Process parameter optimization method and system for selective laser melting copper alloy thin-wall part

Abstract

The invention discloses a technological parameter optimization method and system for a selective laser melting copper alloy thin-wall part. The method comprises the steps that firstly, multiple sets of technological parameters including laser power, the scanning speed, the scanning interval and the powder laying layer thickness are randomly generated; inputting each group of process parameters into a random forest prediction model integrated with physical constraints, and outputting a corresponding wall thickness deviation degree; and finally, based on the wall thickness deviation degree, optimal process parameters are obtained through screening by a particle swarm optimization algorithm. Wherein the random forest prediction model integrated with the physical constraints takes the process parameters and the wall thickness deviation degree as training samples, physical constraint judgment is added in the node splitting process of the decision tree, if the effective energy density of a sample corresponding to a child node exceeds a preset physical feasible interval, splitting is refused or the information gain of the child node is reduced, and if the effective energy density of the sample corresponding to the child node exceeds the preset physical feasible interval, the decision tree node is determined. Therefore, the physical rationality and reliability of model prediction are ensured. According to the method, efficient and accurate optimization of technological parameters is achieved, and the wall thickness forming precision of the copper alloy thin-wall part is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of additive manufacturing technology for copper alloys, and relates to a method and system for optimizing process parameters of thin-walled parts formed by laser selective melting of copper alloys with high thermal conductivity and high reflectivity. Background Technology

[0002] Selective Laser Melting (SLM), a core technology in metal additive manufacturing, directly forms complex structural parts by cladding metal powder layer by layer, demonstrating significant advantages in the manufacture of high-performance thin-walled components in aerospace, medical device, and automotive industries. However, the multi-physics coupling effects involved in SLM forming, such as the dynamic behavior of the molten pool, thermal stress evolution, and powder-laser interaction, make it difficult to stably control the dimensional accuracy of the formed parts. For copper alloys like CuCrZr(Nb) alloys, which have high thermal conductivity, low viscosity, and high laser reflectivity, their inherent low laser energy absorption rate and rapid heat dissipation characteristics significantly shorten the stable existence time of the molten pool, exacerbating the risk of melt collapse and edge warping. Studies have shown that SLM forming of such materials has a strict process window: when the energy input is insufficient, discontinuous fusion of the molten pool is prone to occur, forming spheroidization defects; while excessive energy input leads to overmelting, spattering, and thermal stress accumulation, causing dimensional deviations such as sidewall expansion of thin-walled structures. This means that the forming quality is extremely sensitive to energy input. Especially for thin-walled parts with a thickness of ≤1 mm, the micron-level dimensional deviations will directly affect the fit tolerances between components, severely restricting the application effect of this material in precision assembly scenarios.

[0003] Existing process optimization methods face three major technical bottlenecks: First, traditional single-factor experimental methods are unable to effectively capture the complex nonlinear coupling relationships and synergistic effects among multiple parameters such as laser power, scanning speed, and scanning spacing. Finding the global optimal solution in a high-dimensional parameter space is inefficient and heavily reliant on experience. Second, while pure data-driven machine learning methods (such as neural networks and support vector machines) can fit complex nonlinear relationships, they are severely disconnected from physical mechanisms, resulting in predictive results that lack interpretability and exhibit a "black box" problem, making it impossible to quantify the specific contribution of each process parameter to forming accuracy. Third, while physical simulation-based numerical optimization methods can reveal the dynamic evolution of the molten pool, their simulation accuracy for high-reflectivity copper alloys such as CuCrZr(Nb) is insufficient, and the computational cost is high, making it difficult to meet the rapid iteration needs of industrial sites. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problems of existing processes relying on manual experience to find optimal parameters, being disconnected from physical mechanisms, and having insufficient calculation accuracy for CuCrZr(Nb) alloys with high thermal conductivity and high laser reflectivity. This invention provides a method and system for optimizing process parameters of thin-walled copper alloy parts by laser selective melting.

[0005] To achieve the above objectives, the present invention employs the following technical solution: The first aspect of this invention provides a method for optimizing process parameters of laser selective melting of thin-walled copper alloy parts, comprising the following steps: Based on the actual process parameters, several sets of process parameters are randomly generated; the process parameters include laser power, scanning speed, scanning spacing, and powder layer thickness; Each set of process parameters is input into a random forest prediction model with integrated physical constraints, and the wall thickness deviation corresponding to each set of process parameters is output. Based on the wall thickness deviation corresponding to each set of process parameters, the optimal process parameters are found through particle swarm optimization algorithm. The integrated physical constraint random forest prediction model is trained using process parameters and wall thickness deviation as training set samples. When splitting decision tree nodes, the integrated physical constraint random forest prediction model adds constraint judgment: when splitting decision tree nodes, if the effective energy density of the sample corresponding to the child node exceeds the preset physical feasible range, the splitting of the child node is rejected or the information gain of the child node is reduced.

[0006] Furthermore, for cross-material data, the cross-material dataset is input into a trained random forest prediction model with integrated physical constraints to evaluate the cross-material prediction accuracy. If the cross-material prediction accuracy does not meet the requirements, then the domain adaptive optimization is initiated. The domain adaptive optimization uses DAC coefficients to quantize and correct the cross-material dataset to achieve model transfer.

