Nanosecond laser micro-machining parameter optimization method based on physical information neural network

Abstract

This invention discloses a method for optimizing nanosecond laser micromachining parameters based on a physical information neural network. The method includes: for any combination of machining parameters to be evaluated, obtaining the corresponding predicted machining quality response value using a trained machining quality response prediction model; constructing an objective function with hard constraints based on the obtained predicted machining quality response value; and optimizing the constructed objective function using a Bayesian optimization algorithm to obtain the corresponding optimal combination of machining parameters. This invention can directly output a nanosecond laser micromachining parameter combination that satisfies quality and efficiency constraints while reducing experimental and simulation costs.

Description

Technical Field

[0001] This invention relates to the field of laser processing modeling and intelligent optimization technology, and in particular to a method for optimizing nanosecond laser microprocessing parameters based on physical information neural networks. Background Technology

[0002] In nanosecond laser micromachining, process parameters such as pulse width, spot diameter, laser power, repetition frequency, processing speed, and number of processing passes significantly affect key indicators such as surface roughness (Sa), heat-affected zone (HAZ) thickness, and processing time. In industrial settings, parameter combinations are typically determined using empirical parameter tables and manual trial cuts, which suffers from low efficiency, high cost, and difficulty in balancing quality and efficiency.

[0003] Based on existing experimental data, a more refined and accurate model is developed within the data range to address the intrinsic relationship between processing parameters and process results. Combined with parameter optimization methods, this results in a more versatile, precise, and convenient nanosecond laser micromachining parameter optimization system. This system can significantly reduce the cost of processing experiments and enhance industrial competitiveness. Summary of the Invention

[0004] The purpose of this invention is to provide a nanosecond laser micromachining parameter optimization method with a clear modeling process, explicit handling of engineering hard constraints, and stronger physical rationality and generalization ability. It can directly output the optimal combination of processing parameters that meets the constraints of quality and processing time, reduce experimental costs and improve parameter selection efficiency.

[0005] This invention is based on a neural network prediction model that combines physical information and data-driven approaches, and combines Bayesian optimization to optimize the process parameters of nanosecond laser micromachining (including micro-cutting, microgroove / micro-hole machining, etc.).

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for optimizing nanosecond laser micromachining parameters based on a physical information neural network includes: for any combination of machining parameters to be evaluated, obtaining the corresponding predicted value of machining quality response using a trained machining quality response prediction model; constructing an objective function containing hard constraints based on the obtained predicted value of machining quality response; and optimizing the constructed objective function using a Bayesian optimization algorithm to obtain the corresponding optimal combination of machining parameters.

[0008] Furthermore, this invention provides a method for modeling and optimizing parameters in nanosecond laser micromachining based on a physical information neural network. The method first acquires data samples. Two types of physical extended features are constructed based on the original parameters. The two types of features are normalized, and the normalized parameters are saved. Sa prediction networks and HAZ prediction networks are then established respectively. During network training, in addition to minimizing the prediction error, a monotonicity soft constraint based on the input gradient sign is introduced, and a soft constraint consistent with the physical proxy model is introduced into the HAZ network to improve the physical rationality and generalization ability of the model. Based on this, the user-given upper limits of Sa, HAZ, cutting edge length, and machining time are used as hard constraints. A weighted objective function including roughness, heat-affected zone, and machining time is constructed, and Bayesian optimization driven by the expected improvement class acquisition function is used to iteratively search for the optimal parameter combination in the discrete process parameter space. After optimization, the feasibility of all sampled points is checked, and the parameter combination with the smallest objective function is selected as the final recommended value from the feasible solution set that satisfies all constraints.

[0009] Furthermore, the combination of processing parameters to be evaluated includes one or more of pulse width, spot diameter, laser power, laser frequency, processing speed, and number of processing operations; even further, the combination of processing parameters to be evaluated includes pulse width, spot diameter, laser power, laser frequency, processing speed, and number of processing operations. Similarly, when training the processing quality response prediction model, the input raw data are known existing combinations of processing parameters and the corresponding obtained processing quality response values ​​(preferably surface roughness and heat-affected zone thickness). The processing parameter combination used for training corresponds to the combination of processing parameters to be evaluated, and also includes pulse width, spot diameter, laser power, laser frequency, processing speed, and number of processing operations.

[0010] Furthermore, the machining quality response prediction model includes a surface roughness prediction model and a heat-affected zone thickness prediction model. These two prediction networks are used to obtain the predicted surface roughness value and the predicted heat-affected zone thickness value corresponding to the input machining parameter combination, respectively. The surface roughness prediction model and the heat-affected zone thickness prediction model are set in parallel and are used to obtain the corresponding predicted surface roughness value and heat-affected zone thickness value, respectively.

