Ai-enhanced system and method for alloy development
The system leverages GANs and PINNs to generate synthetic alloy data, optimizing alloy compositions through reinforcement learning and simulation, addressing the inefficiencies of traditional alloy development methods by enhancing data availability and reducing experimental costs.
Patent Information
- Application Number
- PCT/EP2026/062386
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-07
- Filing Date
- 2026-04-29
- Publication Date
- 2026-11-12
AI Technical Summary
Developing new alloys is time-consuming and expensive due to the need for extensive simulation and experimental efforts, and available data is scarce and incomplete, hindering efficient alloy development.
A system utilizing Generative Adversarial Networks (GANs) and Physics-Informed Neural Networks (PINNs) generates synthetic alloy data to complement real data, enhancing the training of a proposal module that suggests new alloy compositions, optimized through reinforcement learning and simulation, with Explainable AI (XAI) for transparency and federated learning for collaborative data sharing.
This approach accelerates alloy development by improving accuracy and reducing reliance on costly experimental datasets, enabling faster and more efficient alloy composition discovery.
Smart Images

Figure EP2026062386_12112026_PF_FP_ABST
Abstract
Description
[0001] AI-ENHANCED SYSTEM AND METHOD FOR ALLOY DEVELOPMENT
[0002] TECHNICAL FIELD
[0003] The present disclosure to a system for developing an alloy, a computer-implemented method for developing an alloy, and a computer-readable storage medium.
[0004] BACKGROUND
[0005] For various devices and machines alloys with specific properties are used. For example, blades in gas turbine engines must resist high temperatures, have a large stiffness, a long lifetime and a low weight. In the automotive and biomedical sectors, for example, other alloys are used, also with fine-tuned properties. Further refinement of these properties can lead to significant improvements, e.g., of the efficiency of a gas turbine engine. Therefore, there is a need for the development of new alloys.
[0006] Nowadays, developing new alloys is very time-consuming and may even take several years for one new alloy. The development of a new alloy typically is also expensive, because high simulation efforts are necessary (e.g., using ICME, Integrated Computational Materials Engineering, ThermoCalc, etc.) and requires a high experimental effort (alloy production, sample generation, testing, etc.).
[0007] There is a need to provide an improved system and method for developing an alloy or at least a useful alternative to known systems or methods for developing an alloy.
[0008] SUMMARY
[0009] In a first aspect, a system for developing an alloy is provided. The system comprises a data-acquisition module, a synthesis module and a proposal module. The data-acquisition module is configured for acquiring known-alloy data (e.g., publicly known) of known alloys. The synthesis module comprises one or more Generative Adversarial Networks (GANs) and / or one or more Physics-Informed Neural Networks (PINNs) and is configured to generate synthetic alloy data. The proposal module comprises a model that is trained with at least a part of the acquired known-alloy data and the synthetic alloy data. The proposal module is configured to generate a proposed new alloy composition (i.e., the composition of a new alloy that is not among the known alloys) and to provide the proposed new alloy composition at an output.
[0010] This is based on the finding that so far available (real) data of known alloys is too scarce and incomplete to serve as a basis for an Al-based development of a new alloy. By training the model not only with the available real data, but also with synthetic data, the gaps in theavailable real data can be filled in a reasonable manner. This combined data set comprising both real and synthetic data, is qualified to train the model such that it is enabled to provide meaningful proposals for new alloys. By leveraging both real and simulated data, the system enhances the accuracy of property predictions while minimizing reliance on costly experimental datasets. Thus, this system allows to speed up the development of new alloys with reduced efforts and, therefore, to improve alloy development.