[0007] Furthermore, the quantization and correction of the cross-material dataset using DAC coefficients is achieved through the following formula:

[0008] Among them, A intra-domain Indicates the laser absorption rate of the material within the domain; A cross-domain λ represents the laser absorptivity of the transdomain material. intra-domain Represents the thermal conductivity and λ of the material within the domain cross-domain This represents the thermal conductivity of the transdomain material; α and β are weighting coefficients.

[0009] Furthermore, the loss function of the integrated physical constraint random forest prediction model is:

[0010]

[0011] Among them, L total It is the total loss function; L MSE It is the mean squared error loss; R physics It is the physical regularization term; γ is the weighting coefficient of the regularization term; E i E is the effective energy density of the i-th sample; min It is the lower limit of the stable range of effective energy density; E max It is the upper limit of the stable range of effective energy density; ASP i It is the aspect ratio of the melt pool for the i-th sample; ASP min It is the stable lower limit of the width-to-depth ratio of the molten pool; ReLU(x)=max(0,x) is the linear rectification function.

[0012] Furthermore, the integrated physical constraint random forest prediction model optimizes the hyperparameters of the random forest through 5-fold cross-validation.

[0013] Furthermore, the effective energy density E effective The calculation formula is:

[0014] in, The absorption rate of the material to laser light. For laser power, For scanning speed, To ensure a thick powder layer.

[0015] Furthermore, the fitness function of the particle swarm optimization algorithm is: w1* +w2* +w3*

[0016] Among them, w1, w2 and w3 are all weighting coefficients; This is for wall thickness deviation; Indicates effective energy density; This indicates the stability of the molten pool.

[0017] Furthermore, the formula for calculating the wall thickness deviation is as follows:

[0018] in, For the design wall thickness, The wall thickness prediction values ​​for this set of process parameters are obtained by the random forest prediction model with integrated physical constraints. The formula for calculating molten pool stability is:

[0019] in, This represents the effective energy density value under the current combination of process parameters; This represents the ideal effective energy density value.

[0020] A second aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for optimizing process parameters of a laser selective melting alloy thin-walled part.

[0021] A third aspect of this invention provides a process parameter optimization system for laser selective melting of thin-walled copper alloy parts, comprising: The data generation module randomly generates several sets of process parameters based on the actual process parameters; the process parameters include laser power, scanning speed, scanning spacing and powder layer thickness. The wall thickness deviation prediction module inputs each set of process parameters into a random forest prediction model with integrated physical constraints, and outputs the wall thickness deviation corresponding to each set of process parameters. The process parameter optimization module, based on the wall thickness deviation corresponding to each set of process parameters, takes the minimum wall thickness deviation and physical mechanism constraints as optimization objectives, and finds the optimal process parameters through particle swarm optimization algorithm; The integrated physical constraint random forest prediction model is trained using process parameters and wall thickness deviation as training set samples. When splitting decision tree nodes, the integrated physical constraint random forest prediction model adds constraint judgment: when splitting decision tree nodes, if the effective energy density of the sample corresponding to the child node exceeds the preset physical feasible range, the splitting of the child node is rejected or the information gain of the child node is reduced.

[0022] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a method for optimizing process parameters in laser selective melting of thin-walled copper alloy parts. This method adds a physical constraint of effective energy density during node splitting in a random forest decision tree. If the effective energy density corresponding to a child node sample exceeds the physically feasible range for "laser melting and forming of copper alloys" (e.g., energy below the melting point leading to incomplete fusion, or energy above the over-melting threshold leading to spheroidization), the split is directly rejected or the information gain of that path is reduced. This design eliminates "non-physical prediction results" at the source, ensuring a high degree of match between the model's output "wall thickness deviation" and the actual forming process. Through a closed-loop process of random parameter generation, rapid model prediction, and particle swarm optimization, parameter selection and optimization can be completed without physical printing. By organically combining the particle swarm optimization algorithm with a random forest prediction model integrating physical constraints, this method can quickly find the optimal combination of process parameters that minimizes the wall thickness deviation from several randomly generated sets of process parameters (within ±5% of the effective energy density). This process significantly shortens the process development cycle for laser selective melting of thin-walled copper alloy parts. In the integrated physical constraint random forest prediction model, the physical constraint on effective energy density is applied throughout the entire decision tree node splitting process. Effective energy density is a key physical parameter in laser selective melting (SSM), directly affecting the morphology of the molten pool, the melting and solidification process of the material, and consequently significantly influencing the wall thickness and quality of thin-walled parts. By constraining the effective energy density, this method ensures the physical feasibility and rationality of the selected process parameters. Traditional process parameter optimization methods typically require extensive experiments to determine the optimal parameters, which not only consumes a significant amount of time but also increases costs related to raw materials, energy, and equipment. This method, however, significantly reduces the number of actual experiments by randomly generating process parameters, using a prediction model for rapid prediction, and employing a particle swarm optimization algorithm for efficient optimization. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is an optimized process parameter diagram for laser selective melting of thin-walled copper alloy parts according to an embodiment of the present invention; Figure 2 Print a physical image for Selective Laser Melting (SLM); Figure 3 This is a microscopic image of the molten pool morphology of a CuCrNb sample in an embodiment of the present invention. Figure 4This is a regression plot of the CuCrZr training set in an embodiment of the present invention; Detailed Implementation To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0025] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0026] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0027] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 This invention proposes a method for optimizing process parameters in laser selective melting of CuCrZr(Nb) alloy thin-walled parts based on physical mechanism constraints, domain adaptive correction coefficients, and particle swarm optimization ensemble algorithm. The aim is to rapidly establish the mapping relationship between process parameters and the forming dimensions of the alloy thin-walled parts through a machine learning model, achieving intelligent iterative optimization of process parameters and significantly improving the dimensional accuracy and stability of the formed parts. The specific steps are as follows: S1. Determine the SLM forming process parameters: The SLM forming process parameters include laser power, scanning speed, and scanning spacing. The laser power for SLM forming of CuCrZr(Nb) alloy thin-walled parts is 330 W~420 W, the scanning speed is 400~1000 mm / s, the scanning spacing is 0.08~0.1 mm, the interlayer rotation angle is 67°, and the powder layer thickness is 0.03 mm. CuCrZr and CuCrNb alloy thin-walled parts are prepared by full factorial experimental design.

