High thermal and electrical conductivity lig laser-induced machine learning regulation method

Abstract

This invention relates to the field of machine learning technology, and particularly to a machine learning control method for LIG laser-induced high thermal conductivity and high electrical conductivity performance. The method includes calculating the contribution weights of processing parameters to dual performance characteristics, identifying key dominant and auxiliary parameters, and generating a training dataset for both performance characteristics. Two machine learning models are trained separately to obtain initial core hyperparameters for the two prediction models, and then the core hyperparameters are tuned. The two prediction models process candidate datasets respectively, outputting thermal conductivity and electrical conductivity prediction sets, and a comprehensive evaluation is performed to obtain recommended laser processing parameters. The contribution weights of key related laser processing parameters are updated based on a comparison between the target dual performance characteristics and the measured dual performance characteristics. The recommended laser processing parameters are then trained and updated again. By training the dual performance relationship model using the contribution weights of processing parameters and optimizing hyperparameters from multiple aspects, global collaborative optimization is achieved, enabling rapid and accurate prediction of the dual performance characteristics of LIG thermal conductivity and electrical conductivity.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method for controlling machine learning induced by LIG lasers with high thermal conductivity and high electrical conductivity. Background Technology

[0002] Currently, the grounding layer in chip packaging faces limitations and challenges in materials, structure, and processes in the pursuit of efficient heat dissipation. Copper is the most commonly used grounding layer material. While copper is an excellent conductor, its thermal conductivity and heat diffusion capabilities present significant bottlenecks for chips with extremely high power densities. When dealing with instantaneous, pulsed power surges, the grounding layer cannot dissipate heat quickly, easily forming localized high-temperature zones. Graphene, as a typical two-dimensional material, possesses excellent thermal and electrical conductivity. Laser-induced graphene, as a novel multifunctional carbon material, has thermal and electrical conductivity as key indicators determining its performance in applications such as thermal management and flexible electronics. Currently, the fabrication of high-performance LIGs heavily relies on a trial-and-error approach to optimize laser processing parameters, which has the following prominent drawbacks: it requires numerous repetitive experiments to screen parameters, consuming huge resources; it is difficult to conduct a global search in a vast parameter space, easily getting trapped in local optima, failing to guarantee optimal performance; and traditional methods cannot clearly analyze the independent and synergistic influence mechanisms of various processing parameters on thermal and electrical conductivity, failing to provide theoretical guidance for performance trade-offs for specific applications.

[0003] Existing technologies include methods for determining laser processing parameters. For example, CN116502455A discloses a method and system for determining process parameters in laser selective melting (LIG) technology. This method trains an SVR model using process parameters and part mechanical properties as the original datasets. However, the variety of process parameters and mechanical properties makes it difficult to accurately determine the correspondence between them, resulting in insufficient precision for high-performance LIG fabrication. Another example is CN118520356A, which discloses an automatic prediction method for laser parameter combinations in laser-induced graphene. This method uses electromagnetic shielding effectiveness as the target input to a serial network to obtain the predicted laser parameter combination corresponding to the target electromagnetic shielding effectiveness data. This method predicts laser parameter combinations based on a single performance characteristic.

[0004] Therefore, it is necessary to provide a method that can efficiently and accurately obtain LIG laser-induced parameters with high thermal conductivity and high electrical conductivity. Summary of the Invention

[0005] The purpose of this invention is to propose a machine learning control method for LIG laser-induced high thermal conductivity and high electrical conductivity. By quantifying and distinguishing the contribution weights of processing parameters to performance, the method differentiates the key dominant parameters and auxiliary parameters, and combines weighted kernel functions and hyperparameter optimization to train the dual performance relationship model, thereby achieving global collaborative optimization and enabling rapid and accurate prediction of the dual performance of LIG thermal conductivity and electrical conductivity.

[0006] To achieve this objective, the present invention adopts the following technical solution:

[0007] A machine learning-based control method for LIG laser-induced high thermal and electrical conductivity lasers includes the following steps:

[0008] S1: Based on the thermal conductivity and electrical conductivity data of LIG samples obtained from different sets of laser processing parameters, construct thermal conductivity datasets and electrical conductivity datasets;

[0009] S2: Calculate the contribution weight of each processing parameter to thermal conductivity and electrical conductivity, classify the processing parameters into key dominant parameters and auxiliary parameters, and generate thermal conductivity training dataset and electrical conductivity training dataset.

[0010] S3: Construct weighted Gaussian kernel functions for the thermal conductivity prediction model and the electrical conductivity prediction model respectively, using the contribution weights of the key dominant parameters. Train the two support vector regression models using the thermal conductivity training dataset and the electrical conductivity training dataset respectively, and obtain the initial core hyperparameters of the thermal conductivity prediction model and the electrical conductivity prediction model.

[0011] S4: The core hyperparameters of the thermal conductivity prediction model and the electrical conductivity prediction model were tuned using the iterative search method and the five-fold cross-validation method, respectively.

[0012] S5: Generate a candidate dataset for laser processing parameters. Input the candidate dataset into the trained thermal conductivity prediction model and electrical conductivity prediction model respectively, and output the thermal conductivity prediction set and electrical conductivity prediction set. Perform a comprehensive evaluation on the data in the thermal conductivity prediction set and electrical conductivity prediction set to obtain recommended laser processing parameters.

[0013] S6: The thermal conductivity and electrical conductivity corresponding to the recommended laser processing parameters are the target dual-performance data. The thermal conductivity and electrical conductivity of the LIG actually processed with the recommended laser processing parameters are measured to obtain the measured dual-performance data. The target dual-performance data are compared with the measured dual-performance data. The measured dual-performance data are incorporated into the training dataset. The contribution weights of the key dominant parameters to thermal conductivity and electrical conductivity are updated. After updating the weight vector in the weighted Gaussian kernel function, the thermal conductivity prediction model and electrical conductivity prediction model are trained again.