[0011] Furthermore, in the optimization stage or the training stage of the processing quality response prediction model: for any combination of input processing parameters, corresponding physical expansions are performed respectively, and then the expanded parameter set is normalized to obtain the input of the processing quality response prediction model.

[0012] The expanded parameter set includes the original input processing parameter combination and the expanded parameters obtained using the physical expansion. The expanded parameters are generally calculated directly based on the corresponding parameters in the original input processing parameter combination.

[0013] Specifically, for the surface roughness prediction model, the extension parameters obtained by the physical extension include pulse overlap rate and power density; for the heat-affected zone thickness prediction model, the extension parameters obtained by the physical extension include energy per unit length and surface energy density.

[0014] For the purposes of the following description, the letters used in this invention are defined as follows:

[0015]

[0016] Furthermore, the physical extension includes:

[0017] (1) Obtain nanosecond laser micromachining data (i.e., any combination of processing parameters to be evaluated (optimization process) or existing historical processing parameter combinations (training process)), for example, the processing parameter combination includes at least 6 process parameters pw, d, P, f rep 、v、n pass and the corresponding processing quality responses Sa and HAZ;

[0018] (2) Based on the above 6 process parameters, construct physical extended input features and form a first input feature vector for Sa prediction (i.e., the extended parameter set as input to the surface roughness prediction model) and a second input feature vector for HAZ prediction (i.e., the extended parameter set as input to the heat-affected zone thickness prediction model).

[0019] The first input feature vector includes [pw, d, P, f] rep 、v、n pass Overlap, I dens ] T The pulse overlap rate is calculated according to the formula shown in the table above, and the overlap is truncated to [0,1]; the power density I dens The calculation formula is shown in the table above.

[0020] The second input feature vector includes [pw, d, P, f] rep 、v、n pass ,EL,F] T The energy per unit length EL and the surface energy density F are calculated according to the formulas shown in the table above.

[0021] When normalizing the expanded parameter set, a dimensional linear normalization method is used for any dimensional feature x. i Press x i '=(x i -x i min ) / (x i max -xi min Map x to the interval [0,1] and set x to 0.1. i min With x i max Saved as a normalization parameter, new input samples are normalized using the same normalization parameter during the optimization phase.

[0022] Preferably, the training steps for the Sa prediction model (i.e., surface roughness prediction model) and the HAZ prediction model (i.e., heat-affected zone thickness prediction model) are as follows:

[0023] (1) Obtain historical nanosecond laser micromachining data (i.e., existing combinations of processing parameters), wherein the data contains at least 6 process parameters pw, d, P, f rep 、v、n pass and the corresponding processing quality responses Sa and HAZ;

[0024] (2) Based on the above 6 process parameters, construct physical extended input features and form a first input feature vector for Sa prediction and a second input feature vector for HAZ prediction respectively;

[0025] (3) Linearly normalize the first input feature vector and the second input feature vector to the [0,1] interval respectively to obtain two sets of normalized samples and save the corresponding normalization parameters;

[0026] (4) Using the obtained normalized samples and the corresponding Sa and HAZ, a first regression neural network is constructed to predict Sa and a second regression neural network is constructed to predict HAZ. The two networks are trained with physical information constraints to obtain the Sa prediction model and the HAZ prediction model.

[0027] As a preferred embodiment, a method for optimizing nanosecond laser micromachining parameters based on a physically constrained neural network includes the following steps:

[0028] (1) Input the nanosecond laser micromachining data to be evaluated (i.e., the combination of processing parameters to be evaluated), which is consistent with the data composition in the training phase; the data contains at least 6 process parameters pw, d, P, and f. rep 、v、n pass ;

[0029] (2) Based on the six process parameters, construct physical extended input features and form the first input feature vector and the second input feature vector corresponding to the data to be evaluated.

[0030] (3) Using the same normalization parameters as in the training phase, the first input feature vector and the second input feature vector are linearly normalized to the [0,1] interval respectively to obtain two sets of normalized samples;

[0031] (4) Input the two sets of normalized samples into the constructed surface roughness prediction network and heat-affected zone thickness prediction model respectively to obtain the corresponding surface roughness prediction value Sa. hat Predicted thickness of the heat-affected zone (HAZ) hat ;

[0032] (5) Obtain the maximum allowable roughness Sa according to the user's processing requirements. req Maximum permissible heat-affected zone (HAZ) req Cutting edge length L and maximum machining time T req ;

[0033] (6) Define the variables to be optimized within the allowable range of process parameters, construct the objective function F, and set the weights w1, w2, and w3 for roughness, heat-affected zone, and processing time, where... ;

[0034] (7) For any combination of parameters to be evaluated, call the Sa prediction model and the HAZ prediction model to obtain Sa. hat With HAZ hat Calculate the processing time T, when Sa hat <Sa req HAZ hat <HAZ req and T <T req When all constraints are satisfied, for Sa hat HAZ hat The objective function value is obtained by normalizing T and weighting it according to the weights; for Sa hat <Sa req HAZ hat <HAZ req and T <T req If at least one constraint is not satisfied, a large penalty strategy is directly adopted, and the target value containing the large penalty term is output to exclude the points that do not satisfy the constraint.

