Additive manufacturing parameter optimization method and system using machine learning, electronic equipment and medium

Abstract

The invention discloses an additive manufacturing parameter optimization method using machine learning. The method comprises the following steps: acquiring a first cladding layer surface topography constructed by powder layers with different powder characteristics under different powder laying parameters, and establishing a first database at least mapping a relationship among the powder laying parameters, the powder characteristics and the first cladding layer surface topography; and according to the first database, a first machine learning model is constructed and trained to generate a first application model used for powder laying parameter prediction, and then the first application model is used for predicting input powder characteristics to output optimal powder laying parameters used for additive manufacturing so as to optimize the additive manufacturing quality.

Description

Technical Field

[0001] The present disclosure relates to the field of additive manufacturing, and more particularly to a method, system, electronic device, and storage medium for optimizing powder spreading parameters using machine learning. Background Art

[0002] Additive manufacturing (AM) encompasses a variety of process types. Among them, powder bed-based additive manufacturing technologies, such as laser powder bed fusion (LPBF) technology, are one of the current research and application hotspots. LPBF manufactures parts with complex geometries by melting and solidifying metal powder layer by layer with a high-energy laser beam.

[0003] Numerous studies have shown that by optimizing the powder spreading parameters, a continuous and uniform powder bed can be obtained, reducing the surface roughness of the powder bed, and thus obtaining a high-quality powder layer. However, to obtain the optimal powder spreading parameters, a large number of powder spreading experiments often need to be carried out. At the same time, the quality of the powder bed not only depends on the powder spreading parameters, but also powder characteristics and powder spreading methods are key factors affecting the quality of the powder bed. With the rapid development of additive manufacturing, the preparation methods of metal powders are also developing towards diversification to meet different needs. Powders prepared by different methods have differences in their particle morphology, particle size distribution and other characteristics. Therefore, it is particularly important to clarify the influence mechanism of powder characteristics, powder spreading methods, and powder spreading parameters on the quality of the powder bed. However, currently, for LPBF-formed parts of the same material, manufacturers all use unified powder spreading parameters for preparation, ignoring the influence of powders with different characteristics on the quality of the powder bed, resulting in a decline in the quality of some powder beds and affecting the quality of the final parts.

[0004] On the other hand, when using powder materials with different characteristics, traditional methods are difficult to fully consider the different performances of these materials in terms of the surface morphology and forming quality of the cladding layer. Material characteristics such as powder preparation process, particle shape, and particle size will significantly affect the printing quality, making the process of comprehensively optimizing the powder spreading parameters complex and challenging.

[0005] To address these problems, introducing intelligent technologies such as machine learning has become the direction of the solution, which can accelerate the parameter optimization speed and improve the stability of the forming results. Summary of the Invention

[0006] The present disclosure proposes a method, system, electronic device, and medium for optimizing additive manufacturing parameters using machine learning to optimize the quality of additive manufacturing.

[0007] In a first aspect, the present disclosure provides a method for optimizing additive manufacturing parameters using machine learning, including: obtaining the surface topography of a first clad layer constructed from powder layers with different powder characteristics under different powder spreading parameters, and establishing a first database that at least maps the relationship between the powder spreading parameters, powder characteristics, and the surface topography of the first clad layer; constructing and training a first machine learning model based on the first database to generate a first application model for predicting powder spreading parameters; and using the first application model to predict the input powder characteristics to output the optimal powder spreading parameters for additive manufacturing.

[0008] Preferably according to the first aspect, the powder spreading parameters include the powder spreading speed and layer thickness.

[0009] Preferably according to the first aspect, the powder characteristics include one or more of the powder manufacturing process, powder material, particle shape, and powder particle size.

[0010] Preferably according to the first aspect, the powder manufacturing process includes one or more categories of atomization, mechanical pulverization, reduction, and electrolysis, wherein the atomization includes gas atomization and plasma rotating electrode atomization.

[0011] Preferably according to the first aspect, the method further includes: capturing images and performing discrete element simulations on the powder pile movement state during the laying process of the powder layer with different powder characteristics under different powder spreading parameters; calibrating the data trend of the discrete element simulation based on the results of the image capture; and correlating the calibrated powder pile movement state of the discrete element simulation with the powder spreading parameters, powder characteristics, and the surface topography of the first clad layer to establish the first database.

[0012] Preferably according to the first aspect, the method further includes: performing discrete element simulations on the surface topography of the first clad layer; and validating the first application model based on the powder pile movement state and the surface topography of the first clad layer obtained from the discrete element simulation.

[0013] Preferably according to the first aspect, the surface topography of the first clad layer is constructed under the same scanning parameters, and the scanning parameters include one or more of laser power, scanning speed, spot diameter, and scanning trajectory.

[0014] Preferably according to the first aspect, the method further includes: obtaining the surface topography of a second clad layer constructed from powder layers with different powder characteristics under different scanning parameters, and establishing a second database that at least maps the relationship between the scanning parameters, powder characteristics, and the surface topography of the second clad layer, wherein the surface topography of the second clad layer is constructed under the optimal powder spreading parameters corresponding to the powder characteristics output from the first application model; constructing and training a second machine learning model based on the second database to generate a second application model for predicting scanning parameters; and using the second application model to predict the input powder characteristics to output the optimal scanning parameters for additive manufacturing.

[0015] Preferably, according to the first aspect, the method further includes: separately establishing unimodal databases corresponding to process parameters, formed surface quality, bead continuity, and mechanical properties, where the process parameters include the powder spreading parameters and scanning parameters, and the formed surface quality includes the surface topography of the first clad layer and the surface topography of the second clad layer; performing early fusion on the unimodal databases of the process parameters and the formed surface quality to form a multimodal database, and respectively constructing a bead continuity prediction model and a mechanical property prediction model according to the unimodal databases of the bead continuity and the mechanical properties; performing late fusion on the prediction results of the bead continuity prediction model and the mechanical property prediction model and extracting multimodal features according to the multimodal database and the late fusion data; respectively constructing sub-prediction models of process parameters and formed surface quality, process parameters and bead continuity, and process parameters and mechanical properties according to the multimodal features; and jointly training the respectively constructed sub-prediction models to construct a comprehensive prediction model for comprehensively predicting and optimizing the material properties of additive manufacturing.

[0016] In a second aspect, the present disclosure provides an additive manufacturing parameter optimization system using machine learning, including: a first clad layer surface topography acquisition module for acquiring the surface topography of a first clad layer constructed from powder layers with different powder characteristics under different powder spreading parameters; a first database construction module for establishing a first database that at least maps the relationship between the powder spreading parameters, the powder characteristics, and the surface topography of the first clad layer; a first model training module for constructing and training a first machine learning model according to the first database to generate a first application model for predicting powder spreading parameters; and a powder spreading parameter prediction module for using the first application model to predict the input powder characteristics to output the optimal powder spreading parameters for additive manufacturing.

[0017] In a third aspect, the present disclosure provides an electronic device, including: at least one processor; at least one memory coupled to the at least one processor and configured to store instructions executed by the at least one processor, which when executed by the at least one processor cause the electronic device to execute the method according to any one of the first aspect.

[0018] In a fourth aspect, the present disclosure provides a computer-readable storage medium storing a computer program, which when executed by a processor implements the method according to any one of the first aspect.

[0019] It should be understood that the above general description and the following detailed description are only exemplary and do not limit the present disclosure. Description of the Drawings

[0020] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate one or more embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure and enable a person of ordinary skill in the relevant art to make and use the present disclosure.

[0021] Figure 1 is a schematic flow diagram of the powder spreading parameter optimization method using machine learning;

[0022] Figure 2 is a schematic diagram of the construction and application of the first machine learning model;

[0023] Figure 3 is the morphological characterization of three types of TC4 powders by scanning electron microscopy (SEM) in the roll powder spreading mode, where (a) is the powder prepared by the GA method, and (b) and (c) are the powders prepared by the PREP method;

[0024] Figures 4 to 6 are the surface morphologies of the formed specimens under different powder spreading parameters, where Figure 4 (a) to (e) are the optical microscope images of the clad layers of GA0-53 powder, Figure 5 (a) to (e) are the optical microscope images of the clad layers of PREP15-45 powder, Figure 6 (a) to (e) are the optical microscope images of the clad layers of PREP15-53 powder;

[0025] Figure 7 is the SVM optimization diagram of the powder spreading parameters for the powder prepared by the GA method;

[0026] Figure 8 is the surface quality morphology diagram of the powder bed of GA powder under different powder spreading parameters;

[0027] Figure 9 is the surface quality morphology diagram of the powder bed of PREP15-45 powder under different powder spreading parameters;

[0028] Figure 10 is the surface quality morphology diagram of the powder bed of PREP15-53 powder under different powder spreading parameters;

[0029] Figure 11 is the schematic flow diagram of the supplementary construction of the first database;

[0030] Figure 12 is the schematic flow diagram of the verification of the first application model;

[0031] Figure 13 is the state of GA powder observed by high-speed photography at different times under different powder spreading speeds;

[0032] Figure 14It is the powder spreading process of PREP15-45 powder under different powder spreading speeds captured by high-speed photography;

[0033] Figure 15 It is the powder spreading process of PREP15-53 powder under different powder spreading speeds captured by high-speed photography;

[0034] Figure 16 It is the schematic diagram of the angle of repose constructed, where (a) is the angle of repose of PREP15-45, and (b) is the angle of repose of PREP15-53;

[0035] Figure 17 It is the spreading process of PREP15-45 powder under different powder spreading parameters;

[0036] Figure 18 It is the surface topography of the powder bed of PREP15-45 simulated by DEM;