[0011] In some embodiments, the model of the proposal module is configured to be triggered by a prompt of a user. The prompt may comprise and / or indicate desired material properties. In some embodiments, the model of the proposal module is configured to generate a proposed new alloy composition in accordance with the desired material properties (e.g., as comprised in the prompt or provided in another way to the model), and to provide the proposed new alloy composition at the output. To give an illustrative example, a user may input a prompt to the model such as: “I need an alloy with the material properties x, y and z” or: “Create a new alloy composition with the material properties indicated in the file located in the database” (the file and database being accessible to the model). Alternatively, the model may simply begin to create a new alloy composition (e.g., in accordance with stored target properties) as soon as the model is executed on a computer. As a further example, an interface (e.g., a GUI) may be provided wherein the user may enter predetermined target material properties into corresponding input items (e.g., including dropdown menus, checkboxes and / or fields accepting numbers) and / or trigger the model to generate an alloy composition, e.g., using a “start” button or the like. The model may send an inquiry to the one or more GANs to generate corresponding alloy data. This data may then, for example, be validated by the one or more PINNs. However, the model may alternatively, or in addition, become active on a regular basis, and / or if new material data becomes available or the like.
[0012] The system may further comprise a simulation module. The simulation module may be configured to simulate material properties of an alloy. By this, the system can determine the properties of proposed new alloys.
[0013] In some embodiments, the model of the proposal module is configured to optimize a new alloy composition by: (a) generating a preliminary alloy composition, (b) providing the preliminary alloy composition to the simulation module, (c) receiving an output of the simulation module, the output being indicative for material properties of the preliminary alloy composition, (d) comparing the material properties of the preliminary alloy composition with predetermined target material properties, and (e) modifying the preliminary alloy composition based on the result of the comparison. In this way, the model is enabled to test the performance of its proposals.
[0014] In some embodiments, the model of the proposal module may be configured to optimize the preliminary alloy composition iteratively. By means of such a fine tuning, an optimizedcomposition can be found.
[0015] In some embodiments, the model of the proposal module may be configured to provide the optimized alloy composition as the proposed new alloy composition if it meets the predetermined target material properties. Thereby, an alloy which at least in simulation has the desired properties can be provided. The target material properties may include thermal resistance, strength, stiffness, specific weight, etc.
[0016] Alternatively, or in addition, in some embodiments, the proposal module or the synthesis module may be configured to provide a synthetic alloy composition of the synthetic alloy data to the simulation module, to receive an output of the simulation module, the output being indicative for material properties of the synthetic alloy composition, and the proposal module may be configured to train the model using the output. This allows to further improve the training of the model.
[0017] The simulation module may comprise a quantum-mechanical simulator and / or a classical material simulator. Thereby, realistic results of the properties can be obtained.
[0018] In some embodiments, the model of the proposal module comprises a reinforcement learning algorithm. This allows to efficiently provide accurate and meaningful new alloy composition proposals. The model (e.g., reinforcement learning model) may learn the relationship between alloy compositions and material properties based on the known-alloy data and the synthetic alloy data. After training, the model may (actively) propose optimized alloy compositions that are predicted to achieve the desired target properties (e.g., creep strength, corrosion resistance and / or high-temperature stability). The model may receive the known-alloy data selected and / or pre-processed by an XAI module (see below) plus synthetic data from the GAN(s) and / or PINN(s). Thus, the model may perform the actual machine learning optimization and propose new alloys.
[0019] In some embodiments, the data-acquisition module may comprise an LLM module (comprising a Large Language Model). The LLM module may be configured to analyze text documents, such as scientific papers, text books and / or patent literature, to extract descriptions of known alloys and to output known-alloy descriptions. This allows to access vast amounts of data, e.g., all alloy-related data publicly available on the internet (or at least a part thereof). In this way, training the model with a dataset that comprises almost the complete knowledge about alloys is provided.
[0020] In some embodiments, the proposal module may comprise an Explainable Al (XAI) module. The XAI module may be configured for selecting alloy compositions from the known-alloy descriptions, for providing reasons for the selection for display, and for outputting data of the selected alloy compositions as the known-alloy data to the model. Therefore, the decisionmaking process can be made comprehensible, and more reliable, for humans. The XAI module may be configured to prepare, filter and / or explain the data for the (e.g., reinforcement learning)model. Specifically, the XAI model may perform a data pre-selection. It may select the most relevant known-alloy data based on target properties (e.g., why nickel, why chromium as an addition). The XAI module may provide human-readable explanations why certain known alloys were selected and why certain material choices are recommended (e.g., “Nickel is selected due to superior heat resistance.”). The XAI module may support the user by making the Al’s reasoning transparent (e.g., shown on a display or dashboard). Thus, the XAI module may prepare meaningful training data and explain the final recommendations but, e.g., does not generate new alloys itself.
