A method and system for predicting IC fuse tuning codes

Abstract

This invention discloses a method and system for predicting IC fuse trimming codes. The method includes: acquiring the test voltage before trimming, the actual trimming code, and the corresponding test voltage after trimming as samples; utilizing the GBDT framework, introducing a mechanism for randomly discarding residual samples and perturbating leaf node outputs to construct an overfitting-resistant GBDT model; and predicting the trimming code using the overfitting GBDT model trained with the applied samples. The trimming code prediction method of this invention can significantly shorten trimming time, reduce resource consumption, and avoid losses caused by trial and error.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit tuning technology, specifically to a method and system for predicting IC fuse tuning codes. Background Technology

[0002] In integrated circuit (IC) manufacturing, laser trimming is a common post-processing adjustment method used to precisely adjust parameters such as voltage, current, or timing of devices. This process typically relies on the difference between the pre-trim parameters (such as voltage, current, and delay) and the target trimming values. By applying trimming codes at specific trimming locations, the parallel channels of the trimming process are altered, allowing the circuit to achieve the desired performance.

[0003] In fuse tuning, parameters such as voltage, current, and delay change irreversibly with variations in tuning position and tuning code. However, current mainstream tuning methods generally rely on global search or binary search to determine the optimal tuning code. These methods require trying every possible code value to obtain the tuned result before making a judgment. Because tuning is irreversible, an incorrect tuning code or position is not only difficult to recover from but may also cause permanent chip damage. Furthermore, to ensure tuning accuracy and yield, this full-exploration search method consumes significant testing time and hardware resources. Even the fastest existing search methods require searching half of the tuning codes, and their efficiency still falls short of modern mass production requirements.

[0004] Therefore, an efficient method for predicting tuning codes is needed that can significantly shorten tuning time, reduce resource consumption, and avoid losses caused by trial and error.

[0005] The invention method disclosed in patent publication number CN115424652A presents a method, apparatus, electronic device, and readable storage medium for tuning tests. It employs a classification model-based tuning test scheme. By collecting historical tuning test data of integrated circuit chips and using it as a classification training sample set, a classification algorithm is used to train a model on the training sample set, resulting in a trained classification tuning model. This model allows the actual test value of the target chip to be input into the model during tuning tests, and corresponding tuning parameters are output based on the chip's location information. The circuit parameters of the tuning circuit on the target chip are then adjusted based on these parameters to match the tuned actual test value with the target tuning value. However, the patent protects only the tuning test method and the tuning file generated by the model, and does not involve tuning code prediction. Summary of the Invention

[0006] The technical problem to be solved by this invention is that traditional global search or binary search methods for determining the optimal modifier code, based on the modifier code index, have long search times and their efficiency is difficult to meet the requirements of modern large-scale mass production.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] An IC fuse trimming code prediction method includes:

[0009] The test voltage before adjustment, the actual adjustment code, and the corresponding test voltage after adjustment are obtained as samples. Based on the GBDT framework, a residual sample random discarding and leaf node output perturbation mechanism are introduced to construct an anti-overfitting GBDT model. The anti-overfitting GBDT model trained with application samples is used to predict the adjustment code.

[0010] Technical Effects: Compared with traditional methods based on modifier code index search, this invention uses the GBDT framework to construct an overfitting-resistant GBDT model for modifier code prediction, resulting in faster speed compared to indexing one by one. Furthermore, by utilizing the traditional GBDT framework and introducing a mechanism for randomly discarding residual samples and perturbing leaf node outputs, overfitting is effectively suppressed, improving the model's robustness and generalization ability.

[0011] In this embodiment, obtaining a GBDT model trained with application samples and resistant to overfitting includes:

[0012] During the training of each regression tree, a portion of the residual sample set in each leaf node is randomly discarded proportionally to obtain the retained sample set.

