Reverse design method for process parameters of additive manufacturing of titanium-based parts based on near-spherical powder

Abstract

The application provides a reverse design method of process parameters of a near-spherical powder-based titanium-based part additive manufacturing process. The reverse design method constructs a powder characteristic parameter-process parameter-material performance gradient boosting decision tree (GBDT) model based on a data set for near-spherical powders with different powder characteristics, and uses a Q-Learning reinforcement learning algorithm to construct an agent model capable of outputting process parameters and powder characteristic parameters. The GBDT model and the agent model are fused and applied, and then the best process path is quickly and accurately designed according to the given performance requirements, solving the problems of long material research and development cycle and high cost in the prior art.

Description

Technical Field

[0001] This invention relates to the field of metal material preparation technology, and specifically to a reverse design method for additive manufacturing process parameters of titanium-based parts based on near-spherical powder. Background Technology

[0002] Compared to traditional forming processes such as forging and casting, titanium and titanium alloy parts manufactured using laser powder bed additive manufacturing exhibit superior mechanical properties and design freedom, showing broad application prospects in strategic emerging industries such as aerospace, defense, biomedicine, and marine equipment, attracting high attention from governments and the scientific and technological industry worldwide. Currently, the raw material powders used in additive manufacturing are mainly spherical powders prepared by gas atomization or plasma filament atomization. However, the generally low yield of fine powders leads to their high price; for example, Ti-6Al-4V powder used in selective laser melting generally costs over 1 million yuan per ton, becoming a major bottleneck restricting technological progress and widespread application in this field. Therefore, in recent years, extensive research has been conducted both domestically and internationally on reducing the cost of titanium powder for additive manufacturing. Previous studies have shown that low-cost, irregularly hydrogenated dehydrogenated (HDH) titanium powder can be used as raw material, and high-temperature ball milling technology can be used to prepare low-cost near-spherical titanium powder that meets the requirements for additive manufacturing powders.

[0003] Compared to spherical powders, near-spherical powders exhibit altered powder properties such as sphericity, surface roughness, flowability, and bulk density. This results in differences in thermal conductivity and laser absorptivity compared to spherical powders, leading to significant variations in the density, surface quality, and mechanical properties of formed parts under the same processing parameters. Existing mature processes are primarily designed for spherical powder raw materials and lack process parameters suitable for near-spherical powders. Therefore, researching the process adaptability of near-spherical powders is crucial for the development of low-cost additive manufacturing of titanium-based parts. Currently, process parameter design and optimization still mainly rely on the traditional "cooking" method. This involves workers designing process parameters based on experience to prepare samples, then repeatedly adjusting the process parameters by analyzing the microstructure and properties of the parts until the target performance is achieved. Furthermore, it is ultimately necessary to explore process parameters suitable for mass production to achieve the goal of high-efficiency and low-cost production.

[0004] Therefore, traditional methods suffer from long cycles and high costs. Furthermore, additive manufacturing is a complex nonlinear process involving the continuous interaction of multiple process parameters, and traditional methods can only design parameters within a relatively narrow range, making it difficult to explore the multi-parameter space flexibly and efficiently. This further hinders the rapid response fabrication of titanium parts that meet target performance requirements. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the main objective of this invention is to provide a reverse design method for additive manufacturing process parameters of titanium-based parts based on near-spherical powder. This design method constructs a GBDT model of powder characteristic parameters, process parameters, and material properties based on a dataset for near-spherical powders with different powder characteristics. It also uses the Q-Learning reinforcement learning algorithm to construct an intelligent agent model that can output process parameters and powder characteristic parameters. By fusing and applying the GBDT model and the intelligent agent model, the optimal process path and powder characteristic parameters can be designed quickly and accurately according to given performance requirements.

[0006] To achieve the above objectives, the present invention provides a reverse design method for additive manufacturing process parameters of titanium-based parts based on near-spherical powder.

[0007] The reverse design method for additive manufacturing process parameters of titanium-based parts based on near-spherical powder includes the following steps:

[0008] S1, Create the dataset;

[0009] S2, Use the dataset to establish the optimal GBDT model for powder characteristic parameters, process parameters, and material properties;

[0010] S3. A smart agent model is constructed using the Q-learning reinforcement learning algorithm. Preset target performance data is input into the smart agent model. The smart agent model is then fused with the GBDT optimal model to obtain the optimal process parameters and powder characteristic parameters.