[0028] S2. Perform SLM forming and data acquisition of CuCrZr(Nb) alloy thin-walled parts: S201. Based on the SLM forming process parameters determined in step one, prepare CuCrZr(Nb) alloy thin-walled parts with different laser powers, scanning speeds and scanning spacings. S202. The wall thickness of the formed thin-walled part is characterized using a micrometer screw gauge. An evaluation index for dimensional accuracy is established based on the deviation between the designed wall thickness and the measured wall thickness. The specific calculation formula is as follows:

[0029] in, For the design wall thickness, This represents the average value of the actual measured wall thickness.

[0030] S203. Metallographic sample preparation was carried out on the thin-walled CuCrZr(Nb) alloy sample. The cross-sectional morphology of the top molten pool was observed using an optical microscope. The width (W) and depth (D) of the molten pool were measured and calculated, and then the width-to-depth ratio (W / D) was calculated.

[0031] S204. Construct a structured dataset: using laser power, scanning speed, and scanning spacing as input features, and wall thickness deviation Δd as the output variable. Data for CuCrZr alloy is used for model training and validation, while data for CuCrNb alloy is specifically used for cross-material migration testing.

[0032] S3. Establish a random forest prediction model with integrated physical constraints: S301. The model is built using the Scikit-learn library. The input variables are laser power, scanning speed, and scanning spacing. The output variable is the wall thickness deviation. S302. Divide the dataset into a training set and a test set in a 7:3 ratio; S303. Use GridSearchCV for hyperparameter space traversal. Optimization scope includes: S303-1, The number of decision trees (n_estimators) is 100~500. Random forests solve the problem of high-dimensional nonlinear modeling by constructing multiple decision trees and using the nonlinear segmentation capability of the tree structure to capture the interaction effects between process parameters. During the growth of a single decision tree, node splitting achieves multi-dimensional parameter space segmentation by traversing feature combinations (such as the joint threshold of laser power and scanning speed), directly quantifying the synergistic or antagonistic effects between parameters. Integrating multiple decision trees reduces model variance through the Bagging strategy and enhances the generalization ability to nonlinear coupling relationships by utilizing a multi-tree voting mechanism. The number of trees is positively correlated with model stability, but the diminishing marginal returns are significant: when the number of trees exceeds a critical value (usually 200~300 trees), the improvement in model accuracy tends to plateau. Predicting the dimensional accuracy of thin-walled CuCrZr alloy parts requires balancing model complexity and computational efficiency. Preliminary experiments show that increasing n_estimators from 100 to 300 reduces the test set RMSE by 29%; from 300 to 500, the reduction is only 5.3%; and after >500, the accuracy shows almost no improvement, while training time increases linearly. Further calculation of the interaction effect strength using OOB (Out-of-Bag) error reveals that the interaction strength index between scanning speed and scanning spacing is 0.324 (a threshold >0.25 is considered strong interaction), confirming significant nonlinear coupling between process parameters. Therefore, setting 100–500 samples to cover the potential optimal range (300±200) can balance interaction effect modeling with industrial real-time requirements.

[0033] S303-2, maximum depth is 10~25. The depth of the decision tree determines the model's ability to fit nonlinear relationships. Too shallow (<10 layers) will lead to underfitting and failure to capture the interaction effects of process parameters; too deep (>25 layers) will easily introduce noise sensitivity and reduce generalization. In the SLM forming of CuCrZr(Nb) alloy, the interaction between laser power and scanning speed can be decomposed into 3rd to 5th order nonlinear terms, requiring a depth of at least 10 layers to characterize. In the measured data, when max_depth=15, the model's prediction error is the smallest (MAPE=4.97%). Therefore, based on the number of features and sample size, pre-training revealed that RMSE>0.15 when the depth is <10, and the RMSE reaches a minimum plateau when the depth is 15~20. Therefore, a depth of 10~25 layers is set to cover the optimal depth range.