[0035] (8) Using the Bayesian optimization algorithm, with the objective function F as the black box function, the Gaussian process surrogate model and the expected improvement class acquisition function are used to iteratively select the next set of parameters to be evaluated until the termination condition is met;

[0036] (9) Perform a feasibility review on all sampling parameter combinations during the Bayesian optimization process, and select those that satisfy Sa. hat ≤Sa req HAZ hat ≤HAZ reqAnd T≤T re The feasible solution set of q is selected, and the parameter combination with the minimum objective function value is selected as the optimal processing parameter combination. If the feasible solution set is empty or there is no feasible solution, the compromise solution with the minimum degree of constraint violation is output and a prompt is given (such as prompting that the current constraint condition is too strict).

[0037] Furthermore, both the surface roughness prediction model and the heat-affected zone thickness prediction model employ fully connected feedforward networks with ReLU activation functions and linear mapping outputs the corresponding predicted processing quality response values. Even further, both the first and second regression neural networks are fully connected feedforward networks, comprising an input layer, two hidden layers, and an output layer; the two hidden layers each contain h1 and h2 neurons with ReLU activation functions, and the output layer outputs a single scalar value via linear mapping.

[0038] More specifically, the number of neurons in the two hidden layers satisfies h1=50~100 (e.g., 64) and h2=50~100 (e.g., 64); the network forward propagation is: h (1) =ReLU(W (1) x'+b (1) ), h (2) =ReLU(W (2) h (1) + b (2) ), y hat =W (3) h (2) +b (3) .

[0039] Furthermore, when training the processing quality response prediction model, physical information constraint training is employed. This physical information constraint training uses monotonic soft constraints, which are implemented by penalizing the sign of the partial derivative of the predicted output with respect to the input features; when the expected value y... hat With a certain input x k When not increasing, the monotonicity penalty term includes When y is expected hat With a certain input x k When not reduced, the monotonicity penalty term includes .

[0040] Furthermore, for the surface roughness prediction model, the input features of its physical information constraints include the number of processing operations; for the heat-affected zone thickness prediction model, the input features of its physical information constraints include the number of processing operations, processing speed, energy per unit length, and surface energy density.

[0041] Specifically, for the Sa prediction network (i.e., the surface roughness prediction model), the number of processing steps n is set. pass The corresponding input channel satisfies The corresponding penalty items are This constraint aims to help the Sa predict the network's response to the number of processing steps n. pass It meets the requirement of monotonically decreasing.

[0042] For the HAZ prediction network (i.e., the heat-affected zone thickness prediction model), the number of processing steps n is set. pass The corresponding input channel satisfies The corresponding penalty items are ; Set the input channels corresponding to EL and F to meet the requirements , The corresponding penalty items are , And set the input channel corresponding to the processing speed v to meet the requirements. The corresponding penalty items are This constraint aims to help the HAZ prediction network predict energy per unit length. Energy density of dough It meets the requirement of monotonically increasing, and the requirement of monotonically decreasing for the number of processing times and processing speed v.

[0043] Furthermore, the physical information constraint training further includes HAZ physical agent consistency soft constraints, which are constructed using the physical agent function g(EL', F') (calculation formula is shown in the table above). In the formula, EL' and F' are normalized features, a and b are weight coefficients, and ε is a very small positive number. The mean square error between the normalized prediction output and g(EL', F') is calculated to obtain the HAZ physical agent consistency penalty term.

[0044] Specifically, when training the processing quality response prediction model, the loss function is:

[0045] Loss=Loss data +λ mono Loss mono +λ phys Loss phys ;

[0046] Loss data The mean square error between the predicted output and the actual output;

[0047] Loss mono This is a monotonic soft constraint penalty term, obtained by averaging the above penalty terms over a small batch of samples;

[0048] Loss phys , where is the HAZ physical agent consistency penalty term, and is the mean square error of the network prediction output after linear normalization within the batch compared to g(EL',F').

[0049] λ mono With λphys Loss mono He Loss phys Weighting coefficients;

[0050] For the surface roughness prediction model λ phys Loss phys exist.

[0051] Different normalization parameters were used for training and inference of the Sa prediction network and the HAZ prediction network to match the differences in the value range of their respective physical extended features.