[0037] Figure 19 It is the spreading process of PREP15-53 powder under different powder spreading parameters;

[0038] Figure 20 It is the surface topography of the powder bed of PREP15-53 simulated by DEM;

[0039] Figure 21 It is the SVM powder spreading parameter optimization diagram constructed when the C value is 1 and its result verification. Among them, (a) is the optimization diagram of GA powder, (b) is the surface topography under the optimal parameters of GA powder, (c) is the surface topography under the recommended parameters of GA powder, (d) is the optimization diagram of PREP15-45 powder, (e) is the surface topography under the optimal parameters of PREP15-45 powder, (f) is the surface topography under the recommended parameters of PREP15-45 powder, (g) is the optimization diagram of PREP15-53 powder, (h) is the surface topography under the optimal parameters of PREP15-53 powder, (i) is the surface topography under the recommended parameters of PREP15-53 powder;

[0040] Figure 22 It is the schematic diagram of the process of optimizing scanning parameters using machine learning;

[0041] Figure 23 It is the schematic diagram of the construction and application of the second machine learning model;

[0042] Figure 24 It is the schematic diagram of the process of the comprehensive prediction and optimization method;

[0043] Figure 25 It is the schematic diagram of the construction process of the comprehensive prediction model;

[0044] Figure 26 It is the block diagram of the parameter optimization system for additive manufacturing using machine learning;

[0045] Figure 27 is a block diagram of an electronic device. Detailed implementation manners

[0046] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments may be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are described so that this disclosure will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art. The features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a better understanding of the embodiments of this disclosure.

[0047] "Additive manufacturing (AM)" as described in this disclosure is a three-dimensional printing technology that builds parts by adding powder materials layer by layer and melting the powder materials using an energy beam (such as a laser or an electron beam). This technology is based on a digital model generated by computer-aided design (CAD) software and precisely controls the melting and solidification processes of each layer of material to gradually build a complete part.

[0048] This disclosure mainly relates to models constructed using machine learning algorithms that learn and improve from "additive manufacturing"-related data to predict and output key information for actual manufacturing processes. This information includes, but is not limited to, powder spreading parameters, scanning parameters, and optimization data related to powder characteristics. Machine learning algorithms analyze data such as powder spreading parameters, powder characteristics, forming surface quality, and clad layer surface topography to establish models to improve the accuracy and efficiency of additive manufacturing. Machine learning algorithms that can be used include, but are not limited to, support vector machines (SVMs), neural networks (such as multi-layer perceptrons and deep neural networks), convolutional neural networks (CNNs), decision trees, random forests, and other supervised or unsupervised learning methods.

[0049] The main body for executing the method steps of the present disclosure is, for example, an AM system including the integration of hardware and software. The AM system includes machine hardware for additive manufacturing, a computer control system for controlling the additive manufacturing process, and various sensors and monitoring devices. The computer control system is the core of the AM system, which is used to process the conversion from the CAD model to the actual printing path, adjust the printing parameters in real time, and execute the construction and optimization of the model through the interface with the external data processing module. The AM system for actual manufacturing can automatically call the trained model to predict and optimize the printing parameters. Through the parameters output by the model, the AM system can adjust and optimize the printing process. In addition, the AM system can continuously monitor and record various types of data during the printing process, update the existing model, and enhance its accuracy and adaptability.

[0050] Implementation manners of the powder spreading parameter optimization method

[0051] Figure 1 It is a schematic flow diagram of the powder spreading parameter optimization method using machine learning. Figure 2 It is a schematic diagram of the construction and application of the first machine learning model. The method 10 provided by the present disclosure has steps 11 to 13 in some embodiments.

[0052] 11. Obtain the surface topography of the first cladding layer constructed by powder layers with different powder characteristics under different powder spreading parameters, and establish a first database 101 that maps the relationship between the powder spreading parameters, powder characteristics, and the surface topography of the first cladding layer.

[0053] The powder spreading parameters described herein mainly include the powder spreading speed and layer thickness. The powder spreading speed refers to the speed at which the powder spreading device (scraper or roller) moves on the building platform, which affects the uniformity and density of the formed powder layer. The layer thickness refers to the thickness of the powder layer laid on the building platform each time powder is spread, which affects the stacking quality of the cladding layer, the internal porosity, and the mechanical properties of the final part.

[0054] In some embodiments, the powder spreading parameters may also include other relevant data that the powder spreading device has when performing the powder spreading operation. For example, the vibration parameters formed by applying vibration to the powder during the powder spreading process, including the frequency, amplitude, and application method of vibration (such as the interval time for applying vibration), etc.

[0055] The powder characteristics described herein affect the laying quality of the powder layer, the forming quality of the cladding layer, and the performance of the final part during the additive manufacturing process. The powder characteristics determine the fluidity and stacking behavior of the powder during laying and affect its melting characteristics and the densification of the cladding layer under the action of the energy beam. In various embodiments of the present disclosure, the main powder characteristics include, but are not limited to, one or more of the powder manufacturing process, powder material, particle shape, and powder particle size.

[0056] Taking the powder manufacturing process as an example, different powder manufacturing processes can produce powders with different shapes, particle size distributions, and purities, which in turn directly affect the cladding quality in additive manufacturing. The powder manufacturing processes described in this article include various different types, such as, but not limited to, atomization, mechanical comminution, reduction, electrolysis, and the carbonyl process.

[0057] In addition, each of the above powder manufacturing process categories can also have different subcategories. Taking atomization as an example, atomization can be further divided into multiple subcategories, such as gas atomization (GA) method, plasma rotating electrode process (PREP) method, and water atomization method. The GA method involves using high-pressure gas (such as argon or nitrogen) to impact molten metal to form spherical particles, which is suitable for manufacturing powders with high purity and good fluidity. The PREP method involves centrifugally throwing molten metal by the centrifugal force of a rotating electrode to form dense powder particles, which is suitable for the production of high-performance alloys. Water atomization involves using high-pressure water jets to cool and disperse the metal. The powder produced has a lower cost, but the particle shape is not as uniform as that of gas atomization, and the surface is prone to irregularities. It should be understood that the selection of each powder manufacturing process parameter will affect the particle shape and particle size distribution of the final powder. For example, the pressure and nozzle diameter of the GA method will change the size and morphology of the particles, and the rotation speed and temperature of the PREP method will affect the density and uniformity of the particles.

[0058] The surface topography of the first cladding layer described in this article refers to the microscopic / macroscopic structure formed after melting and solidifying the laid powder layer by an energy beam. It can reflect the physical changes experienced by the powder during melting and cooling. This topography can be obtained, for example, using instruments such as a 3D profilometer and / or an optical microscope. By measuring the cladding layer, the 3D profilometer can provide data such as surface roughness, flatness, and height changes. Using an optical microscope to detect the surface topography of the cladding layer can understand the microstructure after powder melting, such as the distribution of grains, porosity, surface cracks, and their morphological characteristics. Combining these two detection methods can comprehensively describe and quantify the topography characteristics of the cladding layer. By obtaining the surface topography data formed under different experimental conditions and combining the powder spreading parameters and powder characteristics in the additive manufacturing process, a first database 101 is established.

[0059] The first database 101 stores data in three dimensions: powder spreading parameters, powder characteristics, and the surface topography of the first cladding layer, and is used to provide data sources for the first machine learning model 102 so that the model can learn and analyze the relationships among the powder spreading parameters, powder characteristics, and the first surface topography. In addition, during the establishment of the database, data cleaning and preprocessing are also involved, including removing noise, handling missing values, and standardization operations to ensure the quality of the input data.

[0060] The surface morphology of the first cladding layer is obtained by constructing with powder layers having different powder characteristics under different powder spreading parameters. Changes in powder spreading parameters (such as powder spreading speed and layer thickness) and the diversity of powder characteristics (such as powder manufacturing process, material composition, particle shape, and particle size) will affect the powder stacking state, melting behavior, and solidification structure after cooling during the additive manufacturing process. In actual operation, by changing the powder spreading parameters, the powder is laid on the building platform to form a powder bed. At the same time, different powder characteristics, such as powder manufacturing methods (such as GA method, PREP method, etc.), material compositions (such as nickel-based alloys, titanium alloys), particle shapes (spherical, oval, irregular), and particle size distributions, affect the fluidity and compactness of the powder layer, and thus affect the flatness, porosity, and solidification structure of the cladding layer surface. In the experiments carried out under different powder spreading parameters and powder characteristics, the surface of the cladding layer is detected and analyzed after each building step to obtain its morphology data to obtain the surface morphology of the first cladding layer.

[0061] 12. Construct and train the first machine learning model 102 based on the first database 101 to generate the first application model 103 for predicting powder spreading parameters.

[0062] By using the first database 101 to construct and train the first machine learning model 102. The goal of training the model is to enable the machine learning algorithm to extract key features and potential laws from between different powder spreading parameters, powder characteristics, and their corresponding first cladding layer surface morphologies. During the construction and training process, the first database 101 can be divided into a training set and a validation set. The training set is used for the learning of the first machine learning model 102, enabling it to gradually adjust its internal parameters to optimize the prediction ability by analyzing the relationships between the powder spreading parameters, powder characteristics, and first cladding layer surface morphology therein; the validation set is used to evaluate the generalization ability of the first machine learning model 102 to avoid overfitting or underfitting of the model. The goal of model training is to enable the first machine learning model 102 to capture the influence of different powder spreading parameters and powder characteristics on the formation of the cladding layer, so as to have accurate parameter prediction ability in subsequent applications. After training is completed and verified, the first application model 103 for predicting powder spreading parameters can be generated. The first application model 103 is the specific implementation of the training model and can receive external inputs and output prediction results.