[0034] S303-3, Minimum number of samples for feature splitting (min_samples_split) is 2~15. This parameter controls the minimum number of samples for node splitting; the smaller the value, the more complex the tree. For small sample datasets (n<1000), setting it to 2~20 can avoid over-subdivision. In this experiment, the sample size of the CuCrZr thin-walled parts is small. If min_samples_split<2, it will lead to overfitting of leaf nodes; >15 will fail to capture high gradient changes. Therefore, exponential interval sampling (2,5,10,15) is adopted to prioritize covering the subdivision interval (2~10) that is sensitive to the dynamics of the copper alloy molten pool, while including a conservative segmentation threshold (15).

[0035] S303-4. The minimum number of samples required for each leaf node (min_samples_leaf) is 1 to 4. This parameter prevents local overfitting caused by insufficient leaf node samples. For regression problems, it is usually set to 1 to 5 to balance smoothness and fluctuation suppression. CuCrZr alloy dimensional deviations exhibit significant process fluctuations, requiring leaf nodes to retain a small number of samples to characterize random errors. However, min_samples_leaf > 4 will result in over-smoothing, failing to resolve the difference between scan spacings of 0.08 mm and 0.1 mm. Therefore, an integer from 1 to 4 is chosen, with a focus on 1 and 3, and the optimal value is determined through grid search.

[0036] S304. During the training of the random forest, the effective energy density E effective The node splitting is embedded in the model as a core physical criterion. Specifically, when splitting at each node of the decision tree, the E0 of all samples of the candidate child nodes is calculated. effective Value, if it exceeds 15% of the sample E effective The value exceeds the feasible window determined by the material properties [E] min E max If the split direction is rejected, or a penalty factor (such as 0.7) is multiplied in its information gain, thereby guiding the model to prioritize physically consistent split paths.

[0037] Add a physical regularization term to the loss function:

[0038]

[0039] Among them, L total It is the total loss function; L MSE It is the mean squared error loss; R physics It is the physical regularization term; γ is the weighting coefficient of the regularization term; E i E is the effective energy density of the i-th sample; min It is the lower limit of the stable range of effective energy density; E maxIt is the upper limit of the stable range of effective energy density; ASP i It is the aspect ratio of the melt pool for the i-th sample; ASP min It is the lower bound of the weld pool width-to-depth ratio; ReLU(x) = max(0,x) is the linear rectification function, which only applies when the prediction violates the physical constraints (E0). effective Penalties will only be incurred when the range is exceeded or the ASP is too high.

[0040] S305. Model Training and Validation: The hyperparameters of the random forest, including the number of decision trees, maximum depth, and feature subset ratio, were optimized using 5-fold cross-validation to ensure model generalization ability. To address the issue of process parameter fluctuations in industrial scenarios, ±5% random perturbation of laser power / scanning speed was introduced into the test set, and validation was performed using CuCrZr alloy powder from different batches and suppliers. Based on feature importance analysis, the contribution of parameters such as laser power and the interaction term between scanning speed and scanning spacing to dimensional deviations was quantified, overcoming other interpretability challenges of the model. Finally, the model accuracy was evaluated using the root mean square error (RMSE) and mean absolute percentage error (MAPE) of the test set, calculated using the following formulas: The formula for root mean square error (RMSE) is:

[0041] The formula for Mean Absolute Percentage Error (MAPE) is:

[0042] in, This is a predicted value; This is the actual value.

[0043] S4. Cross-Material Transfer and Domain Adaptive Optimization: In this step, cross-material transfer learning is performed by introducing DAC coefficients to optimize the model's adaptability across different alloy systems. The specific process is as follows: (1) Material Property Embedding Vector Construction: A feature vector containing key physical properties is constructed for each material as additional information for model input. This vector integrates the intrinsic material properties that have a decisive influence on the behavior of the molten pool and the accuracy of the forming dimensions, including thermal conductivity λ and laser absorptivity A. Taking the CuCrZr alloy used in this example as an example, its thermal conductivity λ = 101 W / m·K and laser absorptivity A = 26%, the constructed property embedding vector is [101, 0.26]. This vector is input together with the process parameters during model training and prediction to establish an explicit correlation between material properties and forming accuracy, providing a physical basis for cross-material process transfer.

[0044] (2) DAC coefficient learning: DAC based on sample data is introduced to automatically correct the influence of physical property differences between different materials on the model. The joint optimization method of minimizing the objective loss function and the distribution alignment regularization term is adopted, and model sharing of different alloy systems is achieved through transfer learning.

[0045] (3) DAC calculation formula: The DAC coefficient is obtained through learning based on a small number of experimental samples of the target material. The weight of physical properties on the model prediction is adjusted to ensure cross-domain adaptation of physical characteristics between materials. For example, when DAC>1, it indicates that domain adaptive optimization (fine-tuning) is needed to correct the model to adapt to the differences in properties between materials.

[0046]

[0047] Among them, A intra-domain and A cross-domain λ represents the laser absorption rate of intradomain materials (e.g., CuCrZr) and transdomain materials (e.g., CuCrNb), respectively; intra-domain and λ cross-domain α and β represent the thermal conductivity of materials within and across domains, respectively; α and β are weighting coefficients, which are learned through fine-tuning of the model on cross-domain data and satisfy α+β=1. These coefficients are used to automatically correct for the effects caused by differences in material physical properties when migrating process parameters.

