Dynamic adaptability multi-parameter trimming optimization method and system

Abstract

The invention provides a dynamic adaptive multi-parameter trimming optimization method and system, and the method comprises the steps: obtaining a high-correlation parameter combination through correlation analysis; carrying out combined trimming to obtain a combined trimming parameter group; setting a searched initial vertex according to the combined trimming parameter group, and carrying out initial trimming code combination; the initial trimming code combination is subjected to process deviation function calculation, vertex process deviation values are solved and sorted, and a sorting result is obtained; designing an adaptive search algorithm, based on data driving, fitting a quadratic deviation function, dynamically adjusting a search starting point, searching an optimal trimming parameter set, and obtaining a dynamic centroid; and according to the sorting result and the dynamic centroid, performing reflection, expansion and contraction operations, and iteratively searching an applicable trimming parameter combination. The technical problems that in the trimming process, parameters interfere with one another, the trimming process is tedious, time consumption is large, a globally optimal solution is difficult to find quickly, and dynamic adaptability is lacked are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of integrated circuit testing technology, and in particular to a dynamic adaptive multi-parameter trimming optimization method and system. BACKGROUND

[0002] Process variation refers to the difference between the actual circuit parameters, such as the threshold voltage of a transistor, the resistance value, and the capacitance value, and the ideal design value due to the imperfection of the manufacturing process. With the increasing complexity of integrated circuits (ICs), the fine-tuning of the manufacturing process and the small process variation have a significant impact on the performance and yield of the circuit. In order to ensure that the circuit meets the design specifications and optimizes its performance, the parameters in the circuit must be trimmed to compensate for the process variation in the manufacturing process, so as to ensure that the circuit performance meets the final design requirements.

[0003] In integrated circuits, there is a dependency relationship between multiple trimming parameters, i.e. the adjustment of one parameter may affect the performance of other parameters, thereby causing mutual interference in the trimming process and affecting the final optimization result. The existing trimming method is mainly single-point trimming, i.e. each trimming parameter is adjusted individually. By applying different trimming codes on the automatic test equipment (ATE), the circuit output is observed and the error is calculated until the error is within the tolerance range. This method is simple to operate, but since the trimming is performed one by one, it often needs to re-correct the previous parameter when adjusting one parameter, resulting in a tedious and inefficient trimming process. Therefore, how to efficiently handle the dependency relationship between trimming parameters and design a more intelligent optimization algorithm is a technical problem to be solved in the current integrated circuit trimming field.

[0004] The existing trimming method usually ignores the dependency relationship between multiple trimming parameters, resulting in mutual interference between parameters in the trimming process. Secondly, the method of adjusting one parameter at a time is simple to operate, but since the trimming is performed one by one, when adjusting one parameter, it often needs to re-correct other adjusted parameters, resulting in a tedious trimming process and large time consumption. In addition, the existing method for multi-parameter trimming has a slow convergence speed, is easy to fall into local optimum, is difficult to quickly find the global optimal solution, and lacks dynamic adaptability, and cannot adjust according to the real-time changes of process variation.

[0005] The existing patent application for invention with publication number CN118016132A discloses a chip tuning method, device, system and storage medium. The existing method includes: determining a first target tuning item from at least two tuning items, and determining a second target tuning item from at least two tuning items that satisfies a first condition with the first target tuning item, the first condition being used to indicate that there is a tuning influence between the first target tuning item and the second target tuning item; and tuning the first target tuning item and the second target tuning item based on a test value of the first target tuning item. However, the foregoing prior art obtains a tuning value by judging the interval in which the current parameter is located, such as a high interval or a low interval, and considers the same trend / inverse trend relationship of related parameters to form a tuning strategy and realize linkage tuning, but does not have real-time adaptive capability; lacks modeling of parameter change trend, and one-time tuning may cause excessive or insufficient deviation, and cannot handle nonlinear or non-deterministic parameter coupling.

[0006] The existing patent application for invention with publication number CN119355482A discloses a chip parameter tuning method, device, equipment and medium. The existing method includes: obtaining multiple register data of a to-be-tested chip and corresponding multiple to-be-tuned parameters; wherein the multiple to-be-tuned parameters corresponding to the multiple register data have a change relationship; sorting the multiple to-be-tuned parameters according to a set order to obtain a sorted to-be-tuned parameter list; establishing a linear correspondence relationship based on the sorted to-be-tuned parameter list and a set identification list; wherein the set identification list carries register data; and determining a target tuning parameter according to the linear correspondence relationship and a set standard parameter. However, the foregoing prior art uses a neural network model to perform tuning prediction, inputs a starting parameter and a target value, outputs a predicted tuning value, and writes the tuning value into a chip at one time. However, the foregoing prior art relies on a large number of samples for training, and when there is a large difference between chips and there are insufficient samples, the effect is poor, and the foregoing prior art cannot adapt to customized chips or few-sample problems; the output is a single-step tuning, and lacks a path evaluation and iterative optimization mechanism.

