Laser-induced electrochemical deposition process optimization method and system based on machine learning

Abstract

The invention provides a laser-induced electrochemical deposition process optimization method and system based on machine learning, and the method comprises the steps: collecting historical process parameters of a laser-induced electrochemical deposition process and corresponding deposition effect data, and constructing a training set; training a machine learning model based on the training set, and constructing a performance prediction model for the mapping relationship between the process parameters and the deposition effect; and utilizing the performance prediction model to predict the deposition effect of the target process parameters, taking the expected deposition effect as an optimization target, and utilizing an optimization algorithm to search and reversely derive an optimal process parameter combination in the performance prediction model. According to the technical scheme, the process parameters are reversely solved through the model, and the experiment trial and error cost is greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of laser microfabrication technology, specifically relating to a method and system for optimizing laser-induced electrochemical deposition process based on machine learning. Background Technology

[0002] Laser-induced localized electrochemical deposition (LAD) is an advanced microfabrication method that has shown great potential in microelectronics and MEMS. This method achieves rapid, localized deposition of metal structures through the coupling of the thermal effect of the laser with the electrochemical field. However, its deposition quality and effect are complexly influenced by multiple factors, including laser parameters such as power, scanning speed, and pulse frequency; electrochemical parameters such as current density and pulse duty cycle; and environmental parameters such as electrolyte concentration and temperature. These parameters exhibit strong coupling and nonlinear interactions, making it difficult to accurately obtain the process window through traditional trial-and-error methods or experience-based adjustments. This results in significant problems such as poor deposition consistency, unstable yield, and time-consuming and laborious optimization processes, severely hindering its industrial application and reliability improvement.

[0003] To optimize complex manufacturing processes, machine learning technology has been introduced into the industrial field in recent years. Machine learning models, such as neural networks, random forests, and support vector machines, can learn complex nonlinear mapping relationships from historical data and are hailed as powerful tools for solving multivariate optimization problems. By learning from large amounts of process data, they can establish predictive models between process parameters and performance indicators, thereby achieving intelligent recommendation and optimization of process parameters, potentially significantly reducing the number of experiments and lowering R&D costs. However, directly applying machine learning to the specific scenario of laser-induced electrochemical deposition still faces significant challenges. Most research focuses on the general algorithms themselves, failing to specifically design models and feature engineering tailored to the physicochemical characteristics of this process, resulting in limited model prediction accuracy. Acquiring high-quality experimental data for this process is costly; therefore, how to build high-precision models under small sample conditions and enhance the physical interpretability of the model's decision-making process are urgent problems to be solved.

[0004] Therefore, there is an urgent need for a machine learning-based method and system for optimizing laser-induced electrochemical deposition processes, which can deeply integrate process mechanisms and data-driven models. This would not only enable high-precision prediction of deposition effects but also allow for intelligent reverse calculation and optimization of process parameters, thereby completely solving the problems of blindness and uncertainty in the development of this process. Summary of the Invention

[0005] This invention provides a machine learning-based method and system for optimizing laser-induced electrochemical deposition (LAD) processes. This addresses the problems in LAD localized electrochemical deposition, such as reliance on expert experience, high trial-and-error costs, and difficulty in precisely locking the process window due to complex process parameters and strong coupling. The invention utilizes multiple sets of training data input into a machine learning model to obtain a performance prediction model. Unknown laser parameters are then input into the trained model to predict processing performance. By inversely calculating process parameters from the model, experimental trial-and-error costs are significantly reduced.

[0006] To achieve the above objectives, the present invention provides the following solution: A machine learning-based method for optimizing laser-induced electrochemical deposition processes includes: Historical process parameters and corresponding deposition effect data of laser-induced electrochemical deposition were collected to construct a training set. A machine learning model is trained based on the training set to construct a performance prediction model for the mapping relationship between process parameters and deposition effect. The deposition effect of the target process parameters is predicted using the performance prediction model, and the desired deposition effect is used as the optimization target. The optimal combination of process parameters is then derived by searching the performance prediction model using an optimization algorithm.

[0007] Preferably, the historical process parameters include laser parameters, electrochemical parameters, and environmental parameters; The deposition effect data includes deposition morphology, deposition rate, and dimensional accuracy.