[0048] S5. Based on model training and validation, the PSO algorithm is used to optimize the process parameters in multiple objectives, including minimizing wall thickness deviation and maximizing the rationality of effective energy density.

[0049] Molten pool stability criteria: determined by the molten pool width-to-depth ratio and effective energy density E. effective Whether it is within the preset stable range, excluding parameter combinations that are prone to defects or incomplete fusion; Research has shown that when the width-to-depth ratio (width / depth) of the molten pool is greater than 2, the molten pool is shallow and wide semi-circular, which is considered to be stable in shape and can form a continuous and uniform molten channel. Effective energy density E effective Constraints are calculated and determined based on the following formula:

[0050] Where A is the laser absorptivity of the material, P is the laser power, v is the scanning speed, and h is the thickness of the powder layer.

[0051] Optimization process: Each particle represents a combination of process parameters (laser power, scanning speed, scanning spacing, etc.). In the optimization process, the effective energy density E is introduced. effectiveConstraints and molten pool stability are used as conditions to ensure that the combination of process parameters remains within physically feasible limits. In the fitness function of the particle swarm optimization algorithm, the effective energy density E... effective As a core constraint, its fitness function is designed as: Fitness = w1 * (wall thickness deviation) + w2 * (effective energy density E) effective Penalty term) + w3 * (Molten pool stability penalty term). Where, effective energy density E effective The penalty term is calculated in the E value of the parameter combination. effective The maximum value is taken when the value exceeds the optimal window for the material; otherwise, it is zero. This forces the optimization process to satisfy the energy input criterion determined by the thermal-optical properties of the material while searching for high-precision parameters, thereby enhancing the model's global optimization capability. Particle swarm optimization optimizes the position of each particle in each iteration through backpropagation, ultimately finding the Pareto optimal solution set to obtain the best combination of process parameters for multi-objective optimization.

[0052] Its optimization variables (decision variables) are:

[0053] The objective function is to minimize the weighted sum of the following two metrics:

[0054] Minimize wall thickness deviation as follows:

[0055] Where, d a To design the wall thickness, d m The predicted wall thickness is the value obtained from the prediction model for this set of parameters.

[0056] Rationalized effective energy density E effective The deviation from the ideal interval is:

[0057] Among them, E effective This is the calculated effective energy density value under the current combination of process parameters; E opt The value is the material-specific ideal effective energy density determined based on a large number of experiments, and it is approximately 4.3 J / mm².

[0058] S6. Dimensional Accuracy Prediction and Model Verification of Laser Selective Melting (SLM) Formed Parts. The established random forest prediction model was used to predict the dimensional accuracy of thin-walled CuCrZr(Nb) alloy parts. Five sets of SLM process parameters were randomly input into the random forest prediction model. CuCrZr(Nb) alloy thin-walled part samples were formed according to these parameters. The difference between the predicted and actual measured values ​​of the formed wall thickness deviation was compared to determine the effectiveness and accuracy of the dimensional accuracy prediction model.

[0059] S7. By integrating physical mechanism constraints, domain adaptive correction coefficients, and the feedback mechanism of particle swarm optimization (PSO) algorithm, intelligent optimization of process parameters is achieved, improving the prediction accuracy of cross-material models. Simultaneously, the PSO algorithm optimizes the optimal combination of process parameters through back-iteration of the objective function. Thin-walled parts are re-fabricated using the optimized process parameters, ensuring that the measured wall thickness deviation is controlled within ±10% of the design value, verifying the model's ability to capture the nonlinear compensation effect of parameters.

[0060] In its implementation, this invention first uses a full-factor experimental design to obtain the influence of SLM process parameters such as laser power, scanning speed, and scanning spacing on the dimensional accuracy of CuCrZr alloy thin-walled parts, constructing a structured dataset of process parameters and dimensional accuracy. Based on this dataset, a random forest algorithm is used to establish a predictive model for the dimensional accuracy of CuCrZr alloy forming. During model training, physical mechanisms such as melt pool stability criteria and energy density constraints are embedded, and the contribution weight of each process parameter is quantified through feature importance analysis. Furthermore, the process parameters of CuCrNb material are mapped to the feature space using DAC coefficients to complete cross-material transfer. Finally, the PSO algorithm is integrated for multi-objective parameter optimization, obtaining the globally optimal combination of process parameters under the condition of satisfying process constraints, thereby achieving high-precision forming of CuCrZr(Nb) alloy thin-walled parts.

