Method for optimising the additive manufacture of metal bodies by means of a bayesian optimisation method
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- INST FUER FESTKOERPER & WERKSTOFFORSCHUNG DRESDEN EV
- Filing Date
- 2026-01-27
- Publication Date
- 2026-07-30
Smart Images
Figure EP2026052002_30072026_PF_FP_ABST
Abstract
Description
[0001] Methods for optimizing the additive manufacturing of metallic bodies
[0002] The invention relates to the fields of materials science and mechanical engineering and concerns a method for optimizing the additive manufacturing of metallic bodies. The proposed method can be used, for example, in the laser-based additive manufacturing of components.
[0003] Process control in additive manufacturing (AM) presents a particular challenge due to the complex interplay between process parameters, material properties, and the environmental conditions of the AM equipment. Identifying the input parameters that affect the quality and microstructure of the additively manufactured metallic body during the manufacturing process is challenging and costly, especially with expensive starting materials such as metal powders. The complexity of additive manufacturing of metallic bodies and components precludes holistic analytical modeling and instead necessitates statistical methods and machine learning (ML) techniques for predicting and optimizing the results.
[0004] Current optimization methods for the additive manufacturing of metallic bodies rely heavily on the data and properties of the manufactured body, whereas ML mimics the interrelationships between process parameters to be considered during manufacturing and the desired material properties after additive manufacturing.
[0005] Initial approaches to optimizing process parameters for additive manufacturing are based on data generated from the subsequent characterization of the manufactured samples, which were developed in the scientific literature using ML tools, and are quite simple in procedure.
[0006] Data is collected for one or more materials, consisting of AM process parameter combinations used to manufacture these materials and their corresponding characteristic target properties. A machine learning algorithm is then applied to establish a connection between the process parameters and the target properties, ultimately enabling prediction of the process parameters. This leads to the desired output values, such as maximizing density, phase fraction in the metallic material, or mechanical properties like ductility or yield strength of the metallic body.
[0007] The use of artificial neural networks for optimizing and predicting material properties is well-established. Neural networks are employed to model and optimize the surface roughness of samples produced using laser powder bed fusion (PBF-LB / M). In conjunction with finite element simulations, this can improve process predictions in additive manufacturing. Other algorithms, such as Gaussian elimination, have proven effective in predicting mechanical properties, optimizing the density of PBF-LB / M components, and modeling amorphous volume fractions. While PBF-LB / M remains the dominant metal AM technology, machine learning has also found its way into other AM processes.
[0008] Various machine learning approaches for optimizing AM processes are known from the state of the art.
[0009] US Patent 2019 005 4700 A1 discloses an additive manufacturing process in which at least one initial 3D printing operation is performed to at least partially produce a component. Subsequently, the produced component is measured, the measured dimension is compared with at least one corresponding nominal dimension, and subsequently, at least one regression model is created in response to the at least one comparison to compensate for the observed deviation and improve the manufacturing accuracy.
[0010] US Patent 2019337232 A1 discloses a method for training a machine learning machine to model a physical system. This method includes receiving process data representing measurements of the physical system. The method further includes applying a transformation to the values of at least two variables in the process data to generate a dimensionless parameter with a parameter value corresponding to each measurement of the physical system for the at least two variables. Additionally, the method includes training the machine learning machine using a set of generated training data containing a dimensionless parameter to output a prediction of the value of a physical effect of the physical system for values of variables not included in the process data.The method also includes controlling an additive manufacturing process for the material by adjusting at least one physical property to the value of at least one process variable during the production of an additively manufactured component.
[0011] WO 2020216458 A1 discloses a method and a system for predicting the fatigue life of additively manufactured components, taking local material properties into account. The method and system are used to predict the fatigue life characteristics of an additively manufactured element and comprise a data acquisition step in which multiple data points for the maximum stress versus cycles to failure are recorded for various predefined processing steps of the element; a training step in which a machine learning system is trained with the acquired data; and an evaluation step in which the trained machine learning system is exposed to actual processing steps and used to predict the fatigue characteristics of the element.
[0012] EP 3 651 053 A1 discloses a method for optimizing additive process parameters for an additive manufacturing process, in which the approach is to combine a machine learning model trained to predict properties of a printed sample based on process parameters with an additional element that serves for intelligent planning of experiments in order to improve model fidelity and reduce inherent uncertainty.