[0013] Gaussian noise is added to the residual values ​​of the retained sample set;

[0014] For each leaf node, the final output value is the mean of the retained sample and the residual after adding Gaussian noise;

[0015] After the regression trees are trained, the learning rate is used to control the update magnitude, the output value of each regression tree is updated, and the new prediction probability is calculated by the softmax function for the next round of residual calculation.

[0016] After M iterations, a GBDT model trained on the samples and resistant to overfitting is obtained.

[0017] In this embodiment, the output value of each regression tree is updated using the following formula:

[0018]

[0019] T k (m) (x i )≈r ik (m) ;

[0020] In the formula, F k (m) (xi Let be the score of the k-th class in the m-th iteration, and v be the learning rate. l (k,m) To preserve the mean of the sample and the residuals after adding Gaussian noise, x i Let be the features of the i-th sample, k be the class of each regression tree, m be the iteration number, l be the leaf node, and r be the number of iterations. ik (m) The current error target for sample i in class k is used for the regression tree to learn; T k (m) (x i ) is the regression tree of the k-th class and the m-th round.

[0021] In this embodiment, a fuzzy region for the modifier code is constructed to optimize the modifier code predicted by the GBDT model to combat overfitting.

[0022] In this embodiment, the adjustment code of the GBDT model prediction output to combat overfitting is discriminated and optimized, including:

[0023] Let μ be the average trimming voltage corresponding to the j-th discrete integer trimming code. j If there exists an interval such that for any sample its actual adjustment voltage V∈[μ] j μ j+1 The target precision can be achieved by modifying the code j or j+1 in the interval [μ]. j μ j+1 [This is defined as the initial modifier code ambiguity region;]

[0024] Let μ j+1 >μ j Obtain the center point of the blurred region of the modulated code:

[0025] Using the center point as a reference, the voltage range above and below it by η% is taken as the final ambiguous area of ​​the modulated code:

[0026] When the chip's current actual trimming voltage falls within the ambiguity region of the final trimming code, any adjacent trimming code is used to adjust the voltage within the target range.

[0027] Technical Effects: This invention provides a gradient boosting decision tree multi-classification model that integrates an anti-overfitting mechanism and a fuzzy region discrimination strategy for predicting fuse trimming codes, thereby achieving accurate single-shot trimming code search. By introducing a mechanism of randomly discarding residual samples and Gaussian perturbation of leaf node outputs, overfitting is effectively suppressed, improving the model's robustness and generalization ability. Simultaneously, by identifying fuzzy regions in the trimming codes, the ability to discriminate uncertain samples between adjacent trimming codes is enhanced, significantly reducing the misclassification rate.

[0028] This invention also provides an IC fuse trimming code prediction system, comprising:

[0029] The sample module is used to obtain the test voltage before adjustment, the actual adjustment code, and the corresponding test voltage after adjustment as samples.

[0030] The model building module is used to build an overfit-resistant GBDT model by introducing a residual sample random discarding and leaf node output perturbation mechanism based on the GBDT framework.

[0031] The prediction module is used to predict tuning codes for the overfit-resistant GBDT model trained on the application samples.

[0032] Compared with existing technologies, this invention provides an efficient method for predicting tuning codes, which can significantly shorten tuning time, reduce resource consumption, and avoid losses caused by trial and error. Attached Figure Description

[0033] Figure 1 This is a flowchart of an IC fuse tuning code prediction method according to an embodiment of the present invention.

[0034] Figure 2 This is a flowchart of the IC fuse trimming code prediction method for introducing a trimming code ambiguity region according to an embodiment of the present invention.

[0035] Figure 3 This is a block diagram of an IC fuse trimming code prediction system according to an embodiment of the present invention. Detailed Implementation

[0036] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0037] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0038] Please see Figure 1 As shown, the present invention provides an IC fuse trimming code prediction method, comprising:

[0039] S10: Obtain the test voltage before adjustment, the actual adjustment code, and the corresponding test voltage after adjustment as samples.

