Proxy model construction method and system for metal additive manufacturing heat conduction process

Abstract

The invention discloses an agent model construction method and system for a metal additive manufacturing heat conduction process, and relates to the technical field of heat conduction, and the method comprises the following steps: constructing a metal additive manufacturing heat conduction simulation model by using a moving heat source model; technological parameters with the influence on the metal additive manufacturing heat conduction simulation model larger than a threshold value are selected as input parameters of the agent model, and the depth of the molten pool serves as output response of the agent model; obtaining a construction expression of the GPR agent model and an output analytical expression; and a PVC learning function and a self-adaptive point adding algorithm are defined, the variance of the GPR agent model is obtained through the PVC learning function and the self-adaptive point adding algorithm, and the optimal GPR agent model is obtained when the stopping criterion is met. According to the method, the prediction precision of the proxy model is improved, and the calculation burden of heat conduction simulation is reduced.

Description

Technical Field

[0001] The present invention relates to the field of heat conduction technology, and in particular to a method and system for constructing an agent model of a heat conduction process in metal additive manufacturing. Background Art

[0002] Additive Manufacturing (AM) is a new manufacturing technology based on the principle of discrete-accumulation that has been developed in the past 30 years. It has the advantages of reducing waste, rapid manufacturing, high flexibility, feasibility of complex geometric products, and high design freedom. Among the seven types of AM technologies, laser powder bed melting is considered to be one of the most common processes in direct metal manufacturing because it can promote near-net shape manufacturing with the highest precision. However, compared with traditional subtractive manufacturing, additive manufacturing faces the defects of large dispersion of quality and performance of formed parts, which is specifically manifested in the strong variability of initial defects and mechanical properties of formed parts. This is because metal additive manufacturing is a complex multi-scale, multi-physical field coupling process. Under the environment of rapid cooling and heating, the molten pool and its surrounding areas undergo a process of rapid melting, cooling, and solidification. At this time, microscopic defects such as poor fusion, pores, inclusions, and microcracks are easily generated inside the material, and even macroscopic defects such as warping, deformation, and cracking of printed components occur, thereby reducing the mechanical properties of printed components. Therefore, for such a high-uncertainty process, it is crucial to understand the complex relationship between structure, process and properties generated by the interaction of physical phenomena at multiple time and space scales in the process, which requires a detailed metal additive manufacturing heat conduction simulation model to simulate the complex behavior of the melt pool. However, such modeling usually requires various details such as thermal-fluid dynamics, and even if the model is given reasonable approximations and assumptions, a single simulation calculation of the model is still very expensive.

[0003] In order to improve computational efficiency, Gaussian process regression surrogate models are introduced to reduce the computational burden associated with large-scale simulations. The core idea of ​​surrogate models is to use a relatively simple model to approximate complex functions. This model is able to capture the relationship between input variables and output responses without performing complete complex calculations. Initially, surrogate models were mainly used for simple mathematical function approximations, such as polynomial interpolation. With the advancement of computer technology, more complex surrogate models, such as radial basis functions and kriging, have begun to be applied. In recent years, with the development of machine learning technology, the application scope of surrogate models has been further expanded, including Gaussian process regression and artificial neural networks. The accuracy of surrogate models depends on the selection of surrogate functions, the determination of surrogate function parameters, and the selection of sampling methods. Only with appropriate parameter values ​​can a surrogate model with higher accuracy be obtained. In addition, the initial number of samples when constructing a surrogate model also has a great influence on the accuracy of the surrogate model. Without overfitting, the larger the initial number of samples, the higher the accuracy of the surrogate model.

[0004] It should be pointed out that due to the high time cost of simulating a single metal additive manufacturing heat conduction simulation model, it is impossible to simulate a large number of initial samples for training the proxy model. The GPR proxy model introduced for the proxy simulation model due to the high time cost of the simulation model is not accurate enough, resulting in poor metal additive manufacturing heat conduction simulation results. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and provide a method and system for constructing a proxy model of the heat conduction process in metal additive manufacturing, so as to solve the problem in the prior art that the GPR proxy model used for the proxy simulation model is not accurate enough due to the high time cost of the simulation model, resulting in poor simulation effect of heat conduction in metal additive manufacturing.