[0021] In some embodiments, the XAI module may be configured to extract and / or synthesize information from text documents, e.g., scientific and / or patent literature, analyzed by the LLM module. This allows to support alloy design recommendations.
[0022] In some embodiments, the proposal module comprises a federated-learning module. The federated-learning module may be configured to receive one or more model parameters for the model from one or more external databases and / or to provide one or more model parameters being a result of the training of the model to the one or more external databases. By this, huge amounts of additional training data may be acquired without disclosing the actual details of the alloys which often are confidential.
[0023] In some embodiments, the federated-learning module may be configured to enable multiple institutions to collaboratively train their models without sharing proprietary data. By this, the model (and corresponding model of the other institutions) can be trained more effectively.
[0024] In some embodiments, the system may further comprise an interface for transmitting the proposed new alloy composition to an external additive-manufacturing database. This allows to quickly prepare a physical test of the proposed new alloy in order to verify its simulated performance.
[0025] In some embodiments, the at least one GAN comprises two neural networks, a generator and a discriminator, the generator being trained to generate data similar to real data and the discriminator being trained to discriminate the generated data of the generator from real data. This setup allows to generate highly realistic generated data. The two neural networks may compete with each other in the form of a (e.g., zero-sum) game. The generator’s training objective is to increase the error rate of the discriminator, i.e. , to fool the discriminator by producing generated data that passes the discriminator, i.e., which the discriminator determines not to be synthesized, i.e., to be part of the real data. A given dataset of real data is used as the initial training data for the discriminator. The training comprises presenting the discriminator with samples from the training data until it achieves an acceptable accuracy. The generator is trained based on whether it succeeds in fooling the discriminator.
[0026] Therein, the real data may be at least a part of the known-alloy data. This allows tocreate one or more GANs specialized in alloy data creation.
[0027] In some embodiments, the at least one GAN provides generated data to the at least one PINN. In some embodiments, the at least one GAN provides the generated data which passed the discriminator to the at least one PINN. The at least one PINN may be configured to validate the generated data based on physical laws. The at least one PINN may be further configured to provide, if validated, the generated data as the synthetic alloy data to the proposal module. In this manner, highly realistic alloy property data can be provided that has a high likelihood of containing correct descriptions of alloy compositions. To give an example, two variants of an alloy, both variants comprising the same elements, but with different proportions in two elements, are known with their properties and have been acquired by the data-acquisition module. The GAN and the PINN are able to fill the gap between these known compositions and calculate properties of variants with the proportions varying between the known variants. Optionally, the synthesis module is configured to analyze the acquired known-alloy data provided by the data-acquisition module, to identify gaps within this data, and to provide synthetic alloy data that fills these gaps.
[0028] In a second aspect, a (computer-implemented) method for developing an alloy is provided. The method comprises: acquiring, by a data-acquisition module, known-alloy data of known alloys; generating, by a synthesis module comprising at least one Generative Adversarial Network (GAN), and / or at least one Physics-Informed Neural Network (PINN), synthetic alloy data; and generating, by a proposal module comprising a model trained with the acquired known-alloy data and the synthetic alloy data, a proposed new alloy composition and providing the proposed new alloy composition at an output.
[0029] In some embodiments, the method may comprise training the model with the acquired known-alloy data and the synthetic alloy data to obtain the trained model. In some embodiments, after developing the proposed new alloy having a proposed new alloy composition with the method described above, a method may comprise alloying the proposed new alloy composition to obtain powder of the proposed new alloy composition and / or manufacturing an object (e.g., with a machine for additive manufacturing) using (e.g., powder of) the proposed new alloy composition.
[0030] In a third aspect, a computer-readable storage medium is provided, storing instructions that, when executed by a processor assembly, cause the processor assembly to perform the method described above.
[0031] BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Embodiments will now be described by way of example with reference to the figures in which:FIG. 1 depicts in a schematic manner a trace of a laser beam for melting of powder particles during a laser powder bed fusion process.