[0063] In a specific embodiment, the first machine learning model 102 can be constructed and trained using the support vector machine (SVM) algorithm. The basic principle of SVM is to find an optimal hyperplane in a high-dimensional feature space to effectively separate sample points of different classes. Specifically, the goal of SVM is to find a hyperplane that maximizes the margin from the points closest to it (referred to as support vectors) in the sample set, so that the model has high robustness on the training data and good generalization ability on unknown data. Specifically in this embodiment, the use of SVM is not limited to simple binary classification tasks, but can be extended to multi-classification or regression prediction for handling complex optimization of powder spreading parameters. And SVM can adjust parameters during the training process, such as selecting a suitable kernel function (such as linear kernel, Gaussian kernel or polynomial kernel) and adjusting the penalty parameter to balance margin maximization and misclassification tolerance. The first machine learning model 102 after training, that is, the first application model 103, can be used to predict the optimal powder spreading parameters under different powder property inputs.

[0064] 13. Use the first application model 103 to predict the input powder properties and output the optimal powder spreading parameters for additive manufacturing.

[0065] In the specific implementation process, by inputting the powder properties into the first application model 103, the first application model 103 processes and analyzes the input powder properties using the knowledge and rules obtained during the previous training, so as to predict the optimal powder spreading parameters that match them. In addition, the powder spreading parameter suggestions provided by the first application model 103 can be directly applied to the AM system to achieve automatic adjustment. And after the output of the first application model 103 is verified by practice, it can be continuously fed back into the system to further improve the prediction performance of the first database 101 and the model, so as to achieve self-learning and continuous optimization.

[0066] In this embodiment, the surface morphology of the first cladding layer is constructed under set scanning parameter conditions, especially under the same scanning parameters. The scanning parameters include laser power, scanning speed, and spot diameter. By setting these scanning parameters as constant conditions, the operator can test under different powder spreading parameters and powder properties to observe and record the surface morphology of the first cladding layer generated by these parameter combinations to construct the first database 101.

[0067] In addition, the scanning parameters can also include the scanning trajectory. The scanning trajectory refers to the path of the energy beam moving on the powder bed during the processing. The setting of the scanning trajectory can control the uniform distribution of heat input and reduce thermal stress and deformation. Common scanning paths include straight lines, crosses, grids, and spirals, etc. Each path pattern will have different effects on the surface morphology of the cladding layer.

[0068] Implementation manners of the construction of the first machine learning model 102 in the roll powder spreading mode

[0069] The chemical compositions of the three types of TC4 powders used in this embodiment are shown in Table 1. Among them, GA0-53 is prepared by gas atomization (GA), and PREP15-45 and PREP15-53 are prepared by plasma rotating electrode (PREP) method. Figure 3 This is the morphological characterization of the three types of TC4 powders by scanning electron microscope (SEM) under the roll powder laying method. (a) is the powder prepared by GA method, which is basically spherical particles, but there are a small number of satellite particles. (b) and (c) are the powders prepared by PREP method, with good sphericity, smooth surface and no satellite particles. The particle size distribution of the powders is measured by a laser particle size analyzer, and the particle diameters of the three powders are mainly concentrated below 53μm.

[0070] Table 1 - Chemical compositions (mass fraction %) of three types of TC4 alloy powders

[0071]

[0072] The equipment used in this embodiment is the SLM series laser forming equipment BLT-S210 independently developed by Xi'an BRIGHT Laser Technologies Co., Ltd. The main parameters of the equipment are: laser power, scanning speed, powder laying thickness (layer thickness) and powder laying speed. The uniform experimental design is used to reduce the experimental cost, and five groups of process parameters are set. Among them, the laser power, scanning speed and scanning spacing are fixed, the layer thickness is in the range of 10 - 50μm, and the powder laying speed is in the range of 35 - 275mm / s. According to the uniform experimental design, the five groups of process parameters (NO.1 - 5) and the machine recommended parameters (NO.6) are shown in Table 2 below. According to these four parameters (laser power, scanning speed, layer thickness and powder laying speed), the laser energy density E (J / mm -1 ) input during forming can be calculated by the following formula:

[0073] E = P / (v·h·t)

[0074] In the formula: P is the laser power (W); v is the scanning speed (mm / s); h is the scanning spacing (mm); t is the powder laying thickness (mm). The calculation results of the laser energy density are shown in Table 2.

[0075] Table 2 - SLM forming process parameters

[0076]

[0077]

[0078] The size of the SLM-formed specimen for process parameter optimization is 5*4 mm, the number of scanning layers is 20 layers, and a scanning strategy of rotating 67° between layers is adopted during forming. After the formed TC4 specimen is removed of excess powder by a vacuum cleaner, it is then washed clean with an ultrasonic cleaner. An optical microscope is used to observe the surface morphology of the specimen; a three-dimensional profiler is used to measure the surface roughness and the height of the specimen, and the measurements are taken 3 times in total, and the results are averaged. As Figures 4 to 6 shown, Figures 4 to 6 respectively are the surface morphologies of the formed specimens under different powder spreading parameters, where Figure 4 (a) to (e) of Figure 5 are the optical microscope images of the cladding layers of GA0-53 powder, Figure 6 and (a) to (e) of

[0079] are the optical microscope images of the cladding layers of PREP15-45 powder,

[0080] Table 3—Laser energy density and surface roughness corresponding to the process parameters of GA0-53 specimens

[0081]

[0082] After the preliminary processing of the data (table), R language software was used to perform SVM analysis on 5 groups of data of 3 kinds of powders, and the radial basis function was used as the kernel function to construct a powder spreading parameter optimization model (the first machine learning model 102). Based on the radial basis (RBF) kernel function in the SVM classification model (the first machine learning model 102), the data can be more easily classified or regressed, and the data can be better fitted to avoid overfitting or underfitting. The penalty factor (C) and the kernel function parameter (g) have a great influence on the performance of the classifier. The larger the C value, the stricter the classification and the less errors there are, and the higher the classification accuracy. However, at this time, the generalization ability of SVM is low, and the classification results for data sets with a large number of samples are not ideal; g is a parameter of the RBF function, which determines the distribution of the data after mapping to the new space. The smaller g is, the higher the classification accuracy; the larger g is, the more complex the model is, which will lead to overfitting, the lower the classification accuracy, and the classification accuracy of new samples is also low. Therefore, it is necessary to select appropriate C and g values ​​to prevent overfitting and underfitting while ensuring the accuracy. In order to obtain a more accurate optimization model, it is necessary to study the impact of the cost parameter on the separation boundary of the optimization model by adjusting the cost parameter (CostParameter) to observe the change in its error rate (that is, the number of misclassified data points divided by the total number of data points).

[0083] Figure 7 This is the SVM optimization diagram for powder spreading parameters of powder prepared by GA method. Taking the SVM of GA method as an example, Figure 7 As shown in the figure, with the change of C, the optimal powder spreading parameter range is constantly moving, the optimal layer thickness varies in the range of 9.7μm to 11.5μm, and the powder spreading speed varies in the range of 175-190. When the C value is 5 or 10, the optimal parameter range is stable. The accuracy of the prediction model is expressed by calculating ErrorRate. The lower the value, the higher the accuracy of the prediction result. As shown in Table 4 below, when the value of CostParameter is between 0.1 and 1, its ErrorRate gradually decreases and reaches the minimum value at 1. At this time, even if the C value increases, the ErrorRate value remains stable, so 1 is used as the cost parameter.

[0084] Table 4 - Hyperparameters of the SVM optimization model

[0085]

[0086] Implementation manners of the construction of the first machine learning model 102 in the doctor blade powder spreading mode

[0087] The chemical compositions of the three kinds of TC4 powders used in this embodiment are shown in Table 5 below, wherein GA0-53 is prepared by the GA method, and PREP0-53 and PREP25-75 are prepared by the PREF method.

[0088] Table 5 - Chemical compositions of three kinds of TC4 alloy powders (mass fraction %)

[0089]

[0090] Using the laser forming equipment BLT - S210 mentioned above for forming experiments, the forming parameters according to the uniform experimental design are as shown in Table 6 below:

[0091] Table 6 - SLM forming process parameters

[0092]

[0093] The size of the SLM formed specimen for process parameter optimization is 10 * 10 mm. During forming, a scanning strategy with a 67° rotation between layers is adopted. After forming, a three - dimensional profiler is used to measure the surface roughness and the height of the specimen, and the drainage method is used to measure its density. A total of 3 measurements are taken, and the results are averaged.

[0094] It should be understood that the specific model construction process of this embodiment is similar to the construction process principle of the first machine learning model 102 under the roll powder spreading method, and will not be described in detail here.

[0095] Surface quality analysis of the cladding layer and the powder bed

[0096] Figure 8 are the surface quality morphology diagrams of the powder bed of GA powder under different powder spreading parameters. Refer to Figure 4 (a) - (e) of Figure 8 and (a) - (g) of . When the laser power and scanning speed remain unchanged, as the layer thickness increases, the input laser energy density decreases, and some powders on the scanning path cannot be completely melted, and the melt channels change from continuous, smooth, and neat to distorted and discontinuous. When the layer thickness remains unchanged at 30 μm and the powder spreading speed increases from 50 mm / s to 155 mm / s, although the input laser energy density is the same at this time, the surface roughness of the specimen with a powder spreading speed of 155 mm / s is lower than that of the specimen with 50 mm / s, and the melt channels are continuous with less unmelted powder around. By observing the surface quality of the powder bed under different parameters, it can be seen that due to the increase in the powder spreading speed, the packing density on the powder bed decreases and the number of powders decreases. Therefore, the input laser energy is sufficient to melt most of the powders, the melt channels are continuous and smooth, and the surface roughness of the specimen is lower.