[0061] One embodiment of the present invention provides a method for optimizing process parameters of laser selective melting of thin-walled alloy parts. The specific process is as follows: (1) The alloys used in this implementation process were Cu-0.8Cr-0.2Zr (wt.%) and Cu-3.17Cr-2.48Nb (wt.%) alloys, which were subjected to SLM forming. See [link to relevant documentation]. Figure 2 The steps are as follows: ① The CuCrZr(Nb) alloy powder was dried at 80℃ for 5 hours in a vacuum drying oven; ② Establish a dimensional model of the formed part, with dimensions of 50 mm × 1 mm × 20 mm. Set the SLM process parameters, including laser power, scanning speed, powder layer thickness, scanning spacing, interlayer rotation angle, and scanning strategy. The laser power range is 330-420 W, the scanning speed range is 400-1000 mm / s, the scanning spacing range is 0.08-0.1 mm, the powder layer thickness is 0.03 mm, the interlayer rotation angle is 67°, and the laser scanning strategy is a straight line. The process parameter range is set to ensure the effective energy density E. effective The goal was to promote the formation of stable, shallow, and wide semi-circular molten pools with a width-to-depth ratio >2, falling within the empirically stable range of [2.1, 13.5] J / mm². A total of 168 CuCrZr alloy thin-walled parts and 26 CuCrNb alloy thin-walled parts were prepared using a full factorial experimental design. ③ Import the sliced ​​printing model into the forming equipment and preheat the substrate to 100℃; ④ Take a sample after printing; (2) The wall thickness of the formed thin-walled parts was measured using a micrometer with an accuracy of 0.001 mm. Five points were measured evenly along the length of each sample, and the average value was taken as the measured wall thickness. Metallographic preparation (sampling, mounting, polishing, and etching) was performed on the CuCrZr(Nb) alloy thin-walled part samples. The cross-sectional morphology of the molten pool was observed using an optical microscope, such as... Figure 3 As shown, the molten pool at three different locations was measured for each sample, and the average value was taken as the molten pool morphology characteristic measurement under that process parameter. The measurement results (W, D) and the corresponding process parameters (P, v, h) were recorded together and included in the machine learning dataset.

[0062] (3) Preprocess the machine learning dataset to remove outliers. Based on the SLM forming mechanism and combined with the analysis of the molten pool morphology, samples with obvious process defects (such as spheroidization, lack of fusion, key holes, etc.) were removed. After the above processing, 5 outlier samples were removed from the 168 sets of CuCrZr raw data, and 163 sets of valid data were finally retained for model training. There were no outlier samples in the CuCrNb data, and all 26 sets of data were valid data.

[0063] (4) Establish a random forest prediction model with integrated physical constraints. The construction process includes the following technical details: ① Input layer and feature definition: The input variables are laser power (330-420 W), scanning speed (400-1000 mm / s) and scanning spacing (0.08-0.1 mm), and the output variable is wall thickness deviation Δd (%). ② Divide the data into training and test sets: Divide the input and output CuCrZr machine learning dataset into training and test sets with a ratio of 70% and 30% respectively, and use the training set to train the random forest model; ③ Decision tree generation mechanism: The Random Forest algorithm from the Scikit-learn library is used, which extracts a subset of samples with replacement from the training set through bootstrap sampling. Each decision tree is generated based on the CART algorithm. When splitting a node, in addition to minimizing the mean squared error (RMSE), a physical constraint is added: if more than 15% of the samples after splitting have a lower mean squared error (RMSE), the decision tree is generated based on the mean squared error (RMSE) of the target tree. effective The value exceeds the feasible window determined by the material properties [E] min E max If the split direction is not found, the split direction is rejected, or a penalty factor (such as 0.7) is multiplied in its information gain to guide the model to prioritize physically consistent split paths. The feature space is recursively partitioned until a preset termination condition is reached (such as a maximum depth of 15 layers or a leaf node sample number ≥ 2). When generating a single tree, each node selects the optimal split point from only 2 randomly selected features (in this embodiment, the number of input features is 4) to enhance the model's generalization ability. ④ Hyperparameter Optimization and Cross-Validation: A 5-fold cross-validation method combined with GridSearchCV is used to traverse the hyperparameter space (number of decision trees 100-500, maximum depth 5-20, feature subset ratio 1.0-2.0). Simultaneously, a physical regularization term is introduced into the cross-validation loss function to guide the hyperparameters towards physical consistency. The physical regularization term R... physics Defined as the effective energy density E in the test sample effective Deviation from the stable range [E] min E max The severity of the [data] was assessed. In the total loss function, the weight coefficient γ was determined to be 0.112 through grid search. The results showed an average RMSE of 0.065 ± 0.015, indicating high stability of the model across different data subsets. The optimal parameter combination was finally determined as follows: 200 decision trees, maximum depth 15, minimum number of samples for feature splits 5, and minimum number of samples for leaf nodes 2. ⑤ Feature Importance Analysis: Evaluate the impact of each process parameter on the wall thickness deviation of CuCrZr alloy thin-walled parts. The influence weights (%) show that the interaction effect plays a dominant role, accounting for 47.75%, of which the speed-spacing interaction accounts for 36.04%, meaning that the synergistic change of scanning speed and scanning spacing is the strongest control factor for dimensional accuracy; the nonlinear effect accounts for 27.49%, of which the quadratic term of power accounts for 17.41%, indicating that there is a parabolic relationship between power and wall thickness deviation, i.e., there is an optimal power window, and the quadratic term of scanning speed accounts for 8.95%, indicating that both high-speed and low-speed regions will aggravate dimensional deviation; the main effect distribution accounts for 24.76%, of which power accounts for 16.15%, scanning speed accounts for 7.51%, and spacing accounts for 1.11%. The reason why the main effect of spacing is weak may be that the influence of spacing is completely covered by the interaction term.

