Method of predicting optimized process condition in Laser Powder Bed Fusion

Abstract

A method for optimizing and predicting process conditions used in laser powder bed melting is disclosed. The properties of the alloy powder constituting the data and the process conditions are normalized, and the relative density, which is the target value, is converted into a sigmoid function value. The preprocessed data is used to train a machine learning model. Through this, a set of optimal process conditions for metal powders not used in machine learning can be derived.

Description

Technology Field

[0001] The present invention relates to a high-precision additive manufacturing method (AM) for metal parts, and more specifically, to a method for predicting an optimal process through machine learning in the field of laser powder bed melting. Background Technology

[0002] Laser Powder Bed Fusion (L-PBF) is a selective laser melting technology and is a high-precision additive manufacturing method for producing metal parts with excellent mechanical properties. L-PBF is a technology that uses a laser to stack layers of different types of metals. For example, metal powder is spread onto a bed, melted using a laser, and the molten layer is rapidly cooled to form a single layer. In the next process, another thin layer of metal powder is formed, and the melting and cooling using a laser are repeated. The repetition of the above process is performed until the desired metal part is formed.

[0003] Due to its unique process characteristics, the aforementioned L-PBF is applied to the fabrication of precision components used in the aerospace, automotive, and medical industries, requiring process optimization for new alloy powders. Furthermore, manufacturers of L-PBF equipment propose optimal process conditions for each specific metal powder. These process conditions are obtained through repeated trial and error; while there have been many studies aiming to achieve process optimization through machine learning, they have typically focused on optimizing for a single type of powder.

[0004] Typically, since metal powders in the form of alloys differ in their respective properties or physical characteristics, process conditions suitable for the properties of the metal powder must be applied; otherwise, defects will occur in the metal parts manufactured. Therefore, optimal process conditions need to be determined individually for each type of metal powder. If users require an improvement in the production speed or precision of metal parts, various optimal process conditions are required for the same type of metal powder. In other words, it is highly unreasonable to determine these conditions through repeated trial and error in an environment where diverse process conditions are required.

[0005] Attempts to optimize processes using machine learning are being made to address the aforementioned problems. U.S. Patent Publication No. 2021-0405613 discloses a method for predicting porous appearance occurring during the execution of an L-PBF process. This is accomplished through the training of a machine learning model, where porosity data is acquired via images, and when process conditions are input through the trained model, a porosity map is provided to the user.

[0006] However, the aforementioned patent has limitations in that it applies to only one type of metal powder and requires the user to input process conditions for a specific metal powder. In other words, it remains silent regarding the provision of optimal process conditions for various metal powders. The problem to be solved

[0007] The technical problem that the present invention aims to solve is to predict optimized process conditions for various metal powders regardless of whether they are included in the training of a machine learning model. means of solving the problem

[0008] The present invention, for achieving the aforementioned technical problem, provides a method for predicting an optimal process comprising: a step of training a machine learning model using the properties of a metal powder, process conditions, and relative density; and a step of predicting a set of process conditions having a high relative density when a specific metal powder is used by using the trained machine learning model and random search.

[0009] The above technical problem of the present invention is also achieved by providing a method for predicting an optimal process, comprising the steps of: training a machine learning model using the properties of metal powder, process conditions, and relative density; and predicting a set of process conditions having a high relative density of metal powder not used in training the machine learning model using the trained machine learning model and random search. Effects of the invention

[0010] According to the present invention described above, the properties and process conditions of a metal powder serve as input data, and the correlation between them and the target value, relative density, is confirmed through the training of a machine learning model. In the above process, the properties and process conditions of the metal powder are normalized, and the relative density is converted into a sigmoid function value and used for training the machine learning model. Additionally, the user can predict process conditions with high relative density by inputting the properties of a specific powder they wish to predict. Through this, an optimal set of process conditions is predicted even for powders with compositions not used in the training of the machine learning model. That is, the user inputs the properties of the powder they wish to use and fixed values ​​of process conditions, and can secure an optimal set of process conditions through an already trained machine learning model. Through this, optimal sets of process conditions with high relative density are predicted across various alloys, and the time and cost consumed for optimization by alloy type can be drastically reduced. Brief explanation of the drawing