[0013] DE 102020209 573 A1 discloses a computer-implemented method for operating a laser material processing machine, in which process parameters are varied using Bayesian optimization until a sufficiently good result of the laser material processing is achieved, wherein the Bayesian optimization is carried out using a data-based process model. In a first phase, the data-based process model is trained based on estimated results, while in a second phase, the data-based process model is trained based on the determined result obtained when controlling the laser material processing machine.
[0014] A disadvantage of existing solutions is that, due to the complexity of the process parameters required in additive manufacturing, the desired material properties, and the existing environmental conditions, large amounts of data are needed to provide the appropriate process parameters for production. These data also require complex data processing and are costly to prepare for use in additive manufacturing. Consequently, it is disadvantageous that analytical modeling is not applicable to additive manufacturing, as individual datasets and calibrations must be performed for each production system and each material. This means that the desired relationships between process parameters and material properties of the metallic body produced using AM cannot be provided or transferred from one system to another in a repeatable manner.
[0015] The object of the present invention is to provide a method for optimizing the additive manufacturing of metallic bodies, which is cost-effective and time-saving and enables, with minimal effort, a reliable identification of the process parameters to be applied in order to achieve the desired material properties of the additively manufactured metallic bodies.
[0016] The problem is solved by the technical features according to claim 1. Advantageous embodiments are the subject of the dependent claims, the invention also including combinations of the individual dependent claims in the sense of an AND conjunction, as long as they are not mutually exclusive. The problem is solved by a method for optimizing the additive manufacturing of metallic bodies, in which, under quantified specification of at least one defined material-specific target parameter and at least two different process parameters, at least one metallic test specimen is produced by means of an additive manufacturing process, subsequently at least the mechanical properties of the metallic test specimen are determined, and at least one combined output data set is calculated therefrom, which comprises at least one material-specific target parameter and the associated process parameter of the at least one metallic test specimen.The input data is subsequently transformed into a data acquisition function for optimizing the input dataset. Using machine learning and a Bayesian optimization method, an initial dataset combination of optimized process parameters is calculated for at least one corresponding material-specific target parameter. The Bayesian optimization method is performed iteratively until a predefined value of the at least one material-specific target parameter is reached. If the material-specific target parameter is not reached with the proposed process parameters, at least one data vector is added to the dataset, and a new process parameter combination is proposed using Bayesian optimization. If the proposed process parameter set results in an additively manufactured component with the desired material-specific properties,This optimization method will then be terminated, and subsequently the determined process parameters will be transferred to an additive manufacturing system, and the metallic body will be additively manufactured according to at least one optimized material-specific target parameter.
[0017] In an advantageous embodiment of the method, physical, mechanical, and / or chemical properties are specified as material-specific target parameters, with density, mechanical strength, porosity, surface roughness, and / or ductility being particularly advantageous as material-specific target parameters. It is advantageously possible to define a limited parameter range using Gaussian process regressors and / or neural networks.
[0018] An advantage of this method is that the optimized process parameters x are determined by the acquisition function, where x* = argmax x A(x) with A(x) being the acquisition function and x being a vector that includes the process parameters to be optimized.
[0019] It is particularly advantageous if an information-theoretic entropy search or a knowledge gradient acquisition function according to the equation KG(x) = F[|j*n+1 - p*n% n+i = x] is used as the acquisition function, where p* is the expected best estimate after the n+1-th evaluation of the function of the process parameters x.
[0020] In an advantageous embodiment of the procedure, it may be provided that the optimization is carried out using a batch Bayesian optimization method, in which several sets of process parameters are output and a parallel evaluation and optimization of the material-specific target properties is realized.
[0021] Advantageously, at least two elementary geometric elements are produced based on the overall geometry of the metallic body, subsequently the mechanical properties of the elementary geometric elements are determined, and at least one combined output data set is calculated from the determined values, consisting of at least one material-specific target parameter and the associated process parameters of each element, and subsequently the process parameters for each elementary geometric element are optimized and combined to obtain optimized parameters for the overall geometry, using the acquisition function for multi-objective optimization across multiple properties and geometric configurations.It is particularly advantageous if the geometric parameters of the elementary individual geometries are provided by parallelepipeds, tessellation, mesh decomposition and / or triangulation of the metallic specimen.
[0022] It is also advantageous to incorporate prior knowledge from similar materials or alloys into the optimization process using a Bayesian multi-probability optimization framework, integrating low-reliability data from similar materials into high-reliability experimental data.
[0023] The invention provides a novel method for optimizing the additive manufacturing of metallic bodies, which is cost-effective and time-saving and enables, with minimal effort, a reliable identification of applicable process parameters to achieve the desired material properties of the additively manufactured metallic bodies.