[0007] The prior art patent application document with the publication number CN118914807A discloses a self-adaptive chip parameter tuning method based on convolution operation calculation, which comprises the following steps: collecting time domain response data of a chip under different chip parameter settings; decomposing the time domain response data of the chip into basis functions by using a singular value decomposition method, and reconstructing the time domain response data through convolution operation; comparing the reconstructed time domain response data with preset standard time domain response data, optimizing the chip parameters through an adaptive combined search algorithm, and determining an optimal parameter combination; and performing tuning processing on the chip parameters through a tuning method according to the determined optimal parameter combination. However, the prior art is limited by the initial test combination accuracy, and if the test point distribution is poor, the combination may be selected incorrectly, and the prior art does not have feedback correction capability; the prior art cannot dynamically adjust the path or introduce new tuning points, and has weak expansibility; and the convolution method is only suitable for regular change trends and is difficult to cope with situations with large noise or non-continuous characteristics.

[0008] In summary, the prior art has the technical problems of mutual interference of parameters in the tuning process, complicated tuning process and large time consumption, difficulty in quickly finding a global optimal solution, and lack of dynamic adaptability. SUMMARY

[0009] The technical problem to be solved by the present application is how to solve the technical problems of mutual interference of parameters in the tuning process, complicated tuning process and large time consumption, difficulty in quickly finding a global optimal solution, and lack of dynamic adaptability in the prior art.

[0010] The present application solves the above technical problems by adopting the following technical scheme: a dynamic adaptability multi-parameter tuning optimization method comprises: S1, performing correlation modeling on chip tuning parameters through correlation analysis to obtain a high correlation parameter combination; S2, performing joint tuning according to the high correlation parameter combination to obtain a joint tuning parameter group; S3, setting an initial vertex of a Simplex search according to the joint tuning parameter group, and performing initial tuning code combination; wherein, the Simplex search is dynamically optimized; a quadratic deviation function is fitted through data-driven fitting to capture the change trend of the process deviation value, the best position of the centroid is speculated, and the Simplex search is guided so that the Simplex search is concentrated in the potential optimal solution region; S4, performing process deviation function calculation on the initial tuning code combination, calculating the vertex process deviation value and sorting to obtain a sorting result; S5, designing an adaptive search algorithm, fitting a quadratic deviation function based on data-driven fitting, dynamically adjusting the search starting point, finding the optimal tuning parameter set, and obtaining a Simplex dynamic centroid; S6, according to the sorting result and the Simplex dynamic centroid, reflection, expansion, contraction operation is carried out, and iteration search is carried out on the applicable adjustment parameter combination.

[0011] The application carries out dynamic adaptive multi-parameter adjustment optimization, aims to solve the problems of insufficient dependency processing, low adjustment efficiency and slow convergence speed in the existing adjustment process. By introducing parameter dependency relationship identification and joint adjustment mechanism, using adaptive search algorithm, dynamically adjusting the search starting point, optimizing the search direction, accelerating the convergence of the adjustment process, ensuring that the adjustment process can flexibly cope with the dynamic changes of process deviation, and improving the efficiency and accuracy of the adjustment.

[0012] In a more specific technical solution, in S1, the correlation coefficient between the chip adjustment parameters is calculated by using the Spearman correlation coefficient, a correlation matrix between the parameters is generated, correlation analysis is carried out, the dependency relationship between the chip adjustment parameters is obtained, and a high correlation parameter combination is obtained according to the dependency relationship processing.

[0013] In a more specific technical solution, in S1, the correlation degree between the parameters is determined according to the correlation coefficient between the chip adjustment parameters by using the preset correlation threshold, the high correlation, indirect dependency parameters and the low correlation parameters are obtained, the parameters are grouped based on the correlation matrix between the parameters, the high correlation parameter combination is obtained by combining the high correlation, indirect dependency parameters, and the low correlation parameters are independently adjusted and tested.

[0014] The application adopts Spearman correlation (Spearman), which mainly investigates the strength of the monotonic relationship between variables, and does not depend on the specific distribution form of the data, and is suitable for processing various numerical data. The correlation coefficient between the chip adjustment parameters is calculated by using the Spearman correlation coefficient, which can effectively reveal the dependency relationship between the chip adjustment parameters.

[0015] The application is suitable for any chip type, especially suitable for the scene with limited samples or large chip differences. At the same time, through the feedback of each step of adjustment, it is judged whether the running deviation meets the tolerance requirement, the path correction, dynamic termination and reconstruction are supported, so that the adjustment process has openness, controllability and traceability, and the black box problem of the traditional model prediction method is overcome.

[0016] In a more specific technical solution, in S3, the problem modeling operation is carried out; for the identified joint adjustment parameter combination, the adjustment code corresponding to each chip adjustment parameter is set; Construction Simplex search of vertexes, wherein each vertex represents a combination of different adjustment codes of different chip adjustment parameters in the joint adjustment parameter combination; Based on the current joint trimming parameter set, the actual output of the circuit is calculated, and the process deviation function is set according to the actual output and the expected output: by adjusting the trimming code of the chip trimming parameter, the actual output is close to the expected output, and the minimized process deviation function is obtained .