[0011] Furthermore, the preset target performance data is one or more of density, hardness, tensile strength, and elongation;

[0012] The powder characteristic parameters include, but are not limited to, sphericity and bulk density.

[0013] Furthermore, in step S3, preset target performance data is input into the intelligent agent model, and the intelligent agent model is fused with the GBDT optimal model to obtain optimal process parameters and powder characteristic parameters, specifically including:

[0014] S3-1, Input preset target performance data into the intelligent agent model, and the intelligent agent obtains the initial predicted process parameters and powder characteristic parameters corresponding to the input preset target performance data by interacting with the environment;

[0015] S3-2, Input the initial predicted process parameters and powder characteristic parameters into the GBDT optimal model, output the initial material performance prediction data corresponding to the initial predicted process parameters and powder characteristic parameters, and calculate the distance between the material performance prediction data and the preset target performance data.

[0016] S3-3, the agent selects the initial action at time t based on the distance, the correlation between process parameters, powder characteristic parameters and target performance; wherein the action is the adjustment behavior of process parameters and powder characteristics;

[0017] S3-4, the agent interacts with the environment based on the initial action at time t to obtain a new state and reward R;

[0018] S3-5, the agent obtains a new action guidance policy at the current time t through the new state and reward R. If R>0, it means that the agent's current behavior is beneficial to the result, and the agent will execute the new action. If R≤0, it means that the agent's current behavior is not beneficial to the result, and the agent will return to the original state and repeat S3-4 to S3-5.

[0019] S3-6, the agent continuously repeats steps S3-3 to S3-5 until the number of iterations P reaches the preset threshold. The process parameters and powder characteristics corresponding to the maximum cumulative reward from the current time to the final time are taken as the optimal process parameters and powder characteristics. Where P = 100 to 1000, there are T time points in each iteration. At each time point t, the agent updates once until t = T, at which point the iteration ends.

[0020] Furthermore, the distance between the predicted material properties and the preset target properties is specifically represented as follows:

[0021]

[0022] in, The distance between the predicted value and the target value; y is the target value; y is the predicted value at time T; yt is the predicted value of the process parameters at time t predicted by the GBDT optimal model.

[0023] Furthermore, the reward R is the difference between the reciprocal of the distance at the next time step (t+1) and the reciprocal of the distance at the current time step (t), specifically expressed as:

[0024]

[0025] In the formula, γ is the discount factor for the reward value at the future time step, and γ∈[0,1].

[0026] Furthermore, in step S1, establishing the dataset includes the following steps:

[0027] S1-1: Obtain powder characteristic parameters, process parameters, and material performance data to establish a basic dataset;

[0028] S1-2, Classify the data in the basic dataset, and select the data of the corresponding type as the initial dataset according to the material type and preparation technology;

[0029] S1-3, Normalize the initial dataset to obtain the dataset.

[0030] Furthermore, the material performance data is one or a combination of density, hardness, tensile strength, and elongation;

[0031] Preferably, the material types include, but are not limited to, pure titanium, Ti-6Al-4V, and TA15;

[0032] Preferably, the preparation technology includes, but is not limited to, selective laser melting technology and selective laser sintering technology.

[0033] Furthermore, the dataset is divided into a training set, a validation set, and a test set; establishing the optimal GBDT model for powder characteristic parameters-process parameters-material properties using the dataset includes the following steps:

[0034] S2-1, Use the training set and the validation set to train the model and generate an initial GBDT model of powder characteristic parameters-process parameters-material properties;

[0035] S2-2, Random Search combined with K-Fold Cross Validation algorithm is used to optimize the hyperparameters of the initial GBDT model to obtain the improved GBDT model; where K = 5, 10, and the hyperparameters include the number of trees (n_estimator), learning rate (learning_rate), maximum tree depth (max_depth) and subsample ratio (subsample);

[0036] S2-3, Using the test set, evaluate the prediction performance of the improved GBDT model and optimize its design to obtain the optimal GBDT model.