[0097] Refer to Figure 5(a), under the condition of sufficient laser energy density, the molten tracks are continuous, flat, and smooth, with only a few points showing interrupted molten tracks. Under the same parameters, it is flatter than the GA specimen and there is no phenomenon of coarse grains. Comparing the surface roughness of the two powders, it is found that the surface roughness of the PREP15-45 specimen is only 6.831 μm, which is lower, indicating that the surface quality of the specimen is better under this parameter. This is because under this parameter, the GA powder shows coarse grain phenomenon due to the large input laser energy. Since the volume of particles larger than 53 μm in the PREP15-45 powder is 25%, which is more than that of GA, more energy is required to fully melt it. Therefore, for the PREP15-45 powder, this energy density is matched and there is no obvious over-melting or non-melting phenomenon.

[0098] Figure 9 Figure shows the surface quality morphology of the powder bed of PREP15-45 powder under different powder spreading parameters. Different from the GA0-53 powder, the PREP15-45 powder has a high sphericity and good fluidity, so there is less adhesion between the powders, as Figure 9 shown. As can be seen from Figure 9 (a), due to the low layer thickness and high powder spreading speed, the gap between powder particles is large and the apparent density of the powder bed is small. Compared with the GA powder, the powder bed is more uniform and the size of the voids is smaller. When the layer thickness and powder spreading speed are further increased, the number of particles on the powder bed decreases and the gap between the particles increases. However, different from the GA powder, the PREP15-45 powder particles are evenly distributed, so the quality of the powder bed is good. As the layer thickness increases and the powder spreading speed decreases, the gap between the powder particles decreases and the powder is evenly distributed. At this time, obvious powder adhesion phenomenon appears in the GA powder. This is because some GA powders are satellite-shaped, the particle surface is rough and uneven, and the adhesion ability between the particles is enhanced, resulting in particle adhesion. While the PREP powder surface is smooth, spherical, and has good fluidity, so it is not easy to appear adhesion phenomenon. According to the analysis of Figure 9 (d)-(f), as the layer thickness increases and the powder spreading speed decreases, the gap between the powder particles decreases and the apparent density of the powder bed increases, and vice versa.

[0099] Similar to the powder of PREP15-45, the PREP15-53 powder is also made by the PREP method, so the sphericity of the two powders is the same. However, due to the different particle size distributions, there are certain differences in their fluidity. The angle of repose of the PREP15-53 powder is smaller, and the time required to freely flow through the Hall flowmeter is less, so the fluidity of the powder is better. In order to observe the influence of particle size distribution and fluidity in powder characteristics on the forming quality, the 15-53um powder prepared by the PREP method was used for the above experiments to observe the differences in the surface quality of formed specimens of powders with different characteristics under the same parameters. Compared with GA0-53 and PREP15-45 powders, the particle size of PREP15-53 is larger, and the volume of powder larger than 53um is about 12%.

[0100] As can be seen from Figure 6 (a) of, when the layer thickness is 10um, the molten channels on the specimen surface are continuous and stable, with only a small amount of defects, and the surface roughness is also relatively low at 7.096um, indicating that the surface quality under this parameter is good. However, under this parameter, the surface roughness of the PREP15-45 specimen is even lower, only 6.583um; the surface roughness of the GA0-53 specimen is higher at 8.544um. Although the input laser energy is the same, considering the different powders used and the differences in their powder characteristics, the powder characteristics have a certain influence on the forming quality. The forming quality of the PREP15-45 powder is the best, indicating that under this powder spreading parameter, the input laser energy is moderate, the metal powder can be completely melted, the molten pool is stable, and the molten channels are smooth and flat. For the GA0-53 powder, compared with PREP15-45, its powder particle size is smaller, and the energy required to completely melt all the powder is less. Therefore, the input laser energy for the GA powder is relatively high. Although the metal powder can be completely melted, due to the significant increase in the heat affected zone of the laser, the metal powder has undergone over-melting, and the molten channels in some areas are unstable and distorted, so the surface quality is poor. For the PREP15-53 powder, its powder particle size is larger than that of PREP15-45, so the laser energy is relatively small, and a small number of larger particles cannot be completely melted. Therefore, there is too little liquid metal in the molten channel, and the molten channel is discontinuous; at the same time, the extremely large capillary force and strong convection of the liquid metal cause the molten channel to form a spheroidization phenomenon.

[0101] Figure 10 is the surface quality morphology diagram of the powder bed of PREP15-53 powder under different powder spreading parameters; As can be seen from Figure 10It can be seen from (a)-(c) that for powders with a small layer thickness and a high powder spreading speed, the gaps between powder particles are relatively large, the sizes of the void spots are large and the distribution is uneven. Considering the low layer thickness, the height error at the micron scale of the substrate will affect the powder spreading, resulting in uneven powder spreading. However, as the layer thickness increases, the height error of the substrate has no obvious influence, the number of void spots on the powder bed decreases, the powder particles are evenly distributed and the gaps between the particles become smaller. For a powder bed with a high layer thickness and a high powder spreading speed, compared with other powder beds with a high layer thickness and a low powder spreading speed, the gaps between some of its particles are relatively large and the powder spreading is relatively uneven.

[0102] Compared with GA powder, there is no obvious bonding phenomenon in PREP15-53 powder and the powder is evenly distributed. By comparing the powder characteristics of the two powders, it can be seen that the GA powder particles are satellite-shaped and the surface is rough, while the PREP15-53 powder is spherical and the surface is smooth. Therefore, the PREP15-53 powder has better fluidity, the adhesion force between the powders is poor, and it is not easy to occur powder bonding phenomenon. By comparing the fluidity and sphericity of the two powders under a Hall flowmeter, it can be seen that although the sphericity of the two powders is the same, due to the difference in powder particle size, the time required for the PREP15-53 powder to flow through the Hall flowmeter is less and the angle of repose is also smaller. Therefore, its fluidity is better. By comparing with PREP15-45 powder with the same PREP preparation method but different powder particle sizes, it is found that the powder bed of PREP15-45 has a higher powder bed density while having no obvious void spots under various powder spreading parameters.

[0103] Implementation manners of the simulation of the powder pile movement state and the supplementary construction of the first database 101

[0104] Figure 11 It is a schematic diagram of the supplementary construction process of the first database. The method 10 provided by the present disclosure also has steps 21 to 23 in some embodiments.

[0105] 21. Capture images and perform discrete element simulations on the movement state of the powder pile during the laying process of powder layers with different powder characteristics under different powder spreading parameters.

[0106] Use an experimental device to capture images of the movement state of the powder pile during the powder laying process. The powder pile refers to the accumulation formed by the powder under the action of a scraper or a roller during the powder spreading process, and its movement state reflects the fluidity and distribution characteristics of the powder. Use a high-speed camera to capture the dynamic changes of the powder pile under different powder spreading parameters and different powder characteristics in real time.

[0107] Meanwhile, the movement of the powder bed is simulated by computer using the Discrete Element Method (DEM) to describe the movement behavior of the particle system. By inputting the physical parameters of the powder and the operating parameters of the powder spreading device, the complex particle movement, heat transfer, and the complex interactions between particles-fluid and particles-devices inside the equipment can be accurately simulated to display the movement process of the particle system in real time, and analyze the microscopic force conditions at any moment.

[0108] 22. Calibrate the data trend of the DEM simulation according to the results of image capture.

[0109] Although DEM simulation can provide a detailed prediction of the movement state of the powder bed, due to the possible deviation between the input parameters and calculation assumptions of its model and the actual situation, it is necessary to calibrate it through the experimental image capture results to improve the accuracy of the simulation results. The calibration process, for example, analyzes the differences between the actual powder bed movement images captured by a high-speed camera and the DEM simulation results in key features such as the powder bed movement trajectory, angle of repose, particle distribution, and the uniformity of the powder layer thickness. For example, if the actual image shows a large accumulation at the end of the blade while the simulation data does not show a similar feature, it indicates that the parameters in the DEM model need to be adjusted.

[0110] In addition, after parameter adjustment, the calibration process can also involve comparing the optimized simulation results with the experimental images again to observe whether the DEM simulation data is consistent with the actual powder bed state captured in terms of the accumulation morphology, particle distribution, and dynamic trend. If the deviation between the two is within the allowable range, the calibration is considered complete; otherwise, the parameter adjustment needs to be repeated and further optimized.

[0111] 23. Correlate the movement state of the powder bed in the calibrated DEM simulation with the powder spreading parameters, powder characteristics, and the surface morphology of the first cladding layer to establish the first database 101.

[0112] The calibrated DEM simulation data is integrated into the first database 101 to supplement and improve the mapping relationship between the powder spreading parameters and the powder characteristics.

[0113] Specifically, the calibrated DEM simulation data can be classified according to powder characteristics and powder spreading parameters. Each set of simulation data needs to be labeled with its corresponding input conditions to ensure the traceability of the data. The calibrated powder bed movement state data can also be associated with the actually captured surface topography data of the first cladding layer. For example, if the simulation under a certain set of conditions shows that the powder bed forms a uniform and flat powder spreading layer, and the experimental verification shows that the surface topography has low roughness and high uniformity, then record this corresponding relationship and input it into the first database 101. By integrating the calibrated DEM simulation data, the multi-dimensional data structure of the first database 101 is extended. For example, if the original database only contains the relationship between powder spreading parameters and surface topography, and after supplementation and construction, it also includes the dynamic data of the powder bed movement state, so that the first database 101 can reflect the full process information from the dynamic powder spreading process to the static surface result. This embodiment can also perform quality verification on the integrated first database 101. For example, by randomly extracting data entries and re-entering the DEM simulation conditions, observe whether it can be consistent with the calibrated experimental results.