[0006] The present invention specifically provides the following technical solutions: A method for constructing a proxy model of a heat conduction process in metal additive manufacturing comprises the following steps: Use the moving heat source model to build a metal additive manufacturing heat conduction simulation model; A Gaussian process regression (GPR) proxy model is constructed through input parameters and output responses, and the output response integral form of the GPR proxy model is obtained; wherein the input parameters are process parameters that have an influence on the metal additive manufacturing heat conduction simulation model greater than a threshold, and the output response is the molten pool depth in the metal additive manufacturing heat conduction simulation model; The expected variance contribution PVC learning function and the adaptive point-adding algorithm are obtained, and the point with the largest variance contribution in the output response integral form of the GPR proxy model is obtained through the PVC learning function. The points with the largest variance contribution are added to the training data in sequence through the adaptive point-adding algorithm, and the GPR proxy model is trained until the variance of the GPR proxy model meets the stopping criterion, thereby obtaining the optimal GPR proxy model.

[0007] Preferably, the use of a mobile heat source model to construct a metal additive manufacturing heat conduction simulation model includes: The moving heat source model is used to simulate the heat input of the laser in the additive manufacturing AM process. The specific expression is: ; in, is the material density, is the specific heat, is the current temperature, For time, , and is the coordinate of the reference system, , and for , and Thermal conductivity in the axial direction, The amount of heat generated per unit volume.

[0008] Preferably, the process parameters include: laser power , Printing speed , material parameters and heat source parameters, where material parameters include material absorption rate , thermal conductivity of the material in y direction , thermal conductivity of the material in z direction , where the heat source parameters include the simulated conical heat source height , Simulation cone heat source upper base radius , Simulated conical heat source bottom radius .

[0009] Preferably, constructing a Gaussian process regression (GPR) proxy model through input parameters and output responses includes: For any training data set in the sample space , get and Functional relationship between ;in, Input parameters for the heat transfer simulation model for metal additive manufacturing, is the corresponding molten pool depth, g (·) is the correlation function, i is the current data volume, n is the maximum amount of data; Through a latent function that obeys the prior Gaussian distribution and is in the probability space and observation error Approximately represents the functional relationship; the specific expression is: ; in, Characterize the noise in the data and Cannot be The explanation part, is a set element in the probability space; If the GPR proxy model In the given training set And sample space The mean function is used to obtain and covariance function express, is a random variable, is the mean operation.

[0010] Preferably, the output response integral form of the GPR proxy model is obtained as follows: The output response is of the integral form d The specific expression is: ; in It refers to the variable with respect to the weight density 's points.

[0011] Preferably, the method of obtaining the expected variance contribution PVC learning function and the adaptive point adding algorithm, obtaining the point with the largest variance contribution in the output response integral form of the GPR proxy model through the PVC learning function, and sequentially adding the points with the largest variance contribution to the training data through the adaptive point adding algorithm, comprises: Get the point within the prior range that contributes the most to the variance , through which the PVC learning function is defined , the specific expression is: ; in: Indicated in the weight density Next About The integral of Indicates taking Training set The component is of dimensional column vector, It is The elements are of N×N dimensional matrix, T is the transpose, for N×N Dimensional matrix to calculate the square exponential kernel of the training data; Use the adaptive point addition algorithm to The maximum value points are added to the training data set in sequence To obtain the minimum variance , It is the GPR proxy model.

[0012] Preferably, the stopping criterion is specifically: The coefficient of variation of the integral is defined as the standard deviation divided by the mean, and the stopping criterion of the adaptive point addition algorithm is obtained. The specific expression is: ; make is the stopping criterion, is the mean, The output response is in the form of an integral d The predicted mean.

[0013] The present invention provides a proxy model construction system for a metal additive manufacturing heat conduction process, comprising: A simulation model building module for building a metal additive manufacturing heat conduction simulation model using a moving heat source model; A model definition module is used to construct a Gaussian process regression (GPR) proxy model through input parameters and output responses, and obtain the output response integral form of the GPR proxy model; wherein the input parameters are process parameters that have an impact on the metal additive manufacturing heat conduction simulation model greater than a threshold, and the output response is the molten pool depth in the metal additive manufacturing heat conduction simulation model; The model optimization module is used to obtain the expected variance contribution PVC learning function and the adaptive point-adding algorithm. The point with the largest variance contribution in the output response integral form of the GPR proxy model is obtained through the PVC learning function, and the points with the largest variance contribution are added to the training data in sequence through the adaptive point-adding algorithm. The GPR proxy model is trained until the variance of the GPR proxy model meets the stopping criterion, and the optimal GPR proxy model is obtained.