[0008] Preferably, the machine learning model includes the random forest algorithm and the XGBoost algorithm. The selection of the machine learning model is based on the characteristics of the process parameters and user requirements. The method for constructing a performance prediction model using the random forest algorithm includes: Using the autonomous sampling set of the training set, each decision tree of the random forest algorithm is trained, and a feature subset is selected from all process parameters for optimal splitting to obtain the performance prediction results of each decision tree. The performance prediction results of each decision tree are aggregated using the bagging method to obtain the final performance prediction results, thus completing the construction of the performance prediction model.

[0009] Preferably, the optimization algorithm integrates a set of Bayesian optimization algorithms and genetic algorithms, and combines cross-validation to optimize the hyperparameters of the performance prediction model, thereby deriving the optimal combination of process parameters; wherein, the hyperparameters include the number and maximum depth of decision trees in the random forest; and the learning rate and maximum depth of decision trees in XGBoost.

[0010] This invention also provides a machine learning-based laser-induced electrochemical deposition process optimization system for implementing the method, comprising: The data acquisition module is used to collect historical process parameters and corresponding deposition effect data of laser-induced electrochemical deposition process to build a training set; The model building module is used to train a machine learning model based on the training set and build a performance prediction model for the mapping relationship between process parameters and deposition effect. The performance prediction module is used to predict the deposition effect of the target process parameters using the performance prediction model, and to use the expected deposition effect as the optimization target to search the performance prediction model using an optimization algorithm to deduce the optimal combination of process parameters.

[0011] Preferably, the data acquisition module includes: Historical process parameter acquisition unit, used to collect historical laser parameters, electrochemical parameters and environmental parameters; The deposition effect acquisition unit is used to acquire deposition morphology, deposition rate and dimensional accuracy corresponding to historical process parameters.

[0012] A further implementation involves the model building module, where the machine learning model includes both a random forest algorithm and an XGBoost algorithm. The selection of the machine learning model is based on the characteristics of the process parameters and user requirements. The method for constructing a performance prediction model using the random forest algorithm includes: Using the autonomous sampling set of the training set, each decision tree of the random forest algorithm is trained, and a feature subset is selected from all process parameters for optimal splitting to obtain the performance prediction results of each decision tree. The performance prediction results of each decision tree are aggregated using the bagging method to obtain the final performance prediction results, thus completing the construction of the performance prediction model.

[0013] Preferably, in the performance prediction module, the optimization algorithm integrates a set of Bayesian optimization algorithms and genetic algorithms, and combines cross-validation to optimize the hyperparameters of the performance prediction model, thereby deriving the optimal combination of process parameters; wherein, the hyperparameters include the number and maximum depth of decision trees in the random forest; and the learning rate and maximum depth of decision trees in XGBoost.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention is the first to apply machine learning algorithms to solve the problem of optimizing complex parameters in laser-induced electrochemical deposition processes.

[0015] 2. This invention transforms the traditional "trial and error" experience-based optimization model into a data-driven intelligent optimization model, which significantly reduces process development costs and shortens the R&D cycle.

[0016] 3. This invention can uncover complex parameter interactions and the global optimal process window that are difficult to discover through human experience, significantly improving the consistency and reliability of deposition results.

[0017] 4. This invention, through its parameter inverse calculation function, enables even non-experts in the field to quickly obtain high-quality processing results, thus lowering the technical threshold. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the user interface according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the device structure of an embodiment of the present invention.

[0020] Figure reference numerals: 1-Computer; 2-Laser; 3-Galvanometer system; 4-Focusing lens; 5-Electrolytic cell; 6-Anode; 7-Cathode; 8-3D moving platform; 9-CCD camera; 10-Temperature sensor; 11-Power supply. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1 like Figure 1 The method for optimizing laser-induced electrochemical deposition process based on machine learning is shown, including: S1: Collect historical process parameters and corresponding deposition effect data of laser-induced electrochemical deposition process to construct a training set. A further implementation method involves including historical process parameters such as laser parameters, electrochemical parameters, and environmental parameters. Deposition performance data includes deposition morphology, deposition rate, and dimensional accuracy.

[0024] S2: Train a machine learning model based on the training set to build a performance prediction model for the mapping relationship between process parameters and deposition effects. The machine learning algorithm completes the learning task by building and combining multiple weak learners to achieve significantly better generalization performance and prediction accuracy than a single model.

[0025] A further implementation method involves using machine learning models including random forest and XGBoost algorithms. The selection of the machine learning model is based on the characteristics of the process parameters and user requirements. Specifically, the method for constructing a performance prediction model using the random forest algorithm includes: Using the bootstrap sample from the training set, each decision tree of the random forest algorithm is trained, and a feature subset is selected from all process parameters for optimal splitting to obtain the performance prediction results of each decision tree. During the model training phase, a quality weight factor is assigned to each training sample. w i This factor is positively correlated with the stability and quality of the deposition process represented by the sample.