[0114] Figure 12 It is a schematic diagram of the verification process of the first application model. The method 10 provided by the present disclosure further has steps 31 and 32 in some embodiments.

[0115] 31. Perform discrete element simulation on the surface topography of the first cladding layer.

[0116] The input parameters of the DEM simulation include, for example, powder spreading parameters, powder characteristics, and scanning parameters, etc. The DEM model simulates the whole process of powder from the powder spreading device to the cladding surface by describing the interactions between particles such as collision, adhesion, and rolling. In the simulation, the movement trajectories and final distribution patterns of powder particles can be displayed in real time, so as to generate a three-dimensional prediction image of the surface topography after powder accumulation. In the simulation, the final positions and distributions of the particles determine the initial topography of the first cladding layer, and then the energy beam cladding process is simulated by inputting scanning parameters. The action of the energy beam melts and reorganizes the powder particles, thus forming the surface topography of the first cladding layer. Through simulation calculations, characteristic data such as the surface roughness, porosity, and surface uniformity of the cladding layer are obtained, and these are used as the characterization results of the surface topography. The simulation results will be compared with the real surface topography characteristics obtained from the experimental data to verify the prediction accuracy of the DEM simulation. If the characteristics such as surface roughness and uniformity of the simulation are consistent with the experimental data, it indicates that the simulation process is accurate and reliable. Otherwise, it is necessary to adjust the input parameters or optimize the simulation algorithm.

[0117] 32. Verify the first application model according to the powder bed movement state of the discrete element simulation and the surface topography of the first cladding layer.

[0118] The powder spreading parameters and powder characteristics output by the first application model 103 are optimized prediction values generated based on input conditions. These parameters are input into the DEM model to simulate the powder bed movement state and the surface morphology of the first cladding layer again, and observe whether it is consistent with the actual experimental results. For example, if the optimal powder spreading speed predicted by the first application model 103 is 50 mm / s, then this value is used as the simulation input to analyze whether the powder bed state formed under this parameter is stable and evenly stacked. By comparing the surface morphology of the first cladding layer generated by simulation with the experimental data in the first database 101, verify whether the surface roughness, porosity, and uniformity predicted by the first application model 103 under different parameter combinations meet the expectations.

[0119] If there is a deviation between the powder bed movement state shown by the DEM simulation and the experimental results and the output result of the first application model 103, it is necessary to analyze the error source, such as whether the training data of the first application model 103 is sufficient and whether the input parameters are comprehensive. According to the results of the error analysis, adjust the parameters or expand the data of the first application model 103. For example, if the simulation shows that the prediction error of certain specific powder characteristics is large, it is necessary to add relevant data of powders with this characteristic in the first database 101.

[0120] After the verification is completed, the reliability of the first application model 103 can be evaluated from two aspects: on the one hand, observe whether the predictions of the model for powder spreading parameter optimization and powder characteristic matching meet the experimental requirements. If the accuracy exceeds the set threshold, it is considered that the model verification passes. On the other hand, whether the model can show efficient prediction ability under different material and parameter combinations. For example, when the powder characteristics are switched from spherical particles to flake particles, whether the model can still give reasonable parameter optimization results.

[0121] The powder spreading process involved in the embodiments of the present disclosure can be completed in a powder spreading simulation experimental device, and the specific structure of this powder spreading simulation experimental device can refer to Document CN 116183446A.

[0122] Image capture of the powder pile movement state

[0123] Figure 13 It is to observe the state of GA powder at different times under different powder spreading speeds by high-speed photography. Observe Figure 13 In (a) of, it can be seen that at a speed of 35 mm / s, the powder bed is convex (yellow line). As the scraper moves, only the powder near the bottom of the substrate participates in the spreading process, as shown by the red arrow, while the top of the powder bed (area A) has no obvious change as the roller moves. This is because part of the GA powder is in a satellite shape, and the mechanical interlocking between powder particles hinders the flow of the powder, so only the bottom powder spreads. As the powder spreading speed increases, as Figure 13(b), the powder heap morphology hardly changed significantly, but the slope angle was lower than that at 35 mm / s; when the powder spreading speed continued to increase to 95 mm / s ( Figure 13 (c)), the powder heap morphology began to change, gradually changing from convex upward to convex downward, and some powders began to slowly flow downward from the peak and participated in the spreading process. When the powder spreading speed increased to 155 mm / s ( Figure 13 (d)), the powder heap was significantly convex downward and most of the powders flowed slowly. This is because the relatively fast powder spreading speed caused the powder heap to gather in front of the roller, forming a larger slope angle. As the spreading process proceeded, the powders at the bottom were not enough to spread the entire forming area. As shown in Figure 13 (e), a "notch" was formed due to the lack of powders at the bottom, so the powders at the upper part flowed downward to make up for the notch.

[0124] Figure 14 Figure 10 shows the powder spreading process of PREP15-45 powders at different powder spreading speeds captured by high-speed photography. Figure 14 (a) in Figure 10 is the screenshot of the powder spreading process at different times when the powder spreading speed is 35 mm / s. It can be seen that at time t0, the powders at the bottom spread to the left, and the middle area, i.e., area A, was convex upward. As the spreading process proceeded, at time t1, the convex powders in area A flowed downward along the slope, and the arc of the powder heap was smooth without obvious convexity. For the powders in area B, through high-speed photography observation, there was almost no obvious displacement of these powders, but the roller would squeeze these powders during the movement for spreading. Considering the gap between the glass baffle and the blade, it was impossible to accurately judge the specific movement behavior of these powders, so the subsequent exploration was carried out through discrete element simulation. Different from the spreading state of GA powders under the same parameters, the current powder spreading speed was relatively low, and the morphology of the powder heap was convex upward at the beginning. Considering the poor fluidity of GA powders, only the powders near the bottom flowed, and there was no obvious flow behavior of the middle powders. However, the PREP15-45 powders used in this experiment had good fluidity, so the middle powders flowed along the slope and participated in the spreading process, and thus the morphology of the subsequent powder heap changed. The powders in area C, i.e., the powders at the peak of the powder heap, had no obvious flow behavior at this spreading speed, the same as GA powders. In order to compare the changes in the powder heap morphology at different powder spreading speeds, the spreading process was studied by measuring the change in the inclination angle of the powder heap. For example, at time t2 under this parameter, the inclination angle formed by the powder heap was 35°, and as time passed, the inclination angle increased to 35.5°, with no obvious change. Figure 14(b) shows the spreading process of PREP15-45 powder under the recommended parameter (50 mm / s). At time t0, the powder on the slope of the powder pile is pushed by the roller and flows downward. As powder spreading proceeds, since the powder at the bottom has spread onto the substrate, resulting in the absence of powder in this part, the powder in area D starts to flow downward to replenish the powder at the bottom, and the convex part disappears. Therefore, the inclination angle of the powder pile increases from 31° to 34°. However, the powder in area E, i.e., the powder at the peak, still shows no obvious flow. When the powder spreading speed increases to 95 mm / s, the contour of the powder pile is linear ( Figure 14 (c)), and the inclination angle is relatively low, about 29°. At this time, the powder at the peak also shows no obvious flow.

[0125] Figure 15 These are the powder spreading processes of PREP15-53 powder at different powder spreading speeds captured by high-speed photography. As can be seen from Figure 15 (a), at a powder spreading speed of 35 mm / s, at time t0, the powder in area A aggregates into a convex mass. As the powder spreading process proceeds, this part of the powder disperses and flows downward for spreading (at time t1). Although the adhesion force between the powders will cause some powders to aggregate together briefly, due to good fluidity, as the spreading process proceeds, the powders will disperse, avoiding the appearance of unevenness. However, the powder at the top of the powder pile under this parameter does not show obvious flow behavior (area C), and it may only participate in the powder spreading process when the powder at the bottom is insufficient. Moreover, there is no obvious change in the angle of the powder pile during the spreading process, which is about 28°. As the powder spreading speed increases to 50 mm / s, the powder in area D also shows no obvious movement, but the angle of the powder pile decreases to 27°; when it increases to 95 mm / s, the angle of the powder pile increases to 29°; when it increases to 155 mm / s, the angle of the powder pile decreases again. Considering the errors in the experiment, the change in angle is small, and DEM simulation will be used for subsequent research.

[0126] DEM simulation of the powder pile movement state

[0127] The quality of the powder bed affects the dimensional accuracy and structural integrity of the formed parts. Therefore, optimizing the quality of the powder bed is an important way to obtain high-quality additive manufacturing parts. The influencing factors of the quality of the powder bed mainly include layer thickness, powder spreading speed, powder properties, doctor blade type, etc. During the experiment, high-speed photography can only monitor the shape and movement trend of the powder pile, but cannot analyze the movement and force of the particle system during the powder spreading process. Therefore, a DEM model similar to the powder spreading simulation experimental device is established to analyze the influence of powder properties, layer thickness, and powder spreading speed on the quality of the powder bed.

[0128] Figure 16It is a schematic diagram of the angle of repose for construction. Among them, (a) is the angle of repose of PREP15-45, and (b) is the angle of repose of PREP15-53. In this embodiment, two DEM simulation models are constructed to evaluate the flowability of powders: Hall flowmeter and drum test. Among them, the static and dynamic angles of repose and avalanche angles of the powders are obtained through the Hall flowmeter and drum test to correct the parameters of DEM. The Hall flowmeter model is constructed using UG, and its size is 1 / 30 of the original experiment. 3000 powders are formed in the funnel area (with a pore size of 200um) and naturally fall under the action of gravity to form a powder pile. The drum model is used to evaluate the dynamic flowability of the powder and the change of the avalanche angle, and its size is 1 / 20 of the original experiment and rotates counterclockwise at a speed of 10 rpm. According to the results of the Hall flowmeter, the simulation results of the two powders PREP15-45 and PREP15-53 are in agreement with the actual results, so the spreading experiment is carried out.