[0014] The present invention provides a computer device, including a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of the above-mentioned method for constructing an agent model of a heat conduction process in metal additive manufacturing.

[0015] The present invention provides a storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the above-mentioned method for constructing an agent model of a heat conduction process in metal additive manufacturing are implemented.

[0016] Compared with the prior art, the present invention has the following significant advantages: The present invention defines a GPR proxy model through the molten pool depth in the metal additive manufacturing heat conduction simulation model and the process parameters whose influence on the metal additive manufacturing heat conduction simulation model is greater than a threshold value. When training the GPR proxy model, a PVC learning function is used to obtain the point with the largest variance contribution in the output response integral form of the GPR proxy model, and the points with the largest variance contribution are sequentially added to the training data through an adaptive point adding algorithm to train the GPR proxy model, thereby solving the problem that the time cost of a single metal additive manufacturing heat conduction simulation model simulation in traditional technology is high, resulting in the inability to obtain training samples. While improving the prediction accuracy of the proxy model, the computational burden of the heat conduction simulation is reduced, and the proxy model finally obtained meets the proxy accuracy requirements in practical applications, thereby improving the metal additive manufacturing heat conduction simulation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 : is a diagram of a heat conduction simulation model for metal additive manufacturing established in an embodiment of the present invention; wherein Figure 1 (a) is the geometric structure diagram of the metal additive manufacturing heat conduction simulation model. Figure 1 (b) is a heat representation diagram of the metal additive manufacturing heat conduction simulation model; Figure 2 Graph showing experimental observation results of molten pool depth in an embodiment of the present invention; Figure 3 This is a flow chart of PVC adaptive point addition in an embodiment of the present invention; Figure 4 This is a comparison diagram of the results of the leave-one-out cross-validation method in an embodiment of the present invention; Figure 5 is a convergence diagram of the coefficient of variation in an embodiment of the present invention; Figure 6 It is a diagram of the improvement amount of the improvement points added by the optimization algorithm in the embodiment of the present invention; Figure 7 The present invention is a flow chart of a method for constructing an agent model of a heat conduction process in metal additive manufacturing. DETAILED DESCRIPTION

[0018] The following is a clear and complete description of the technical solutions of the embodiments of the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0019] like Figure 7 As shown, the present invention provides a method for constructing a proxy model of a heat conduction process in metal additive manufacturing, which specifically includes the following steps: Step S1: Use the moving heat source model to build a metal additive manufacturing heat conduction simulation model.

[0020] The metal additive manufacturing heat transfer simulation model includes a moving heat source model to simulate the heat input of the laser in the additive manufacturing AM process, which is represented by the governing expression of the three-dimensional heat transfer process: ; in is the material density, is the specific heat, is the current temperature, For time, , and is the coordinate of the reference system, , and for , and Thermal conductivity in the axial direction, The amount of heat generated per unit volume.

[0021] It contains a moving heat source model for simulating the laser heat input of the additive manufacturing AM process and a uniform cube with effective powder layer material properties and substrate layer material properties to build a metal additive manufacturing heat conduction simulation model.

[0022] In the selection of mobile heat source models, different geometric shapes are often used to simulate the actual shape of the heat source, such as cylinder, hemisphere, double ellipsoid and cone. Here, the cone heat source model is used, and the mathematical expression of the heat source intensity at a certain point can be written as: ; in, and are the radii of the top and bottom of the cone, respectively, and , is the material absorption rate, is the laser power, exp is the exponential function with e as the base, is the radius of the cone at its current height, H is the height of the simulated conical heat source.