[0024] The term additive manufacturing (AM) refers in the broadest sense to all metal AM technologies, including, but not limited to, the PBF-LB / M process with directed energy deposition and material extrusion. The invention proposes the use of so-called Bayesian optimization of AM process parameters. Within the framework of Bayesian optimization, the process parameters are systematically optimized through a series of iterations required for the additive manufacturing of a specific metallic body with the associated material-specific target parameters.
[0025] The technical advantages and effects are achieved through a process for optimizing the additive manufacturing of metallic bodies, which can be used flexibly, especially with different manufacturing systems and changing conditions, such as geometry or the powder material used, in a time- and cost-effective manner.
[0026] In a first step, at least one metallic test specimen is produced using an additive manufacturing process under quantified specification of at least one defined material-specific target parameter and at least two different process parameters, and subsequently at least one material-specific target parameter of the metallic test specimen is determined with reference to the applied process parameters.
[0027] For this purpose, the mechanical, structural, and / or physical properties of at least one metallic test specimen, additively manufactured under various combinations of process parameters, are measured or determined, and at least one combined output data set is created from the measurement results. This combined output data set contains at least two process parameters and at least one associated material-specific target parameter, thus reflecting the performance results under different process and material conditions.
[0028] The additive manufacturing of a metallic test specimen as a first step before the production of the desired metallic body has the technical advantage that the requirements for the respective geometry, the material used and the various differing system parameters of the manufacturing plant can be taken into account individually and with regard to the manufacturing plant used, flexibly, time-savingly and with little effort, and the optimal process parameters can be determined experimentally.
[0029] It is advantageous to define a predefined range of possible process parameters, thereby limiting the parameter space of the AM process to a manageable and relevant dataset—specifically, to the range within which the search for optimal process parameters will take place. Within this range, a set of process parameters is generated to create an initial combined output dataset. Various strategies, including Latin hypercube sampling, can be advantageously employed to ensure a diverse and representative population of output data points with a small data volume.
[0030] The combined output dataset is then transformed into a capture function for optimizing this output dataset. Using machine learning and a Bayesian optimization method, an initial optimized dataset combination of possible process parameters is calculated, corresponding to at least one material-specific target parameter. This approach allows the process parameter optimization problem to be transferred to the capture function optimization problem. The process parameters that maximize the capture function are then selected for evaluation during the next iteration until the optimized process parameter data matches the material-specific target parameter at which the capture function reaches its maximum value.
[0031] It may be advantageous to provide that the optimized process parameters x are determined by the acquisition function x* = argmax x A(x), where A(x) represents the acquisition function and x denotes a vector that includes the process parameters to be optimized.
[0032] It is also conceivable that an information-theoretic entropy search or a knowledge gradient acquisition function according to the equation KG(x) = F[|j*n+1 - p*n% n+i = x] is used as the acquisition function, where p* is the expected best estimate after the n+1-th evaluation of the function of the process parameters x.
[0033] To improve the efficiency of the optimization process, it can be advantageously provided that, in addition to the initial data set, physical data relating to material-related properties that are known per se are also included when applying the Bayesian optimization method. The technical advantage lies in the fact that data from, for example, similar metallic materials with similar chemical compositions that can be used for additive manufacturing exhibit essentially the same physical properties, such as similar thermal conductivity or similar laser beam absorption properties. Consequently, these similar metallic materials are likely to exhibit similar relationships between the defined material-specific target parameters and the corresponding process parameters.
[0034] This approach offers the technical advantage of considering both the process parameters to be optimized and the reliability level of the data. Reliability refers to the accuracy or quality of the data used. The highest level of accuracy corresponds to the most precise data, typically representing the material for which the process parameters are directly optimized. In contrast, lower levels of accuracy represent data from similar materials, which, while less precise, still provide valuable insights and thus continuously improve the optimization process.
[0035] In an advantageous embodiment of the process, a batch Bayes optimization method may be employed. Using this method generates not just one set of optimal process parameters, but several potential sets of optimized process parameters that lead to the desired material-specific target parameter(s) of the additively manufactured metallic body. The optimization method is performed such that in each iteration of the optimization step, the preceding set of process parameter combinations is optimized. The respective set of data combinations are then used as potential candidates for the optimal process parameters to further optimize the material-specific target properties of the additively manufactured metallic bodies.The use of batch Bayes optimization also allows multiple test specimens to be used simultaneously, making the optimization process more efficient and effective.