[0129] Figure 17 It is the spreading process of PREP15-45 powder under different powder spreading parameters. During the powder spreading process, the particle velocity distribution in the powder pile can be divided into three regions: (1) At the bottom of the powder pile, affected by the rough substrate, the particle velocity is the smallest; (2) On the left slope of the powder pile, due to the action of gravity, the particles accelerate and slide down, having the highest velocity; (3) The powders in other regions are the main components of the powder pile, and the velocity is moderate. As the powder spreading speed increases, the slope of the powder pile also gradually increases, and the moving speed of the particles on the slope also increases. And due to the differences in layer thickness and powder spreading speed, the contact force between the roller and the particles is also different. The contact force with a small speed and a large layer thickness is relatively small. Therefore, after the roller spreads, the backward movement speed of the particles is relatively small, and it is easier to stay on the substrate to form a uniform powder spreading layer. Figure 18 It is the surface topography of the powder bed simulated by DEM for PREP15-45. As the layer thickness increases and the powder spreading speed decreases, the powder bed spreads more evenly. And as the speed increases or the layer thickness decreases, there will be more "void spots" in the powder bed.

[0130] Figure 19 It is the spreading process of PREP15-53 powder under different powder spreading parameters. Different from PREP15-45 powder, on the left slope of the PREP15-53 powder pile, the flow velocity of the particles is relatively large because the angle of repose of PREP15-53 powder is smaller and the flowability of its powder particles is better. Figure 20It is the surface morphology of the powder bed in the DEM simulation of PREP15-53. Compared with the surface morphology of the PREP15-45 powder bed, the powder spreading is uneven and there are many large-sized voids. The reason is that the PREP15-53 powder contains more large-sized particles. When entering the powder spreading gap, the large-sized particles require more time and a larger contact surface with the roller, blocking the spreading of small-sized particles. For the large-sized particles, the layer thickness is small, and the roller has a certain dragging effect on the large-sized particles, increasing the particle speed and making it difficult to form a uniform powder layer on the substrate.

[0131] Therefore, both powder properties and powder spreading parameters have a certain impact on the quality of the powder bed. It is necessary to find the corresponding powder spreading parameters for different powders to obtain a uniform powder bed.

[0132] Surface quality verification of the optimal powder spreading parameter specimen

[0133] Figure 21 It is the SVM powder spreading parameter optimization diagram constructed when the C value is 1 and its result verification. Among them, (a) is the optimization diagram of GA powder, (b) is the surface morphology under the optimal parameters of GA powder, (c) is the surface morphology under the recommended parameters of GA powder, (d) is the optimization diagram of PREP15-45 powder, (e) is the surface morphology under the optimal parameters of PREP15-45 powder, (f) is the surface morphology under the recommended parameters of PREP15-45 powder, (g) is the optimization diagram of PREP15-53 powder, (h) is the surface morphology under the optimal parameters of PREP15-53 powder, and (i) is the surface morphology under the recommended parameters of PREP15-53 powder.

[0134] The optimal powder spreading parameters of GA powder predicted by the first application model 103 based on SVM are a layer thickness of 10um and a powder spreading speed of 185mm / s; the best powder spreading parameters of PREP15-45 are a layer thickness of 15um and a powder spreading speed of 50mm / s; the best powder spreading parameters of PREP15-53 powder are a layer thickness of 20um and a powder spreading speed of 65mm / s. Specimens are formed according to the predicted optimal powder spreading parameters and the machine recommended parameters (layer thickness 30um, powder spreading speed 50mm / s).

[0135] Refer to Figure 21For (b) and (c), the melt tracks of the specimens under the optimal parameters are continuous and smooth, with less unmolten powder around and no obvious unmolten defects. At the same time, the surface roughness of the formed specimens was measured using a 3D profiler, and it was found that the surface roughness of the optimal powder spreading parameters was low and the surface quality was better. While for the specimens with the recommended parameters, the melt tracks are discontinuous, and there is a large amount of unmolten powder accumulated at the discontinuous points, so the surface quality is poor and the surface roughness is high. As can be seen from Table 7 below, under the optimal powder spreading parameters, the input laser energy density is high, and the powder can be completely melted, so there are no obvious unmolten defects. While under the recommended parameters, the input laser energy density is relatively low, which is not enough to melt all the powder, resulting in the formation of unmolten defects and discontinuous melt tracks.

[0136] Table 7 - Optimal process parameters and recommended parameters for SLM forming of 3 kinds of powders

[0137]

[0138] Refer to Figure 21 For (e) and (f), the melt track morphology under the optimal powder spreading parameters is regular and continuous, without obvious high and low changes, and the surface roughness is only 6.058 μm, and the surface quality of the formed specimens is excellent. But for the specimens with the recommended parameters, the melt tracks are discontinuous, and the high and low fluctuations of the melt tracks are more obvious, and there are unmolten defects. This is because the input laser energy is high under the optimal parameters, and the powder is sufficient to be completely melted. However, compared with GA powder, the layer thickness of the optimal powder spreading parameters of PREP15 - 45 powder is larger and the powder spreading speed is lower. By comparing the particle size distributions of the two powders, it can be seen that the particle size distribution of PREP15 - 45 powder is concentrated in the range of 15 - 53, and the particle size is smaller, and the energy required for its melting is less. While in GA powder, the powder larger than 53 μm accounts for about 8%, and more laser energy is required for its complete melting. Therefore, GA powder requires a smaller layer thickness and a larger powder spreading speed.

[0139] Refer to Figure 21 For (h) and (j), the melt tracks under the optimal powder spreading parameters are smooth and neat, and there are occasionally small areas with unmolten defects, and the surface roughness is low. While for the specimens with the recommended parameters, the melt track morphology is irregular and discontinuous, there are many unmolten defects, and spheroidization occurs. Comparing with PREP15 - 45 powder made by the same method, the optimal layer thickness and powder spreading speed of PREP15 - 53 powder are both smaller than those of PREP15 - 45 powder. This is because PREP15 - 53 powder contains more large - sized particles. On the one hand, more laser energy is required to melt the large - sized particles. On the other hand, the large - sized particles affect the spreading process of the powder bed, causing phenomena such as particle dragging when the powder spreading speed is fast. Therefore, a lower powder spreading speed is required to make the powder bed spread evenly.

[0140] By comparing the surface quality of the formed specimens under the optimal powder spreading parameters and the machine-recommended parameters, it can be found that the quality of the specimens optimized by the support vector machine is better. And through the quality of the powder bed, it can be seen that the powder characteristics have a certain impact on both the powder bed quality and the forming quality. Therefore, for powders with different characteristics, matching powder spreading parameters are required to obtain better powder bed quality. And this study proves that it is indeed feasible to predict the optimal powder spreading parameters through uniform experiments combined with machine learning.

[0141] Implementation manners of the scanning parameter optimization method

[0142] Figure 22 It is a schematic flow chart of the optimization method for scanning parameters using machine learning. Figure 23 It is a schematic diagram of the construction and application of the second machine learning model. The method 10 provided by the present disclosure has steps 41 to 43 in some embodiments.

[0143] 41. Obtain the surface morphology of the second cladding layer constructed by powder layers with different powder characteristics under different scanning parameters, and establish a second database 104 that maps the relationship between the scanning parameters, powder characteristics, and the surface morphology of the second cladding layer. The surface morphology of the second cladding layer is constructed under the optimal powder spreading parameters corresponding to the powder characteristics output by the first application model 103.

[0144] Specifically, in order to achieve the goal of obtaining the surface morphology of the second cladding layer constructed using different powder characteristics under different scanning parameters, it is necessary to use the optimal powder spreading parameters output by the first application model 103 and combine different powder characteristics for experiments. In the experiments, the scanning parameters include laser power, scanning speed, spot diameter, and scanning trajectory, etc. Through the experiments, the melting process of the powder layer and the final surface morphology characteristics under different scanning conditions can be obtained, and these data are recorded at the same time. After the experiments, use equipment such as a three-dimensional profiler, optical microscope, or scanning electron microscope to detect the surface morphology of the second cladding layer, obtain important characteristics such as surface roughness, microstructure, and porosity, and establish a relationship between these surface morphology data and the corresponding scanning parameters and powder characteristics to form the second database 104.

[0145] 42. Construct and train a second machine learning model 105 based on the second database 104 to generate a second application model 106 for scanning parameter prediction.

[0146] By cleaning, preprocessing, and performing feature engineering on the data collected in the second database 104, the complex relationships between the scanning parameters, powder properties, and the surface morphology of the cladding layer are extracted. Then, algorithms such as support vector machine (SVM), decision tree, random forest, or neural network are used to train the second machine learning model 105, whose goal is to predict which combination of scanning parameters can obtain the best surface morphology of the second cladding layer under given powder properties. During the training process, methods such as cross-validation are used to optimize the parameters of the model to ensure its accuracy for practical applications. After training is completed, the obtained second application model 106 can be used for further prediction and optimization.

[0147] 43. Use the second application model 106 to predict the input powder properties to output the optimal scanning parameters for additive manufacturing.