[0017] The application does not rely on fixed trend preset rules, but generates a differentiated trimming path according to the real-time running parameters, target values, parameter boundaries and historical trimming results of each chip. Feedback detection is performed after each step of trimming, and a dynamic decision is made on whether to continue, adjust the direction or terminate. Compared with the one-time trimming scheme, the application can effectively avoid over-adjustment or insufficient adjustment, improve trimming stability and consistency, and has higher adaptability and flexibility.

[0018] In a more specific technical solution, in S4, for the initial trimming code combination, the process deviation value of the vertex is calculated according to the output corresponding to the first vertex and the expected output: the process deviation value combination is obtained by combining the process deviation values of the respective vertices through the initial trimming code combination, and the sorting operation is performed according to the process deviation value combination.

[0019] In a more specific technical solution, in S5, the quadratic deviation function is fitted, the mode of change of the vertex process deviation value with the parameter is predicted according to the current data point, and the centroid calculation is performed, wherein the position of the Simplex dynamic centroid in the Simplex search is obtained by solving the minimum point of the fitted quadratic deviation function.

[0020] The application captures the change trend of the vertex process deviation value by fitting the quadratic deviation function, and speculates the best position of the centroid, so as to effectively guide the search to focus on the potential optimal solution area, thereby optimizing the reflection, expansion and contraction operations.

[0021] The quadratic deviation function fitting process of the application is simple, and can accurately capture the trend of change of the vertex process deviation value with the parameter, which helps to reduce the number of iterations and quickly guide the search towards the optimal value.

[0022] The position of the centroid is obtained by solving the minimum point of the fitted quadratic deviation function, which not only considers the influence weight of the current point, but also considers the distribution trend of the entire parameter space, which helps to jump out of the local optimum and approach the global optimal solution.

[0023] The application can more efficiently search the multi-parameter space and find the optimal trimming parameter combination. The dynamic calculation of the centroid avoids the low efficiency and local minimum value problem that may be encountered in the traditional method, thereby improving the convergence speed and trimming accuracy of the algorithm.

[0024] In a more specific technical solution, in the contraction operation of S6, if the reflection or expansion does not realize the improved solution operation, the contraction operation is performed, and the contraction point is calculated according to the Simplex dynamic centroid, the worst point in the sorting result and the contraction coefficient; if the objective function value of the contraction point is better than that of the worst point, the worst point is updated to the contraction point.

[0025] In a more specific technical solution, in S6, when the reflection, expansion and contraction operations do not realize the process deviation value improvement operation, the reduction operation is performed to narrow the search range, so that all the vertices except the optimal point in the sorting result are reduced in distance to the optimal point.

[0026] In a more specific technical solution, in S6, the reflection point is calculated, and if the objective function value of the reflection point is better than that of the optimal point in the sorting result, the expansion point is calculated according to the Simplex dynamic centroid, the worst point in the sorting result and the expansion coefficient , the search area is expanded, and if the objective function value of the expansion point meets the preset applicable condition, the worst point is updated to the expansion point.

[0027] The application dynamically constructs the adjustment path by combining point-by-point adjustment and real-time feedback, does not depend on the global characteristics of the combination space, but implements fine adjustment for specific chip states. Each step determines whether to continue or give up according to the adjustment result, which improves robustness and avoids error accumulation. At the same time, historical adjustment experience is used to improve reasoning accuracy, which is more stable and reliable when dealing with complex and irregular trends.

[0028] In a more specific technical solution, the dynamic adaptive multi-parameter adjustment optimization system comprises: A correlation analysis module is used to perform correlation modeling on chip adjustment parameters through correlation analysis to obtain a high-correlation parameter combination. A joint adjustment module is used to perform joint adjustment according to the high-correlation parameter combination to obtain a joint adjustment parameter group, and the joint adjustment module is connected with the correlation analysis module. A dynamic optimization Simplex search module is used to set the initial vertex of the Simplex search according to the joint adjustment parameter group to perform initial adjustment code combination, wherein the Simplex search is dynamically optimized, a quadratic deviation function is fitted through data driving to capture the change trend of the process deviation value, the best position of the centroid is speculated, the Simplex search is guided, and the Simplex search is concentrated in the potential optimal solution region, and the dynamic optimization Simplex search module is connected with the joint adjustment module. A deviation calculation and sorting module is used to calculate the process deviation function of the initial adjustment code combination, calculate the vertex process deviation value and sort to obtain a sorting result, and the deviation calculation and sorting module is connected with the dynamic optimization Simplex search module. The quadratic deviation function fitting module is used to design an adaptive search algorithm, fit a quadratic deviation function based on data driving, dynamically adjust a search starting point, find an optimal set of trimming parameters, and obtain a Simplex dynamic centroid. The iterative search module is used to perform reflection, expansion and contraction operations according to the sorting result and the Simplex dynamic centroid, and iteratively search for an applicable trimming parameter combination. The iterative search module is connected with the quadratic deviation function fitting module and the deviation computer sorting module.