[0052] Furthermore, the objective function is obtained by summing the weighted surface roughness term, heat-affected zone thickness term, and processing time term; the surface roughness term is obtained by multiplying the normalized value of the predicted surface roughness value with its corresponding weight coefficient; the heat-affected zone thickness term is obtained by multiplying the normalized value of the predicted heat-affected zone thickness value with its corresponding weight coefficient; and the processing time term is obtained by multiplying the normalized value of the processing time calculated from the current parameter combination with its corresponding weight coefficient.

[0053] Specifically, the objective function F is constructed as F = w1 × Sa norm +w2×HAZ norm +w3×T norm Sa norm HAZ norm T norm This represents the result after mapping to [0, 1] within a fixed physical range. The range of process parameters is determined with reference to the initial micromachining data.

[0054] Furthermore, the Bayesian optimization algorithm used treats the objective function as a deterministic black-box function, adopts the "expected-improvement-plus" acquisition function, and sets each variable to be optimized as an integer or discrete variable to match the equipment resolution or process step size.

[0055] Furthermore, when using the Bayesian optimization algorithm for optimization:

[0056] If at least one of the predicted surface roughness, predicted heat-affected zone thickness, or processing time does not meet the constraint condition of being less than the maximum value, a large penalty strategy is directly adopted to output a larger target value.

[0057] When the predicted surface roughness, predicted heat-affected zone thickness, and processing time all meet the constraint condition of being less than the maximum value, the target value is calculated according to the objective function.

[0058] Furthermore, candidate parameter combinations are obtained using the Bayesian optimization algorithm. The feasibility of all sampled parameter combinations during the Bayesian optimization process is then reviewed, and those satisfying Sa are selected.hat ≤Sa req HAZ hat ≤HAZ req And T≤T req The feasible solution set is used to select the parameter combination with the minimum objective function value as the optimal processing parameter combination. If the feasible solution set is empty, the compromise solution with the minimum degree of constraint violation (i.e. the minimum penalty term) is output and a prompt is given.

[0059] A nanosecond laser micromachining parameter optimization system for implementing any of the methods described above includes: a data processing module, a feature construction module, a neural network training module, an objective function evaluation module, a Bayesian optimization module, and a result output module; each module is configured to execute the corresponding steps of the methods described above.

[0060] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to perform any of the methods described above.

[0061] This invention can directly output nanosecond laser micromachining parameter combinations that meet quality and efficiency constraints while reducing testing and simulation costs. Attached Figure Description

[0062] The technical solution of the present invention is explained more clearly with reference to the accompanying drawings:

[0063] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0064] Figure 2 A schematic diagram of the Sa prediction network structure (fully connected feedforward network);

[0065] Figure 3 A schematic diagram of the HAZ prediction network structure (fully connected feedforward network);

[0066] Figure 4 A schematic diagram of the training process / loss function constrained by physical information (data item + monotonicity item + physical surrogate item);

[0067] Figure 5 A schematic diagram of the Bayesian optimization iterative search and feasible optimal solution selection process. Detailed Implementation

[0068] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are used to explain the present invention, and not to limit the scope of protection of the present invention.

[0069] See Figure 1 :

[0070] Step 1: Sample set construction and physical extended feature calculation:

[0071] This data is based on nanosecond laser micromachining data obtained from experiments under various process parameters. Each set of data includes: pulse width pw (in nanoseconds), spot diameter d (in mm). Figure 2 and Figure 3 The corresponding values ​​are D), laser power P (in W), and laser frequency f. rep (unit: kHz), processing speed v (unit: mm / s), Figure 2 and Figure 3 (corresponding to V) and the number of processing times n pass (Dimensionless integer), and the corresponding surface roughness Sa (in nm) and heat-affected zone thickness HAZ (in μm). The optimization objective is to optimize the surface roughness Sa within the user-given maximum surface roughness. req Maximum heat-affected zone thickness HAZ req and maximum processing time T req Within the requirements, and combined with the given shape data of the tool, a parameter combination is obtained that minimizes the objective function F of the evaluation index.

[0072] Data augmentation and expansion were performed on the original samples to obtain a sample set with 100 effective samples.

[0073] The range of values ​​for each parameter is defined based on the sample set as follows:

[0074]

[0075] A dataset is constructed based on existing basic process parameters and considered quality responses. Simultaneously, different physical extension features are constructed for different processing responses.

[0076] (1) Sa extended features: Calculate the pulse overlap rate (Overlap) Figure 2 and Figure 3 (Corresponding to Overlap)

[0077] .

[0078] When v / ((f) rep ×10 3 When )×d) is greater than 1 or less than 0, overlap is truncated to [0,1] to avoid non-physical values;

[0079] And calculate the power density I dens :

[0080] .