[0037] Furthermore, in steps S2-3, the coefficient of determination R is used. 2 The mean absolute percentage error (MAPE) is used to determine whether the GBDT improved model has reached the required model accuracy to complete the optimization design.

[0038] When the coefficient of determination R 2 The model accuracy is achieved when the mean absolute percentage error (MAPE) is greater than or equal to a preset threshold and the preset threshold is less than or equal to a preset threshold, thus obtaining the optimal GBDT model.

[0039] When the coefficient of determination R 2<Preset threshold or mean absolute percentage error (MAPE)> Preset threshold, reset the RS algorithm to optimize the hyperparameter range and iteration number (n_iter) and continue to optimize the model.

[0040] Furthermore, the determination coefficient R 2 The preset thresholds for mean absolute percentage error (MAPE) are 0.95 and 5%, respectively.

[0041] Preferably, the optimization ranges for the number of trees (n_estimator), learning rate (learning_rate), maximum tree depth (max_depth), and subsample ratio (subsample) in the hyperparameters are 1~1000, 0.01~0.5, 1~10, and 0~1, respectively, and the number of iterations (n_iter) is 200~1500.

[0042] Furthermore, the amount of data in the training set accounts for 40-90% of the total amount of data in the dataset, the amount of data in the validation set accounts for 5-30% of the total amount of data in the dataset, and the amount of data in the test set accounts for 5-30% of the total amount of data in the dataset, and the sum of the percentages of the training set, the validation set, and the test set is always 100%.

[0043] This invention has the following advantages:

[0044] 1) Compared with the traditional "cooking" method, the machine learning method used in this invention makes full use of the existing large amount of additive manufacturing product production and research data. Without having to understand the internal mechanism of additive manufacturing technology, it establishes an implicit relationship model between powder characteristic parameters, additive manufacturing process parameters and part performance, thereby significantly reducing the R&D cycle and cost of near-spherical powder additive manufacturing adaptability process.

[0045] 2) The process parameter reverse design method proposed in this invention, which integrates GBDT model and reinforcement learning, can find the best path to achieve the target performance for a given material property by intelligently optimizing additive manufacturing process parameters and powder property parameters, thereby realizing rapid response preparation of additively manufactured titanium-based parts oriented to performance requirements; compared with traditional parameter optimization model, it has stronger transfer ability and better parameter recommendation effect, and is applicable to most preparation technologies.

[0046] 3) The distance function proposed in this invention calculates the cumulative distance from the initial time to a certain time, which fully considers the error between the current time and the previous time, and helps to find the optimal process path; the proposed reward function fully considers the error between the previous and next time, and obtains a denser learning reward information, which helps to strengthen the accurate optimization of the learning process.

[0047] 4) Compared with spherical titanium powder, the near-spherical powder preparation equipment and process used in this invention are simple, low in cost, and have a high yield of nearly 100%. The powder also has good flowability, which meets the requirements for additive manufacturing powder and can meet the demand for low cost of additive manufacturing titanium-based parts. Attached Figure Description

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0049] Figure 1 A flowchart of the reverse design method provided in the embodiments of the present invention;

[0050] Figure 2 This is a schematic diagram of the reverse design method provided in the embodiments of the present invention. Detailed Implementation

[0051] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0052] This invention conducts a study on the process adaptability of near-spherical titanium powder laser powder bed melting technology, in order to establish an efficient reverse design method for low-cost additive manufacturing process parameters of titanium-based parts with target performance, thereby achieving efficient and low-cost preparation of additively manufactured titanium-based parts with given target performance.

[0053] This invention presents a reverse design method for additive manufacturing process parameters aimed at specific performance targets. Taking any given material property as the target, it outputs predicted process parameters and powder characteristic parameters, which can then be used as the optimal process and powder characteristic parameters. Specifically, an intelligent agent model is integrated with a GBDT optimal model. Before outputting the predicted process parameters, the system automatically inputs them into the GBDT optimal model, outputs predicted material property values, calculates the distance between these predicted values ​​and the target material property values, and obtains adjustment strategies for the process and powder characteristic parameters. This process is iterated repeatedly until the maximum number of iterations P is reached. The process and powder characteristic parameters corresponding to the maximum cumulative reward are then output as the optimal process and powder characteristic parameters, achieving intelligent and precise control of the part's performance.