[0014] Furthermore, in step S1, thermal conductivity dataset and electrical conductivity dataset are constructed based on laser processing experiments. Each set of data in the thermal conductivity dataset takes processing parameters as input and thermal conductivity as the output index. Each set of data in the electrical conductivity dataset takes processing parameters as input and electrical conductivity after logarithmic transformation as the output index.

[0015] The processing parameters include laser power, scanning speed and filling spacing, spot diameter and pulse frequency.

[0016] Furthermore, in step S2, the contribution weight of each processing parameter to thermal conductivity and electrical conductivity is calculated by analysis of variance. Among them, the contribution weight of laser power, scanning speed and filling spacing is greater than or equal to 15%, which are key dominant parameters, while spot diameter and pulse frequency are auxiliary parameters.

[0017] Furthermore, the method for calculating the contribution weights of processing parameters to thermal conductivity and electrical conductivity using analysis of variance is as follows:

[0018] Using a single parameter in the processing parameters as a factor, thermal conductivity and electrical conductivity are used as output variables respectively;

[0019] The mean square value and error mean square value are calculated based on the sum of squares of each factor. Then, the F value is calculated. Finally, the contribution of each factor to thermal conductivity and electrical conductivity are calculated separately. , among which, SS 因子 It is the sum of squares of the factors, SS 误差 It is the sum of squared errors. It is the total sum of squares of the factors;

[0020] The calculated laser power, scanning speed, and fill spacing all have significant weights in contributing to thermal conductivity and electrical conductivity, making laser power, scanning speed, and fill spacing the key processing parameters.

[0021] Furthermore, in step S3,

[0022] The contribution weights of the key dominant parameters obtained from analysis of variance to thermal conductivity or electrical conductivity are incorporated into the weighted Euclidean distance to construct a weighted Gaussian kernel function:

[0023] The contribution weights of key dominant parameters are used as weighting coefficients w=[w p ,w s ,w h ], where w p w s w h These are the original proportional values ​​of the contribution weights of laser power, scanning speed, and fill spacing, respectively.

[0024] The weighted distance is defined as:

[0025] ;

[0026] in, and These are the key dominant parameter vectors of the two input samples. and These represent the laser power values ​​of the two input samples, respectively. and These represent the scanning speed values ​​for the two input samples, respectively. and These represent the padding spacing values ​​for the input samples;

[0027] Constructing a weighted Gaussian kernel function:

[0028] ;

[0029] in, It is the bandwidth parameter of the kernel function, and the complete input feature vector of the model. Includes all processing parameters, i.e. ,in For the key dominant parameter vector, =[spot diameter, pulse frequency];

[0030] During training, RMSE was used as the core regression performance evaluation metric. The hyperparameters of the SVR model were the penalty factor C and the tolerance. ;

[0031] Where n is the total number of samples. For the model to the first The predicted value for each sample. For the first The true value of each sample;

[0032] Set the number of training iterations for the model, compare the RMSE values ​​of the trained models after different number of iterations, and select the training model with the smallest RMSE value.

[0033] Furthermore, the core hyperparameter tuning method in step S4 includes:

[0034] Kernel function parameters K, penalty factor C, and tolerance for thermal conductivity prediction models and electrical conductivity prediction models. Each iteration search is performed with a fixed step size. The prediction error is calculated using a five-fold cross-validation method in each iteration search. After the iteration search is completed, the optimal hyperparameter combination for electrical conductivity and the optimal hyperparameter combination for thermal conductivity are determined.

[0035] The mean absolute error (MAPE) was used as the quality index of the regression model to evaluate the trained thermal conductivity prediction model and electrical conductivity prediction model, respectively.

[0036] The five-fold cross-validation method is as follows:

[0037] The thermal conductivity training dataset and the electrical conductivity training dataset are each divided into five similar-sized, non-overlapping subsets.

[0038] Each of the five subsets is used as a validation set, and the remaining subsets are used as the training set. The SVR model is trained with the core hyperparameters obtained in S3. Then, the prediction error RMSE is calculated using the validation set to obtain five different prediction errors RMSE.

[0039] Calculate the average of the five prediction errors RMSE, and use the average RMSE to evaluate the performance of the current core hyperparameters.

[0040] Furthermore, step S5 includes:

[0041] The value range of key dominant parameters and several fixed values ​​of auxiliary parameters are preset. A candidate dataset of laser processing parameters is generated within the value range by using an ergonomic grid search algorithm. The complete candidate dataset is formed by connecting the auxiliary parameters.

[0042] Input the complete parameter candidate dataset containing key dominant parameters and auxiliary parameters into the thermal conductivity prediction model and electrical conductivity prediction model trained in step S4, respectively, and output the predicted value for each parameter candidate data point to obtain the thermal conductivity prediction set and electrical conductivity prediction set;

[0043] The predicted values ​​of the two performance metrics are evaluated using a comprehensive scoring function:

[0044] ;

[0045] in, It is the normalized thermal conductivity. It is the normalized conductivity;

[0046]

[0047]

[0048] in, This is a predicted value for thermal conductivity. A is the electrical conductivity value, A is the approximate upper limit of thermal conductivity, and B is the predicted upper limit of electrical conductivity.

[0049] All candidate parameter combinations are globally sorted based on their comprehensive performance scores, and the parameter combination with the highest comprehensive performance score is selected as the optimal solution for dual-performance collaboration.

[0050] Furthermore, step S6 includes:

[0051] S61: The recommended laser processing parameters and measured dual performance data are used as new samples and added to the thermal conductivity training dataset and electrical conductivity training dataset respectively to obtain the amplified thermal conductivity training dataset and amplified electrical conductivity training dataset. The contribution weight of each processing parameter to thermal conductivity and electrical conductivity is recalculated.