[0081] This yields the 8-dimensional first input feature vector x. Sa :

[0082] .

[0083] (2) HAZ extended characteristics: Calculate the energy per unit length EL, EL=P / v; calculate the single pulse energy E pulse :

[0084] ;

[0085] And calculate the surface energy density F:

[0086] .

[0087] This yields the 8-dimensional second input feature vector x. HAZ :

[0088] .

[0089] The modeling sample set is constructed from the first input feature vector and the second input feature vector. The constructed modeling sample set is then normalized to obtain the normalized sample set x. Sa 'and x HAZ For any dimension of laser processing parameters and physical extension characteristics x i Map to the [0,1] interval using the following formula to complete the normalization process:

[0090] .

[0091] In the formula, x i ' is the normalized result of the i-th group of data in the modeling sample set, x i It is the value of the i-th group of processing data in the modeling sample set before normalization, x i max and x i min Let x be the maximum and minimum values ​​in the i-th group of data in the sample set to be normalized and modeled, respectively, and let x be... i max and x i min Saved as a normalization parameter, the newly input sample data is normalized using the same normalization parameter during the optimization phase.

[0092] The normalized modeling sample set x is obtained from step one. Sa '(the first input feature vector after normalization) and x HAZ (The normalized second input feature vector), and save the maximum and minimum values ​​corresponding to the samples before normalization as normalization parameters. In the optimization stage, the new input is normalized with the same normalization parameters.

[0093] Step 2: Training the Sa prediction network with physical information constraints:

[0094] A first regression neural network was constructed for the Sa prediction task. The network adopts a fully connected feedforward structure, such as... Figure 2 As shown: The input layer has 8 nodes (corresponding to pw, d, P, f respectively). rep ,v,n pass overlap, I dens (Normalized parameter input); the first hidden layer contains 64 neurons with ReLU activation function; the second hidden layer contains 64 neurons with ReLU activation function; the output layer has 1 node and uses linear mapping output to obtain the surface roughness prediction value Sa. hat .

[0095] In actual training:

[0096] For the previously normalized dataset x Sa The entire sample is randomly divided into a training set, a validation set, and a test set. For example, the training set accounts for 70%, the validation set accounts for 15%, and the test set accounts for 15%.

[0097] During the training of the first regression neural network, the Adam optimization algorithm with mini-batch gradient descent is used to update the network parameters. The initial learning rate is set to 0.005, the maximum number of training epochs can be set to 2000, and the batch size is 64. A learning rate decay strategy is adopted, in which the learning rate is multiplied by 0.96 every 50 epochs (representing a complete cycle) to improve the convergence stability in the later stages.

[0098] To prevent overfitting, an early stopping strategy based on the validation set error is introduced: if the validation set mean squared error does not increase by more than a threshold of 10 within 60 consecutive epochs... -6 Training is stopped at this point, and the network parameters are rolled back to the optimal value at the time of the validation set error.

[0099] See Figure 4 The training loss function corresponding to the first regression neural network is Loss. Sa It consists of data terms and monotonic physical constraint terms:

[0100] .

[0101] Loss Sa,data for Sa hat Mean squared error between the true Sa and the actual Sa; Loss Sa,mono λ is the monotonic soft constraint penalty term corresponding to the first regression neural network. Sa,mono For Loss Sa,mono The corresponding weighting coefficients.

[0102] In the first regression neural network corresponding to the monotonicity constraint, the gradient of the network output with respect to the input is obtained using automatic differentiation. The number of processing times n pass For example, it is expected that, all other things being equal, increasing the number of machining operations should not lead to an increase in surface roughness (or remain unchanged under specific process conditions), therefore applying... The corresponding penalty term uses the ReLU penalty function, which can be written as: The average value within the batch is then used to obtain the corresponding loss. Sa,mono . λ Sa,mono For Loss Sa,mono The corresponding monotonicity constraint weight coefficient is set to 10. -3 The magnitude was adjusted according to data fitting and physical consistency requirements. Using surface roughness Sa as the prediction target, the constructed physical information constrained neural network was trained and validated. Results show that the network exhibits good prediction accuracy and stability on the training, validation, and test sets. Specifically, at the optimal training epoch (634) with the minimum mean square error on the validation set, the mean square errors of the model on the training, validation, and test sets were 124.08, 15.87, and 36.50, respectively, with corresponding root mean square errors of 4.90, 3.98, and 6.04. The normalized root mean square error (NRMSE) based on the sample value range was less than 4%, and the mean absolute percentage error (MAPE) was less than 3%. Furthermore, the correlation coefficient R between the predicted results and the actual measured values ​​was greater than 0.98 on the training, validation, and test sets. (Coefficient of determination R...) 2 The values ​​of 0.97 on the training, validation, and test sets all indicate that the Sa prediction model has high fitting accuracy, good generalization ability, and stable prediction performance, which can meet the requirements for roughness prediction accuracy in the subsequent process parameter optimization stage.