[0064] (5) Model validation and robustness testing, see Figure 4 : ① Test set accuracy verification: The CuCrZr model achieved RMSE=0.05 and MAPE=4.88% on the independent test set, which meets the accuracy requirements of claim 5, indicating that the model has good accuracy. The obtained model is used as the prediction model for forming size accuracy. ② Process disturbance test: ±5% random disturbance of laser power / scanning speed was introduced into the test data, and three batches of CuCrZr powder from different suppliers (composition deviation ≤0.5 wt%) were used for verification. The model prediction error remained stable within ±6% (MAPE≤5.8%), which proved that the method is robust to parameter fluctuations and material batch differences.

[0065] ③ Physical consistency verification: Perform posterior analysis on all prediction results in the test set and calculate the corresponding effective energy density E. effective It was found that 98.5% of the predicted samples had an E effective The value falls within a reasonable range of [2.1, 13.3] J / mm², which is significantly better than the baseline model without physical constraints, proving the effectiveness of the physical constraints.

[0066] (6) Cross-material model transfer: ① Direct transfer performance analysis: When the CuCrNb dataset is input into the trained CuCrZr model, the mean absolute percentage error (MAPE) of the test set is 14.57%, which is greater than 10%, indicating a certain "domain shift". This suggests that the difference in thermophysical properties between the two alloys (CuCrZr: λ=101 W / mK, A=26%; CuCrNb: λ=143 W / mK, A=33%) is not negligible. ② DAC coefficient quantifies material differences: Based on the DAC coefficient formula, the DAC coefficient (α=0.831, β=0.169) is introduced to quantify the difference in thermal conductivity and absorptivity. After correction, DAC=0.831*(143 / 101)+0.169*(26 / 33)≈1.31>1, indicating that there are certain differences in physical properties between the two materials. Therefore, domain adaptive optimization (fine-tuning) is required to adapt to the differences between materials.

[0067] ③ Statistical Distribution Alignment: Z-score standardization is used to align the statistical distribution of CuCrNb data to the feature space of CuCrZr, reducing distribution differences. The specific steps are: calculate the mean μ and standard deviation σ of CuCrZr (on the training set); standardize the CuCrNb data using Z-score standardization.

[0068] ④ Domain Adaptive Fine-tuning: 90% of the decision tree structure of the CuCrZr model was frozen, and 10% of the decision tree parameters were fine-tuned. The complexity was controlled by early stopping to avoid overfitting in the CuCrNb domain. The PSO algorithm was combined to optimize the adaptability of cross-material process parameters. Finally, the MAPE of the CuCrNb test set was reduced to 8.76%, which shows the significant effect of the integration of domain adaptation and PSO in cross-material migration.

[0069] (7) Determining the effectiveness and accuracy of the prediction model: Five sets of process parameters were randomly selected and input into the CuCrZr model to predict the wall thickness deviation of thin-walled parts for these five sets of process parameters. Then, thin-walled alloy parts of CuCrZr and CuCrNb were printed according to the corresponding process parameters. The average difference between the predicted and actual measured values ​​of the wall thickness deviation for these five sets of process parameters was calculated to verify the effectiveness and accuracy of the model. Finally, the average difference between the predicted and actual measured values ​​of the wall thickness deviation for CuCrZr was 3.23±0.33%, and the average difference between the predicted and actual measured values ​​of the wall thickness deviation for CuCrNb was 5.36±0.54%.

[0070] (8) Based on the introduction of physical constraints and DAC correction, the PSO algorithm is used for multi-objective optimization. According to the objective function formula of claim 8, the optimization objective is x=[P, v, h], and the constraints are: 330≤P≤420, 400≤v≤1000, 0.08≤h≤0.1, and the effective energy density E must be satisfied. effectiveThe aspect ratio of the molten pool is within the empirical range (physical constraint). PSO parameters are set as follows: particle number 50, maximum iterations 200, and inertia weight decreasing linearly from 0.8 to 0.3. Rapid optimization yields the optimal SLM process parameter combination for thin-walled CuCZr(Nb) alloy parts. CuCrZr(Nb) alloy thin-walled part samples are printed according to the optimized SLM process parameters, and the wall thickness deviation is obtained. It is found that the wall thickness deviation of the CuCrZr sample after optimization is 2.70%, and that of the CuCrNb sample is 3.87%. Furthermore, the effective energy density E corresponding to all optimized parameters is... effective All results were stable within the empirical range, achieving high-precision forming of thin-walled CuCrZr(Nb) alloy parts.

[0071] See Figure 1 One embodiment of the present invention provides a process parameter optimization system for laser selective melting of thin-walled copper alloy parts, comprising: The data generation module randomly generates several sets of process parameters based on the actual process parameters; the process parameters include laser power, scanning speed, scanning spacing and powder layer thickness. The wall thickness deviation prediction module inputs each set of process parameters into a random forest prediction model with integrated physical constraints, and outputs the wall thickness deviation corresponding to each set of process parameters. The process parameter optimization module, based on the wall thickness deviation corresponding to each set of process parameters, takes the minimum wall thickness deviation and physical mechanism constraints as optimization objectives, and finds the optimal process parameters through particle swarm optimization algorithm; The integrated physical constraint random forest prediction model is trained using process parameters and wall thickness deviation as training set samples. When splitting decision tree nodes, the integrated physical constraint random forest prediction model adds constraint judgment: when splitting decision tree nodes, if the effective energy density of the sample corresponding to the child node exceeds the preset physical feasible range, the splitting of the child node is rejected or the information gain of the child node is reduced.