[0052] S62: Compare the changes in the contribution weights of each parameter before and after the update, reclassify the key dominant parameters and auxiliary parameters, and update the weight vector in the weighted Gaussian kernel function;

[0053] S63: Using the expanded thermal conductivity training dataset and the expanded electrical conductivity training dataset, the thermal conductivity prediction model and the electrical conductivity prediction model are trained respectively by updating the weighted Gaussian kernel function, in order to maintain the optimization of the support vector coefficients by the hyperparameters; the updated thermal conductivity prediction model and the electrical conductivity prediction model are used to process the candidate dataset again to obtain the updated recommended laser processing parameters;

[0054] S64: Determine the difference between the two recommended laser processing parameters. If the difference is less than the process allowable error, output the final optimized parameters. If the difference is greater than the process allowable error, repeat steps S61-S63.

[0055] A method for inducing LIG lasers with high thermal conductivity and high electrical conductivity includes the following steps:

[0056] The preset range of values ​​for processing power, scanning speed, and fill spacing is used to generate a candidate dataset of laser processing parameters within the range of values.

[0057] Input the candidate parameter datasets into the thermal conductivity prediction model and electrical conductivity prediction model obtained above, respectively, and output the predicted value for each candidate parameter data point to obtain the thermal conductivity prediction set and electrical conductivity prediction set;

[0058] All candidate parameter combinations are globally sorted based on the comprehensive performance score, and the parameter combination with the highest comprehensive performance score is selected as the optimal solution for dual performance collaboration.

[0059] By inputting the optimal solution for dual-performance synergy into the laser processing system, LIG ​​products are obtained by processing the substrate surface.

[0060] Furthermore, the substrate has a thickness of Polyimide paper, the resulting LIG product is used as a flexible grounding layer.

[0061] The technical solution provided by this invention may include the following beneficial effects:

[0062] This invention quantifies and distinguishes key dominant parameters from auxiliary parameters through analysis of variance, enabling the model to focus on core influencing factors. The constructed weighted Gaussian kernel function embeds the physical statistical weights of key parameters into the model structure, improving learning efficiency and prediction accuracy. A fixed or finite value search strategy is adopted for auxiliary parameters, significantly reducing the dimensionality and computational complexity of the optimization problem while ensuring process integrity. Simultaneously, iterative search and five-fold cross-validation are used to fine-tune the core hyperparameters, efficiently determining the optimal hyperparameters of the model. Combined with global scanning of the candidate dataset, the optimal process window can be quickly and accurately located, achieving laser-induced high thermal conductivity and high electrical conductivity LIGs. This invention also updates the contribution weights of processing parameters by comparing the target dual properties with measured dual properties, further improving the accuracy of laser processing parameter recommendations. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating a machine learning control method according to an embodiment of the present invention;

[0064] Figure 2 This is a graph showing the hyperparameter search results for different iterations of the thermal conductivity model in this invention.

[0065] Figure 3 This is a diagram illustrating the hyperparameter optimization process of the thermal conductivity prediction model in this invention.

[0066] Figure 4 This is a diagram illustrating the hyperparameter optimization process of the conductivity prediction model in this invention.

[0067] Figure 5 This is a fitting graph of the predicted and actual values ​​of LIG thermal and electrical conductivity for 100 sets of samples and 100 sets of actual experimental samples. Detailed Implementation

[0068] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the present invention.

[0069] An embodiment of the present invention provides a machine learning control method for LIG laser-induced high thermal conductivity and high electrical conductivity, comprising the following steps:

[0070] S1: Based on the thermal conductivity and electrical conductivity data of LIG samples obtained from different sets of laser processing parameters, construct thermal conductivity datasets and electrical conductivity datasets;

[0071] S2: Calculate the contribution weight of each processing parameter to thermal conductivity and electrical conductivity, classify the processing parameters into key dominant parameters and auxiliary parameters, and generate thermal conductivity training dataset and electrical conductivity training dataset.

[0072] S3: Construct weighted Gaussian kernel functions for the thermal conductivity prediction model and the electrical conductivity prediction model respectively, using the contribution weights of the key dominant parameters. Train the two support vector regression models using the thermal conductivity training dataset and the electrical conductivity training dataset respectively, and obtain the initial core hyperparameters of the thermal conductivity prediction model and the electrical conductivity prediction model.

[0073] S4: The core hyperparameters of the thermal conductivity prediction model and the electrical conductivity prediction model were tuned using the iterative search method and the five-fold cross-validation method, respectively.

[0074] S5: Generate a candidate dataset for laser processing parameters. Input the candidate dataset into the trained thermal conductivity prediction model and electrical conductivity prediction model respectively, and output the thermal conductivity prediction set and electrical conductivity prediction set. Perform a comprehensive evaluation on the data in the thermal conductivity prediction set and electrical conductivity prediction set to obtain recommended laser processing parameters.

[0075] S6: The thermal conductivity and electrical conductivity corresponding to the recommended laser processing parameters are the target dual-performance data. The thermal conductivity and electrical conductivity of the LIG actually processed with the recommended laser processing parameters are measured to obtain the measured dual-performance data. The target dual-performance data are compared with the measured dual-performance data. The measured dual-performance data are incorporated into the training dataset. The contribution weights of the key dominant parameters to thermal conductivity and electrical conductivity are updated. After updating the weight vector in the weighted Gaussian kernel function, the thermal conductivity prediction model and electrical conductivity prediction model are trained again.

[0076] This scheme quantifies the contribution weight of each process parameter, clearly identifying the most critical processing parameters affecting both thermal and electrical conductivity properties, providing a clear theoretical basis and data support for the targeted design and performance control of LIGs. Iterative search and five-fold cross-validation methods are used to fine-tune the core hyperparameters, efficiently determining the optimal hyperparameters for the model. Combined with global scanning of the candidate dataset, the optimal process window can be quickly and accurately located. This scheme obtains dual-property prediction datasets using both thermal and electrical conductivity prediction models, comprehensively evaluates and obtains recommended laser processing parameters, and comprehensively considers the impact of processing parameters on dual properties, thereby achieving laser-induced high thermal and electrical conductivity LIGs.

[0077] In another aspect of the invention, the contribution weights of the processing parameters are updated by comparing the target dual properties with the measured dual properties, thereby further improving the accuracy of laser processing parameter recommendations.