[0029] Compared with the prior art, the present application has the following advantages: The present application performs dynamic adaptive multi-parameter trimming optimization, aiming to solve the problems of insufficient dependency processing, low trimming efficiency and slow convergence speed in the existing trimming process. By introducing parameter dependency relationship identification and joint trimming mechanism, an adaptive search algorithm is used to dynamically adjust the search starting point, optimize the search direction, accelerate the convergence of the trimming process, and ensure that the trimming process can flexibly respond to the dynamic changes of process deviation, thereby improving the efficiency and accuracy of trimming.

[0030] The present application uses Spearman correlation to mainly investigate the strength of the monotonic relationship between variables, without relying on the specific distribution form of the data, and is suitable for processing various numerical data. The Spearman correlation coefficient is used to calculate the correlation coefficient between the chip trimming parameters, which can effectively reveal the dependency relationship between the chip trimming parameters.

[0031] The present application fits a quadratic deviation function through data driving, captures the trend of the vertex process deviation value, and speculates the best position of the centroid, so as to effectively guide the search to focus on the potential optimal solution region, thereby optimizing the reflection, expansion and contraction operations.

[0032] The quadratic deviation function fitting process of the present application is simple, and can accurately capture the trend of the vertex process deviation value with parameter changes, which helps to reduce the number of iterations and quickly guide the search towards the optimal value.

[0033] The position of the centroid is obtained by solving the minimum value point of the fitted quadratic deviation function, which takes into account the influence weight of the current point and also considers the distribution trend of the entire parameter space, which helps to jump out of the local optimum and approach the global optimal solution.

[0034] The present application can more efficiently search the multi-parameter space and find the optimal trimming parameter combination. The dynamic calculation of the centroid avoids the inefficiency and local minimum problem that may be encountered in traditional methods, thereby improving the convergence speed and trimming accuracy of the algorithm.

[0035] The present application solves the technical problems of parameter interference in the trimming process, complicated trimming process and large time consumption, difficulty in quickly finding a global optimal solution, and lack of dynamic adaptability in the prior art. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the basic steps of the dynamic adaptive multi-parameter tuning and optimization method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram illustrating the specific implementation steps of the iterative search for the optimal combination of adjustment parameters in the reflection, expansion, and contraction operations of Embodiment 1 of the present invention. Figure 3 This is a comparison chart of the average number of iterations for code finding in sample analysis of Embodiment 1 of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1 like Figure 1 As shown, the dynamic adaptive multi-parameter tuning and optimization method provided by this invention includes the following basic steps: S1. Based on the combination of highly correlated parameters, perform joint adjustment to obtain a joint adjustment parameter set; In this embodiment, a dependency analysis of the tuning parameters is performed; specifically, correlation analysis is used to model the correlation of the chip tuning parameters and identify the dependencies between the parameters. In this embodiment, process deviation is the difference between circuit parameters, such as the threshold voltage, resistance, and capacitance values ​​of transistors, and their ideal design values ​​due to imperfections in the manufacturing process. Process deviation is one of the key factors affecting circuit performance and yield. To compensate for process deviation, the trimming operation optimizes circuit performance by adjusting the variable parameters in the integrated chip's circuitry, restoring the circuit to within the design specifications.

[0039] In the joint tuning of multiple parameters of an integrated chip, the circuit parameters are not independent but dependent on each other. If these dependencies are ignored, the adjustment of certain circuit parameters may interfere with each other during the tuning process, thus affecting the optimization results. In this embodiment, the optimization of joint tuning of multiple parameters is the goal.

[0040] In this embodiment, Spearman correlation coefficient is used to calculate the correlation coefficient between chip tuning parameters, generate a correlation matrix between parameters, perform correlation analysis, obtain the dependency relationship between chip tuning parameters, and process the dependency relationship to obtain a combination of highly correlated parameters.

[0041] In the process of the foregoing tuning parameter dependency analysis, evaluating the correlation between variables is a key link for understanding the tuning parameter dependency. In the embodiment, the correlation analysis method includes a Chi-Square Test, a Pearson Correlation Coefficient, a Spearman's Rank Correlation Coefficient, and a Kendall's Tau-b Correlation Coefficient. Unlike the rest of the methods, the Spearman's Rank Correlation Coefficient examines the strength of the monotonic relationship between variables, and does not depend on the specific distribution form of the tuning parameters, and is suitable for processing numerical values and data in various tuning parameters. Therefore, the present application intends to use the Spearman's Rank Correlation Coefficient to calculate the correlation coefficient between parameters, and effectively reveal the dependency relationship between the tuning parameters.