[0054] The reverse design method for additive manufacturing process parameters of titanium-based parts based on near-spherical powder in this invention includes the following steps:

[0055] S1, Create a dataset; the dataset is divided into a training set, a validation set, and a test set.

[0056] The training set is used for training the GBDT model; the validation set is used to control the training process to prevent overfitting of the GBDT model; and the test set is used to test the model's accuracy. The training and validation sets are used during model training, while the test set is used for model optimization and improvement after training is complete.

[0057] In embodiments of the present invention, the amount of data in the training set accounts for 40-90% of the total amount of data in the dataset, the amount of data in the validation set accounts for 5-30% of the total amount of data in the dataset, and the amount of data in the test set accounts for 5-30% of the total amount of data in the dataset, and the sum of the percentages of the training set, validation set, and test set is always 100%.

[0058] The creation of the dataset in this invention includes the following steps:

[0059] S1-1: Obtain powder characteristic parameters, process parameters, and material performance data from materials databases, publicly published high-quality literature, and experimental data to establish a basic dataset;

[0060] S1-2, classify the data in the basic dataset, and select the data of the corresponding type as the initial dataset according to the material type and preparation technology;

[0061] S1-3, Normalize the initial dataset to obtain the dataset.

[0062] In embodiments of the present invention, the material performance data are one or a combination of density, hardness, tensile strength, and elongation.

[0063] In embodiments of the present invention, powder characteristic parameters include, but are not limited to, sphericity and bulk density.

[0064] In embodiments of the present invention, the material types include, but are not limited to, pure titanium, Ti-6Al-4V, and TA15.

[0065] In the embodiments of the present invention, the preparation techniques include, but are not limited to, selective laser melting and selective laser sintering.

[0066] S2, using the dataset, establish the optimal GBDT model for powder characteristic parameters, process parameters, and material properties, specifically including the following steps:

[0067] S2-1 uses the training set and validation set to train the model and generate the initial GBDT model of powder characteristic parameters-process parameters-material properties.

[0068] In an embodiment of the present invention, the validation set control training process is achieved by adopting an early stopping strategy, that is, training is stopped after the validation set error has decreased continuously for a set number of times, which can effectively prevent overfitting and save training time.

[0069] In embodiments of the present invention, the training set contains multiple sets of training data, and each set of training data includes powder characteristic parameters, process parameters, and corresponding material property data. Using the process parameter data and powder characteristic data from the training set as input, and the material property data as output, the values ​​of hyperparameters n_estimator, learning_rate, max_depth, and subsample are set to establish an initial GBDT model of powder characteristic parameters-process parameters-material properties.

[0070] S2-2 uses Random Search combined with K-Fold Cross Validation algorithm to optimize the hyperparameters of the initial GBDT model, resulting in an improved GBDT model; where K = 5, 10, and the hyperparameters include the number of trees (n_estimator), learning rate (learning_rate), maximum tree depth (max_depth), and subsample ratio (subsample).

[0071] It should be noted that the value of K can be 5 or 10. Of course, the value of K can be designed according to actual needs and is not specifically limited.

[0072] S2-3, using the test set to evaluate the prediction effect and optimize the design of the GBDT improved model, and obtain the best GBDT model; the test set contains multiple sets of test data, and each set of test data includes powder characteristic parameters, process parameters (for better differentiation, denoted as the second process parameters) and corresponding material performance data (denoted as the second material performance data).

[0073] In embodiments of the present invention, optimizing the GBDT improved model using a test set includes the following steps:

[0074] S2-3-1, Input the test set process parameters (second process parameters) into the GBDT improved model to obtain the corresponding material property second prediction data;

[0075] S2-3-2, using the second predicted data of the material properties and the second material property data, the coefficient of determination R is calculated. 2 and mean absolute percentage error (MAPE);

[0076] S2-3-3, using the coefficient of determination R2 The mean absolute percentage error (MAPE) is used to determine whether the GBDT improved model has reached the required model accuracy to complete the optimization design.