[0103] Step two is used to complete the training of the first regression neural network.

[0104] Step 3: Training the HAZ prediction network with physical information constraints:

[0105] A second regression neural network is constructed for the HAZ prediction task. (Reference) Figure 3 As shown, the network structure is the same as the Sa prediction network: the input layer has 8 nodes (corresponding to Pw, d, P, f respectively). rep ,v,n pass The input parameters (EL and F after normalization) are two ReLU hidden layers with 64 neurons each. The output layer (including one node) linearly outputs the predicted thickness of the heat-affected zone (HAZ). hat .

[0106] Normalized data x from the model input HAZ '.

[0107] See Figure 4 The training loss function corresponding to the second regression neural network is Loss. HAZ It consists of data items, monotonicity physical constraint items, and physical proxy consistency items:

[0108] .

[0109] Loss HAZ,data For HAZ hat Mean squared error between the true HAZ and the actual HAZ; Loss HAZ,mono Loss is the monotonic soft constraint penalty term corresponding to the second regression neural network. HAZ,phys λ represents the physical surrogate consistency loss, which is the mean square error between the predicted output after linear normalization within the batch and g(EL',F'); g(EL',F') is the surrogate function constructed from the normalized features EL' and F', the formula of which will be discussed later; HAZ,phys With λ HAZ,phys Loss HAZ,mono Loss HAZ,phys The corresponding weighting coefficients.

[0110] Loss HAZ,mono By applying a gradient sign penalty, the following monotonic trend is applied: increasing the number of processing iterations should not lead to an increase in the expected HAZ, therefore, the following is applied... The corresponding penalty item is: An increase in EL indicates an increase in energy input per unit length, with the expected HAZ remaining constant. Corresponding punishment An increase in F indicates an increase in surface energy density, and the expected HAZ remains unchanged. Corresponding punishment Increasing the processing speed v usually reduces heat input, and the expected HAZ does not increase. Corresponding punishment The above penalty terms are averaged over a small batch of samples to obtain the Loss. HAZ,mono , and press λ HAZ,mono Weighted. λ HAZ,mono Take 10 -2 The magnitude will be adjusted based on actual data. Loss HAZ,phys This is used to further guide HAZ predictions to satisfy the simplified physical surrogate model. The surrogate function g(EL', F') is constructed using normalized features EL' and F':

[0111] ,

[0112] ε is a very small positive number;

[0113] EL' refers to the normalized data corresponding to the energy EL per unit length;

[0114] F' refers to the normalized data corresponding to the surface energy density F;

[0115] a and b are parameters that are set manually; in this example, a = 0.6 and b = 0.4.

[0116] The mean squared error of the network prediction output after in-batch linear normalization is calculated by comparing it with g(EL', F'), and this error is used as the loss. HAZ,phys , and according to λ HAZ,phys Weighted; λ HAZ,phys Take 10 -3 The scale is designed to maintain the "lightly constrained" characteristics, allowing the model to focus primarily on data fitting with secondary consideration of physical consistency.

[0117] Using the thickness of the heat-affected zone (HAZ) as the prediction target, the constructed physical information-constrained neural network was trained and validated. Training results show that the model exhibits high prediction accuracy and good generalization performance on the training, validation, and test sets. At the optimal training epoch (893) with the minimum mean squared error on the validation set, the mean squared errors of the model on the training, validation, and test sets are 2.31, 2.73, and 5.52, respectively, with corresponding root mean square errors of 1.52, 1.65, and 2.35. The NRMSE (Number of Rectified Mean Squares) after normalization based on the sample range is below 1.7%, and the mean absolute percentage error (MAPE) is below 1.4%. Furthermore, the correlation coefficient R between the predicted results and the measured values ​​is greater than 0.99 on the training, validation, and test sets. (Coefficient of Determination R0) 2 The fact that the HAZ prediction model is greater than 0.97 in the training, validation, and test sets indicates that it has strong stability and generalization ability while ensuring prediction accuracy, and can effectively support the subsequent selection of processing parameters and constraint evaluation based on Bayesian optimization.

[0118] After training is complete, the Sa neural network prediction model (i.e., the first regression neural network) and the HAZ neural network prediction model (i.e., the second regression neural network) and their respective normalized parameters are saved as model files for use in the subsequent Bayesian optimization stage to avoid repeated training.

[0119] Example 4: Parameter Selection and Feasible Solution Screening in Bayesian Optimization

[0120] See Figure 5 After completing the training of the Sa and HAZ neural network prediction models, the process parameters are optimized and selected.