[0042] In the embodiment, a preset correlation threshold is used to determine the degree of correlation between the tuning parameters according to the correlation coefficient between the tuning parameters of the integrated chip, to obtain high-correlation, indirect-dependence parameters and low-correlation parameters, and to perform parameter grouping based on the correlation matrix between parameters, to jointly obtain a high-correlation parameter combination for the high-correlation, indirect-dependence parameters, and to perform independent tuning test on the low-correlation parameters.

[0043] In the embodiment, the Spearman's Rank Correlation Coefficient between two tuning parameters, for example, A and B, is calculated to generate a correlation matrix between parameters of the tuning parameters, and the correlation coefficient r The correlation coefficient can be calculated by the following formula:

[0044] wherein, is the rank difference of the i th sample in two tuning parameters, is the number of samples. The sample is a tuning code.

[0045] In the embodiment, the dependency relationship between parameters is identified by calculating the Spearman's Rank Correlation Coefficient between the tuning parameters. High correlation, for example, indicates that the correlation between the tuning parameters is strong, low correlation, for example, indicates that the correlation between the tuning parameters is weak. Based on the correlation matrix between parameters, the tuning parameters are grouped: the parameters with high correlation and indirect dependence are combined into a group for joint tuning, and the low-correlation parameters are independently tuned and tested.

[0046] ​S2, according to the high correlation parameter combination, joint adjustment is performed to obtain a joint adjustment parameter combination; In this embodiment, the high correlation parameter combination obtained according to the parameter correlation matrix is combined for joint adjustment. S3, according to the joint adjustment parameter combination, an initial vertex of Simplex search is set, and initial adjustment code combination is performed; wherein, the Simplex search is dynamically optimized; a quadratic deviation function is fitted by data driving, a change trend of the process deviation value is captured, an optimal position of the centroid is inferred, the Simplex search is guided, and the Simplex search is concentrated in a potential optimal solution region. In this embodiment, the Simplex search is based on dynamic optimization; specifically, the Simplex algorithm is used for the search strategy of the multi-parameter adjustment process. The traditional centroid calculation removes the worst point and calculates the average position, and the average position is used as the starting point of the search to provide a reasonable direction for the next search. However, this method only considers the average position of all points, which may lead to excessive dependence on poor performing points. The present application aims to effectively guide the search to concentrate in the potential optimal solution region by fitting a quadratic deviation function by data driving, capturing the change trend of the process deviation value, inferring the optimal position of the centroid, and optimizing the reflection, expansion and contraction operations.

[0047] In this embodiment, a problem modeling operation is performed; for the identified joint adjustment parameter combination, the adjustment code corresponding to each chip adjustment parameter is set. Construction Simplex search of vertices, wherein each vertex represents a combination of different adjustment codes of different chip adjustment parameters in the joint adjustment parameter combination. Based on the current joint adjustment parameter combination, the actual output of the circuit is calculated, and the process deviation function is set according to the actual output and the expected output: by adjusting the adjustment code of the chip adjustment parameter, the actual output is close to the expected output, and the minimized process deviation function is obtained .

[0048] In this embodiment, a problem modeling operation is performed; specifically, for the identified joint adjustment parameter combination, for example: , the adjustment code corresponding to each chip adjustment parameter is set , construction Simplex of vertices, the vertices are respectively represented as: …, , wherein each vertex represents a combination of different adjustment codes of different chip adjustment parameters in the joint adjustment parameter combination, for example: the three adjustment parameters of vertex may use the adjustment codes as follows: .n representing the number of vertices, P representing the joint trimming parameter set, X representing the vertex, m representing the trimming code.

[0049] set the current joint trimming parameter set i the process deviation function wherein, is the actual output calculated based on the current trimming parameter set i , such as: the working frequency of the system, power, is the expected output, by adjusting the trimming code of the trimming parameter, so that the actual output of the circuit is as close as possible to the expected output, that is, the process deviation function is minimized. O representing the output.

[0050] S4, the process deviation function is calculated for the initial trimming code combination, the vertex process deviation value is calculated and sorted, and the sorting result is obtained; In this embodiment, for the initial trimming code combination, the process deviation function is calculated according to the output corresponding to the first vertex and the expected output, and the vertex process deviation value is obtained: by combining the respective vertex process deviation values of the initial trimming code combination, the process deviation value combination is obtained, and the sorting operation is performed according to the process deviation value combination; In this embodiment, for the initial joint trimming parameter set corresponding to the vertex …, , the process deviation function is calculated as follows:

[0051] wherein, representing the output corresponding to the trimming code corresponding to each chip trimming parameter; according to the vertex …, the respective process deviation values are sorted, for example: the worst point , the second worst point , the best point .