[0148] By inputting the powder properties into the second application model 106, the second application model 106 processes and analyzes the input powder properties using the knowledge and rules obtained during the previous training, thereby predicting the optimal scanning parameters that match them. In addition, the powder spreading parameter suggestions provided by the second application model 106 can be directly applied to the AM system for automatic adjustment. And after the output of the second application model 106 is verified by practice, it can be continuously fed back into the system to further improve the prediction performance of the second database 104 and the model, thereby achieving self-learning and continuous optimization. In this way, not only can the selection efficiency of the scanning parameters in the additive manufacturing process be improved, but also the desired quality results can be obtained for each manufacturing. With the continuous optimization of the machine learning model and data accumulation, the prediction ability of the system will also be continuously enhanced, thereby providing a more accurate parameter optimization scheme in the actual production process.

[0149] To enable the second machine learning model 105 to adapt to more types of powder properties and process scenarios, in some possible implementation manners, a dynamic optimization mechanism can also be introduced, that is, by collecting and feeding back the data in the actual manufacturing process to the second database 104 in real time, the system can achieve adaptive learning. Specifically, in each additive manufacturing process, the optimal scanning parameters output by the second application model 106 are used for actual production. After production is completed, the surface morphology of the second cladding layer is detected again by a high-precision measuring device to obtain the actual surface morphology data. Then, these data are compared with the model prediction results, the source of the deviation is analyzed, and the new data are input into the second database 104 to update the training set. During the dynamic optimization process, the model will resample the existing training samples and retrain the second machine learning model 105 in combination with the new data, so that its weights can more accurately reflect the latest production conditions and powder properties.

[0150] Implementation manners of the comprehensive prediction and optimization of material properties

[0151] Figure 24 It is a flowchart of the comprehensive prediction and optimization method. Figure 25 It is a flowchart for constructing the comprehensive prediction model. The method 10 provided by the present disclosure has steps 51 to 55 in some embodiments.

[0152] 51. Separate single-modal databases corresponding to process parameters, forming surface quality, melt channel continuity, and mechanical properties are established, such as Figure 25 the process parameter database 107, the forming surface quality database 108, the melt channel continuity database 109, and the mechanical property database 110 shown. Among them, the process parameters include the powder spreading parameters and scanning parameters described above, and the forming surface quality includes the surface topography of the first cladding layer and the second cladding layer described above.

[0153] The process parameter database 107 includes two main categories: powder spreading parameters and scanning parameters. The powder spreading parameters include powder spreading speed, powder layer thickness, vibration parameters, etc. The scanning parameters cover laser power, scanning rate, spot diameter, scanning path, etc. The establishment of the process parameter database 107 requires recording multiple sets of process parameter combinations and their corresponding experimental results through experiments to cover a sufficiently wide process space.

[0154] The forming surface quality database 108 records the detailed information of the surface topography of the cladding layer during the additive manufacturing process, including the characteristic data of the surfaces of the first cladding layer and the second cladding layer. It is obtained by surface measurement equipment such as a three-dimensional profiler, a scanning electron microscope, or an optical microscope, and the indicators include surface roughness, particle uniformity, micro-defects (such as pores and cracks), etc.

[0155] The melt channel continuity database 109 is used to characterize the internal quality and integrity of the melt channel, and is the core data source for evaluating the microstructure and properties of the formed material. The melt channel continuity can be obtained through sectioning experiments combined with high-resolution microscopy techniques. The detection content includes the morphological consistency of the melt channel, cross-sectional characteristics, and its distribution law inside the component. These data are quantified by image analysis methods, such as calculating the penetration index of the melt channel, the bonding strength between melt channels, and the proportion of discontinuous regions caused by parameter fluctuations. In addition, special attention should be paid to the melt channel behavior under extreme parameter combinations when establishing the melt channel continuity database 109, such as the over-melting phenomenon caused by high power and low scanning rate or the under-melting problem caused by low power and high scanning rate, so as to comprehensively reflect the variation law of melt channel continuity.

[0156] The mechanical property database 110 records the macroscopic mechanical property data of the formed material, such as tensile strength, yield strength, ductility, and hardness. These property indicators can be obtained through standardized mechanical testing methods. For example, a tensile testing machine is used to measure strength and ductility, and a Vickers hardness tester is used to evaluate the hardness of the material. The mechanical properties are not only affected by process parameters but also closely related to the characteristics of the powder material and the forming surface quality.

[0157] When constructing the above single-modal database, it is also necessary to unify the data recording standards. For example, adopt internationally common material characterization methods and process parameter description formats to ensure the compatibility and integratability between different databases. In addition, for the possible differences in data dimensions for different characteristic indicators (for example, surface quality is mainly in the form of images, and mechanical properties are mainly in the form of numerical values), standardized format conversion and storage methods need to be adopted.

[0158] 52. Early fusion is performed on the single-modal databases of process parameters and forming surface quality (process parameter database 107 and forming surface quality database 108) to form a multi-modal database 111, and a melt channel continuity prediction model 112 and a mechanical property prediction model 113 are respectively constructed based on the single-modal databases of melt channel continuity and mechanical properties (melt channel continuity database 109 and mechanical property database 110).

[0159] 53. Late fusion is performed on the prediction results of the melt channel continuity prediction model 112 and the mechanical property prediction model 113, and multi-modal features are extracted based on the multi-modal database 111 and the late fusion data 114.

[0160] Early fusion is mainly applied to process process parameter data and forming surface quality data that are closely related to the additive manufacturing process. This data reflects the correlation between manufacturing process conditions and part surface defects. By performing early fusion on the process parameter database 107 and the forming surface quality database 108 to form a multi-modal database 111, the process parameters and surface topography information are integrated into a unified data dimension. The core purpose of this fusion strategy is to achieve deep association between data before feature extraction, enabling the model to simultaneously capture the complex influence mechanism of process conditions on defect features. For example, by integrating scanning parameters (such as laser power, spot diameter) and surface topography features (such as roughness, micro-defects), it can be found how specific process combinations affect the surface quality and uniformity of the cladding layer. The advantage of early fusion is that the data has completed preliminary information integration before being input into the machine learning model, thus significantly reducing the redundancy between features and improving the subsequent model's ability to analyze complex non-linear relationships, especially suitable for situations where there is a strong correlation and mutual complementarity between multi-modal data.

[0161] Different from early fusion, the late fusion strategy targets the data of melt channel continuity and mechanical properties. Since melt channel continuity and mechanical properties involve the internal structural integrity of materials and macroscopic mechanical behavior respectively, and they have different physical characteristics and influencing factors, it is suitable to build separate models to fully explore their respective characteristic laws. Based on the melt channel continuity database 109, a melt channel continuity prediction model 112 is constructed to describe the morphological consistency, penetrability of the melt channel and the response relationship with process parameters. At the same time, a mechanical property prediction model 113 is constructed using the mechanical property database 110 to capture the correlation between the mechanical properties (such as tensile strength, ductility) of the formed part and the microstructure. In the late fusion stage, the output results of the separately and independently constructed melt channel continuity prediction model 112 and mechanical property prediction model 113 are integrated into the late fusion data 114. This not only retains the high-precision analysis of the characteristics of each model in its respective field, but also further explores the potential relationship between melt channel continuity and mechanical properties, such as how melt channel defects affect the load-bearing capacity of parts, etc.

[0162] After completing early fusion and late fusion, combined with the multi-modal database 111 and the late fusion data 114, multi-modal feature extraction is carried out. Through advanced feature extraction algorithms (such as principal component analysis, autoencoders, etc.), multi-dimensional raw data can be converted into key features reflecting the deep associations between data. In addition to retaining the independent contributions of process parameters, formed surface quality, melt channel continuity and mechanical properties, it can also capture the complex non-linear interaction relationships between them.

[0163] 54. According to the multi-modal features, sub-prediction models 115 of process parameters and formed surface quality, sub-prediction model 116 of process parameters and melt channel continuity, and sub-prediction model 117 of process parameters and mechanical properties are respectively constructed.

[0164] In this embodiment, three sub-prediction models are respectively constructed according to the multi-modal features, namely, the sub-prediction model 115 of process parameters and formed surface quality, the sub-prediction model 116 of process parameters and melt channel continuity, and the sub-prediction model 117 of process parameters and mechanical properties. Each sub-prediction model aims at the key target performance in the additive manufacturing process to achieve more targeted prediction and correlation analysis.

[0165] Specifically, the process parameter and formed surface quality sub-prediction model 115 mainly predicts the influence of process parameters (such as powder spreading speed, layer thickness, laser power, scanning speed, etc.) on the formed surface quality. Based on the multi-modal features generated by early fusion, the model can capture the specific effects of process parameters on the surface morphologies of the first and second clad layers. During the model construction process, regression algorithms (such as support vector regression) can be used to extract the corresponding relationships between process parameters and surface morphology features (such as roughness, ripple depth). Secondly, deep learning methods (such as convolutional neural networks) are combined to further explore non-linear feature interactions, such as how laser energy density affects the formation of microscopic surface defects, so as to improve the model's prediction ability under complex process conditions and provide guidance for the control of surface defects. The process parameter and melt track continuity sub-prediction model 116 mainly focuses on how process parameters affect the continuity and integrity of the melt track. During model training, the geometric parameters of the melt track (such as width, penetrability) in the multi-modal features can be used as response variables, and algorithms such as gradient boosting decision tree are used to fit their relationships with process parameters. To enhance the model's sensitivity to melt track defects, feature data based on finite element simulation can also be added to supplement the quantitative relationship between the process and the melt track from the perspective of physical mechanisms. The process parameter and mechanical property sub-prediction model 117 mainly reveals the influence of process parameters on the macroscopic mechanical properties (such as tensile strength, ductility, hardness) of the formed part. During the model construction process, the microstructure information (such as grain size, directionality) and mechanical property data in the multi-modal features can be combined, and algorithms such as deep neural networks are used to explore how process parameters affect mechanical behavior by changing the microstructure of the material. Through multi-level feature learning to identify the contribution degree of key process variables to the target performance, it provides a basis for the directional optimization of achieving the target performance.