[0121] Users input engineering constraints and decision preferences based on actual processing requirements. The ranges of each parameter are referenced from the dataset, with the maximum permissible roughness Sa. req Range (e.g., 100–300 nm), maximum permissible heat-affected zone thickness HAZ req The range (e.g., 100–900 μm), within the allowable range, is set by the user-input Sa. req and HAZ req Input the cutting edge length L (e.g., 1–10 mm) matched with the machining tool, and the maximum machining time T. req (e.g., 1–100 s), and the weights w1, w2, w3 corresponding to surface roughness, allowable heat-affected zone thickness, and processing time, and normalized to sum to 1.

[0122] Define the variable to be optimized and its value range to match the device resolution, for example:

[0123] ;

[0124] ;

[0125] ;

[0126] ;

[0127] ;

[0128] .

[0129] For any candidate parameter combination x(pw, d, P, f) rep ,v,n pass First, construct the corresponding x according to the formula in Example 1. Sa With x HAZ And obtain x according to their respective normalization parameters. Sa 'with x HAZ ', Input the trained neural network model (i.e., the Sa neural network prediction model and the HAZ neural network prediction model) to obtain the predicted surface roughness value Sa. hat Predicted maximum heat-affected zone thickness HAZ hat Processing time is calculated as T = (L × n) pass ) / v calculation.

[0130] Construct a large penalty objective function F with hard constraints:

[0131] When Sa hat >Sa req or HAZ hat >HAZ reqOr T>T req When this happens, a large penalty strategy is used to output a larger target value to exclude points that do not meet the constraints.

[0132] When all constraints are met, the predicted values ​​of the quality indicators are normalized using a fixed physical range. For example, let Sa... min =100、Sa max =500, then the normalized value of the predicted surface roughness, Sa norm The clip() function is used to calculate the following:

[0133] ;

[0134] Let HAZ min =50, HAZ max =200, then the normalized value HAZ of the predicted thickness of the heat-affected zone is... norm Calculate using the following formula:

[0135] .

[0136] Based on the input L and the current optimization parameter n pass The normalized value T of the processing time obtained from v norm The theoretical minimum processing time can be used. (Corresponding to a maximum speed of 5mm / s) and the user limit T req Perform mapping:

[0137] ;

[0138] ε also represents a very small positive number;

[0139] The final objective function is:

[0140] .

[0141] A Bayesian optimization algorithm (implemented in this embodiment using the `bayesopt` function in MATLAB) is used to iteratively minimize F(x) (the objective function F) and search for the optimal combination of process parameters. A surrogate model is established using a Gaussian process (GP) to approximate F(x), and the acquisition function is the expected-improvement-plus (EI+). The search logic of `bayesopt` can be broken down into four iterative steps:

[0142] Step 1: Initialize sampling: First, select several points in the parameter space to perform real evaluation and obtain a batch of (x, F(x)) data;

[0143] Step 2: Fitting the surrogate model (Gaussian process GP): Fitting GP with all sampling points to obtain the predicted distribution for any candidate x: mean μ(x) (expected target value) and variance σ²(x) (uncertainty);

[0144] Step 3: Select the next point using the sampling function, evaluate the actual value and update: Submit the selected x to your objective function to calculate the actual F(x), add it to the dataset and repeat steps 2 and 3 until the upper limit of the number of evaluations is reached or the convergence condition is met.

[0145] The optimal combination of processing parameters is obtained by searching within the candidate process parameter space using the Bayesian optimization algorithm described above. The optimal combination of processing parameters obtained at this time is the candidate combination of sampling point data.

[0146] To ensure the final results are directly usable in engineering, a feasibility check was performed on the bayesopt sampling points: after Bayesian optimization, Sa was recalculated for all sampling points. hat HAZ hat With T, and filter to satisfy The feasible solution set is determined; the normalization index and F-value are recalculated within the feasible solution set, and the combination with the smallest F-value is selected as the final recommended parameter combination. If the feasible solution set is empty, the compromise solution with the smallest penalty term is output, and a message is provided indicating that the current constraint combination may not be simultaneously satisfied.

[0147] In one embodiment, the user inputs processing requirements and constraints, including the maximum allowable surface roughness Sa, through a parameter optimization system. max (corresponding to Sa) req Maximum heat-affected zone width (HAZ) max (corresponding to HAZ) req ), blade length L, and maximum machining time T max (corresponding to T) req The system sets weighting coefficients for surface roughness, heat-affected zone, and processing time based on actual processing preferences. Upon receiving user-input constraints (Sa... max =200nm, HAZ max =200μm, L=2 mm, T max After configuring the parameters (w1=0.6, w2=0.3, w3=0.1) and weights (w1=0.6, w2=0.3, w3=0.1), an objective function is constructed based on the trained Sa neural network prediction model and HAZ neural network prediction model. A Bayesian optimization algorithm is then used to iteratively search within the discrete process parameter space. The optimization results provide a set of optimal process parameter combinations that satisfy all constraints.