[0011] FIG. 1 is a flowchart for deriving optimal process conditions for alloy powder according to a preferred embodiment of the present invention. FIG. 2 is a flowchart illustrating the steps of training the machine learning model of FIG. 1 according to a preferred embodiment of the present invention. Figure 3 is a graph showing the sigmoid function used for the preprocessing of relative density according to a preferred embodiment of the present invention. FIG. 4 is a flowchart illustrating a method for predicting process conditions having a high relative density when using a specific powder of FIG. 1 according to a preferred embodiment of the present invention. FIG. 5 is a table showing the sets of process conditions predicted according to Prediction Example 1 of the present invention and the relative densities measured through experiments. Figure 6 is a table showing the sets of process conditions predicted according to Prediction Example 2 of the present invention and the relative densities measured through experiments. FIG. 7 is a table showing the sets of process conditions predicted according to Prediction Example 3 of the present invention and the relative densities measured through experiments. Specific details for implementing the invention

[0012] The present invention is susceptible to various modifications and may take various forms; therefore, specific embodiments are illustrated in the drawings and described in detail in the text. However, this is not intended to limit the invention to the specific disclosed forms, and it should be understood that the invention includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention. Similar reference numerals have been used for similar components in the description of each drawing.

[0013] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which the present invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this application.

[0014] Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the attached drawings.

[0016] Examples

[0017] In this invention, L-PBF is used for the manufacture of metal parts, and optimal process conditions for various types of metal powders are predicted upon the application of L-PBF. The relative density resulting from these optimal process conditions serves as a criterion for evaluating the quality of the resulting metal part. Specifically, a low relative density indicates the presence of a large number of pores or voids within the metal part; this acts as a crack initiation point within the layered metal part and degrades the mechanical properties of the metal part. Therefore, it is desirable for the metal part to have a high relative density.

[0018] The present invention provides a machine learning-based process optimization method applicable to various alloy powders. The machine learning model of the present invention combines a specific metal powder, considering the properties of the metal powder, with process conditions, and predicts the relative density of the resulting metal part. To this end, the machine learning model receives a sufficient amount of data as input to learn the interaction between the properties of the metal powder and the process conditions.

[0019] In the present invention, the properties of the metal powder refer to the material characteristics of the alloy powder, such as reflectance, thermal conductivity, specific heat capacity, mass density, and melting point, and the process conditions are the laser power, scan speed, hatch distance, and layer thickness set in the L-PBF equipment.

[0020] In this invention, the correlation between the properties of metal powder, process conditions, and relative density can be established through a machine learning model trained. Additionally, when the properties of a metal powder selected by a user are input, process conditions and relative density can be derived through random search, and through an optimization process, process conditions for the user-input metal powder having the highest relative density can be derived.

[0021] In particular, multiple sets of process conditions may be provided for the metal powder input by the user. Additionally, a target value, namely relative density, is provided in combination with each set of process conditions. The set of process conditions consists of the aforementioned laser power, scan speed, hatch distance, and layer thickness. That is, the relative density based on the set of process conditions for the selected metal powder is provided.

[0023] FIG. 1 is a flowchart for deriving optimal process conditions for alloy powder according to a preferred embodiment of the present invention.

[0024] Referring to FIG. 1, the method for deriving optimal process conditions of the present invention includes the step of training a machine learning model (S100) and the step of predicting process conditions having a high relative density when using a specific metal powder (S200).

[0025] First, the machine learning model is trained, and the target value, relative density, is learned through the input of various properties of the metal powders and process conditions.

[0026] In the present invention, the properties of metal powders and process conditions serve as inputs during the training of the machine learning model, and the relative density of the metal part serves as the target value. Through the training of the machine learning model, the correlations between the properties of the metal powders, process conditions, and relative density are learned through regression analysis.

[0027] In addition, optimal process conditions can be derived using a trained machine learning model. To derive or predict optimal process conditions, the user inputs the properties of the metal powder and derives the process conditions through random search.

[0028] The derived process conditions are input into a trained machine learning model along with the properties of the metal powder entered by the user, and the relative density is derived. The input of the process conditions and the properties of the metal powder into the machine learning model and the derivation of the relative density through the above random search can be performed until there is no change in the relative density according to the optimal process conditions.

[0029] First, the step of training a machine learning model includes collecting various information about the powder and performing data preprocessing on the collected information. Subsequently, training of the machine learning model is performed using the preprocessed data.