[0077] When the coefficient of determination R 2 The model accuracy is achieved when the mean absolute percentage error (MAPE) is ≥ the preset threshold and the GBDT optimal model is obtained.

[0078] When the coefficient of determination R 2 <Preset threshold or mean absolute percentage error (MAPE)> Preset threshold, reset the range of RS algorithm optimization hyperparameters (n_estimator, learning_rate, max_depth, subsample) and iteration number (n_iter) and continue to step S2 for model optimization.

[0079] In an embodiment of the present invention, R 2 The preset thresholds for MAPE can be 0.95 and 5%, respectively.

[0080] In this invention, the coefficient of determination R is used. 2 The mean absolute percentage error (MAPE) is used to determine whether the GBDT improved model has achieved the required model accuracy. 2 The value is in the range of 0 to 1, and the closer it is to 1, the higher the model accuracy. The smaller the MAPE value, the higher the model accuracy.

[0081] S3. A smart agent model is constructed using the Q-learning reinforcement learning algorithm. Pre-set target performance data is input into the smart agent model, which is then fused with the GBDT optimal model to obtain the optimal process parameters and powder characteristic parameters. Specifically, this includes:

[0082] S3-1, Input the preset target performance data into the intelligent agent model. The intelligent agent obtains the initial predicted process parameters and powder characteristic parameters corresponding to the input preset target performance data by interacting with the environment.

[0083] S3-2: Input the initial predicted process parameters and powder characteristic parameters into the GBDT optimal model, and output the material property prediction data corresponding to the initial predicted process parameters and powder characteristic parameters.

[0084] S3-3, the agent selects its initial action at time t based on environmental observations, namely, the correlation between the distance, process parameters, powder characteristic parameters and target performance; where the action is the adjustment behavior of process parameters and powder characteristic data.

[0085] S3-4, the agent interacts with the environment based on the initial action at time t to obtain a new state and reward R;

[0086] S3-5, the agent obtains a new action guidance policy at the current time t through the new state and reward R. If R>0, it means that the agent's current behavior is beneficial to the result, and the agent will execute the new action. If R≤0, it means that the agent's current behavior is not beneficial to the result, and the agent will return to the original state and repeat S3-4 to S3-5.

[0087] S3-6, the agent continuously repeats steps S3-3 to S3-5 until the number of iterations P reaches the preset threshold. The process parameters and powder characteristics corresponding to the maximum cumulative reward from the current moment to the final moment are taken as the optimal process parameters and powder characteristics parameters. Where P = 100 to 1000, there are T moments in each iteration. At each moment t, the agent will update once until t = T, at which point the current iteration ends.

[0088] In an embodiment of the present invention, the distance between the predicted material performance data and the preset target performance data is specifically represented as follows:

[0089]

[0090] in, The distance between the predicted value and the target value; y is the target value; y is the predicted value at time T; yt is the predicted value of the process parameters at time t predicted by the GBDT optimal model.

[0091] In an embodiment of the present invention, the reward R is the difference between the reciprocal of the distance at the next time step (t+1) and the reciprocal of the distance at the current time step (t), specifically expressed as:

[0092]

[0093] γ is a discount factor for the reward value at future time steps, γ∈[0,1].

[0094] The following detailed description of the reverse design method for additive manufacturing process parameters of titanium-based parts based on near-spherical powder, through specific embodiments, will be provided in this invention.

[0095] Example 1:

[0096] Taking the selective laser melting (SLM) forming of Ti-6Al-4V parts with a density of 99.9% as an example, this paper illustrates the prediction / design of printing process parameters and the selection of powder raw materials with certain characteristics. The specific steps are as follows:

[0097] 1) Establish the initial dataset: Extract powder characteristic data, selective laser melting technology process parameters, and Ti-6Al-4 part performance data from public literature, experimental and production data to establish the initial dataset.

[0098] 2) Normalization: The powder characteristic data, process parameters and part performance data in the initial dataset are normalized (0~1) to obtain the powder characteristic and process parameter-Ti-6Al-4 part performance dataset. The dataset is divided into training set, validation set and test set. The data volume in the training set accounts for 90% of the total data volume of the dataset, and the data volume in the validation set and test set each accounts for 5% of the total data volume of the dataset.