[0040] In this embodiment, the test voltage (current, delay, frequency, etc.) before adjustment, the actual adjustment code, and the corresponding test voltage (current, delay, frequency, etc.) after adjustment are extracted from the actual fuse adjustment test results.

[0041] S20 utilizes the GBDT framework, introduces a mechanism for randomly discarding residual samples and perturbing leaf node outputs, and constructs a GBDT model resistant to overfitting.

[0042] In this embodiment, the core idea of ​​GBDT (Gradient Boosting Decision Tree) is to train a new weak learner at each step through sequential iteration to correct the residuals of the previous step. Furthermore, the leaf nodes and regression tree together determine the prediction method and structure of the GBDT model. The residual samples refer to the set of residual samples formed by the GBDT algorithm calculating the residuals of all samples in each iteration.

[0043] S21. Before the overfitting-resistant GBDT model has undergone any training, it does not yet possess any discriminative ability. At this point, the initial prediction score for each category can be uniformly set to 0, i.e., the regression tree can be initialized.

[0044]

[0045] In the formula, The output value of the initialized regression tree, x i Let be the feature of the i-th sample, and k be the category. Where , i, i∈{1,2,…,n}, and k, k∈{1,2,…,K}.

[0046] S22, since the regression tree is initialized with all output values ​​set to 0, the initial predicted probability obtained after applying softmax is uniformly distributed as follows:

[0047]

[0048] In the formula, The initial prediction probability.

[0049] S23. To train the model and gradually improve its ability to predict the correct class, a suitable loss function needs to be constructed. The cross-entropy loss function is adopted as follows:

[0050]

[0051] Among them, y ik ∈{0,1} is a single effective code, and y is the true class of sample i when k. ik =1, otherwise 0. Where l(m) is the loss function for the m-th iteration. Let be the predicted probability of sample i in the m-th iteration on class k.

[0052] S24. To minimize the loss function, a gradient boosting method is used. In the m-th iteration, the residual is defined as the current error target for each sample with respect to each class, which is then used by the regression tree to learn:

[0053]

[0054] In the formula, r ik (m) Let be the current error target for sample i in class k, for the regression tree to learn.

[0055] S25. In each iteration, a regression tree is trained for each category to fit the residual value of that category, which is then used to correct the output value of the regression tree in subsequent iterations.

[0056]

[0057] T k (m) (x i )≈r ik (m) ;

[0058] In the formula, F k (m) (x i Let be the score of the k-th class in the m-th iteration, v∈(0,1] be the learning rate, and T be the learning rate. k (m) Let be the regression tree of the k-th class and the m-th round.

[0059] S26. In order to improve the robustness of the model and introduce randomness, the residual samples within the leaf nodes are processed during the training process of each regression tree.

[0060] S261, random discarding: For the residual sample set in each leaf node, a portion of samples are randomly discarded proportionally ρ∈[0,1]. This strategy not only effectively reduces the model's dependence on individual samples but also introduces structural diversity, thereby reducing the possibility of overfitting and obtaining a retained sample set. The retained sample set is as follows:

[0061]

[0062] In the formula, To preserve the sample set, ρ represents the discard ratio. Let l be the residual sample set, and l be a leaf node, where l∈{1,2,...,L}.

[0063] S262 adds Gaussian noise to the residual values ​​of the retained sample set, improving the model's robustness to noisy samples while maintaining the validity of the residual information. It is defined as follows:

[0064] r′ ik =r ik +ε,ε∈N(0,σ noise 2 );

[0065] In the formula, r′ ik r is the residual value after Gaussian noise. ik Let ε be the residual value in the retained sample set, σ be the noise sampled from the normal distribution, and σ be the residual value. noise N(0,σ) is the noise standard deviation used to control the disturbance intensity. noise 2 () represents normally distributed noise.

[0066] S263, for each leaf node, the final output value is the mean of the residuals after retaining the sample and adding noise, representing the correction value of that leaf node to the regression tree output value:

[0067]

[0068] In the formula, v l (k,m) The mean of the retained sample and the residual after adding Gaussian noise.