[0036] The proposed batch-Bayes optimization advantageously allows for the definition of a limited parameter range for the at least two different process parameters for the additive manufacturing of the at least one metallic specimen by employing Gaussian process regressors and / or neural networks. This, in conjunction with the machine learning used, creates a link between the process parameters to be provided and the selected material-specific target parameters. The optimization results predict both the mean and the standard deviation of the material-specific target properties of the additively manufactured metallic specimens.
[0037] To achieve the best possible combination of required process parameters and material-specific target parameters of the additively manufactured metallic body, the Bayesian optimization method is to be repeated iteratively until at least one predefined value of the material-specific target parameter with the associated process parameters is reached.
[0038] In an advantageous embodiment of the proposed optimization method, the optimization can be applied to additively manufactured metallic specimens with complex geometries. For this purpose, at least two elementary geometric elements are additively manufactured based on the geometry of the metallic body. Subsequently, at least the mechanical properties of these elementary geometric elements are determined, and at least one combined output dataset is calculated from the determined values. This output dataset consists of at least one material-specific target parameter and the corresponding process parameters for each element. The process parameters for each elementary geometric element are then optimized and combined to obtain optimized parameters for the overall geometry, using the acquisition function for multi-objective optimization across multiple properties and geometric configurations.This advantageous approach is implemented by applying the Bayesian optimization method after the additive manufacturing of the elementary geometric elements, using the geometric data of at least two formed elementary geometries of the metallic specimen. This improves the optimization method by eliminating the need to perform the optimization based on complex geometries of the metallic specimens. Instead, the metallic specimens are decomposed into simplified individual geometries, which are oriented in different directions and, for example, examined and characterized at different angles of inclination relative to the process build plate.
[0039] The technical benefit of this approach lies in the ability to simulate various thermal conditions encountered during the additive manufacturing process, such as those occurring in metallic bodies of differing geometric complexity. The aim is to identify the optimal process and material parameters for each elementary geometry and to determine the element-based process parameters required for the additive manufacturing of the complex metallic bodies as a whole. In essence, this involves decomposing the complex geometry into several simpler geometries and, once the corresponding optimal process parameters are identified, optimizing the process parameters for additive manufacturing of arbitrarily complex geometries.Advantageously, the provision of the geometric data of the elementary individual geometries can be achieved through parallelepipeds, tessellation, mesh decomposition and / or triangulation of the metallic body.
[0040] In a final step according to the invention, the determined and optimized process parameters are transferred to an additive manufacturing system, and the metallic body is produced in combination with at least one material-specific target parameter. The additive manufacturing process of the metallic body is thus carried out with previously optimized process parameters that lead to achieving at least one desired material-specific target property of the metallic body. Material-specific target parameters can include density, mechanical strength, porosity, surface roughness, and / or ductility.
[0041] In summary, the technical advantages and effects of the novel process for optimizing the additive manufacturing of metallic bodies consist of the following:
[0042] - for each individual additive manufacturing system with a minimal data set, an iterative optimization method is provided, in which the size of the initial data set is significantly reduced and the associated work and resource expenditure for its creation is considerably decreased,
[0043] - enables efficient use of the physical properties of metallic test specimens in combination with optimized data analysis,
[0044] optimized process parameters are provided more efficiently and cost-effectively overall
[0045] - by gradually narrowing down the optimal process parameters, it is ensured that resources are used effectively,
[0046] - through the possibility of using substitute models that generate a Gaussian distribution, both the mean and the standard deviation of the target properties are captured, leading to higher reliability of the optimization results,
[0047] - several possible process parameter combinations are provided for the optimized material-specific target parameters, enabling a high degree of flexibility in the additive manufacturing process, and
[0048] - the process parameters can be individually optimized and adjusted to the complex component geometries of the metallic bodies to be manufactured.
[0049] The invention is explained in more detail below using an exemplary embodiment. The accompanying figure and table illustrate this.
[0050] Table 1 shows an initial input data set for the subsequent Bayesian optimization procedure and the predicted batch of process parameters, and
[0051] Figure 1 schematic representation of formed elementary individual geometries of a metallic test specimen
[0052] Example of implementation
[0053] The aim is to optimize three process parameters required for the additive manufacturing of a metallic body—namely, laser power, exposure speed, and hatching spacing—in order to precisely achieve the density of a metallic component manufactured using a laser beam melting process. Based on the Latin hypercube sampling algorithm, combinations of process parameters are statistically selected. The process begins with the experimental generation of an initial dataset of relative density measurements of additively manufactured metallic specimens produced under various parameter conditions. Three specimens are produced for each process parameter combination, with each combination yielding one experimental data point.This initial dataset consists of seven experimental data points proposed using Latin hypercube sampling, where each point represents a combination of laser power, exposure speed, and hatching spacing.