[0052] S5, design an adaptive search algorithm, based on data driving, fit a quadratic deviation function, dynamically adjust the search starting point, find the optimal trimming parameter set, and obtain the Simplex dynamic centroid; In this embodiment, the quadratic deviation function is fitted, the mode of the vertex process deviation value changing with the parameter is predicted according to the current data point; the centroid calculation is performed, wherein the position of the Simplex dynamic centroid in the Simplex search is obtained by solving the minimum point of the fitted quadratic deviation function; In this embodiment, a quadratic deviation function is fitted based on the vertex corresponding to the current data, and the dynamic centroid of the Simplex is obtained; In this embodiment, a quadratic deviation function is fitted; specifically, due to each joint adjustment parameter group i process deviation value This only describes the deviation at that vertex and cannot fully reflect the changing trend of the process deviation value in the current parameter space. Therefore, this invention fits a quadratic deviation function. It predicts the pattern of process deviation values ​​as parameters change using current data points.

[0053]

[0054] in, f Indicates the reflection parameters. Denotes the first fitting coefficient. This represents the second fitting coefficient. Represents the third fitting coefficient. Indicates the first i Individual adjustment code. Indicates the first j Individual adjustment code.

[0055] The quadratic deviation function fitting is relatively simple and can capture the trend of process deviation values ​​changing with parameters well. This helps to reduce the number of iterations and quickly guide the search toward the optimal value.

[0056] In this embodiment, the centroid is calculated. Specifically, the traditional centroid calculation method is to take the average of all vertices except the worst point. However, in actual optimization, especially in multidimensional parameter space, simple average calculation cannot fully reflect the complex error distribution and interdependent relationships.

[0057] This invention solves for the fitted quadratic deviation function. The minimum point determines the location of the centroid. This approach considers both the influence weight of the current point and the distribution trend of the entire parameter space, which helps the search process escape local optima and approach the global optimum.

[0058] S6. Based on the sorting results and the dynamic centroid of the Simplex, perform reflection, expansion, and contraction operations, and iteratively search for suitable adjustment parameter combinations.

[0059] In this embodiment, based on the sorting results and the calculation of the centroid, reflection, expansion, contraction, and reduction operations can be performed to optimize the search process; like Figure 2 As shown, the specific implementation steps of the aforementioned step S6 include, but are not limited to: S61. Initialize the vertices of the Simplex.X 1, X 2,..., X n+1 , calculate each process deviation value ; S62, sort vertices X h , X s ,..., and calculate dynamic centroid X centroid ; S63, calculate reflection point X r and reflection point X r corresponding objective function f(X r ) ; Reflection: calculate reflection point , update search direction by reflecting the worst point to the other side of the centroid:

[0060] where, is the centroid, is the worst point, is the reflection coefficient, usually 1.

[0061] S64, determine whether to meet: ; In this embodiment, if the reflection point X r objective function f(X r ) is less than the objective function of the optimal point , replace the worst point with the reflection point; S65, if not, determine whether to meet: ; S66, if yes, perform contraction operation: calculate contraction point X c and contraction point X c corresponding objective function f(X c ) ; In the contraction operation, if reflection or expansion does not achieve improved solution operation, perform contraction operation, calculate contraction point according to Simplex dynamic centroid, worst point in the sorting result and contraction coefficient; if the objective function value of the contraction point is better than the worst point, update the worst point to the contraction point; Contraction: if reflection or expansion fails to improve the solution effectively, contraction operation is performed to calculate a contraction point :

[0062] wherein, is a contraction coefficient, which can be set as, for example, 0. If the objective function value of the contraction point is better than that of the worst point, the worst point is updated to be the contraction point; S67, determine whether the following condition is satisfied: f(X c ) f(X h ) ; wherein, f(X c ) represents the objective function of the contraction point X c ; f(X h ) represents the objective function of the worst point X h ; S68, if yes, all the vertices are contracted to the optimal point; In the present embodiment, when the reflection, expansion and contraction operations fail to achieve the process deviation value improvement operation, a reduction operation is performed to narrow the search range, so that all the vertices except the optimal point in the sorting result are contracted to the optimal point:

[0063] wherein, represents the i-th vertex i , is the optimal point, is a reduction coefficient, which can be set as, for example, ; S69, when the following condition is not satisfied: , an expansion operation is performed to calculate an expansion point e and X e f(X ; ) represents the objective function of the optimal point ; the reflection point is calculated, if the objective function value of the reflection point is better than that of the optimal point in the sorting result, the expansion point is calculated according to the Simplex dynamic centroid, the worst point in the sorting result and the expansion coefficient, the search region is expanded, and if the objective function value of the expansion point satisfies the preset applicable condition, the worst point is updated to be the expansion point;​​ Expansion: if the objective function value of the reflection point is better than the optimal point, calculate the expansion point , further expanding the search region:

[0064] wherein is the expansion coefficient, which can be set as, for example: If the objective function value of the expansion point is better, update the worst point as the expansion point; S610, determine whether to meet: f(X e ) f(X r ) ; wherein, f(X e ) denotes the objective function of the expansion point X e . S611, if yes, replace the worst point X h . S612, if not, replace the worst point with the reflection point.

[0065] When the process deviation value is within the preset tolerance range or the maximum number of iterations is reached, the algorithm terminates and outputs the final tuning parameter combination.