[0023] like Figure 1 As shown in this work, COMSOL 5.6 Multiphysics commercial software was used for macro-scale finite element thermal analysis, and the model output was the molten pool size, specifically the molten pool depth. The established simulation model consists of two parts: powder and solid matrix. The upper part is the powder layer, the material used is stainless steel 17PH-4PH, and the size is ; The lower part is a solid matrix, made of low carbon steel, with a size of In addition, to speed up the simulation, only half of the model geometry needs to be constructed and symmetry constraints are set on the symmetry plane, since the scanning area is symmetrical with respect to the vertical plane containing the laser beam motion line.

[0024] In addition, considering the heat loss of the upper surface of the powder layer to the ambient air through radiation and convection, the convection heat coefficient and radiation heat coefficient are defined as 15 and 0.5 respectively. The ambient temperature is set to 293K and the preheating temperature is set to 353K. In addition, the minimum unit is 20 in the laser beam action area. The grid elements are used for grid division, and other areas are filled with standard size units. After simulation test, the calculation time required is about two hours.

[0025] Step S2: construct a Gaussian process regression (GPR) proxy model through input parameters and output responses, and obtain the integral form of the output response of the GPR proxy model; wherein the input parameters are process parameters whose influence on the metal additive manufacturing heat conduction simulation model is greater than a threshold, and the output response is the molten pool depth in the metal additive manufacturing heat conduction simulation model.

[0026] Proxy model input parameters, select the process parameters that have a greater impact on the simulation model output results: laser power , Printing speed , Material parameters: Material absorption rate , thermal conductivity of the material in y direction , thermal conductivity of the material in z direction And heat source parameters: simulated conical heat source height , Simulation cone heat source upper base radius , Simulated conical heat source bottom radius The output of the surrogate model responds to the selected melt pool depth.

[0027] Determine the model input parameters and model output response. The input parameter list is as follows: Table 1 Input parameter settings

[0028] The established metal additive manufacturing heat conduction simulation model is macroscopic, and the model output response is selected as the depth of the molten pool after the laser melts the powder. ,Right now For experimental observations of the molten pool depth, refer to Figure 2 .

[0029] GPR proxy model construction expression and output analytical expression: for any training data set in the sample space , assuming the output and Functional relationship between Yes, get and Functional relationship between ;in, Input parameters for the heat transfer simulation model for metal additive manufacturing, is the corresponding molten pool depth, g (·) is the correlation function, i is the current data volume, n is the maximum amount of data; it is distributed by a priori Gaussian distribution and in the probability space The latent function and observation error Approximately, .in The noise used to characterize the data and Cannot be represented by GPR model The explanation part is usually assumed to be an independent random variable obeying a Gaussian distribution, that is, , assuming the data is noise-free, is a set element in the probability space. The input corresponds to the input parameters of the metal additive manufacturing heat conduction simulation model, including power speed, material parameters and heat source parameters, a total of eight (see Table 1 for details); the output corresponds to the molten pool depth.

[0030] Under the above assumptions, if the GPR proxy model In the given training set And sample space trained in , then it can be obtained by its mean function and covariance function The only structure is, is a random variable, is the mean operation.

[0031] The analytical expression of the mean function of the GPR proxy model (taking the linear mean function as an example) is: , and the analytical expression of the covariance function (taking the square exponential kernel as an example) is: ,in, and is the regression coefficient, For the i Input parameters for a metal additive manufacturing heat transfer simulation model, After training the GPR proxy model, the predicted mean can be output through the GPR proxy model. , prediction variance and the predicted covariance , the analytical expressions of the three are: ; ; ; in Indicates taking Training set The component is of dimensional column vector, It is The elements are of dimensional matrix, i.e. the variance of the training set, Input parameters of the metal additive manufacturing heat conduction simulation model, representing the points to be predicted, for The mean function of the training set, is the squared exponential kernel of the random variable in the integral, To express Training set The component is of dimensional column vector.

[0032] The output response integral form that makes the GPR proxy model more predictive and global is: .

[0033] in It refers to the variable with respect to the weight density 's points.

[0034] Analytically derive the mean of the GPR model for this integral prediction and variance Right now: ; ; in, yes In the weight density and The integral with respect to two parameters is is the variance, To express Training set The component is of dimensional column vector.