[0054]
[0055] Table 1
[0056] According to Table 1 above, a sample is produced for each process parameter set using laser beam melting, and the respective relative density is determined using the Archimedean method and confirmed by additional p-computed tomography measurements. This initial data set is used to initiate the optimization process. For optimization, a batch Bayesian optimization with a capture function is performed, thereby transforming the optimization problem for the required process parameters into a capture optimization problem. In a first optimization step, a set of possible process parameter set combinations is determined that improves the target property, namely the required density of the metallic body. Once the capture function is calculated, three sets of optimal process parameters are provided, representing a potentially optimal combination of laser power, exposure rate, and hatching spacing.This group is selected simultaneously to allow for multiple experiments to be conducted concurrently. The selection of the three sets of optimal process parameters is achieved by considering both uncertain regions of the parameter space and concentrating on regions that yield optimal results. In this example, the next group is selected based on its highest knowledge gradient values to allow for further experimentation if necessary. These candidates are then experimentally evaluated, and their results are fed back into the dataset. This ensures that each iterative optimization result, with its predicted process parameters and material-specific target parameters, is updated and improved in the current or subsequent iterations. This iterative process continues until the optimal process parameters are determined.In the present example, two iterations were required to determine the optimal processing conditions that yield samples with a relative density of over 99.2%.
Claims
Patent claims 1. A method for optimizing the additive manufacturing of metallic bodies, in which at least one metallic test specimen is produced using an additive manufacturing process under quantified specification of at least one defined material-specific target parameter and at least two different process parameters, subsequently at least the mechanical properties of the metallic test specimen are determined and at least one combined output data set is calculated from it, which includes at least one material-specific target parameter and the associated process parameters of the at least one metallic test specimen, subsequently the output data are transformed into an acquisition function for optimizing the output data set.wherein, by using machine learning with the application of a Bayesian optimization method, a first data set combination of optimized process parameters is calculated for at least one associated material-specific target parameter, wherein the Bayesian optimization method is carried out iteratively until a predefined value of the at least one material-specific target parameter is reached, and subsequently the determined process parameters are transferred to an additive manufacturing system and the metallic body is produced in combination with at least one material-specific target parameter.
2. The method according to claim 1, wherein physical, mechanical and / or chemical values are specified as material-specific target parameters.
3. Method according to claim 2, wherein the density, mechanical strength, porosity, surface roughness and / or ductility are specified as material-specific target parameters.
4. The method of claim 1, wherein a limited parameter range is specified by the use of Gaussian process regressors and / or neural networks.
5. Method according to claim 1, wherein the optimized process parameters x are determined by the acquisition function, where x* = argmax x A(x) with A(x) represents the acquisition function and x denotes a vector comprising the process parameters to be optimized.
6. Method according to claim 5, wherein the acquisition function is an information-theoretic entropy search or a knowledge gradient acquisition function according to the equation KG(x) = £"[p* n +i P*n% n+1 = x] is used, where p* is the expected best estimate after the n+1 -th evaluation of the function of the process parameter x.
7. Method according to claim 1, wherein the optimization is carried out using a batch Bayesian optimization method in which multiple sets of process parameters are output and a parallel evaluation and optimization of the material-specific target properties is realized.
8. The method of claim 1, wherein at least two elementary geometric elements are additively manufactured based on the overall geometry of the metallic body, subsequently at least the mechanical properties of the elementary geometric elements are determined, and at least one combined output data set is calculated from the determined values, consisting of at least one material-specific target parameter and the associated process parameters of each element, and subsequently the process parameters for each elementary geometric element are optimized and combined to obtain optimized parameters for the overall geometry, wherein the acquisition function for multi-target optimization across multiple properties and geometric configurations is used.
9. Method according to claim 8, wherein the geometric parameters of the elementary individual geometries are provided by parallelepipeds, tessellation, mesh decomposition and / or triangulation of the metallic specimen.
10. Method according to claim 1, wherein prior knowledge from similar materials or alloys is incorporated into the optimization process using a Bayesian multi-probability optimization framework, integrating low-reliability data from similar materials into high-reliability experimental data.