[0030] FIG. 2 is a flowchart illustrating the steps of training the machine learning model of FIG. 1 according to a preferred embodiment of the present invention.

[0031] Referring to FIG. 2, information necessary for training a machine learning model is collected (S110).

[0032] The above information includes the properties of the metal powder, process conditions, and relative density. That is, at least one process condition associated with the properties of one type of metal powder and the relative density according to said process condition constitute the information, and the collected information includes process conditions and relative densities for various types of metal powders. However, the information required for training does not include the composition of the metal powder, but includes the properties of the metal powder, the process conditions associated with said properties, and the relative density of the product associated therewith.

[0033] The properties of the metal powder are the reflectance, thermal conductivity, specific heat capacity, melting point, and density of the metal powder, and the process conditions are the laser power, the laser scan speed, the hatch distance, and the thickness of the layer. However, the designer for training the machine learning model may vary the properties of the metal powder and may add to or change the process conditions described above.

[0034] The properties of metal parts fabricated using L-PBF are significantly affected by the presented process conditions. An inappropriate combination of the above process conditions may result in the formation of pores or a lack of fusion, leading to the problem of voids forming within the metal parts. In particular, voids or pores act as crack initiation points within metal parts having a layered structure. That is, high porosity within the metal parts degrades the mechanical properties of the metal parts.

[0035] In particular, the relative density of metal parts can vary depending on process conditions, and even when using the same material, the relative density changes depending on the influence of process conditions. For example, laser energy density is defined by the following mathematical formula 1.

[0036]

[0037] In the above mathematical formula 1, Ed is the laser energy density (J / mm²). 3), P is laser power (W), v is scan speed (mm / s), h is hatch spacing (mm), and t is layer thickness (mm).

[0038] In the above mathematical formula 1, if the laser energy density is insufficient, the fusion of the metal material is insufficient and pores are generated, and if the laser energy density is excessive, pores are generated due to vaporization. Therefore, depending on the metal powder used, the process conditions for achieving high relative density need to be adjusted.

[0039] In addition, even if the same process conditions are applied, the relative density may vary depending on the properties of the metal powder used as the material.

[0040] The properties of the metal powder also significantly affect relative density. Specifically, as thermal conductivity decreases, relative density increases; this implies that lower thermal conductivity concentrates energy in the laser-irradiated area, allowing for uniform melting. Conversely, if thermal conductivity increases, the laser energy density required for uniform melting increases.

[0041] As mentioned above, process conditions and the properties of the metal powder affect the relative density.

[0042] In this invention, process conditions, properties of metal powder, and relative density are collected by referring to previous research results and numerous papers.

[0043] In particular, relative density can be measured through image analysis without using the traditional Archimedes method. The image analysis is derived using the area of ​​pores appearing through the cross-section of a metal part. For example, the relative density follows Equation 2 below.

[0044]

[0045] Next, the properties and process conditions of the collected metal powder are normalized using the mean and standard deviation, and the target value, relative density, is converted into a sigmoid function value using a sigmoid function (S120).

[0046] In this embodiment, data preprocessing refers to normalization using mean values ​​and standard deviations for process conditions and the properties of metal powder, and is an operation of converting the target value, relative density, into a sigmoid function value. That is, the relative density of a metal part manufactured by applying process conditions to a metal powder having the corresponding properties is investigated, and these are preprocessed into a normal distribution and a sigmoid function value.

[0047] The sigmoid function for relative density follows Equation 3 below.

[0048]

[0049] In the above mathematical equation 3, y is the sigmoid function value preprocessed with the sigmoid function for relative density, x represents relative density, and x st represents the reference value of relative density. The reference value of relative density is named the reference relative density.

[0050] Figure 3 is a graph showing the sigmoid function used for the preprocessing of relative density according to a preferred embodiment of the present invention.

[0051] Referring to Fig. 3, when the reference relative density is set to 98%, the sigmoid function value is 0.5 when the relative density is 98%, the sigmoid function value approaches 0 when the relative density is less than 98%, and the sigmoid function value approaches 1 when the relative density exceeds 98%. In other words, the closer the sigmoid function value is to 1, the higher the relative density of the metal part exceeds 98%.

[0052] Again, referring to FIG. 2, a machine learning model is trained using the preprocessed data (S130).