[0069] S264, After the regression trees are trained, the learning rate is used to control the update magnitude, and the output value of each regression tree is updated:

[0070] F k (m) (x i ) = F k (m-1) (x i )+v·f k (m) (x i );

[0071] Among them, f k (m) (x i ) = v l (k,m) v is the learning rate, controlling the update magnitude in each round, and F is the learning rate. k (m) (x i Update the output value for each regression tree.

[0072] S265, the new prediction probabilities are calculated again using the softmax function for the next round of residual calculation:

[0073]

[0074] In the formula, Let x be the probability predicted in the m-th iteration, i.e., in the m-th round, the probability of sample x. i The probability of predicting it as class k.

[0075] S266, after M iterations, the final model output is:

[0076]

[0077] That is, in the formula, A well-trained, overfit-resistant GBDT model for the samples. This represents the probability predicted during the Mth iteration.

[0078] S30 predicts the trimming code for the overfit-resistant GBDT model trained on the application samples. Specifically, it selects the class with the highest final softmax probability as the classification result of the overfit-resistant GBDT model, enabling the prediction of the trimming code value using the test voltage before trimming and the corresponding test voltage after trimming.

[0079] Please see Figure 2 As shown, in one embodiment of the present invention, the IC fuse trimming code prediction method further includes: S40, constructing a trimming code ambiguity region, performing discrimination optimization on the trimming code predicted by the anti-overfitting GBDT model, enhancing the discrimination ability of uncertain samples between adjacent trimming codes, and significantly reducing the false positive rate. During fuse trimming, the average trimming voltage ranges corresponding to some adjacent trimming codes overlap to a certain extent. When the current chip voltage falls within this overlapping range, using any adjacent trimming code may achieve adjustment within the target voltage range, thereby meeting the trimming accuracy requirements; this range is the ambiguity region. Specifically, it includes:

[0080] S41, let μ be the average trimming voltage corresponding to the j-th discrete integer trimming code. j If there exists an interval such that for any sample its actual adjustment voltage V∈[μ] j μ j+1 The target precision can be achieved by modifying the code j or j+1 in the interval [μ]. j μ j+1 [This is defined as the initial modifier code ambiguity region.]

[0081] S42, let μ j+1 >μ j Obtain the center point of the blurred region of the modulated code:

[0082] S43, taking the center point as a reference, the voltage range above and below it by η% is taken as the final ambiguous region of the modulated code, then this region: Further simplification:

[0083] S44: When the current actual adjustment voltage of the chip falls within the ambiguity region of the final adjustment code, any adjacent adjustment code is used to adjust the voltage within the target range.

[0084] It should be noted that, in this field, adjustment codes typically refer to codes or parameters used for equipment debugging or configuration. Adjustment codes are usually grouped, with 0-15 forming one group and 16-31 forming another. Adjustment codes 0-15 are typically used for basic parameter adjustments, such as basic calibration of frequency and voltage. Adjustment codes 16-31 are typically used for advanced parameter adjustments or reserved bits, such as temperature compensation and nonlinear correction. Each adjustment code corresponds to a specific adjustment step size. Adjustment codes 0-15 and 16-31 are located in two different linear adjustment step size intervals. Therefore, there is a step size abrupt change point between adjustment codes 15 and 16, and this position does not constitute a continuous ambiguity area; therefore, it is not included in the ambiguity region of the adjustment codes.

[0085] Please see Figures 1 to 3 As shown, the present invention also provides an IC fuse tuning code prediction system, comprising:

[0086] The sample module is used to obtain the test voltage before adjustment, the actual adjustment code, and the corresponding test voltage after adjustment as samples.

[0087] The model building module is used to build an overfit-resistant GBDT model by introducing a residual sample random discarding and leaf node output perturbation mechanism based on the GBDT framework.

[0088] The prediction module is used to predict tuning codes for the overfit-resistant GBDT model trained on the application samples.