[0078] Specifically, in step S1, a thermal conductivity dataset and an electrical conductivity dataset are constructed based on the laser processing experiment. Each set of data in the thermal conductivity dataset uses processing parameters as input and thermal conductivity as the input. As the output metric, each data set in the conductivity dataset takes processing parameters as input and outputs the conductivity after logarithmic transformation. For output metrics;

[0079] The processing parameters include laser power, scanning speed and filling spacing, spot diameter and pulse frequency.

[0080] As an example, parameters for laser processing experiments are set based on experience, such as laser power of 1.325~12.96W, scanning speed of 100~300mm / s, and fill spacing of 0.01~0.03mm. The purpose of performing a logarithmic transformation on the conductivity is to convert it into a comparable range.

[0081] In step S2 of the present invention, the contribution weight of each processing parameter to thermal conductivity and electrical conductivity is calculated by analysis of variance. Among them, the contribution weight of laser power, scanning speed and filling spacing is greater than or equal to 15%, which are key dominant parameters, while spot diameter and pulse frequency are auxiliary parameters.

[0082] By accurately quantifying the contribution values ​​of processing parameters, a significant systematic correlation between input and output variables was pre-verified before being incorporated into a complex machine learning model. This laid a solid physical and statistical foundation for the subsequent prediction model. By comparing the variance analysis results of thermal conductivity and electrical conductivity, the specificity of each parameter's impact on different performance characteristics was revealed, thereby identifying key relevant processing parameters.

[0083] Specifically, the method for calculating the contribution weights of processing parameters to thermal conductivity and electrical conductivity using analysis of variance is as follows:

[0084] Each processing parameter was treated as a factor, with thermal conductivity and electrical conductivity as output variables. A multifactor ANOVA model was used to evaluate the significant impact of each factor on thermal conductivity and electrical conductivity. The model included the sum of squares (SS), degrees of freedom (df), mean square (MS), F-value, and P-value for each factor.

[0085] Calculate the mean square value and error mean square value using the sum of squares of each factor, and calculate the mean square value of the factors ( ) and the mean square value of the error ( The calculation formula is:

[0086]

[0087]

[0088] in, It is the sum of squares of the factors, reflecting the variation caused by different levels of the factors. This represents the degrees of freedom of the factor, typically the number of levels minus 1. It is the sum of squared errors, reflecting the variation caused by random errors. It represents the degrees of freedom of error, which is usually the total sample size minus the degrees of freedom occupied by the factors.

[0089] Then calculate the F-value. The larger the F-value, the more significant the factor's influence on the output variable. In analysis of variance (ANOVA), the P-value is used to determine whether a factor is significant. The P-value is determined by the F-statistic and its two degrees of freedom. MATLAB outputs precise P-values ​​after performing ANOVA. A significance level is typically set. =0.05. The effect of the factor is significant when P < 0.05. The effect of the factor is not significant when P ≥ 0.05.

[0090] Finally, the contributions of the factors to thermal conductivity and electrical conductivity were calculated separately: ,in, It is the sum of squares of the factors.

[0091] The calculated laser power, scanning speed, and fill spacing all have significant weights in contributing to thermal conductivity and electrical conductivity, making laser power, scanning speed, and fill spacing the key processing parameters.

[0092] In this embodiment of the invention, the machine learning model is an SVR model based on a weighted Gaussian kernel function. Traditional unweighted Gaussian kernel functions use standard Euclidean distance, which implicitly assumes that all input features have equal importance to the model output, failing to distinguish the differences in performance impact of various process parameters. The weighted Gaussian kernel function of this invention, through the aforementioned weighted distance, directionally amplifies the influence of parameters with high contribution weights (such as laser power) in similarity measurement, while suppressing fluctuations caused by parameters with low contribution weights (such as fill spacing). This results in higher resolution and sensitivity of the model along the key parameter dimensions in the feature space.

[0093] Specifically, in step S3, the contribution weights of the key dominant parameters obtained from variance analysis to thermal conductivity or electrical conductivity are introduced into the weighted Euclidean distance to construct a weighted Gaussian kernel function:

[0094] The contribution weights of key dominant parameters are used as weighting coefficients. ,in , , These are the original proportional values ​​of the contribution weights of laser power, scanning speed, and fill spacing, respectively.

[0095] The weighted distance is defined as:

[0096] ;

[0097] in, and These are the key dominant parameter vectors of the two input samples. and These represent the laser power values ​​of the two input samples, respectively. and These represent the scanning speed values ​​for the two input samples, respectively. and These represent the padding spacing values ​​for the input samples;

[0098] Constructing a weighted Gaussian kernel function:

[0099] ;

[0100] in, It is the bandwidth parameter of the kernel function, and the complete input feature vector of the model. Includes all processing parameters, i.e. ,in For the key dominant parameter vector, =[spot diameter, pulse frequency]; in the weighted Gaussian kernel function In the calculation, only the key dominant parameter vector is used. and Calculate the weighted distance based on its contribution weights, and use the auxiliary parameter vector. It does not participate in the similarity measurement calculation of the kernel function; its influence is learned and represented through global hyperparameters during the model training process.

[0101] This weighting mechanism produces the following specific effects during model training: From the initial stages of training, the model can focus on the core mapping relationship between key dominant parameters and performance, reducing the exploration of noise or secondary relationships, thus achieving lower prediction errors (e.g., RMSE) with fewer iterations. The model's decision-making process explicitly aligns with known physical and statistical laws, reducing the risk of erroneous generalization due to overfitting of secondary or random variations, and improving the model's predictive reliability on unknown parameter combinations. During hyperparameter tuning (e.g., iterative search) and global parameter optimization (e.g., grid search), the weighted kernel function guides the optimization algorithm to perform a fine search in high-potential regions dominated by key parameters, significantly improving global optimization efficiency.

[0102] During training, RMSE (Root Mean Square Error) is used as the core regression performance evaluation metric. The hyperparameters of the SVR model are the penalty factor C and the tolerance. ;

[0103] Where n is the total number of samples. For the model to the first The predicted value for each sample, For the first The true value of each sample.