[0026] in, S i C is the process stability index corresponding to this sample. i It is the consistency coefficient corresponding to this sample. α and β This is a weighting coefficient that adjusts the importance of the two indicators. When constructing each decision tree, the criteria for node splitting will consider the sample weights. Ultimately, when making voting predictions in the forest, the votes for each tree will be weighted according to the model confidence obtained from training on this set of weighted samples, rather than simply one tree, one vote. Its prediction function can be expressed as: in, The total number of decision trees, Indicates the first Trees for input process parameter vectors The predicted output.

[0027] The performance prediction results of each decision tree are aggregated using the bagging method to obtain the final performance prediction result, thus completing the construction of the performance prediction model. This improved random forest algorithm transforms from a general data learner into a learner that focuses more on high-quality and stable deposition process data in the LECD process. This makes the model's prediction results more inclined towards stable and reliable process regions, significantly improving the robustness and safety of the model in actual control.

[0028] S3: Use the performance prediction model to predict the deposition effect of the target process parameters, and use the expected deposition effect as the optimization target. Use the optimization algorithm to search in the performance prediction model and deduce the optimal combination of process parameters.

[0029] The XGBoost algorithm also integrates multiple decision trees, but these trees are built sequentially, with each tree learning and correcting the prediction residuals of the previous tree. It employs an additive model and a forward stepwise algorithm. The model's objective function includes not only a loss function measuring prediction accuracy but also a regularization term to control model complexity and prevent overfitting. Its objective function can be expressed as: in, It is a loss function. It is the first The regularization term for the tree, L P It is an introduced physical rule penalty item. These are physical constraint weighting coefficients, used to balance data fitting and physical consistency. L P Two constraint mechanisms can be used, namely energy constraint. and boundary constraints Based on the setting that the deposition rate should monotonically increase with increasing laser power and approach zero when the power is zero, the design... Based on the fact that the deposition rate cannot be negative, the design... : Where P represents the actual laser power, This represents the predicted deposition rate. The objective function is minimized using gradient descent, with new trees added incrementally to fit the negative gradient (i.e., the residual approximation). The final prediction is a weighted sum of all tree predictions, expressed as: in, Represents an independent decision tree. This is the function space consisting of all possible decision trees. This improvement combines the advantages of data-driven and physical model-driven approaches. Introducing physical constraints prevents the model from making predictions that violate physical laws, improving the model's generalization ability. In regions with scarce training data or high noise, physical rules can play a guiding role, enhancing the model's reliability.

[0030] A further implementation method involves integrating Bayesian optimization and genetic algorithms into an optimization algorithm. The training dataset is divided into training and validation sets. Combined with cross-validation, the genetic algorithm performs global training to find a simple model while ensuring process stability. Based on this, Bayesian optimization is then used to precisely optimize the predicted morphology quality, thereby deriving the optimal combination of process parameters. The hyperparameters include the number and maximum depth of decision trees in the random forest, and the learning rate and maximum depth of decision trees in XGBoost.

[0031] This embodiment provides specific implementation steps for utilizing different machine learning models.

[0032] Methods for optimizing laser-induced electrochemical deposition processes using the XGBoost algorithm include: Collect 100 sets of previous laser-induced electrochemical deposition experimental data from the laboratory. Input features include: laser power, scanning speed, pulse frequency, and number of scans. Output labels are: deposition linewidth and surface roughness rating.

[0033] The XGBoost algorithm was used, with 90% of the data as the training set and 10% as the test set for training. The resulting model achieved a prediction accuracy R for line width. 2 It reached 0.89.

[0034] With the goal of "line width of 40 μm and surface roughness rating of A", Bayesian optimization was used for inverse optimization. The recommended parameters of the model are: laser power = 7.1 W, scanning speed = 500 mm / s, pulse frequency = 1000 kHz, and number of scans = 3.

[0035] Using the recommended parameters, laser-induced electrochemical deposition experiments were performed on silicon surfaces. The average linewidth of the deposited structure was 41.1 μm, and the surface was smooth and burr-free, achieving the expected goal.