[0099] 3) Establish the initial GBDT model of powder properties and selective laser melting process parameters - Ti-6Al-4 part performance: The process parameters and powder properties parameters in the training set and validation set are used as input data, and the titanium alloy part performance data are used as output data. The hyperparameters n_estimator(200), learning_rate(0.02), max_depth(5), and subsample(1) are set for model training.

[0100] Then, Random Search combined with K-Fold Cross Validation algorithm was introduced to automatically optimize the above hyperparameters, improve the model accuracy and generalization ability, and obtain the GBDT improved model; where K=10, the hyperparameters n_estimator, learning_rate, max_depth and subsample optimization range in Random Search algorithm were set to 1~1000, 0.01~0.5, 1~10 and 0~1 respectively, and the number of iterations n_iter was 400.

[0101] 4) The accuracy of the GBDT improved model is evaluated using a test set. Specifically, the second process parameters and second powder property data of each test set are used as input data to obtain the second predicted density data. Then, the coefficient of determination R is calculated using this second predicted density data and the density data (corresponding to the second material property data) in each test set. 2 =0.94 (not meeting the ≥0.95 condition), therefore the hyperparameters n_estimator, learning_rate, max_depth, and subsample optimization range were reset to 200–1000, 0.01–0.5, 3–10, and 0–1 respectively, the number of iterations n_iter was set to 600, and the model was trained and tested to obtain the coefficient of determination R. 2 =0.98 and mean absolute percentage error (MAPE) = 2.1% (satisfying R 2 The optimal GBDT model was obtained under the condition that the value is ≥0.95 and MAPE ≤5%.

[0102] 5) Using a density of 99.9% as the target performance data, the best path to achieve the target performance is output by combining the agent model based on Q-learning reinforcement learning algorithm with the above-mentioned GBDT optimal model, thereby realizing intelligent and precise control of the performance of the part. The maximum number of iterations is P=800.

[0103] Examples 2 and 3 used the same preparation techniques and raw materials as Example 1, only the target performance was different. The optimal process parameters and powder characteristics obtained in Examples 1 to 3 are shown in Table 1.

[0104] Table 1. Summary of optimal process parameters corresponding to the target performance in Examples 1-3.

[0105]

[0106] It should be noted that the term "comprising" and any variations thereof in the specification and claims of this invention are intended to cover non-exclusive inclusion, for example, including a series of components that are not necessarily limited to those explicitly listed, but may include other components that are not explicitly listed or that are inherent to the component.

[0107] The descriptions using terms such as "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0108] Furthermore, the technical solutions of the various embodiments can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0109] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A reverse design method for additive manufacturing process parameters of titanium-based parts based on near-spherical powder, characterized in that, Includes the following steps: S1, Create a dataset; the dataset is divided into a training set, a validation set, and a test set; S2, using the dataset, establish the optimal GBDT model for powder characteristic parameters, process parameters, and material properties, including the following steps: S2-1, Use the training set and the validation set to train the model and generate an initial GBDT model of powder characteristic parameters-process parameters-material properties; S2-2, Random Search combined with K-Fold Cross Validation algorithm is used to optimize the hyperparameters of the initial GBDT model to obtain the improved GBDT model; where K=5, 10, and the hyperparameters include the number of trees, learning rate, maximum tree depth and subsampling ratio; S2-3, Using the test set, evaluate and optimize the prediction performance of the improved GBDT model to obtain the optimal GBDT model; S3, an agent model is constructed using the Q-learning reinforcement learning algorithm. Preset target performance data is input into the agent model, and the agent model is fused with the GBDT optimal model to obtain optimal process parameters and powder characteristic parameters; wherein... Inputting preset target performance data into the intelligent agent model, and then fusing the intelligent agent model with the GBDT optimal model to obtain optimal process parameters and powder characteristic parameters, specifically includes: S3-1, Input preset target performance data into the intelligent agent model, and the intelligent agent obtains the initial predicted process parameters and powder characteristic parameters corresponding to the input preset target performance data by interacting with the environment; S3-2, Input the initial predicted process parameters and powder characteristic parameters into the GBDT optimal model, output the initial material performance prediction data corresponding to the initial predicted process parameters and powder characteristic parameters, and calculate the distance between the material performance prediction data and the preset target performance data. S3-3, the agent selects the initial action at time t based on the distance, the correlation between process parameters, powder characteristic parameters and target performance; wherein the action is the adjustment behavior of process parameters and powder characteristics; S3-4, the agent interacts with the environment based on the initial action at time t to obtain a new state and reward R; S3-5, the agent obtains a new action guidance policy at the current time t through the new state and reward R. If R>0, it means that the agent's current behavior is beneficial to the result, and the agent will execute the new action. If R≤0, it means that the agent's current behavior is not beneficial to the result, and the agent will return to the original state, repeating S3-4~S3-5. S3-6, the agent continuously repeats steps S3-3 to S3-5 until the number of iterations P reaches the preset threshold. The process parameters and powder characteristics corresponding to the maximum cumulative reward from the current moment to the final moment are taken as the optimal process parameters and powder characteristics parameters. Where P = 100~1000, there are T moments in each iteration. At each moment t, the agent updates once until t = T, at which point the iteration ends.