[0104] The number of model training iterations is set, and the RMSE values ​​of the trained models after different number of iterations are compared. The training model with the smallest RMSE value is selected. The RMSE value intuitively represents the overall deviation between the model's predicted values ​​and the actual values. The smaller the RMSE, the higher the model's prediction accuracy and the stronger its robustness. This scheme uses Bayesian optimization to train the SVR model, realizing autonomous optimization of hyperparameter combinations. It overcomes the subjectivity and computational complexity of traditional parameter tuning methods, explicitly transforming the relationship between processing parameters and performance into a mathematical mapping. This provides a reliable digital twin platform for subsequent global optimization of process parameters, significantly reducing experimental verification costs. The complex nonlinear relationship between process parameters and material properties is successfully quantified into a predictive model with clear engineering accuracy, laying a theoretical foundation for the precise control of material properties.

[0105] The core hyperparameter tuning method in step S4 includes:

[0106] Kernel function parameters K, penalty factor C, and tolerance for thermal conductivity prediction models and electrical conductivity prediction models. Each iteration search is performed with a fixed step size. The prediction error is calculated using a five-fold cross-validation method in each iteration search. After the iteration search is completed, the optimal hyperparameter combination for electrical conductivity and the optimal hyperparameter combination for thermal conductivity are determined.

[0107] The five-fold cross-validation method is as follows:

[0108] The thermal conductivity training dataset and the electrical conductivity training dataset are each divided into five similar-sized, non-overlapping subsets.

[0109] Each of the five subsets is used as a validation set, and the remaining subsets are used as the training set. The SVR model is trained with the core hyperparameters obtained in S3. Then, the prediction error RMSE is calculated using the validation set to obtain five different prediction errors RMSE.

[0110] Calculate the average RMSE of the five prediction errors to evaluate the performance of the current core hyperparameters. This average RMSE reflects the model's stability and generalization ability across different subsets of data; a smaller value indicates better model performance and less susceptibility to overfitting.

[0111] Finally, the mean absolute error (MAPE) was used as the quality index of the regression model to evaluate the trained thermal conductivity prediction model and electrical conductivity prediction model.

[0112] As an example, 100 samples were selected from both the trained thermal conductivity prediction model and electrical conductivity prediction model to compare the deviations between the predicted and actual values ​​of LIG thermal conductivity and electrical conductivity with 100 actual experimental samples. The mean absolute error (MAPE) was used as the quality metric for the regression model.

[0113] .

[0114] Within a certain processing parameter space exploration range, this invention uses a prediction model of thermal conductivity and electrical conductivity to output a predicted dataset of thermal conductivity and electrical conductivity, and then uses the geometric mean method to select the optimal parameter points.

[0115] Step S5 includes:

[0116] The value range of key dominant parameters and several fixed values ​​of auxiliary parameters are preset. A candidate dataset of laser processing parameters is generated within the value range by using an ergonomic grid search algorithm. The complete candidate dataset is formed by connecting the auxiliary parameters.

[0117] Input the complete parameter candidate dataset containing key dominant parameters and auxiliary parameters into the thermal conductivity prediction model and electrical conductivity prediction model trained in step S4, respectively, and output the predicted value for each parameter candidate data point to obtain the thermal conductivity prediction set and electrical conductivity prediction set;

[0118] The predicted values ​​of the two performance metrics are evaluated using a comprehensive scoring function:

[0119] ;

[0120] in, It is the normalized thermal conductivity. It is the normalized conductivity;

[0121]

[0122]

[0123] in, This is a predicted value for thermal conductivity. A is the electrical conductivity value, A is the approximate upper limit of thermal conductivity, and B is the predicted upper limit of electrical conductivity.

[0124] All candidate parameter combinations are globally sorted based on their comprehensive performance scores, and the parameter combination with the highest comprehensive performance score is selected as the optimal solution for dual-performance collaboration.

[0125] In this embodiment of the invention, a candidate dataset of laser processing parameters is generated within a range of values ​​using an ergonomic grid search algorithm. Each candidate data point in the dataset contains laser power, scanning speed, and fill spacing. Therefore, dense, gridded sampling is performed within a preset parameter space, ultimately generating a candidate set containing tens to hundreds of thousands of virtual parameter combinations (P, V, S). This invention, combined with global scanning in a three-dimensional parameter space, enables rapid and accurate location of the optimal process window.

[0126] In one embodiment of the present invention, step S6 includes S61-S64.

[0127] S61: The recommended laser processing parameters and measured dual performance data are used as new samples and added to the thermal conductivity training dataset and electrical conductivity training dataset respectively to obtain the amplified thermal conductivity training dataset and amplified electrical conductivity training dataset. The contribution weight of each processing parameter to thermal conductivity and electrical conductivity is recalculated using analysis of variance.

[0128] S62: Compare the changes in the contribution weights of each parameter before and after the update, reclassify the key dominant parameters and auxiliary parameters, and update the weight vector in the weighted Gaussian kernel function. Specifically, if the updated contribution weight of an auxiliary parameter (spot diameter, pulse frequency) exceeds a set threshold (e.g., 15%), it is upgraded to a key dominant parameter; conversely, if the contribution weight of the original key dominant parameter drops below the threshold, it is downgraded to an auxiliary parameter. Generally, parameter classification remains stable. Extract the weight values ​​of the key dominant parameters from the complete weight vector and update the weight coefficients of the weighted Gaussian kernel function. The weight changes of auxiliary parameters are used for process stability monitoring but do not directly participate in the construction of the weighted kernel function.

[0129] S63: Using expanded thermal conductivity and electrical conductivity training datasets, the thermal conductivity prediction model and electrical conductivity prediction model are trained respectively using an updated weighted Gaussian kernel function. The training method optimizes the support vector coefficients while preserving the hyperparameters, i.e., preserving the model's hyperparameters, such as kernel function parameters K, penalty factor C, and tolerance. Without changing the parameters, only the support vector coefficients are re-optimized with fewer iterations to avoid overfitting. The updated thermal conductivity and electrical conductivity prediction models are then used to process the candidate dataset again to obtain updated recommended laser processing parameters.