[0035] At the given parameters Under the condition of standard normal distribution, the integral of the square exponential kernel and With an analytical expression: ; ; Among them, vec is a vector representation, It means constructing a column vector with diagonal elements. represents the inverse of the sum of the given parameters, I Indicates the corresponding variable. The calculation of depends on the selected mean function and the weight density of the parameters. For low-dimensional problems, a constant mean function can be applied, and then , when the parameter When the weight density of follows the standard normal distribution, The integral of is: ; In the case of high-dimensional nonlinearity, the constant mean function often cannot meet the nonlinear requirements of the model, so the linear mean function is used. , when the parameter When the weight density of follows the standard normal distribution, The integral is as follows: ; Define the global integral , the stopping criterion is: ; Step S3: define the expected variance contribution PVC (Posterior Variance Contribution) learning function and the adaptive point adding algorithm, obtain the variance and mean of the GPR proxy model through the PVC learning function and the adaptive point adding algorithm, and obtain the optimal proxy model when the stopping criterion is met.

[0036] According to step S2, the predicted mean of the probability integral is equal to the integral of the mean function plus the improvement learned from the training data , the predicted variance of the probability integral is equal to the covariance function of the probability integral minus , which is a measure of the reduction in variance of the integral learned from the training data, and can be interpreted as the remaining epistemic uncertainty in the integral after the information learned from the training data. Therefore, in order to minimize the epistemic uncertainty of the probability integral, we need to find the variance of the predicted epistemic uncertainty The point with the largest contribution within the prior range Therefore, the following PVC learning function is defined through this point: ; in: Indicated in the weight density Next About This formula is actually to take the derivative of the variance of the integral in step S3 to obtain the point with the largest contribution to the variance of the integral. Indicates taking Training set The component is of dimensional column vector, It is The elements are of N×N dimensional matrix, T is the transpose, for N×N dimensional matrix to compute the squared exponential kernel of the training data.

[0037] The adaptive point adding algorithm is: The maximum value points are added to the training data set in sequence , so that the variance Get the most effective reduction, It is the GPR proxy model.

[0038] The stopping criterion of the algorithm is: define the coefficient of variation of the integration as the standard deviation divided by the mean, that is: ,make is the stopping criterion, The output response is in the form of an integral d The predicted mean, is the mean value. You can specify a number yourself. In addition, in practical applications, when the initial sample size is small, although the integral result does not converge, it is still possible to meet this stopping criterion. This can be avoided by delaying the judgment, that is, the algorithm can be set to complete only when the stopping condition is met for multiple consecutive times (for example, three times).

[0039] The leave-one-out cross-validation method was used to verify the goodness of fit of the optimized GPR proxy model accuracy, and the coefficient of determination was used for measurement.

[0040] The leave-one-out cross-validation method takes 30% of the input samples of the iteration sample points as the test set of the GPR proxy model to test the prediction accuracy of the GPR proxy model on non-training points.

[0041] according to Figure 3 The process shown is as follows: First, the other six parameters are randomly sampled in the form of uniform distribution under the power and speed specified in COMSOL to obtain 24 sets of input-output initial training samples. That is, the number of initial samples is defined N 0 and stopping criteria , using sampling method to randomly generate N 0 training points (i.e. sample points) and calculate the corresponding response function values ​​as the initial training set , use the training set to train the GPR model and calculate and . Use the 24 groups of samples as the initial training set to construct an eight-dimensional GPR proxy model, in which the proxy model uses a linear mean function and a square exponential kernel. After the proxy model is trained, the PVC learning function is used to find the first global optimization point (maximum point). After obtaining the first optimization point, substitute it into the metal additive manufacturing heat conduction simulation model for simulation calculation. After obtaining the simulation results, add the optimization point to the GPR proxy model training set. Repeat this work until the predicted coefficient of variation reaches the set delayed convergence judgment condition (stop criterion), and use the leave-one-out cross-validation method for cross-validation to determine the improvement effect of the model accuracy.

[0042] Here we set the initial training point (initial sample point) , stopping criterion , set the delayed convergence judgment to 4. It should be noted that when performing the optimization algorithm to improve the proxy accuracy of the GPR proxy model, it is necessary to convert the extracted sample points that obey the uniform distribution to the standard normal distribution interval.