2. The reverse engineering method as described in claim 1, characterized in that, The preset target performance data is one or a combination of density, hardness, tensile strength, and elongation.

3. The reverse engineering method as described in claim 1, characterized in that, The powder characteristic parameters include sphericity and bulk density.

4. The reverse engineering method as described in claim 1, characterized in that, The distance between the predicted material properties data and the preset target properties data is represented as: in, The distance between the predicted value and the target value; The target value; y for T The predicted value at any given time; For predictions made using the best GBDT model t The predicted values ​​of the process parameters at each time point.

5. The reverse engineering method as described in claim 1, characterized in that, The reward R is the difference between the reciprocal of the distance at the next time step (t+1) and the reciprocal of the distance at the current time step (t), specifically expressed as: In the formula, γ is the discount factor for the reward value at the future time step, and γ∈[0,1].

6. The reverse engineering method as described in claim 1, characterized in that, In step S1, establishing the dataset includes the following steps: S1-1: Obtain powder characteristic parameters, process parameters, and material performance data to establish a basic dataset; S1-2, Classify the data in the basic dataset, and select the data of the corresponding type as the initial dataset according to the material type and preparation technology; S1-3, Normalize the initial dataset to obtain the dataset.

7. The reverse engineering method as described in claim 6, characterized in that, The material performance data is one or a combination of density, hardness, tensile strength, and elongation.

8. The reverse engineering method as described in claim 5, characterized in that, The material types include pure titanium, Ti-6Al-4V, and TA15.

9. The reverse engineering method as described in claim 6, characterized in that, The preparation techniques include selective laser melting and selective laser sintering.

10. The reverse engineering method as described in claim 1, characterized in that, In steps S2-3, the coefficient of determination R is used. 2 The mean absolute percentage error (MAPE) is used to determine whether the GBDT improved model has reached the preset accuracy to complete the optimization design. When the coefficient of determination R 2 The model accuracy is achieved when the mean absolute percentage error (MAPE) is greater than or equal to a preset threshold and the preset threshold is less than or equal to a preset threshold, thus obtaining the optimal GBDT model. When the coefficient of determination R 2 <Preset threshold or mean absolute percentage error (MAPE)> Preset threshold, reset RS algorithm to optimize the hyperparameter range and iteration number and continue to optimize the model.

11. The reverse engineering method as described in claim 10, characterized in that, The determination coefficient R 2 The preset threshold distributions for the mean absolute percentage error (MAPE) are 0.95 and 5%.

12. The reverse engineering method as described in claim 1, characterized in that, The optimization ranges for the number of trees, learning rate, maximum tree depth, and subsampling ratio in the hyperparameters are 1~1000, 0.01~0.5, 1~10, and 0~1, respectively, and the number of iterations is 200~1500.

13. The reverse engineering method as described in claim 1, characterized in that, The training set accounts for 40-90% of the total data volume of the dataset, the validation set accounts for 5-30% of the total data volume of the dataset, and the test set accounts for 5-30% of the total data volume of the dataset. The sum of the percentages of the training set, the validation set, and the test set is always 100%.

Images