[0053] The machine learning model is input with the properties of normalized metal powder and detailed process conditions, and through machine learning, the distribution of the relative density of the manufactured metal parts around a reference relative density is verified. This is a type of regression analysis in which the correlation between the input values—metal powder properties and process conditions—and the target value—relative density—is implemented using a normal distribution and a sigmoid function.

[0054] The machine learning model of the present invention is a regression model that converts a target value into a sigmoid function value and then predicts the converted sigmoid function value. This configuration differs from existing regression models that directly predict the target value and classification models that predict the target value converted into binary.

[0055] In terms of the data structure of the present invention, regression models tend to underestimate predicted values ​​when the target value exceeds the reference value, and classification models have the disadvantage of overestimating predicted values ​​when the target value exceeds the reference value. However, the machine learning model of the present invention mitigates the problems of the two existing models by normalizing the target value using a sigmoid function.

[0056] The type or composition of the metal powder is not input during the training process of the machine learning model. However, the properties, process conditions, and relative density of the metal powder corresponding to physical properties are trained.

[0057] In addition, a gradient boosting model is used as one example of a machine learning model in this invention.

[0058] Referring again to FIG. 1, after training a machine learning model, process conditions having a high relative density when using a specific powder are predicted using the machine learning model and random search (S200).

[0059] FIG. 4 is a flowchart illustrating a method for predicting process conditions having a high relative density when using a specific powder of FIG. 1 according to a preferred embodiment of the present invention.

[0060] Referring to FIG. 4, the user inputs the properties of the metal powder, whether the process conditions are fixed and the fixed value, and other options (S210).

[0061] The properties of the metal powder are thermal conductivity, specific heat capacity, reflectance, mass density, and melting point, and the reflectance can be derived based on the input composition. For example, the compositional formula of the metal powder entered by the user is A a B b C c If so, the reflectance of the metal powder follows the following mathematical formula 4.

[0062]

[0063] In the above mathematical formula 4, R is the reflectance of the input metal powder, and a, b, and c represent the fractions of elements A, B, and C, respectively. In addition, R A is the reflectance (%) of element A, R B is the reflectance (%) of element B, and R C represents the reflectance (%) of element C.

[0064] The reason for deriving reflectance using Equation 4 is that the reflectance of physical materials in an alloy state can be measured differently depending on the surface condition. Therefore, calculating the fraction of constituent elements that make up the alloy and the reflectance of those elements arithmetically can improve the reliability of the data. In addition, since the reflectance of various metal powders is partially omitted and disclosed during the information collection process, Equation 4 above can be used to supplement this.

[0065] The above process conditions can be fixed within the range where training data exists through machine learning. For example, the laser power can be in the range of 50W to 400W, the scan speed in the range of 100 mm / s to 4500 mm / s, the layer thickness in the range of 0.025 mm to 0.12 mm, and the hatch gap in the range of 0.08 mm to 0.2 mm. Additionally, the user can fix the layer thickness of the metal part to be manufactured. Under the condition where the layer thickness is fixed, multiple process conditions capable of achieving the same layer thickness can be derived by exploring other process conditions.

[0066] If the user does not fix a specific process condition in the process condition set, process condition sets within the trained range are randomly generated in the subsequent process.

[0067] Additionally, other options refer to the number of process condition sets M and the maximum number of iterations N calculated from a single prediction.

[0068] Next, a set of process conditions is randomly generated based on the properties of the input metal powder and other options (S220).

[0069] In the present invention, the term "process condition set" refers to a combination of each process condition, namely laser power, scan speed, hatch spacing, and layer thickness, and indicates that the combination thereof forms a single unit.

[0070] 2M process condition sets formed through random search are generated, which is twice the number desired by the user. In the above step, if there is a fixed value among the process conditions, other process condition values ​​are randomly generated centered on the fixed process condition value, and the number of process condition sets generated becomes 2M, which is twice the number of process condition sets entered by the user.

[0071] However, the number of randomly generated process condition sets may be selectively generated within a limit that exceeds the number of process condition sets desired by the user. That is, if the number of process condition sets generated through random search is greater than M, it does not deviate from the spirit of the present invention.

[0072] In addition, the range of explored process conditions is limited to within the range where training data from machine learning exists.