[0130] S64: Determine the difference between the two obtained recommended laser processing parameters. If the difference is less than the process tolerance, output the final optimized parameters. If the difference is greater than the process tolerance, repeat steps S61-S63. Specifically, a difference less than the process tolerance means that the numerical differences between the new round of recommended parameters and the previous round of recommended parameters in laser power, scanning speed, and fill spacing are less than a pre-set error threshold, such as a laser power tolerance of 0.2W. The repetition of step S6 treats the new round of recommended parameters generated after the model update as a new, unverified process scheme, and re-performs the physical verification process of "experimental preparation—performance measurement—comparative evaluation," forming a closed-loop iterative optimization process.

[0131] Accordingly, the present invention also provides a method for inducing LIG lasers with high thermal conductivity and high electrical conductivity, comprising the following steps:

[0132] The preset range of values ​​for processing power, scanning speed, and fill spacing is used to generate a candidate dataset of laser processing parameters within the range of values.

[0133] Input the candidate parameter datasets into the thermal conductivity prediction model and electrical conductivity prediction model obtained above, respectively, and output the predicted value for each candidate parameter data point to obtain the thermal conductivity prediction set and electrical conductivity prediction set;

[0134] All candidate parameter combinations are globally sorted based on the comprehensive performance score, and the parameter combination with the highest comprehensive performance score is selected as the optimal solution for dual performance collaboration.

[0135] By inputting the optimal solution for dual-performance synergy into the laser processing system, LIG ​​products are obtained by processing the substrate surface.

[0136] As an example, the substrate has a thickness of Polyimide paper, the resulting LIG product is used as a flexible grounding layer.

[0137] The polyimide paper-based LIG product obtained by the above method has high thermal conductivity and high electrical conductivity, making it suitable for flexible grounding electrodes and applicable to products such as flexible displays, wearable devices (such as smartwatch wristbands, smart glasses temples, and smart clothing), flexible sensors, and system-in-package structures.

[0138] The following specific embodiments further illustrate the machine learning control method induced by high thermal conductivity and high electrical conductivity LIG lasers according to the present invention.

[0139] S1: With laser power Scanning speed The fill spacing is 0.01, 0.02, and 0.03 mm, based on thermal conductivity. Japanese Classics Transformed conductivity As output metrics, a dataset of thermal and electrical conductivity values ​​for samples processed using different processing parameters was collected through 100 laser processing experiments. Each of the 100 experiments corresponds to 100 thermal conductivity values ​​and 100 electrical conductivity values. No normalization is required for the data in this dataset.

[0140] S2: Analysis of variance was used to screen the key processing parameters related to the thermal and electrical conductivity of LIG. The results showed that laser power, scanning speed, and fill spacing contributed 60.2%, 20.5%, and 19.3% to thermal conductivity, respectively, and 45.8%, 27.3%, and 26.9% to electrical conductivity, respectively.

[0141] S3: Using thermal conductivity and electrical conductivity training datasets, support vector machine (SVR) models were trained separately using Bayesian optimization. The number of iterations for different training models was investigated, referring to... Figure 2 The training models with 50, 100, 150, and 200 iterations were compared. Among these four iteration counts, for the thermal conductivity prediction model, the training model with 70 iterations had the smallest RMSE value. Therefore, the result of the training model after 70 iterations was selected as the initial result for the hyperparameter search. For the electrical conductivity prediction model, the optimal number of iterations was 140. Initial values ​​for the hyperparameters were found: for thermal conductivity C = 5.3174, =0.02196, for conductivity: C=7.4859, .

[0142] S4: The core hyperparameters were tuned using an iterative search method and a five-fold cross-validation method.

[0143] The method of five-fold cross-validation is as follows:

[0144] First, the 100 datasets are randomly and evenly divided into 5 mutually exclusive subsets of similar size, with 20 datasets in each subset. Then, 5 rounds of cross-validation are performed.

[0145] Round 1: Select subset 1 as the validation set, and the remaining 4 subsets (subsets) Combine them into the training set. Use the hyperparameter combination of the current iteration. Train an SVR model on the training set, then calculate the prediction error using the validation set to obtain a... .

[0146] Round 2: Select subset 2 as the validation set, and the remaining subset (1, 3, 4, 5) as the training set. Train the model again, and compute on the validation set. .

[0147] Repeat this process until each subset has been used as the validation set exactly once. A total of 5 RMSE values ​​are obtained. arrive Calculate the average of these 5 RMSE values ​​and use it as the performance evaluation score for the current hyperparameter combination.

[0148] During the iterative search, five-fold cross-validation is performed when scanning each core hyperparameter to obtain the average RMSE value corresponding to that parameter.

[0149] Reference Figure 3 For the thermal conductivity prediction model, the kernel function parameter K, penalty factor C, and tolerance are respectively determined. Optimization:

[0150] Optimization of the kernel function parameter K: Perform 101 iterations within the range of 0 to 10 with a step size of 0.1. The minimum value RMSE (2.171) was found when K=1.6. Therefore, K=1.6 is determined to be the optimal value.

[0151] Optimization of the penalty factor C: Perform 1001 iterations within the range of 0 to 100 with a step size of 0.1. The minimum value (RMSE) of 1.633 was found when C = 85.7. Therefore, C = 85.7 is determined to be the optimal value.

[0152] Tolerance Optimization: Perform 101 iterative searches within the range of 0 to 10 with a step size of 0.1. The minimum value RMSE was found to be 2.597. Therefore, the result is determined. This is the optimal value.

[0153] Similarly, refer to Figure 4 The conductivity prediction model was modified by adjusting the kernel function parameters K, penalty factor C, and tolerance. After optimization and iterative search, the optimal hyperparameter combination for conductivity was determined to be: kernel function parameter K=3.9, penalty factor C=56.8, tolerance... .