[0166] 55. The separately constructed sub-prediction models 115, 116, and 117 are jointly trained to construct an integrated prediction model 118 for comprehensively predicting and optimizing the material properties of additive manufacturing.

[0167] After constructing three sub-prediction models, this embodiment integrates the sub-prediction models into a comprehensive prediction model 118 through a joint training method. The goal of joint training is to utilize the characteristics of different sub-models to construct a global model that can balance prediction accuracy and generalization ability. The first step of joint training is to share parameters or fuse features among the sub-prediction models. For example, using the Multi-Task Learning (MTL) framework, share the feature extraction layer of process parameters in the same network, and optimize the output layers for the forming surface quality, bead continuity, and mechanical properties respectively. Through this sharing mechanism, different sub-models can learn from each other and improve the overall prediction ability. In addition, a weight dynamic adjustment mechanism can be introduced to balance the influence of sub-models according to the priority of target performance or the complexity of data features. For example, in some scenarios, bead continuity may be more critical than surface quality, and the weight allocation algorithm will prioritize optimizing the contribution of the relevant model. In actual training, a distributed computing framework is used to optimize the joint model in parallel. Through a hierarchical gradient update strategy, ensure that the independent optimization goals of each sub-model are not interfered by the global optimization process, and at the same time use the global loss function (such as weighted mean square error) to evaluate and feedback the overall performance of the model. And a cross-validation mechanism is introduced to avoid the decline of prediction accuracy caused by uneven data distribution or model overfitting. During the validation process, the integrated model can continuously adjust the parameter relationship between its sub-models to improve the prediction ability.

[0168] Additive manufacturing parameter optimization system

[0169] It should be understood that all the above-mentioned processes and steps can be implemented by configuring the functional modules of the additive manufacturing parameter optimization system. Although the specific content of the method has been described in detail, in order to more clearly understand the core part of the additive manufacturing parameter optimization system involved in the present disclosure, the following will briefly introduce its main components in combination with Figure 26 and briefly introduce its main components, Figure 26 is a block diagram of an additive manufacturing parameter optimization system using machine learning. The system 60 involved in the present disclosure includes the following modules.

[0170] The first clad layer surface topography acquisition module 61 is used to acquire the surface topography of the first clad layer constructed from powder layers with different powder characteristics under different powder spreading parameters.

[0171] The first database construction module 62 is used to establish a first database that at least maps the relationship among the powder spreading parameters, powder characteristics, and the surface topography of the first clad layer.

[0172] The first model training module 63 is used to construct and train a first machine learning model according to the first database to generate a first application model for predicting powder spreading parameters.

[0173] The powder spreading parameter prediction module 64 is configured to predict the input powder characteristics using the first application model to output the optimal powder spreading parameters for additive manufacturing.

[0174] It should be noted that although several modules for the system 60 are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more of the above-described modules may be embodied in one module. Conversely, the features and functions of one module described above may be further divided and embodied by multiple modules.

[0175] Electronic device

[0176] Figure 27 is a block diagram of an electronic device. In some embodiments, the electronic device 70 includes a processor 71 and a memory 72 (where the number of the processor 71 and the memory 72 may be one or more). The memory 72 is coupled to the processor 71 and is configured to store instructions executed by the processor 71, and when the instructions are executed by the processor 71, the electronic device 70 is caused to execute the method 10 described in any of the foregoing.

[0177] Specifically, the processor 71 communicates with the memory 72. The memory 72 may include a read-only memory and a random access memory, and provides instructions and data to the processor 71. In addition, a part of the memory 72 may further include a non-volatile random access memory (NVRAM). In the memory 72, operation instructions, executable modules, data structures, or subsets thereof, or even extended sets thereof are stored. These operation instructions cover various operations for implementing various operations. The method 10 described in the embodiments of the present disclosure may be applied to or implemented by the processor 71. The processor 71 may be any suitable computer processor, such as a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), etc. In the embodiments of the present disclosure, the processor 71 is configured to execute each step of the method 10.

[0178] Computer-readable storage medium

[0179] In some embodiments, the present disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor 71, implements the method 10 described in any one of the foregoing. Specifically, a computer-readable storage medium refers to a medium that can be read by a computer system, such as a hard disk, a solid-state drive, an optical disc, a flash drive, etc. In some embodiments of the present disclosure, the computer-readable storage medium stores a set of computer programs, and these programs are executed by the processor 71 to implement the various steps and functions described in the method 10. These computer programs may include an operating system, embedded software, application programs, etc., for controlling and managing the process of the method. By reading and executing the programs stored on the computer-readable storage medium, the computer system can effectively implement the method 10 of the present disclosure.

[0180] Computer program product

[0181] In some embodiments, the present disclosure also provides a computer program product, which includes computer-executable instructions that, when executed by a processor, cause the computer to implement the method 10 described in any one of the foregoing. A computer program product is a product storing computer-executable instructions, and its purpose is to implement the various steps and functions described in the method 10 when executed by the processor 71 of a computer system. The computer-executable instructions may include an operating system, application programs, embedded software, etc., to control and manage the process of the method. By using such a computer program product, a user can execute the method 10 of the present disclosure on a computer system.

[0182] Those skilled in the art will readily conceive of other embodiments of the present disclosure after considering the specification and practicing the content disclosed herein. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include the common general knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and embodiments are only to be considered as exemplary, and the present disclosure is not limited to the precise structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. An optimization method for additive manufacturing parameters using machine learning, comprising: Obtaining the surface topography of the first clad layer constructed from powder layers with different powder characteristics under different powder spreading parameters, and establishing a first database that at least maps the relationships among the powder spreading parameters, powder characteristics, and the surface topography of the first clad layer; Constructing and training a first machine learning model based on the first database to generate a first application model for predicting powder spreading parameters; and Using the first application model to predict the input powder characteristics to output the optimal powder spreading parameters for additive manufacturing.

2. The method according to claim 1, wherein, The powder spreading parameters include powder spreading speed and layer thickness.

3. The method according to claim 1, wherein, The powder characteristics include one or more of powder manufacturing process, powder material, particle shape, and powder particle size.

4. The method according to claim 3, wherein, The powder manufacturing process includes one or more categories of atomization, mechanical pulverization, reduction, and electrolysis, where the atomization includes gas atomization and plasma rotating electrode atomization.

5. The method according to claim 1, further comprising: Capturing images and performing discrete element simulations on the powder pile movement state during the laying process of the powder layers with different powder characteristics under different powder spreading parameters; Calibrating the data trend of the discrete element simulation according to the results of the image capture; And Corresponding the calibrated powder pile movement state of the discrete element simulation with the powder spreading parameters, powder characteristics, and the surface topography of the first clad layer to establish the first database.

6. The method according to claim 5, further comprising: Performing discrete element simulations on the surface topography of the first clad layer; And Validating the first application model according to the powder pile movement state of the discrete element simulation and the surface topography of the first clad layer.

7. The method according to any one of claims 1 to 6, wherein The surface topography of the first clad layer is constructed under the same set of scanning parameters, and the scanning parameters include one or more of laser power, scanning speed, spot diameter, and scanning trajectory.

8. The method according to claim 7, further comprising: Obtaining the surface topography of the second clad layer constructed from powder layers with different powder characteristics under different scanning parameters, and establishing a second database that at least maps the relationships among the scanning parameters, powder characteristics, and the surface topography of the second clad layer, where the surface topography of the second clad layer is constructed under the optimal powder spreading parameters corresponding to the powder characteristics output by the first application model; Constructing and training a second machine learning model based on the second database to generate a second application model for predicting scanning parameters; and Using the second application model to predict the input powder characteristics to output the optimal scanning parameters for additive manufacturing.

9. The method according to claim 8, further comprising: Respectively establishing unimodal databases corresponding to process parameters, formed surface quality, melt channel continuity, and mechanical properties, where the process parameters include the powder spreading parameters and scanning parameters, and the formed surface quality includes the surface topography of the first clad layer and the surface topography of the second clad layer; Performing early fusion on the unimodal databases of the process parameters and formed surface quality to form a multimodal database, and respectively constructing a melt channel continuity prediction model and a mechanical property prediction model based on the unimodal databases of the melt channel continuity and mechanical properties; Perform late fusion on the prediction results of the melt channel continuity prediction model and the mechanical property prediction model, and extract multi-modal features based on the multi-modal database and the late fusion data; Construct sub-prediction models of process parameters and formed surface quality, process parameters and melt channel continuity, and process parameters and mechanical properties respectively according to the multi-modal features; and Jointly train the respectively constructed sub-prediction models to construct a comprehensive prediction model for comprehensively predicting and optimizing the material properties of additive manufacturing.

10. An additive manufacturing parameter optimization system using machine learning, comprising: A first clad layer surface topography acquisition module for acquiring the surface topography of a first clad layer constructed from powder layers with different powder characteristics under different powder spreading parameters; A first database construction module for establishing a first database that at least maps the relationship between powder spreading parameters, powder characteristics, and the surface topography of the first clad layer; A first model training module for constructing and training a first machine learning model according to the first database to generate a first application model for predicting powder spreading parameters; and A powder spreading parameter prediction module for predicting the input powder characteristics using the first application model to output the optimal powder spreading parameters for additive manufacturing.

11. An electronic device, comprising: At least one processor; At least one memory, the at least one memory being coupled to the at least one processor and configured to store instructions executed by the at least one processor, the instructions when executed by the at least one processor causing the electronic device to perform the method according to any one of claims 1 to 9.

12. A computer-readable storage medium storing a computer program, the computer program when executed by a processor implementing the method according to any one of claims 1 to 9.

Images