[0089] In this embodiment, obtaining a GBDT model trained with application samples and resistant to overfitting includes:

[0090] During the training of each regression tree, a portion of the residual samples in each leaf node are randomly discarded proportionally to obtain the retained sample set.

[0091] Gaussian noise is added to the residual values ​​of the retained sample set.

[0092] For each leaf node, the final output value is the mean of the retained sample and the residual after adding Gaussian noise.

[0093] After the regression trees are trained, the learning rate is used to control the update magnitude, the output value of each regression tree is updated, and the new prediction probability is calculated using the softmax function for the next round of residual calculation.

[0094] The output value of each regression tree is updated using the following formula:

[0095] F k (m) (x i ) = F k (m-1) (x i )+v·f k (m) (x i );

[0096] f k (m) (x i ) = v l (k,m) ;

[0097] In the formula, F k (m) (x i ) represents the update value of the output value for each regression tree, where v is the learning rate. l (k,m) To preserve the mean of the sample and the residuals after adding Gaussian noise, x i Let be the features of the i-th sample, k be the category of each regression tree, m be the iteration number, and l be the leaf node.

[0098] After M iterations, a GBDT model trained on the samples and resistant to overfitting is obtained.

[0099] In this embodiment, a fuzzy region for the modifier code is constructed to optimize the modifier code predicted by the GBDT model to combat overfitting. Specifically, the optimization of the modifier code predicted by the GBDT model to combat overfitting includes:

[0100] Let μ be the average trimming voltage corresponding to the j-th discrete integer trimming code. j If there exists an interval such that for any sample its actual adjustment voltage V∈[μ] j μ j+1 The target precision can be achieved by modifying the code j or j+1 in the interval [μ]. j μ j+1 [This is defined as the initial modifier code ambiguity region.]

[0101] Let μ j+1 >μ j Obtain the center point of the blurred region of the modulated code:

[0102] Using the center point as a reference, the voltage range above and below it by η% is taken as the final ambiguous area of ​​the modulated code:

[0103] When the chip's current actual trimming voltage falls within the ambiguity region of the final trimming code, any adjacent trimming code is used to adjust the voltage within the target range.

[0104] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0105] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for predicting IC fuse trimming codes, characterized in that, include: The test voltage before adjustment, the actual adjustment code, and the corresponding test voltage after adjustment are obtained as samples. Based on the GBDT framework, a GBDT model resistant to overfitting is constructed by introducing a mechanism for randomly discarding residual samples and perturbing leaf node outputs, including: The initial prediction score for each category is uniformly set to 0, the regression tree is initialized, and the softmax function is applied to obtain the initial prediction probability. Construct a loss function that can progressively enhance the ability to predict the correct category; Using the gradient boosting method, in the first... In each iteration, the residual is defined as the current error target for each sample for each class, which is used by the regression tree to learn and minimize the loss function; In each iteration, a regression tree is trained for each category to fit the residual value of that category, which is then used to correct the output value of the regression tree in subsequent iterations. During the training of each regression tree, the residual samples within the leaf nodes are processed as follows: For each regression tree, a portion of the residual samples in each leaf node are randomly discarded proportionally to obtain a retained sample set; Gaussian noise is added to the residual values ​​of the retained sample set; for each leaf node, the final output value is the mean of the retained samples and the residuals after adding Gaussian noise; after the regression trees are trained, the learning rate is used to control the update magnitude, updating the output values ​​of each regression tree, and a new prediction probability is calculated using the softmax function for the next round of residual calculation; after M iterations, a well-trained, overfit-resistant GBDT model is obtained. For the overfit-resistant GBDT model trained on application samples, predict the modifier code; among which, is the number of iterations; M is the maximum number of iterations.

2. The IC fuse trimming code prediction method according to claim 1, characterized in that, The output values ​​of each regression tree are updated using the following formula: ； ； In the formula, For the first During the nth iteration Class score, For learning rate, For the first Features of each sample For each regression tree category, Leaf node For the sample In category The current error target is used for regression tree learning; For the first Class 1 The cycle of the regression tree.

