] 5. The present invention can significantly improve the production efficiency, product quality and profit level of the silicon purification process, effectively promote the intelligent upgrading of the silicon purification industry, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of a method according to an embodiment of the present invention.

[0020] Figure 2 It is a data processing flow chart of the silicon purification process parameter optimization model in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0022] like Figure 1 As shown, this embodiment discloses a method for intelligent optimization of silicon material purification process, comprising the following steps: Step 1: Optimize the silicon material purification process flow based on the target drive, divide the various silicon material purification process flows in the existing industrial production into several categories according to the grade of silicon material output by each silicon material purification process flow in the existing industrial production and the final purification purity, and sort and weight the silicon material purification process flows of each category according to their importance, thereby obtaining a process flow sample; In this embodiment, the existing silicon material purification process flow is divided into K categories according to the grade of the output silicon material and the final purification purity. The first category of silicon material purification process flow represents the optimal process, corresponding to the highest grade and purity, and has the largest weight during training. The kth category of silicon material purification process flow corresponds to the suboptimal process, with slightly lower product grade and purity, and the smallest training weight.

[0023] In this embodiment, the silicon material purification process of each category is then ranked according to the importance. Although the importance is different, all silicon material purification process flows need to be included in the training to expand the sample size, increase data diversity, and fully characterize the complex relationship between process parameters and targets.

[0024] Therefore, this embodiment uses reasonable classification and weight design, taking into account global optimization and local focus, and finally obtains process flow samples. By improving sample utilization efficiency, increasing diversity, mining parameter connections, indicating optimization directions, and forming a closed-loop iteration, the silicon purification process is continuously optimized.

[0025] Step 2: Use a deep learning algorithm to construct a silicon purification process parameter optimization model, and use the process parameters of each sub-process of each silicon material purification process in the process sample in step 1 to train the silicon purification process parameter optimization model, so that the silicon purification process parameter optimization model can learn to maximize the differences between different silicon material purification process flows, and minimize the differences in process parameters within each sub-process in each silicon material purification process flow.

[0026] In this embodiment, the silicon purification process parameter optimization model includes a first Transformer module, a global average pooling layer, a pooling operation layer, and a second Transformer module that introduces a relative position bias.

[0027] like Figure 2 As shown, in this embodiment, the data processing process during the training of the silicon purification process parameter optimization model is as follows: Assume that each silicon material purification process is a sequence of m sub-processes, each sub-process has a specific set of process parameters, and assume that the number of process parameters of the i-th sub-process is Ni, i=1, 2, 3, 4, ... m. Through end-to-end training, the silicon purification process parameter optimization model automatically learns and optimizes the parameter combination of each sub-process, realizing parameter optimization of the entire process.

[0028] In the silicon purification process parameter optimization model, the first Transformer module is first used to perform x Transformers on the process parameters of the i-th sub-process to perform feature extraction and information fusion. The first Transformer module outputs a tensor tensor_i with a shape of (b, Ni, c), where b is the training batch size and c is the dimension of parameter encoding. Then, the second dimension Ni of tensor tensor_i is globally averaged pooled through the global average pooling layer, and the feature representation of each process parameter is compressed into a vector of length c, thereby obtaining a tensor local_tensor_i with a shape of (b, 1, c). Therefore, for each m-channel sub-process in the silicon material purification process, the first Transformer module outputs tensors tensor_1, tensor_2, ..., tensor_m for a total of m-channel sub-processes; For m-way sub-processes, the tensors tensor_1, tensor_2, ..., tensor_m of the total m-way sub-processes output by the first Transformer module are pooled through the pooling operation layer and then concatenated in the second dimension, thereby obtaining a tensor global_tensor with a shape of (b, m, c). The tensor global_tensor represents the overall feature representation of the m-way process in the current batch. Then, the second Transformer module performs y times of feature extraction and information fusion on the tensor global_tensor. The position encoding of the second Transformer module introduces a relative position bias, thereby performing y times of Transformer stacking on the tensor global_tensor through the second Transformer module to complete feature extraction and information fusion, and obtains a tensor tensor_y with a shape of (b, m, c). The tensor tensor_y represents the final feature representation of the m-way process after information interaction and fusion.

[0029] In this embodiment, the loss function of the silicon purification process parameter optimization model training is constructed based on the contrastive learning concept, including global difference loss and local consistency loss. The calculation process of the global difference loss is as follows: Perform global average pooling on tensor tensor_y to obtain the global average pooling processing result refined_global_tensor; then calculate the mean square error MSE between refined_global_tensor and tensor global_tensor as the global difference loss loss_global. The global difference loss loss_global reflects the difference of overall process characteristics under different optimization levels.

[0030] The calculation process of local consistency loss of each sub-process is as follows: Calculate the mean square error MSE between the tensor local_tensor_i and its corresponding tensor tensor_i as the local consistency loss loss_i of the i-th sub-process. loss_i reflects the consistency and tightness of the process parameters within the i-th sub-process, i=1, 2, 3, 4, ...m.