[0043] Finally, 28 optimization points were added and the loop was exited. Therefore, 10 groups of samples were extracted as test samples without specifying power and speed parameters, and substituted into the metal additive manufacturing heat conduction simulation model for calculation to perform leave-one-out cross validation. The coefficient of determination value was used as an indicator. To measure, The definition of is: ; in, is the fitted value, is the true observed value, is the average of the true observations, i is the current data volume, n is the maximum amount of data, The closer the value is to 1, the better the regression model fits the observed value; conversely, The smaller the value of is, the worse the regression model fits the observed value. Figure 4 The calculated square error is 6.3451 and the coefficient of determination is 0.9760, which is better than 0.9265 before optimization. Therefore, the optimized GPR proxy model has a better effect in fitting the metal additive manufacturing heat conduction simulation model.

[0044] in accordance with Figure 5 The convergence diagram of the coefficient of variation in the algorithm flow shows that false convergence occurs at the sixth point, which also illustrates the importance of delayed convergence. Figure 6 The improvement amount of each improvement point is iteratively added during the optimization algorithm. It can be seen that the overall trend of the improvement amount is decreasing, but it is not always the case in each adjacent step. This is normal because in some important local areas without training data, the GPR proxy model may give a less biased prediction. It can also be seen that the PVC learning function Small values ​​of do not necessarily mean that the change in the integrated posterior estimate is small.

[0045] Based on the above method, the present invention provides a proxy model construction system for the heat conduction process of metal additive manufacturing, including: a simulation model construction module, a model definition module and a model optimization module.

[0046] Among them, the simulation model construction module is used to use the mobile heat source model to construct a metal additive manufacturing heat conduction simulation model; the model definition module is used to construct a Gaussian process regression GPR proxy model through input parameters and output responses, and obtain the output response integral form of the GPR proxy model; wherein the input parameters are process parameters whose influence on the metal additive manufacturing heat conduction simulation model is greater than the threshold, and the output response is the molten pool depth in the metal additive manufacturing heat conduction simulation model; the model optimization module is used to obtain the expected variance contribution PVC learning function and the adaptive point addition algorithm, and the point with the largest variance contribution in the output response integral form of the GPR proxy model is obtained through the PVC learning function, and the points with the largest variance contribution are added to the training data in turn through the adaptive point addition algorithm, and the GPR proxy model is trained until the variance of the GPR proxy model meets the stopping criterion to obtain the best GPR proxy model.

[0047] The present invention also provides a computer device, including a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of a method for constructing an agent model of a heat conduction process in metal additive manufacturing.

[0048] According to the disclosed embodiments, a computing device may communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth communications, etc.), or with any device (e.g., routers, modems, etc.) that enables a computing device to communicate with one or more other computing devices.

[0049] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for constructing an agent model of a heat conduction process in metal additive manufacturing are implemented.

[0050] According to the disclosed embodiments, the storage medium may be a non-volatile computer-readable storage medium, such as but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, the storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0051] The above content is a further detailed description of the present invention in combination with a specific preferred embodiment. For technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for constructing a proxy model of a heat conduction process in metal additive manufacturing, characterized in that: include: Use the moving heat source model to build a metal additive manufacturing heat conduction simulation model; A Gaussian process regression (GPR) proxy model is constructed through input parameters and output responses, and the output response integral form of the GPR proxy model is obtained; wherein the input parameters are process parameters that have an influence on the metal additive manufacturing heat conduction simulation model greater than a threshold, and the output response is the molten pool depth in the metal additive manufacturing heat conduction simulation model; The expected variance contribution PVC learning function and the adaptive point-adding algorithm are obtained, and the point with the largest variance contribution in the output response integral form of the GPR proxy model is obtained through the PVC learning function. The points with the largest variance contribution are added to the training data in sequence through the adaptive point-adding algorithm, and the GPR proxy model is trained until the variance of the GPR proxy model meets the stopping criterion, thereby obtaining the optimal GPR proxy model.