[0073] Next, the properties of the input metal powder and the generated set of process conditions are input into a trained machine learning model, and a sigmoid function value is calculated (S230).

[0074] The above sigmoid function value is a value converted from the relative density of the metal part, and the closer it is to 1, the higher the relative density.

[0075] The properties of the metal powder entered by the user and the set of randomly generated process conditions to be input into the machine learning model are normalized using the mean and standard deviation used in the normalization process of the machine learning step, and then input into the model.

[0076] Next, sets of similar process conditions are removed, and M sets of process conditions are selected from the input process condition sets in which the sigmoid function value is close to 1 (S240).

[0077] That is, if the sum of the differences between sets of normalized process conditions is less than a specific real difference threshold value m, the set of process conditions with a small relative density converted into a sigmoid function value, which is the model output value, is removed. The determination of the difference value between sets of normalized process conditions follows Equation 5 below.

[0078]

[0079] In the above mathematical formula 5, a i represents the normalized process conditions of process set A, and b irepresents the normalized process conditions of process set B. The difference reference value m can have a relatively large value if the user wants to search a wide range of process conditions, and a relatively small value if the user wants to search a narrow range of process conditions.

[0080] For example, if the difference reference value m is 1 and arbitrary process conditions A and B satisfy the above mathematical formula 5, and the sigmoid function value of process condition A is greater than the sigmoid function value of process condition B, then process condition B is deleted.

[0081] The above process involves selecting a process condition with a high relative density among sets of similar process conditions and deleting sets of similar process conditions with a low relative density during the prediction process.

[0082] After a randomly generated set of similar process conditions is removed, M process condition sets are selected starting from those having the highest sigmoid function value among the remaining process condition sets, and the remaining process condition sets are removed.

[0083] Since the machine learning model receives the properties of the metal powder and 2M randomly generated process conditions as input, it can select a process condition with a sigmoid function value close to 1 through the above operation. Through this, a set of M process conditions entered by the user is selected.

[0084] Next, it is determined whether the arbitrary formation of process conditions and the selection of M process conditions have reached the maximum number of repetitions N (S250).

[0085] If the formation of a random process set has not reached the maximum number of repetitions N, a new process condition set with the same number as the number of process condition sets removed in the removal step of the process condition set described above is randomly generated (S260).

[0086] In the process condition set removal step, M inputs are selected, and since the remaining process condition sets are removed, new ones are created equal to the number of removed process condition sets and can be compared with each other.

[0087] After randomly generating a new set of process conditions, the process of steps S230 to 250 is repeated.

[0088] Additionally, when it is determined that the formation of an arbitrary set of process conditions and the selection of M sets of process conditions have reached a maximum number of repetitions N, the procedure is terminated and the selected set of process conditions is finally selected (S270).

[0089] Additionally, depending on the embodiment, a step (S255) of determining whether the current process condition set selected after step S250 is identical to the past process condition set previously selected may be added. If it is determined that it is identical to the process condition set selected in the past step, the procedure is terminated and this is selected as the final process condition set (S270).

[0090] Additionally, depending on the form of implementation, step S255 may proceed prior to step S250.

[0091] In other words, if the process condition set selected in the current step is identical to the process condition set selected in the previous step, the procedure is terminated, and the selected process condition set is confirmed as the final process condition set. If the process condition sets are not identical in the comparison, it is determined whether the number of random generation and selection operations for the process condition sets has reached the maximum number of iterations N.

[0092] When the maximum number of iterations N is reached, the procedure is terminated, and the set of process conditions selected at the current step is finally confirmed. If the maximum number of iterations N is not reached, the step of randomly generating a new set of process conditions is performed.

[0093] Additionally, a person skilled in the art may perform steps S250 and S255 simultaneously depending on the embodiment. That is, if the process condition set selected in the previous step and the process condition set selected in the current step are identical, and it is determined that the number of iterations has reached the maximum number of iterations N, the process condition set selected in the current step is confirmed as the final process condition set. If the maximum number of iterations N has not been reached, the operation of randomly generating a new process condition set is performed, and the operation of selecting the process condition set is repeated.

[0095] Training example of a machine learning model

[0096] 2,167 process conditions for 49 types of powders were collected from prior research literature. In addition, properties and relative densities of metal powders, in addition to process conditions, were also collected from prior research literature.