[0031] Finally, the loss function of this embodiment is the weighted sum of the global difference loss loss_global and the local consistency losses loss_1, loss_2, ..., loss_m of each sub-process. The loss function Loss calculation formula is as follows: Loss = α * loss_global + Σ(loss_1, loss_2, ..., loss_m) Where: α is the balance factor, which is used to control the weight between global difference and local consistency.

[0032] In this embodiment, the process parameters of each sub-process of each silicon material purification process in the process flow sample in step 1 are used to train the silicon purification process flow parameter optimization model end-to-end. In each training batch, the number of samples is b, and each process parameter is encoded as a vector of length c. By minimizing the loss function Loss, the silicon purification process flow parameter optimization model can learn to maximize the differences between different processes and minimize the differences in the internal parameters of each process, thereby achieving the purpose of comprehensive optimization.

[0033] In addition, in this embodiment, when training the silicon purification process parameter optimization model, the back propagation algorithm and optimizer (such as Adam) are used to update the model parameters, and learning rate scheduling, early stopping and other strategies are used to improve the training effect. Through extensive hyperparameter tuning experiments, this embodiment will find the optimal model configuration, including the number of Transformer layers x and y, encoding dimension c, batch size b, etc.

[0034] After the model training converges, an excellent performance silicon purification process parameter optimization model is obtained. This silicon purification process parameter optimization model can be used to guide actual production and provide process engineers with reliable parameter adjustment suggestions, thereby improving the product quality and yield of silicon purification. Step 3: Input the process parameters of all silicon material purification processes in existing industrial production into the silicon purification process parameter optimization model trained in step 2, and perform iterative optimization through the silicon purification process parameter optimization model to find the optimal parameter combination for each silicon material purification process. The specific process is as follows: (3.1) Input the process parameters of all silicon material purification process flows in existing industrial production into the silicon purification process flow parameter optimization model trained in step 2 for inference. By comparing the inter-class mean square errors corresponding to different silicon material purification process flows, determine which silicon material purification process flow process parameter combinations need to be further optimized; for the silicon material purification process flow that needs to be optimized, use the intra-class mean square error to determine the specific process parameters that need to be optimized; (3.2) After determining the silicon material purification process and corresponding process parameters that need to be optimized, use the silicon purification process parameter optimization model to make small changes to the process parameters, and then re-input the modified process parameters into the silicon purification process parameter optimization model, and compare the mean square errors corresponding to the process parameters before and after the modification to determine the effect and direction of the adjustment; If the mean square error decreases, keep the current adjustment direction and continue to optimize; if the mean square error increases, first determine whether the mean square error decreases continuously for n times and then suddenly increases. If so, it is considered that the parameters of the previous step have reached the optimal point and stop adjusting. If not, reverse the adjustment direction and continue to search for the optimal parameter combination.

[0035] (3.3) Iterate the process of adjusting the process parameters and comparing the mean square error in step (3.2) until any of the following two iterative optimization termination conditions is met: (1) If the mean square error decreases n times continuously and then suddenly increases, the parameters of the previous step are considered to be optimal; (2) If the mean square error still does not reach the preset convergence threshold after m iterations, the current optimal parameters are considered to be the global optimal point.

[0036] Step 4: Input the optimal parameter combination of each silicon material purification process obtained in step 3 into the digital twin system, and use the digital twin system to control the actual production process to run with the optimal parameter combination.

[0037] As a virtual mirror of the actual production process, the digital twin system updates its own settings according to the optimized optimal parameter combination, and controls the actual silicon material purification process so that it runs with the optimal parameter combination.

[0038] At the same time, the digital twin system records and displays the original silicon material grade, optimized process parameters, and the purity of the purified silicon, presenting these key data in a visual way to provide intuitive decision-making references for process engineers. The digital twin system also provides accurate suggestions for parameter adjustment of the actual silicon material purification process based on the guidance of the optimization model.

[0039] Furthermore, the optimal parameter combination obtained in step 4 is also re-input into the silicon purification process parameter optimization model to start a new round of iterative optimization.

[0040] Therefore, this embodiment builds a closed-loop intelligent optimization system through the collaborative work of the silicon purification process parameter optimization model and the digital twin system. The system can continuously optimize the parameters of the silicon purification process and dynamically adjust the optimization strategy according to actual production data, ultimately achieving improved production efficiency and product quality.

[0041] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0042] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.