[0066] Through the foregoing improvement, the multi-parameter space can be searched more efficiently to find the optimal tuning parameter combination. The dynamic calculation of the centroid avoids the inefficiency and local minimum problem that may be encountered in the traditional method, thereby improving the convergence speed and tuning accuracy of the algorithm.

[0067] In the present embodiment, in order to evaluate the performance of the traditional method and the dynamic adaptive multi-parameter tuning optimization method provided by the present application in terms of efficiency, the average number of iterations for finding codes of 30 groups of samples was analyzed. The traditional method directly uses Nelder-Mead (simplex search method) for searching. The essence of this traditional method is to find the optimal solution through a geometric shape deformation, such as reflection, expansion, and contraction. Although this method is simple and easy to implement, it has obvious limitations: the centroid calculation only uses simple arithmetic mean and fails to utilize the trend information of the objective function, which is prone to fall into local optimum in high-dimensional complex space, resulting in low search efficiency. For example, Figure 3 ​As shown, the iteration number of the conventional method is an orange column, and the iteration number of the method of the present application is a green column. The present application exhibits a lower iteration number on most samples. The average iteration number of the conventional method is 12.70, and the average iteration number of the method of the present application is 6.37. The average iteration number of the present application is reduced by 50% compared with the conventional method. The overall performance of the present application is significantly better than that of the conventional method.

[0068] Compared with the conventional method, the improvement of the present application mainly embodies in two aspects. First, the parameter dependence relationship identification is introduced before the search starts. For example, the strong correlation between chip trimming parameters can be quantified and identified by Spearman correlation analysis, so that the chip trimming parameters with strong correlation are divided into a group for joint and collaborative optimization, thereby improving the pertinence of the search from the root cause. The present application adopts a dynamic fitting centroid strategy in the search process, instead of the traditional static centroid calculation. The present application fits a deviation trend function through historical sample data, and then predicts and locates the centroid point closer to the global optimal region. This makes the subsequent reflection, expansion and other search operation directions more accurate, not only accelerating the convergence, but also having the strong ability to jump out of the local optimum, effectively overcoming the inherent defects of the conventional method.

[0069] In summary, the present application performs dynamic adaptive multi-parameter trimming optimization, aiming to solve the problems of insufficient dependence processing, low trimming efficiency and slow convergence speed in the existing trimming process. By introducing parameter dependence relationship identification and joint trimming mechanism, and using an adaptive search algorithm, the search starting point is dynamically adjusted, the search direction is optimized, the convergence of the trimming process is accelerated, and it is ensured that the trimming process can flexibly cope with the dynamic changes of process deviation, thereby improving the efficiency and accuracy of trimming.

[0070] The Spearman correlation adopted by the present application mainly investigates the strength of the monotonic relationship between variables, and does not depend on the specific distribution form of the data, and is suitable for processing various numerical data. The Spearman correlation coefficient is used to calculate the correlation coefficient between chip trimming parameters, which can effectively reveal the dependence relationship between chip trimming parameters.

[0071] The present application fits a quadratic deviation function through data driving, captures the change trend of the vertex process deviation value, and speculates the best position of the centroid, so as to effectively guide the search to focus on the potential optimal solution region, thereby optimizing the reflection, expansion and contraction operations.

[0072] The fitting process of the quadratic deviation function of the present application is simple, and the trend of the vertex process deviation value with the change of the parameter can be accurately captured, which helps to reduce the iteration number and quickly guide the search to move towards the optimal value.

[0073] The position of the centroid is obtained by solving the minimum point of the fitted quadratic deviation function, which considers the influence weight of the current point and the distribution trend of the whole parameter space, and is helpful to jump out of the local optimum and approach the global optimum.

[0074] The application can search the multi-parameter space more efficiently and find the optimal tuning parameter combination.

[0075] The application solves the technical problems of mutual interference of parameters in the tuning process, complicated tuning process and large time consumption, difficulty in quickly finding the global optimal solution and lack of dynamic adaptability in the prior art.

[0076] The above examples are only used to illustrate the technical solutions of the application, but not to limit it; although the application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A dynamic adaptive multi-parameter trim optimization method, characterized in that, The method comprises: S1, correlation modeling of chip trimming parameters is performed through correlation analysis to obtain a high correlation parameter combination; S2, joint trimming is performed according to the high correlation parameter combination to obtain a joint trimming parameter group; S3, an initial vertex of Simplex search is set according to the joint trimming parameter group, and an initial trimming code combination is performed; wherein, the Simplex search is dynamically optimized; a quadratic deviation function is fitted through data driving to capture the change trend of process deviation values, to speculate the best position of the centroid, to guide the Simplex search, so that the Simplex search is focused on the potential optimal solution area; S4, process deviation function calculation is performed on the initial trimming code combination, the vertex process deviation values are sorted to obtain a sorting result; S5, an adaptive search algorithm is designed, a quadratic deviation function is fitted based on data driving, the search starting point is dynamically adjusted, the optimal trimming parameter set is searched, and a Simplex dynamic centroid is obtained; S6, according to the sorting result and the Simplex dynamic centroid, reflection, expansion and contraction operations are performed, and an applicable trimming parameter combination is iteratively searched.