[0033] FIG. 2 depicts a schematic view of an example machine for additive manufacturing.
[0034] FIG. 3 depicts a schematic illustration of components of the machine of FIG. 2.
[0035] FIG. 4 depicts a system for developing an alloy.
[0036] FIG. 5 depicts a method for developing an alloy and for manufacturing an object using the alloy.
[0037] DETAILED DESCRIPTION
[0038] Aspects and embodiments of the present disclosure will now be discussed with reference to the accompanying figures. Further aspects and embodiments will be apparent to those skilled in the art.
[0039] One additive manufacturing process is Laser Powder Bed Fusion (LPBF). Using the LPBF process, it is possible to produce high-precision metal parts layer by layer using a laser. Such metal parts can be used, e.g., for blades in gas turbine engines or the like. By creating an internal structure of such blades, a lightweight design with a large stiffness is possible, to name just an example.
[0040] FIG. 1 shows how one layer of a manufactured part (herein also generally referred to as “object”) is created in various additive manufacturing techniques such as LPBF. Here, an even layer of metal powder of a given alloy is deposited on a surface. A laser beam B, strong enough to melt the metal powder, is focused on the powder surface. Using an adjustable optical assembly, the focused laser beam B spot is scanned along a pattern of vectors V. The laser beam B has a laser power L that determines time needed to melt a given area of the metal powder. This time results in a velocity of the laser beam B spot with which the spot can be moved to reliably melt all metal powder within a given area. The size of this area determines the hatch distance H and the possible layer thickness T. The velocity S with the resulting path length on one layer for the laser determines the time necessary for one layer. The layer thickness T, the times for the layers and the number of layers to obtain the desired shape and size of the manufactured object determine the total time to manufacture one object. So, while objects with very complex shapes can be produced by LPBF with a very high precision, the manufacturing time is substantially longer when compared to other manufacturing techniques. Therefore, it is not only desirable to decrease the time necessary to produce one object, but also to minimize defects that render an object useless as much as possible.
[0041] FIGS. 2 and 3 depict a machine 1 for additive manufacturing. The machine 1 is a laser powder bed fusion, LPBF, machine 1, however, it is noted that other additive manufacturing processes are conceivable, too, and the LPBF process is described by way of example, only.The machine 1 comprises a case 10 which defines a build chamber 100 therein. Inside the build chamber 100, a build plate 19 is arranged on which objects O (see FIG. 3) can be manufactured.
[0042] FIG. 3 shows the laser 12 that emits the laser beam B and an optical assembly 18 which focuses the beam over a surface of the build plate 19 and moves it as described above with reference to FIG. 1. The optical assembly may comprise one or more movable mirror(s) and / or one or more lenses.
[0043] As depicted in FIG. 3, the build plate 19 can be moved downwards. As soon as the laser 12 has melted all powder P within the current top layer which shall solidify to build the object O, the laser is stopped, the build plate 19 is moved downwards by a distance corresponding to the layer thickness T. Then, a recoater 16 spreads new powder P evenly over the previous layer. The powder P is provided by a powder supply 17. The recoater 16 is, e.g., an elongate bar which is displaced parallel to the build plate 19 top surface.
[0044] On the case 10, within a wall 101 of the case 10, a window 11 is mounted. Outside of the case 10, the laser 12 is arranged (see FIG. 3) to emit a laser beam B through the window 11 into the build chamber 100.
[0045] A control unit 14 (see FIG. 2) is configured to control the parts of the machine 1, in particular the laser 12, the optical assembly 18, the build plate 19.