Claims

1. A method for intelligent optimization of silicon material purification process, characterized in that: The following steps are involved: Step 1: According to the grade of silicon material output by each silicon material purification process in existing industrial production and the final purification purity, the various silicon material purification process flows in existing industrial production are divided into several categories, and the silicon material purification process flows of each category are sorted and weighted according to their importance, thereby obtaining a process flow sample; Step 2: Use a deep learning algorithm to construct a silicon purification process parameter optimization model, and use the process parameters of each sub-process of each silicon material purification process in the process sample in step 1 to train the silicon purification process parameter optimization model, so that the silicon purification process parameter optimization model can learn to maximize the differences between different silicon material purification process flows and minimize the differences in process parameters within each sub-process in each silicon material purification process flow; Step 3: Input the process parameters of all silicon material purification process flows in existing industrial production into the silicon purification process flow parameter optimization model trained in step 2, and perform iterative optimization through the silicon purification process flow parameter optimization model to find the optimal parameter combination for each silicon material purification process flow; Step 4: Input the optimal parameter combination of each silicon material purification process obtained in step 3 into the digital twin system, and control the actual production process through the digital twin system to run with the optimal parameter combination.

2. The method for intelligent optimization of silicon material purification process according to claim 1, characterized in that: In step 2, the silicon purification process parameter optimization model includes a first Transformer module, a global average pooling layer, a pooling operation layer, and a second Transformer module that introduces a relative position bias. The data processing process during the training of the silicon purification process parameter optimization model is as follows: Assume that each silicon material purification process has m sub-processes, and the number of process parameters of the i-th sub-process is Ni; First, the first Transformer module is used to extract features and fuse information of the process parameters of the i-th sub-process, and the output is a tensor tensor_i with a shape of (b, Ni, c), where b is the training batch size and c is the dimension of parameter encoding; then, the second dimension Ni of tensor tensor_i is globally averaged pooled through the global average pooling layer, and the feature representation of each process parameter is compressed into a vector of length c, thereby obtaining a tensor local_tensor_i with a shape of (b, 1, c); Therefore, for each m-channel sub-process in the silicon material purification process, the first Transformer module outputs tensors tensor_1, tensor_2, ..., tensor_m of a total of m-channel sub-processes; The first Transformer module outputs a total of m sub-process tensors tensor_1, tensor_2, ..., tensor_m, which are pooled through the pooling operation layer and then concatenated in the second dimension to obtain a tensor global_tensor with a shape of (b, m, c); then the second Transformer module with relative position bias is used to perform y times of feature extraction and information fusion on the tensor global_tensor to obtain a tensor tensor_y with a shape of (b, m, c).

3. The method for intelligent optimization of silicon material purification process according to claim 2, characterized in that: In step 2, the loss function during training adopts the weighted sum of the global difference loss and the local consistency loss of each sub-process, where the local consistency loss calculation process of each sub-process is as follows: Calculate the mean square error (MSE) between the tensor local_tensor_i and its corresponding tensor tensor_i as the local consistency loss loss_i of the i-th sub-process; The calculation process of global difference loss is as follows: Perform global average pooling on tensor tensor_y to obtain the global average pooling processing result refined_global_tensor; then calculate the mean square error MSE between refined_global_tensor and tensor global_tensor as the global difference loss loss_global.

4. The method for intelligent optimization of silicon material purification process according to claim 2, characterized in that: In step 2, a back propagation algorithm and an optimizer are used during training to update the parameters of the silicon purification process parameter optimization model.

5. The method for intelligent optimization of silicon material purification process according to claim 1, characterized in that: Step 3 process is as follows: (3.1) Input the process parameters of all silicon material purification process flows in existing industrial production into the silicon purification process flow parameter optimization model trained in step 2 for inference. By comparing the inter-class mean square errors corresponding to different silicon material purification process flows, determine which silicon material purification process flow process parameter combinations need to be further optimized; for the silicon material purification process flow that needs to be optimized, use the intra-class mean square error to determine the specific process parameters that need to be optimized; (3.2) After determining the silicon material purification process and corresponding process parameters that need to be optimized, use the silicon purification process parameter optimization model to make small changes to the process parameters, and then re-input the modified process parameters into the silicon purification process parameter optimization model, and compare the mean square errors corresponding to the process parameters before and after the modification to determine the effect and direction of the adjustment; If the mean square error decreases, keep the current adjustment direction and continue to optimize; if the mean square error increases, first determine whether the mean square error decreases continuously for n times and then suddenly increases. If so, it is considered that the parameters of the previous step have reached the optimal point and the adjustment is stopped. If not, reverse the adjustment direction and continue to search for the optimal parameter combination. (3.3) Iterate the process of adjusting the process parameters and comparing the mean square error in step (3.2) until the iterative optimization termination condition is met.

6. The method for intelligent optimization of silicon material purification process according to claim 1, characterized in that: In step (3.3), the iteration is stopped when any of the following two iterative optimization termination conditions is met: (1) If the mean square error decreases n times continuously and then suddenly increases, the parameters of the previous step are considered to be optimal; (2) If the mean square error still does not reach the preset convergence threshold after m iterations, the current optimal parameters are considered to be the global optimal point.

Citation Information

Cited By

  • Battery reclaimed material separation and purification process optimization method based on swarm intelligence algorithm

    CN121583399A

  • Battery recycling material separation and purification process optimization method based on swarm intelligence algorithm

    CN121583399B