[0036] The method for optimizing laser-induced electrochemical deposition (LAD) using the random forest algorithm includes the same steps as applying the XGBoost algorithm, except that the optimization objective is set to "maximize deposition rate". The model recommends another set of parameters: laser power = 7.5 W, scanning speed = 1000 mm / s, pulse frequency = 1000 kHz, and 2 scans. Verification showed that the deposition rate under these parameters was 200% higher than in Example 1, demonstrating the adaptability of the method to different optimization objectives. The model's prediction accuracy R for linewidth on the test set was [not specified]. 2 It reached 0.91.

[0037] Example 2 This invention also provides a machine learning-based laser-induced electrochemical deposition process optimization system for implementing the method of Example 1, comprising: The data acquisition module is used to collect historical process parameters and corresponding deposition effect data of laser-induced electrochemical deposition process to build a training set; The model building module is used to train machine learning models based on the training set and build performance prediction models for the mapping relationship between process parameters and deposition effects. The performance prediction module is used to predict the deposition effect of the target process parameters using the performance prediction model, and uses the expected deposition effect as the optimization target to search in the performance prediction model using the optimization algorithm to deduce the optimal combination of process parameters.

[0038] The present invention also designs a human-computer interaction interface for inputting desired targets, displaying prediction results and recommended parameters.

[0039] A further implementation method includes a data acquisition module comprising: Historical process parameter acquisition unit, used to collect historical laser parameters, electrochemical parameters and environmental parameters; The deposition effect acquisition unit is used to acquire deposition morphology, deposition rate and dimensional accuracy corresponding to historical process parameters.

[0040] A further implementation method involves using machine learning models including random forest and XGBoost algorithms. The selection of the machine learning model is based on the characteristics of the process parameters and user requirements. Specifically, the method for constructing a performance prediction model using the random forest algorithm includes: Using the autonomous sampling set of the training set, each decision tree of the random forest algorithm is trained, and a feature subset is selected from all process parameters for optimal splitting to obtain the performance prediction results of each decision tree. The performance prediction results of each decision tree are aggregated using the bagging method to obtain the final performance prediction results, thus completing the construction of the performance prediction model.

[0041] A further implementation method involves an optimization algorithm in the performance prediction module that integrates a set of Bayesian optimization algorithms and genetic algorithms, combined with cross-validation, to optimize the hyperparameters of the performance prediction model and deduce the optimal combination of process parameters. The hyperparameters include the number and maximum depth of decision trees in the random forest, and the learning rate and maximum depth of decision trees in XGBoost.

[0042] The software of the system described in this invention is developed based on Python. Its core model building and performance prediction modules integrate various machine learning algorithm libraries, enabling flexible construction and training of different prediction models based on data characteristics and user selections. The parameter optimization and inverse calculation module integrates optimizers such as Bayesian optimization and genetic algorithms for parameter search within the trained model.

[0043] Example 3 like Figure 3 As shown, this embodiment also provides a machine learning-based laser-induced electrochemical deposition process optimization device, including a computer 1, a laser 2, a galvanometer system 3, a focusing lens 4, an electrolytic cell 5, an anode 6, a cathode 7, a three-dimensional moving platform 8, a CCD camera 9, a temperature sensor 10, and a power supply 11.

[0044] The computer 1 is externally connected to a laser 2, a CCD camera 9, a temperature sensor 10, and a power supply 11. The computer 1 sends signals to the laser 2 and the power supply 11, and the CCD camera 9 and the temperature sensor 10 send signals to the computer.

[0045] Electrolytic cell 5 is placed on three-dimensional moving platform 8. Anode 6 and cathode 7 are placed inside electrolytic cell 5. Anode 6 is connected to the positive terminal of power supply 11, and cathode 7 is connected to the negative terminal of power supply 11. Computer 1 is operated, power supply 11 is turned on, and laser 2 starts working. After the laser beam is focused onto cathode 7 by galvanometer system 3 and focusing lens 4, the temperature of the electrolyte in electrolytic cell 5 rises. Temperature sensor 10 detects the signal and inputs it into computer 1 to collect data. CCD camera 9 takes pictures and uploads them in real time to the data management module of computer 1 to build a historical database for model training. Computer 1 trains a machine learning prediction model. The desired deposition performance indicators are input into computer 1. The performance prediction module calls the built-in optimization algorithm to search within the machine learning model, calculates the optimal parameter combination, and then automatically sends it to the hardware execution unit via the equipment control module to drive the equipment to complete the processing. Simultaneously, new data during the processing is collected and fed back for continuous model optimization (see the input interface for reverse calculation). Figure 2 ).