[0154] Based on the hyperparameter optimization and adjustment results of the above thermal conductivity prediction model and electrical conductivity prediction model, referring to Figure 5 One hundred samples were selected from the trained thermal and electrical conductivity models to compare the predicted and actual values ​​of LIG thermal and electrical conductivity with 100 actual experimental samples. The MAPE of the thermal conductivity prediction model was [value missing]. In the conductivity prediction model, MAPE is... This indicates that the two models are in an ideal fit state.

[0155] S5: The range of processing parameter space exploration is selected as follows: laser power (P) from 1.325W to 12.96W, scanning speed (V) is... to The scanning interval (S) is 0.01 mm, 0.02 mm, or 0.03 mm; dense, gridded sampling is performed within this parameter space. A candidate set containing tens of thousands to hundreds of thousands of virtual parameter combinations (P, V, S) is generated, namely the (laser power, scanning speed, scanning interval) candidate set.

[0156] The candidate dataset is input into the trained thermal conductivity prediction model and electrical conductivity prediction model, respectively. The thermal conductivity model outputs a predicted thermal conductivity value for each point. The conductivity model outputs a predicted conductivity value for each point. .

[0157] Map the predicted values ​​of both performance metrics to a comparable range:

[0158]

[0159]

[0160] 25 is slightly higher than the maximum value measured in the dataset of 100 thermal conductivity samples. A regularized approximate upper bound. All resistance values ​​in the dataset were... Transformation, the overall value after the transformation is Similarly, within the range, the ideal performance target value for LIG conductivity is 6.

[0161] All candidate parameter combinations are globally sorted based on their comprehensive performance scores, and the parameter combination with the highest comprehensive performance score is selected as the optimal solution for dual-performance collaboration: Power. ,speed ,spacing .

[0162] S6: Experimentally verified that the process parameters for the globally optimal thermal conductivity and electrical conductivity points selected are: laser power. Scanning speed Scanning spacing The predicted thermal conductivity is The predicted conductivity is Experiments were conducted under these processing parameters, and the corresponding actual thermal conductivity and electrical conductivity were measured and calculated. By comparison, the error between the predicted thermal conductivity and the experimentally measured value was [missing value]. The error between the predicted and experimentally measured values ​​of conductivity is... .

[0163] Step S6 confirmed the high accuracy and generalization ability of the SVR model in predicting the thermal and electrical conductivity of LIG; it also verified that the constructed multi-objective comprehensive scoring function can effectively quantify and synergistically optimize multiple performance indicators, achieving a precise mapping from theoretical optimality to actual manufacturability. This demonstrates that the process points determined by this framework possess high engineering feasibility and stability.

[0164] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0165] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0166] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a multimedia terminal device (which may be a mobile phone, computer, television receiver, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0167] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A machine learning control method induced by a LIG laser with high thermal conductivity and high electrical conductivity, characterized in that, Includes the following steps: S1: Based on the thermal conductivity and electrical conductivity data of LIG samples obtained from different sets of laser processing parameters, construct thermal conductivity datasets and electrical conductivity datasets; S2: Calculate the contribution weight of each processing parameter to thermal conductivity and electrical conductivity, classify the processing parameters into key dominant parameters and auxiliary parameters, and generate thermal conductivity training dataset and electrical conductivity training dataset. S3: Construct weighted Gaussian kernel functions for the thermal conductivity prediction model and the electrical conductivity prediction model respectively, using the contribution weights of the key dominant parameters. Train the two support vector regression models using the thermal conductivity training dataset and the electrical conductivity training dataset respectively, and obtain the initial core hyperparameters of the thermal conductivity prediction model and the electrical conductivity prediction model. S4: The core hyperparameters of the thermal conductivity prediction model and the electrical conductivity prediction model were tuned using the iterative search method and the five-fold cross-validation method, respectively. S5: Generate a candidate dataset for laser processing parameters. Input the candidate dataset into the trained thermal conductivity prediction model and electrical conductivity prediction model respectively, and output the thermal conductivity prediction set and electrical conductivity prediction set. Perform a comprehensive evaluation on the data in the thermal conductivity prediction set and electrical conductivity prediction set to obtain recommended laser processing parameters. S6: The thermal conductivity and electrical conductivity corresponding to the recommended laser processing parameters are the target dual-performance data. The thermal conductivity and electrical conductivity of the LIG actually processed with the recommended laser processing parameters are measured to obtain the measured dual-performance data. The target dual-performance data are compared with the measured dual-performance data. The measured dual-performance data are incorporated into the training dataset. The contribution weights of the key dominant parameters to the thermal conductivity and electrical conductivity are updated. After updating the weight vector in the weighted Gaussian kernel function, the thermal conductivity prediction model and electrical conductivity prediction model are trained again.

2. The method according to claim 1, characterized in that, In step S1, thermal conductivity dataset and electrical conductivity dataset are constructed based on laser processing experiments. Each set of data in the thermal conductivity dataset takes processing parameters as input and thermal conductivity as the output index. Each set of data in the electrical conductivity dataset takes processing parameters as input and electrical conductivity after logarithmic transformation as the output index. The processing parameters include laser power, scanning speed and filling spacing, spot diameter and pulse frequency.

3. The method according to claim 2, characterized in that, In step S2, the contribution weight of each processing parameter to thermal conductivity and electrical conductivity is calculated by analysis of variance. Among them, the contribution weight of laser power, scanning speed and filling spacing is greater than or equal to 15%, which are the key dominant parameters, while spot diameter and pulse frequency are auxiliary parameters.

4. The method according to claim 3, characterized in that, The method for calculating the contribution weights of processing parameters to thermal conductivity and electrical conductivity using analysis of variance is as follows: Using a single parameter in the processing parameters as a factor, thermal conductivity and electrical conductivity are used as output variables respectively; Calculate the mean square value and error mean square value using the sum of squares of each factor, and then calculate the F value. Finally, calculate the contribution of each factor to thermal conductivity and electrical conductivity respectively. , among which, SS 因子 It is the sum of squares of the factors, SS 误差 It is the sum of squared errors. It is the total sum of squares of the factors; The calculated laser power, scanning speed, and fill spacing all have significant weights in contributing to thermal conductivity and electrical conductivity, making laser power, scanning speed, and fill spacing the key processing parameters.