[0097] The collected data, including process conditions and metal powder properties, were normalized through preprocessing, and the relative density data were converted into sigmoid function values ​​before training a gradient boosting model. The above model is one of the machine learning algorithms.

[0098] SHAP (SHapley Additive exPlanation) analysis was conducted to verify the relationship between the input process conditions and metal powder properties and the target value, relative density. Through this, the interaction between the input and the target value was represented as a SHAP score, and the contribution of each input data point to the target value was calculated.

[0099] Through training, a correlation was established between the properties of metal powder, process conditions, and relative density, and the evaluation results showed that the accuracy of the machine learning model was 87%.

[0101] Prediction Example 1 of Optimal Process Conditions

[0102] In the training example of the above machine learning model, the optimal process conditions for the alloy powder not used in training are predicted.

[0103] Predicts optimal process conditions for Stainless Steel 316L (STS 316L) powder. Composition of STS 316L Fe 0.6875 Ni 0.1146 Cr 0.1833 Mo 0.0146 Through this, the reflectance was calculated to be 65.33%, and with a reflectance of 65.33%, thermal conductivity of 21.4 W / mK, specific heat capacity of 0.5 J / gK, and mass density of 8 g / cm³ 3 , enter the melting point 1375 ℃.

[0104] Since the average diameter of the powder is 46 µm, the layer thickness in the process conditions cannot be less than 46 µm, so the layer thickness is fixed at 50 µm, and other options are entered as the number of process condition sets is 12 and the maximum number of iterations is 100,000.

[0105] Random generation and selection operations of process condition sets are performed, and 12 types of optimal process condition sets are derived through the machine learning model disclosed in the training example above. To verify the accuracy of the derived optimal process condition sets, metal parts are manufactured using the derived process condition sets, and relative densities are measured and compared.

[0106] FIG. 5 is a table showing the sets of process conditions predicted according to Prediction Example 1 of the present invention and the relative densities measured through experiments.

[0107] Referring to FIG. 5, 12 sets of optimal process conditions are shown. The layer thickness is fixed at 50 μm through input, and the list of process condition sets is arranged in order of sigmoid function values ​​being closest to 1.

[0108] In addition, the column labeled relative density represents the value obtained by feeding the corresponding metal powder into the L-PBF equipment according to the derived sets of process conditions, forming an actual alloy film, and then measuring the relative density using a microscope. The relative density shows a high value exceeding 98%, and it is confirmed that when the sigmoid function value is high, the actual measured relative density is also high. Through this, the accuracy of this machine learning model is confirmed.

[0110] Prediction Example 2 of Optimal Process Conditions

[0111] In the training example of the above machine learning model, the optimal process conditions for the alloy powder not used in training are predicted.

[0112] Predict the optimal process conditions for AlSi10Mg powder. Composition of the corresponding metal powder Al 0.9045 Si 0.0916 Mg 0.0039 The reflectance was calculated to be 67.07%, and the values ​​are reflectance 67.07%, thermal conductivity 146 W / mK, specific heat capacity 0.88 J / gK, and mass density 2.67 g / cm³. 3 , enter a melting point of 570 ℃.

[0113] Since the average diameter of the powder is set to 46 µm, the layer thickness in the process condition cannot be less than 46 µm, so the layer thickness in the process condition is fixed at 50 µm, and other options are entered as the number of process condition sets to 12 and the maximum number of iterations to 100,000.

[0114] Random generation and selection operations of process condition sets are performed, and 12 types of optimal process condition sets are derived through the machine learning model disclosed in the above training example. To verify the accuracy of the derived optimal process condition sets, metal parts are fabricated using the derived process condition sets, and relative densities are measured and compared.

[0115] Figure 6 is a table showing the sets of process conditions predicted according to Prediction Example 2 of the present invention and the relative densities measured through experiments.

[0116] Referring to FIG. 6, 12 sets of optimal process conditions are disclosed, and the layer thickness is fixed at 50 μm. The various sets of process conditions are arranged in order of having high sigmoid function values. Additionally, the relative density is a value measured for samples formed to a thickness of 50 μm by applying the derived 12 sets of process conditions to the metal powder.

[0117] The measured relative densities show high values ​​of over 98%, and it is confirmed that the higher the sigmoid function value, the higher the relative density in the actual samples.