[0072] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device; it can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that more specific examples (a non-exhaustive list) of the computer-readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0073] Computer-readable storage media also include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0074] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0075] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the process parameter optimization method for laser selective melting of thin-walled copper alloy parts in the above embodiments.

[0076] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., 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 process parameters of laser selective melting of a thin-walled copper alloy part, characterized in that, The method comprises the following steps: a plurality of sets of process parameters are randomly generated according to actual process parameters; the process parameters comprise laser power, scanning speed, scanning interval and powder layer thickness; each set of process parameters is input into an integrated physical constraint random forest prediction model to output a wall thickness deviation degree corresponding to each set of process parameters; optimal process parameters are found through a particle swarm optimization algorithm based on the wall thickness deviation degree corresponding to each set of process parameters; the integrated physical constraint random forest prediction model is trained by using process parameters and wall thickness deviation degrees as training set samples; when a decision tree node is split, a constraint condition is added for judgment: if the effective energy density of the samples corresponding to the child node exceeds a preset physically feasible interval, the splitting of the child node is refused or the information gain of the child node is reduced.

2. The process parameter optimization method for laser selective melting of thin-walled alloy parts according to claim 1, wherein, For cross-material data, the cross-material data set is input into the trained integrated physical constraint random forest prediction model to evaluate cross-material prediction accuracy; if the cross-material prediction accuracy does not meet the requirements, domain adaptive optimization is started; the domain adaptive optimization quantifies and corrects the cross-material data set by using DAC coefficients to realize model migration.

3. The process parameter optimization method for laser selective melting of thin-walled alloy parts according to claim 1, wherein, The quantification and correction of the cross-material data set by using DAC coefficients are realized by the following formula: where A intra-domain represents the laser absorption of the in-domain material; A cross-domain represents the laser absorption of the cross-domain material; λ intra-domain represents the thermal conductivity of the in-domain material and λ cross-domain represents the thermal conductivity of the cross-domain material; and α and β are weight coefficients.

4. The process parameter optimization method for selective laser melting of thin-walled alloy parts according to claim 1, wherein, The loss function of the integrated physical constraint random forest prediction model is: where L total is the total loss function; L MSE is the mean squared error loss; R physics is the physical regularizer; γ is the weight coefficient of the regularizer; E i is the effective energy density of the i-th sample; E min is the lower bound of the stable interval of the effective energy density; E max is the upper bound of the stable interval of the effective energy density; ASP i is the width-to-depth ratio of the molten pool of the i-th sample; ASP min is the lower bound of the stable interval of the width-to-depth ratio; ReLU(x) = max(0, x) is the linear rectifier function.

5. The process parameter optimization method for selective laser melting of thin-walled alloy parts according to claim 1, wherein, The integrated physical constraint random forest prediction model optimizes the hyperparameters of the random forest through 5-fold cross-validation.

6. The process parameter optimization method for selective laser melting of thin-walled alloy parts according to claim 1, wherein, The effective energy density E effective The formula for calculating is: wherein, is the absorption of the material to the laser, is the laser power, is the scanning speed, is the powder layer thickness.

7. The process parameter optimization method for laser selective melting of thin-walled alloy parts according to claim 1, wherein, The fitness function of the particle swarm optimization algorithm is: w1* +w2* +w3* wherein w1, w2 and w3 are weight coefficients; for wall thickness deviation; represents effective energy density; represents molten pool stability.

8. The process parameter optimization method for laser selective melting of thin-walled alloy parts according to claim 7, wherein, The calculation formula of the wall thickness deviation is: wherein, is the design wall thickness, is the wall thickness prediction value under the set of process parameters derived from the integrated physics-constrained random forest prediction model; The calculation formula of the molten pool stability is: wherein is the effective energy density value for the current process parameter combination; is the ideal effective energy density value.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the process parameter optimization method for laser selective melting of alloy thin-walled parts according to any one of claims 1-8. 10.A system for optimizing process parameters of selective laser melting of a thin-walled copper alloy part, characterized in that, It comprises: a data generation module that randomly generates a plurality of sets of process parameters according to actual process parameters; the process parameters comprise laser power, scanning speed, scanning interval and powder layer thickness; a wall thickness deviation degree prediction module that inputs each set of process parameters into an integrated physical constraint random forest prediction model to output a wall thickness deviation degree corresponding to each set of process parameters; a process parameter optimization module that finds optimal process parameters through a particle swarm optimization algorithm based on the wall thickness deviation degree corresponding to each set of process parameters, with minimum wall thickness deviation degree and physical mechanism constraint as optimization objectives; the integrated physical constraint random forest prediction model is trained by using process parameters and wall thickness deviation degrees as training set samples; when a decision tree node is split, a constraint condition is added for judgment: if the effective energy density of the samples corresponding to the child node exceeds a preset physically feasible interval, the splitting of the child node is refused or the information gain of the child node is reduced.

Images