5. The method according to claim 1, characterized in that, In step S3, the contribution weights of the key dominant parameters obtained from variance analysis to thermal conductivity or electrical conductivity are introduced into the weighted Euclidean distance to construct a weighted Gaussian kernel function: The contribution weights of key dominant parameters are used as weighting coefficients w=[w p ,w s ,w h ], where w p w s w h These are the original proportional values ​​of the contribution weights of laser power, scanning speed, and fill spacing, respectively. The weighted distance is defined as: ； in, and These are the key dominant parameter vectors of the two input samples. and These represent the laser power values ​​of the two input samples, respectively. and These represent the scanning speed values ​​for the two input samples, respectively. and These represent the padding spacing values ​​for the input samples; Constructing a weighted Gaussian kernel function: ； in, It is the bandwidth parameter of the kernel function, and the complete input feature vector of the model. Includes all processing parameters, i.e. =[ ; ],in For the key dominant parameter vector, =[spot diameter, pulse frequency]; During training, RMSE was used as the core regression performance evaluation metric. The hyperparameters of the SVR model were the penalty factor C and the tolerance. ; Where n is the total number of samples, Predicted i This represents the model's prediction for the i-th sample. i This represents the true value of the i-th sample. Set the number of training iterations for the model, compare the RMSE values ​​of the trained models after different number of iterations, and select the training model with the smallest RMSE value.

6. The method according to claim 5, characterized in that, The core hyperparameter tuning method in step S4 includes: Kernel function parameters K, penalty factor C, and tolerance for thermal conductivity prediction models and electrical conductivity prediction models. Each iteration search is performed with a fixed step size. The prediction error is calculated using a five-fold cross-validation method in each iteration search. After the iteration search is completed, the optimal hyperparameter combination for electrical conductivity and the optimal hyperparameter combination for thermal conductivity are determined. The mean absolute error (MAPE) was used as the quality index of the regression model to evaluate the trained thermal conductivity prediction model and electrical conductivity prediction model, respectively. The five-fold cross-validation method is as follows: The thermal conductivity training dataset and the electrical conductivity training dataset are each divided into five similar-sized, non-overlapping subsets. Each of the five subsets is used as a validation set, and the remaining subsets are used as the training set. The SVR model is trained with the core hyperparameters obtained in S3. Then, the prediction error RMSE is calculated using the validation set to obtain five different prediction errors RMSE. Calculate the average of the five prediction errors RMSE, and use the average RMSE to evaluate the performance of the current core hyperparameters.

7. The method according to claim 3, characterized in that, Step S5 includes: The value range of key dominant parameters and several fixed values ​​of auxiliary parameters are preset. A candidate dataset of laser processing parameters is generated within the value range by using an ergonomic grid search algorithm. The complete candidate dataset is formed by connecting the auxiliary parameters. Input the complete candidate dataset containing key dominant parameters and auxiliary parameters into the thermal conductivity prediction model and electrical conductivity prediction model trained in step S4, respectively, and output the predicted value for each candidate data point to obtain the thermal conductivity prediction set and electrical conductivity prediction set; The predicted values ​​of the two performance metrics are evaluated using a comprehensive scoring function: ； in, It is the normalized thermal conductivity. It is the normalized conductivity; ； ； in, This is a predicted value for thermal conductivity. A is the electrical conductivity value, A is the approximate upper limit of thermal conductivity, and B is the predicted upper limit of electrical conductivity. All candidate parameter combinations are globally sorted based on their comprehensive performance scores, and the parameter combination with the highest comprehensive performance score is selected as the optimal solution for dual-performance collaboration.

8. The method according to claim 1, characterized in that, Step S6 includes: S61: The recommended laser processing parameters and measured dual performance data are used as new samples and added to the thermal conductivity training dataset and electrical conductivity training dataset respectively to obtain the amplified thermal conductivity training dataset and amplified electrical conductivity training dataset. The contribution weight of each processing parameter to thermal conductivity and electrical conductivity is recalculated. S62: Compare the changes in the contribution weights of each parameter before and after the update, reclassify the key dominant parameters and auxiliary parameters, and update the weight vector in the weighted Gaussian kernel function; S63: Using the expanded thermal conductivity training dataset and the expanded electrical conductivity training dataset, the thermal conductivity prediction model and the electrical conductivity prediction model are trained respectively by updating the weighted Gaussian kernel function, in order to maintain the optimization of the support vector coefficients by the hyperparameters; the updated thermal conductivity prediction model and the electrical conductivity prediction model are used to process the candidate dataset again to obtain the updated recommended laser processing parameters; S64: Determine the difference between the two recommended laser processing parameters. If the difference is less than the process allowable error, output the final optimized parameters. If the difference is greater than the process allowable error, repeat steps S61-S63.

9. A method for inducing LIG lasers with high thermal conductivity and high electrical conductivity, characterized in that, Includes the following steps: The preset range of values ​​for processing power, scanning speed, and fill spacing is used to generate a candidate dataset of laser processing parameters within the range of values. Input the candidate parameter datasets into the thermal conductivity prediction model and electrical conductivity prediction model obtained by any one of claims 1-8, respectively, and output the predicted value for each candidate parameter data point to obtain the thermal conductivity prediction set and electrical conductivity prediction set; All candidate parameter combinations are globally sorted based on the comprehensive performance score, and the parameter combination with the highest comprehensive performance score is selected as the optimal solution for dual performance collaboration. By inputting the optimal solution for dual-performance synergy into the laser processing system, LIG ​​products are obtained by processing the substrate surface.

10. The method according to claim 9, characterized in that, The substrate is a 50-150 μm thick polyimide paper, and the resulting LIG product is used as a flexible grounding layer.

Images