[0046] The process of executing an optimization method based on an optimization device includes: CCD camera 9 acquires images of sedimentary morphology, temperature sensor 10 acquires environmental parameters, and transmits them to the database of computer 1; Computer 1 records the laser parameters set by laser 2 and the electrochemical parameters set by power supply 11 at that time, and saves them in the database; Computer 1 invokes its built-in machine learning algorithm, reads the constructed historical database, and trains the model; When the system generates a new, untested set of process parameters, these parameters can be combined and input into a pre-trained model. Computer 1 runs the model and quickly calculates the predicted deposition effect; On the interface of computer 1 (see) Figure 2 Enter the desired deposition effect in the input field; Computer 1 uses the performance prediction model as the objective function, runs an optimization algorithm, and automatically searches for the optimal combination of process parameters; The optimal parameter combination obtained by reverse calculation is sent to hardware execution units such as laser 2, galvanometer system 3, power supply 11 and three-dimensional moving platform 8 to complete the deposition processing task.

[0047] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A machine learning-based method for optimizing laser-induced electrochemical deposition processes, characterized in that, include: Historical process parameters and corresponding deposition effect data of laser-induced electrochemical deposition were collected to construct a training set. A machine learning model is trained based on the training set to construct a performance prediction model for the mapping relationship between process parameters and deposition effect. The deposition effect of the target process parameters is predicted using the performance prediction model, and the desired deposition effect is used as the optimization target. The optimal combination of process parameters is then derived by searching the performance prediction model using an optimization algorithm.

2. The method according to claim 1, characterized in that, The historical process parameters include laser parameters, electrochemical parameters, and environmental parameters; The deposition effect data includes deposition morphology, deposition rate, and dimensional accuracy.

3. The method according to claim 1, characterized in that, The machine learning models include the random forest algorithm and the XGBoost algorithm. The selection of the machine learning model is based on the characteristics of the process parameters and user requirements. Specifically, the method for constructing a performance prediction model using the random forest algorithm includes: Using the autonomous sampling set of the training set, each decision tree of the random forest algorithm is trained, and a feature subset is selected from all process parameters for optimal splitting to obtain the performance prediction results of each decision tree. The performance prediction results of each decision tree are aggregated using the bagging method to obtain the final performance prediction results, thus completing the construction of the performance prediction model.

4. The method according to claim 3, characterized in that, The optimization algorithm integrates Bayesian optimization and genetic algorithms, and combines cross-validation to optimize the hyperparameters of the performance prediction model, thereby deriving the optimal combination of process parameters. The hyperparameters include the number and maximum depth of decision trees in the random forest, and the learning rate and maximum depth of decision trees in XGBoost.

5. A machine learning-based laser-induced electrochemical deposition process optimization system, used to implement the method of claim 1, characterized in that, include: The data acquisition module is used to collect historical process parameters and corresponding deposition effect data of laser-induced electrochemical deposition process to build a training set; The model building module is used to train a machine learning model based on the training set and build a performance prediction model for the mapping relationship between process parameters and deposition effect. The performance prediction module is used to predict the deposition effect of the target process parameters using the performance prediction model, and to use the expected deposition effect as the optimization target to search the performance prediction model using an optimization algorithm to deduce the optimal combination of process parameters.

6. The system according to claim 5, characterized in that, The data acquisition module includes: Historical process parameter acquisition unit, used to collect historical laser parameters, electrochemical parameters and environmental parameters; The deposition effect acquisition unit is used to acquire deposition morphology, deposition rate and dimensional accuracy corresponding to historical process parameters.

7. The system according to claim 6, characterized in that, In the model building module, the machine learning model includes the random forest algorithm and the XGBoost algorithm. The machine learning model is selected based on the characteristics of the process parameters and user requirements. The method for constructing a performance prediction model using the random forest algorithm includes: Using the autonomous sampling set of the training set, each decision tree of the random forest algorithm is trained, and a feature subset is selected from all process parameters for optimal splitting to obtain the performance prediction results of each decision tree. The performance prediction results of each decision tree are aggregated using the bagging method to obtain the final performance prediction results, thus completing the construction of the performance prediction model.

8. The system according to claim 7, characterized in that, In the performance prediction module, the optimization algorithm integrates a set of Bayesian optimization algorithms and genetic algorithms, and combines cross-validation to optimize the hyperparameters of the performance prediction model, thereby deriving the optimal combination of process parameters. The hyperparameters include the number and maximum depth of decision trees in the random forest, and the learning rate and maximum depth of decision trees in XGBoost.

Images