[0148]

[0149] The corresponding predicted processing results are a surface roughness Sa of approximately 148.78 nm, a heat-affected zone width HAZ of approximately 108.74 μm, and a processing time of approximately 3.5 s, all of which meet the quality and efficiency constraints set by the user. This verifies that the method of the present invention can directly output a combination of processing parameters that is engineering-feasible and satisfies multiple objective trade-offs under user interaction conditions.

[0150] The above are merely preferred embodiments of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing parameters of nanosecond laser microfabrication based on a physical information neural network, characterized in that, include: For any combination of processing parameters to be evaluated, the corresponding processing quality response prediction value is obtained using the trained processing quality response prediction model. Based on the obtained processing quality response prediction values, an objective function containing hard constraints is constructed. The Bayesian optimization algorithm is then used to optimize the constructed objective function to obtain the corresponding optimal processing parameter combination. 2.The physical information neural network-based nanosecond laser micro-machining parameter optimization method according to claim 1, wherein, The processing quality response prediction model includes a surface roughness prediction model and a heat-affected zone thickness prediction model. The two prediction networks are used to obtain the predicted surface roughness value and the predicted heat-affected zone thickness value corresponding to the combination of input processing parameters, respectively. 3.The physical information neural network-based nanosecond laser micro-machining parameter optimization method according to claim 2, wherein, In the optimization phase or the training phase of the processing quality response prediction model: for any combination of input processing parameters, corresponding physical expansions are performed, and then the expanded parameter set is normalized to obtain the input of the processing quality response prediction model.

4. The physical information neural network-based nanosecond laser micro-machining parameter optimization method according to claim 3, characterized in that, For the surface roughness prediction model, the expanded parameters obtained by the physical expansion include pulse overlap rate and power density; for the heat-affected zone thickness prediction model, the expanded parameters obtained by the physical expansion include energy per unit length and surface energy density. 5.The physical information neural network based nanosecond laser micro-machining parameter optimization method according to claim 3, wherein, When training the processing quality response prediction model, physical information constraint training is adopted, which adopts monotonic soft constraint and / or physical surrogate model consistency soft constraint; the monotonic soft constraint is implemented by penalizing the sign of the partial derivative of the predicted output with respect to the input features; the physical surrogate model consistency soft constraint is implemented by normalizing the mean square error between the predicted output and its corresponding physical surrogate function.

6. The physical information neural network-based nanosecond laser micro-machining parameter optimization method according to claim 5, characterized in that, When training the processing quality response prediction model, the loss function is: ； The mean square error between the predicted output and the actual output; This is a soft constraint penalty term for monotonicity; For HAZ physical agent consistency penalty items; and They are respectively what Weighting coefficients; For surface roughness prediction models It does not exist.

7. The method for optimizing nanosecond laser micromachining parameters based on a physical information neural network according to claim 2, characterized in that, Both the surface roughness prediction model and the heat-affected zone thickness prediction model adopt a fully connected feedforward network with ReLU as the activation function and a linear mapping output corresponding to the predicted processing quality response value.

8. The method for optimizing nanosecond laser micromachining parameters based on a physical information neural network according to claim 2, characterized in that, The objective function is obtained by summing the weighted surface roughness term, heat-affected zone thickness term, and processing time term; the surface roughness term is obtained by multiplying the normalized value of the predicted surface roughness value with its corresponding weight coefficient; the heat-affected zone thickness term is obtained by multiplying the normalized value of the predicted heat-affected zone thickness value with its corresponding weight coefficient; the processing time term is obtained by multiplying the normalized value of the processing time calculated by the current parameter combination with its corresponding weight coefficient.

9. The method for optimizing nanosecond laser micromachining parameters based on a physical information neural network according to claim 8, characterized in that, When using the Bayesian optimization algorithm for optimization: If at least one of the predicted surface roughness, predicted heat-affected zone thickness, or processing time does not meet the constraint condition of being less than the maximum value, a large penalty strategy is directly adopted to output a larger target value. When the predicted surface roughness, predicted heat-affected zone thickness, and processing time all meet the constraint condition of being less than the maximum value, the target value is calculated according to the objective function.

10. The method for optimizing nanosecond laser micromachining parameters based on a physical information neural network according to any one of claims 1 to 9, characterized in that, The combination of processing parameters to be evaluated includes one or more of the following: pulse width, spot diameter, laser power, laser frequency, processing speed, and number of processing operations.

Images