[0119] Prediction Example 3 of Optimal Process Conditions

[0120] In the training example of the above machine learning model, the optimal process conditions for the alloy powder not used in training are predicted.

[0121] Predicts optimal process conditions for Fe60Co15Ni15Cr10 MEA powder. Composition of Fe60Co15Ni15Cr10 MEA Fe 0.6 Co 0.15 Ni 0.15 Cr 0.1 Through this, the reflectance was calculated to be 66.15%, and with a reflectance of 66.15%, thermal conductivity of 13.66 W / mK, specific heat capacity of 0.46 J / gK, and mass density of 7.91 g / cm³ 3 , enter the melting point 1461.75 ℃.

[0122] Since the average diameter of the powder is 46 µm, the layer thickness in the process conditions cannot be less than 46 µm, so the layer thickness in the process conditions is fixed at 50 µm, and as other options, the number of process condition sets is 12 and the maximum number of iterations is 100,000.

[0123] Random generation and selection operations of process condition sets are performed, and 12 types of optimal process condition sets are derived through the machine learning model disclosed in the above training example. To verify the accuracy of the derived optimal process condition sets, metal parts are fabricated using the derived process condition sets, and relative densities are measured and compared.

[0124] FIG. 7 is a table showing the sets of process conditions predicted according to Prediction Example 3 of the present invention and the relative densities measured through experiments.

[0125] Referring to Fig. 7, 12 sets of process conditions are derived. As with other prediction examples, the sets of process conditions are sorted in order of highest sigmoid function values.

[0126] In addition, the relative density is the value obtained by forming a 50 μm layer using the derived set of process conditions and measuring the relative density. It is confirmed that the higher the sigmoid function value of the set of process conditions, the higher the measured relative density. Through this, it can be seen that the machine learning model trained according to the present invention has high accuracy.

[0128] According to the present invention described above, the properties and process conditions of a metal powder serve as input data, and the correlation between them and the target value, relative density, is confirmed through the training of a machine learning model. In the above process, the properties and process conditions of the metal powder are normalized, and the relative density is converted into a sigmoid function value and used for training the machine learning model. Additionally, the user can predict process conditions with high relative density by inputting the properties of a specific powder they wish to predict. Through this, an optimal set of process conditions is predicted even for powders with compositions not used in the training of the machine learning model. That is, the user inputs the properties of the powder they wish to use and fixed values ​​of process conditions, and can secure an optimal set of process conditions through an already trained machine learning model. Through this, optimal sets of process conditions with high relative density are predicted across various alloys, and the time and cost consumed for optimization by alloy type can be drastically reduced. Explanation of the symbols delete