2. The method for constructing a proxy model for a heat conduction process in metal additive manufacturing according to claim 1, characterized in that: The method of using a mobile heat source model to construct a metal additive manufacturing heat conduction simulation model includes: The moving heat source model is used to simulate the heat input of the laser in the additive manufacturing AM process. The specific expression is: ； in, is the material density, is the specific heat, is the current temperature, For time, , and is the coordinate of the reference system, , and for , and Thermal conductivity in the axial direction, The heat generated per unit volume inside.

3. The method for constructing a proxy model for a heat conduction process in metal additive manufacturing according to claim 1, characterized in that: The process parameters include: laser power , Printing speed , material parameters and heat source parameters, where material parameters include material absorption rate , thermal conductivity of the material in y direction , thermal conductivity of the material in z direction , where the heat source parameters include the simulated conical heat source height , Simulated conical heat source upper base radius , Simulated conical heat source bottom radius .

4. The method for constructing a proxy model for a heat conduction process in metal additive manufacturing according to claim 1, characterized in that: The method constructs a Gaussian process regression (GPR) proxy model through input parameters and output responses, including: For any training data set in the sample space , get and Functional relationship between ;in, Input parameters for the heat transfer simulation model for metal additive manufacturing, is the corresponding molten pool depth, g (·) is the correlation function, i is the current data volume, n is the maximum amount of data; Through a latent function that obeys the prior Gaussian distribution and is in the probability space and observation error Approximately represents the functional relationship; the specific expression is: ； in, Characterize the noise in the data and Cannot be The explanation part, is a set element in the probability space; If the GPR proxy model In the given training set And sample space Obtained through training, using the mean function and covariance function express, is a random variable, is the mean operation.

5. The method for constructing a proxy model for a heat conduction process in metal additive manufacturing according to claim 4, characterized in that: The output response integral form of the GPR proxy model is obtained as follows: The output response is of the integral form d The specific expression is: ； in It refers to the variable with respect to the weight density 's points.

6. A method for constructing a proxy model for a heat conduction process in metal additive manufacturing according to claim 5, characterized in that: The method of obtaining the expected variance contribution PVC learning function and the adaptive point adding algorithm, obtaining the point with the largest variance contribution in the output response integral form of the GPR proxy model through the PVC learning function, and sequentially adding the points with the largest variance contribution to the training data through the adaptive point adding algorithm, includes: Get the point within the prior range that contributes the most to the variance , through which the PVC learning function is defined , the specific expression is: ； in: Indicated in the weight density Next About The integral of Indicates taking Training set The component is of dimension column vector, It is The elements are of N×N dimensional matrix, T is the transpose, for N×N Dimensional matrix to calculate the square exponential kernel of the training data; Use the adaptive point addition algorithm to The maximum value points are added to the training data set in sequence To obtain the minimum variance , It is the GPR proxy model.

7. A method for constructing a proxy model for a heat conduction process in metal additive manufacturing according to claim 6, characterized in that: The stopping criteria are specifically: Define the coefficient of variation of the integral as the standard deviation divided by the mean, and obtain the stopping criterion of the adaptive point addition algorithm , the specific expression is: ； make is the stopping criterion, is the mean, The output response is in the form of an integral d The predicted mean.

8. A proxy model building system for the heat conduction process of metal additive manufacturing, characterized in that: include: A simulation model building module for building a metal additive manufacturing heat conduction simulation model using a moving heat source model; A model definition module is used to construct a Gaussian process regression (GPR) proxy model through input parameters and output responses, and obtain the output response integral form of the GPR proxy model; wherein the input parameters are process parameters that have an impact on the metal additive manufacturing heat conduction simulation model greater than a threshold, and the output response is the molten pool depth in the metal additive manufacturing heat conduction simulation model; The model optimization module is used to obtain the expected variance contribution PVC learning function and the adaptive point-adding algorithm. The point with the largest variance contribution in the output response integral form of the GPR proxy model is obtained through the PVC learning function, and the points with the largest variance contribution are added to the training data in sequence through the adaptive point-adding algorithm. The GPR proxy model is trained until the variance of the GPR proxy model meets the stopping criterion, and the optimal GPR proxy model is obtained.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of a method for constructing an agent model of a heat conduction process in metal additive manufacturing as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a method for constructing an agent model of a heat conduction process in metal additive manufacturing according to any one of claims 1 to 7 are implemented.

Images