3. The IC fuse trimming code prediction method according to claim 1, characterized in that, A fuzzy region for the modifier code is constructed to optimize the modifier code predicted by the GBDT model to combat overfitting.

4. The IC fuse trimming code prediction method according to claim 3, characterized in that, Discriminative optimization of the modifier code in the predicted output of the GBDT model to combat overfitting includes: Record No. j The average trimming voltage corresponding to each discrete integer trimming code is: If there exists an interval such that the actual adjustment voltage for any sample... All can be adjusted by modifying the code. j or j+1 To achieve the target accuracy, this interval Defined as the initial modifier ambiguity region; set up Obtain the center point of the blurred region of the modulated code: ; Using the center point as a reference, take its upper and lower parts. The voltage range is used as the final modulated code ambiguity area: ; When the chip's current actual trimming voltage falls within the ambiguity region of the final trimming code, any adjacent trimming code is used to adjust the voltage within the target range.

5. An IC fuse trimming code prediction system, characterized in that, include: The sample module is used to obtain the test voltage before adjustment, the actual adjustment code, and the corresponding test voltage after adjustment as samples. The model building module, based on the GBDT framework, introduces a mechanism for randomly discarding residual samples and perturbing leaf node outputs to construct an overfitting GBDT model, including: The initial prediction score for each category is uniformly set to 0, the regression tree is initialized, and the softmax function is applied to obtain the initial prediction probability. Construct a loss function that can progressively enhance the ability to predict the correct category; Using the gradient boosting method, in the first... In each iteration, the residual is defined as the current error target for each sample for each class, which is used by the regression tree to learn and minimize the loss function; In each iteration, a regression tree is trained for each category to fit the residual value of that category, which is then used to correct the output value of the regression tree in subsequent iterations. During the training of each regression tree, the residual samples within the leaf nodes are processed as follows: For each regression tree, a portion of the residual samples in each leaf node are randomly discarded proportionally to obtain a retained sample set; Gaussian noise is added to the residual values ​​of the retained sample set; for each leaf node, the final output value is the mean of the retained samples and the residuals after adding Gaussian noise; after the regression trees are trained, the learning rate is used to control the update magnitude, updating the output values ​​of each regression tree, and a new prediction probability is calculated using the softmax function for the next round of residual calculation; after M iterations, a well-trained, overfit-resistant GBDT model is obtained. The prediction module is used to predict the modifier code for the overfit-resistant GBDT model trained on the application samples; among which, is the number of iterations; M is the maximum number of iterations.

6. The IC fuse trimming code prediction system according to claim 5, characterized in that, The output values ​​of each regression tree are updated using the following formula: ； ； In the formula, For the first During the nth iteration Class score, For learning rate, For the first Features of each sample For each regression tree category, Leaf node For the sample In category The current error target is used for regression tree learning; For the first Class 1 The cycle of the regression tree.

7. The IC fuse trimming code prediction system according to claim 5, characterized in that, A fuzzy region for the modifier code is constructed to optimize the modifier code predicted by the GBDT model to combat overfitting.

8. The IC fuse trimming code prediction system according to claim 7, characterized in that, Discriminative optimization of the modifier code in the predicted output of the GBDT model to combat overfitting includes: Record No. j The average trimming voltage corresponding to each discrete integer trimming code is: If there exists an interval such that the actual adjustment voltage for any sample... All can be adjusted by modifying the code. j or j +1 achieves the target precision, then this interval Defined as the initial modifier ambiguity region; set up Obtain the center point of the blurred region of the modulated code: ; Using the center point as a reference, take its upper and lower parts. The voltage range is used as the final modulated code ambiguity area: ; When the chip's current actual trimming voltage falls within the ambiguity region of the final trimming code, any adjacent trimming code is used to adjust the voltage within the target range.

Images