Claims

Claim 1 A method for predicting an optimal process, comprising: a step of training a machine learning model using the properties of a metal powder, process conditions, and relative density; and a step of predicting a set of process conditions when a specific metal powder is used using the trained machine learning model and random search, wherein the step of predicting the set of process conditions includes: a step of inputting the properties of the metal powder desired by the user, the number of sets of process conditions, whether the process conditions are fixed, and a fixed process condition value; a step of randomly generating sets of process conditions based on the properties of the metal powder desired by the user; a step of inputting the properties of the metal powder desired by the user and the randomly generated sets of process conditions into the trained machine learning model to derive a sigmoid function value in which the relative density is transformed for each randomly generated set of process conditions; and a step of selecting the sets of process conditions according to the order in which the sigmoid function value approaches 1, wherein the selected sets of process conditions are selected as many times as the number of sets of process conditions. Claim 2 A method for predicting an optimal process according to claim 1, wherein the step of training the machine learning model comprises: collecting information consisting of the properties of the metal powder, the process conditions, and the relative density required for training the machine learning model; performing data preprocessing on the collected information; and training the machine learning model using the preprocessed data. Claim 3 A method for predicting an optimal process according to paragraph 2, wherein the data preprocessing is characterized in that the properties of the metal powder and the process conditions form a normal distribution using the mean and standard deviation, and the relative density is converted into a sigmoid function value. Claim 4 A method for predicting an optimal process according to claim 3, characterized in that the sigmoid function value is determined by the following formula.[Formula] In the above formula, y is the sigmoid function value preprocessed with the sigmoid function for the relative density, x represents the relative density, and x st represents the reference value of the above relative density. Claim 5 A method for predicting an optimal process according to paragraph 2, wherein the relative density is measured through image analysis, and the image analysis is derived using the area of ​​pores appearing through the cross-section of a metal part. Claim 6 A method for predicting an optimal process according to claim 2, wherein the properties of the metal powder are reflectance, thermal conductivity, specific heat capacity, melting point, and mass density, and the process conditions are laser power, scan speed, hatch spacing, and layer thickness. Claim 7 delete Claim 8 A method for predicting an optimal process according to claim 1, wherein the step of inputting the desired metal powder properties, the number of process condition sets, whether the process conditions are fixed, and the fixed process condition values ​​further includes inputting the maximum number of iterations as an additional option. Claim 9 A method for predicting an optimal process according to claim 8, characterized in that, in the step of randomly generating the process condition set, the number of randomly generated process condition sets exceeds the number of process condition sets. Claim 10 A method for predicting an optimal process according to claim 9, wherein the step of selecting process condition sets comprises: a step of removing process condition sets with a low sigmoid function value by comparing similar process condition sets among the randomly generated process condition sets; and a step of selecting process condition sets in order of highest sigmoid function value among the remaining process condition sets after removing similar process condition sets. Claim 11 A method for predicting an optimal process according to claim 8, characterized by including: a step of determining whether the number of times the process condition sets are randomly generated has reached the maximum number of iterations after the step of selecting the process condition sets; and a step of randomly generating a new process condition set if the maximum number of iterations has not been reached. Claim 12 A method for predicting an optimal process according to claim 11, further comprising the step of determining the selected set of process conditions as the final set of process conditions when the maximum number of iterations is reached. Claim 13 A method for predicting an optimal process according to claim 11, further comprising, after the step of determining whether the maximum number of iterations has been reached, a step of determining whether the process condition set selected in the previous step and the process condition set selected in the current step are identical, and if the two process condition sets are identical, determining the process condition set selected in the current step as the final process condition set. Claim 14 A method for predicting an optimal process, comprising: a step of training a machine learning model using the properties of metal powder, process conditions, and relative density; and a step of predicting a set of process conditions of metal powder not used in training the machine learning model using the trained machine learning model and random search, wherein the step of predicting the set of process conditions comprises: a step of inputting the properties of metal powder not used in training, whether the process conditions are fixed, the number of process condition sets, and a maximum number of iterations; a step of randomly generating the set of process conditions based on the properties of metal powder not used in training, whether the process conditions are fixed, and the number of process condition sets; a step of inputting the properties of metal powder not used in training and the randomly generated set of process conditions into the machine learning model and deriving a sigmoid function value; a step of selecting the set of process conditions equal to the number of process condition sets according to the order in which the sigmoid function value approaches 1; a step of determining whether the selection of the set of process conditions has reached the maximum number of iterations; and a step of confirming the selected set of process conditions as the final set of process conditions when the maximum number of iterations has been reached. Claim 15 delete Claim 16 A method for predicting an optimal process according to claim 14, characterized in that if the selection of the above process condition set does not reach the above maximum number of iterations, a new process condition set is randomly generated, and the new process condition set is input into the machine learning model to generate the sigmoid function anew. Claim 17 A method for predicting an optimal process according to claim 16, characterized in that the number of newly generated process condition sets is equal to the number of process condition sets not selected in the selection step of the process condition sets. Claim 18 A method for predicting an optimal process according to claim 14, characterized in that the number of the arbitrarily generated process condition sets is greater than the number of process condition sets entered by the user. Claim 19 A method for predicting an optimal process according to claim 14, wherein the step of selecting the process condition set comprises: a step of selecting the process condition set according to the order in which the sigmoid function value approaches 1 by comparing similar process condition sets among the randomly generated process condition sets; and a step of selecting the process condition set according to the order in which the sigmoid function value approaches 1 among the remaining process condition sets after the comparison of the similar process condition sets. Claim 20 A method for predicting an optimal process according to claim 14, wherein the step of training the machine learning model comprises: collecting information consisting of the properties of the metal powder, the process conditions, and the relative density required for training the machine learning model; performing data preprocessing in which the properties of the metal powder and the process conditions are normalized and the relative density is converted into a sigmoid function value; and training the machine learning model using the preprocessed data.

Images