2. The dynamic adaptive multi-parameter retuning optimization method of claim 1, wherein, In S1, Spearman correlation coefficient is used to calculate the correlation coefficient between the chip trimming parameters, a parameter correlation matrix is generated, the correlation analysis is performed to obtain the dependency relationship between the chip trimming parameters, and the high correlation parameter combination is obtained according to the dependency relationship.

3. The dynamic adaptive multi-parameter retuning optimization method of claim 2, wherein, A preset correlation threshold is used to determine the correlation degree between parameters according to the correlation coefficient between the chip trimming parameters, high correlation, indirect dependency parameters and low correlation parameters are obtained, parameter grouping is performed based on the parameter correlation matrix, the high correlation and the indirect dependency parameters are combined to obtain the high correlation parameter combination, and the low correlation parameters are independently trimmed and tested.

4. The dynamic adaptive multi-parameter retuning optimization method of claim 1, wherein, In S3, a problem modeling operation is performed; for the identified joint trimming parameter group, a trimming code corresponding to each chip trimming parameter is set; constructing the Simplex search of vertices, wherein each of the vertices represents a combination of different trim codes of different chip trim parameters in the joint trim parameter set; Based on the current set of joint trim parameters, the actual output of the circuit is calculated, and based on the actual output and the desired output, a process bias function is set: by adjusting the trim code of the chip trim parameters such that the actual output approaches the desired output, the process bias function is minimized .

5. The dynamic adaptive multi-parameter retuning optimization method of claim 1, wherein, In S4, according to the output corresponding to the first vertex and the expected output, the process deviation function calculation is performed on the initial trimming code combination to obtain the vertex process deviation value; the vertex process deviation values of the initial trimming code combination are combined to obtain a process deviation value combination, and the sorting operation is performed according to the process deviation value combination.

6. The dynamic adaptive multi-parameter retuning optimization method of claim 1, wherein, In S5, the quadratic deviation function is fitted, the mode of change of the vertex process deviation value with the parameter is predicted according to the current data point; the centroid calculation is performed, wherein, the position of the Simplex dynamic centroid in the Simplex search is obtained by solving the minimum point of the fitted quadratic deviation function.

7. The dynamic adaptive multi-parameter retuning optimization method of claim 1, wherein, In the contraction operation of S6, if the reflection or the expansion does not realize the improved solution operation, the contraction operation is performed, and a contraction point is calculated according to the Simplex dynamic centroid, the worst point in the sorting result and a contraction coefficient; If the objective function value of the contraction point is better than that of the worst point, the worst point is updated to the contraction point.

8. The dynamic adaptive multi-parameter retuning optimization method of claim 1, wherein, In the S6, when the reflection, the expansion, and the contraction operation do not achieve the process bias value improvement operation, then a reduction operation is performed to reduce the search range, and all vertices in the sorting result except the optimal point are reduced in distance to the optimal point.

9. The dynamic adaptive multi-parameter retuning optimization method of claim 1, wherein, In the S6, a reflection point is obtained, and if a target function value of the reflection point is better than an optimal point in the sorting result, an expansion point is calculated according to the Simplex dynamic centroid, the worst point in the sorting result, and an expansion coefficient , the search region is expanded, and if a target function value of the expansion point satisfies a preset applicable condition, the worst point is updated as the expansion point.

10. A dynamically adaptive multi-parameter trim optimization system, characterized by, The system comprises: a correlation analysis module configured to perform correlation modeling on chip trimming parameters through correlation analysis to obtain a high-correlation parameter combination; a joint trimming module configured to perform joint trimming according to the high-correlation parameter combination to obtain a joint trimming parameter group, wherein the joint trimming module is connected to the correlation analysis module; a dynamic optimization Simplex search module configured to set initial vertices of Simplex search according to the joint trimming parameter group to perform initial trimming code combination, wherein the Simplex search is dynamically optimized, a quadratic bias function is fitted through data-driven fitting to capture a change trend of a process bias value, an optimal position of a centroid is speculated to guide the Simplex search, and the Simplex search is concentrated in a potential optimal solution region, wherein the dynamic optimization Simplex search module is connected to the joint trimming module; a bias computer sorting module configured to perform process bias function calculation on the initial trimming code combination to calculate vertex process bias values and sort them to obtain a sorting result, wherein the bias computer sorting module is connected to the dynamic optimization Simplex search module; a quadratic bias function fitting module configured to design an adaptive search algorithm, fit a quadratic bias function based on data-driven fitting, dynamically adjust a search starting point, find an optimal trimming parameter set, and obtain a Simplex dynamic centroid; an iterative search module configured to perform reflection, expansion, and contraction operations according to the sorting result and the Simplex dynamic centroid to iteratively search for an applicable trimming parameter combination, wherein the iterative search module is connected to the quadratic bias function fitting module and the bias computer sorting module.

Images