[0046] The selection of the alloy for the powder P is crucial for the properties of the object O. For example, if the object O will be a fan, compressor or turbine blade of a gas turbine engine, a nickel alloy can be used, which may further comprise aluminum, chromium tungsten and / or additional elements (e.g., five additional elements or even more). The exact composition, i.e., which elements are combined in the alloy and to which percent by weight each of the elements is present determines the performance of the alloy. Further, for many parameters such as stiffness, heat resistance, corrosion resistance, hardness etc. certain target values may be defined. And the selection of the most important target parameters, and their values, strongly depend on the use case. In the example of a turbine blade, the heat resistance is of specific importance, whereas for a compressor blade, the specific weight may be more relevant than the heat resistance, at least when compared to the turbine blade. Other applications, e.g., in the aerospace, automotive, energy, marine propulsion, wind energy or biotechnology sectors may have still other target performances. Due to the number of different target performances in the various use cases, the mere number of possible elements in the alloy and their percentage of the total weight, developing an alloy and choosing the best alloy for the given use case, e.g., for choosing the powder P for use in the machine 1, traditionally would be a very difficult and time-consuming process.
[0047] Here, a system 2 for developing new alloys is used which speeds up this process and allows improved results in a much shorter timeframe.The system 2 is illustrated in FIG. 4 and comprises various modules or blocks. These modules can be embodied as hardware and / or software. In one example, all components of the system 2 are embodied as a software which is stored in a memory assembly 25 of the system 2 (see, e.g., FIG. 2), e.g., as instructions that, when executed by a processor assembly 26 of the system 2, perform a method of developing an array, e.g., as illustrated in FIG. 5. The memory assembly 25 and the processor assembly 26 of the system 2 may be part of a desktop computer, an Al server, a computing farm, a high-performance computing cluster, a distributed arrangement of computers, a cloud-based data infrastructure, etc. Alternatively, or in addition, the modules of the system 2 may be embodied at least partially as hardware. The modules of the system 2 may be located at one place, or at least partially remote from one another. That is, one module may be located and / or processed on one computer which is remote from another module and / or the computer that module is running on. However, the modules 2 are communicatively connected with one another and, thereby, form the system 2.
[0048] Here, the system 2 comprises a data-acquisition module 20 (DAQ) for acquiring known-alloy data of known alloys. The known-alloy data is in the form text documents 30 and available on the internet 3. The data-acquisition module 20 may have access (wired and / or wireless) to the internet 3 in order to search for and read the text documents 30. The text documents 30 may comprise scientific publications, patent publications and / or webpages or even other types of documents.
[0049] The data-acquisition module 20 may use a search engine and / or a crawler or the like to search and find text documents 30 having a content that is related to alloys. Further, the data-acquisition module 20 may have access to scientific databases such as to journals and / or to preprint servers and / or the like. In the present example, the data-acquisition module 20 comprises a Large Language Model, LLM 200. Various suitable LLMs are available on the market which are suitable (e.g., with a corresponding training) for this task. The LLM 200 is configured to (and used by the system 2 to) extract (e.g., large amounts of) data from the text documents 30. Here, the LLM 200 is trained to material science and metallurgy. The LLM 200 is configured for extracting and synthesizing knowledge around the world. The LLM 200 is trained in various languages, so that it is able to construct an alloy database which includes almost the entire publicly available human knowledge about alloys. In particular, the specific alloy compositions (ingredients and their contribution by weight) with their respective properties, such as stiffness, melting point, elasticity, strength and / or heat resistance, and / or other parameters.
[0050] However, even the entire publicly available human knowledge about alloys is incomplete and sparse. Therefore, the system 2 generates additional synthetic alloy data using a synthesis module 21.
[0051] The synthesis module 21 comprises a Generative Adversarial Network, GAN 210 (or aplurality of GANs 210) and a Physics-Informed Neural Network, PINN 211 (or a plurality of PINNs 211). For example, multiple GAN-PINN pairs may be applied. To ensure comprehensive and accurate synthetic data generation for alloy development, the system may employ multiple Generative Adversarial Networks (GANs) and multiple Physics-Informed Neural Networks (PINNs) instead of relying on a single model. Each GAN is specialized to generate synthetic data focusing on specific aspects of alloy properties or different scales of material behavior, such as atomic structure, microstructural features, or macroscopic mechanical properties. Correspondingly, each PINN is tailored to validate the synthetic data against distinct physical laws relevant to the targeted property, such as thermodynamic stability, phase transformation kinetics, or mechanical stress responses. The use of multiple GANs and PINNs provides specialization and accuracy: By dividing tasks among specialized GANs and PINNs, the system achieves higher accuracy in representing and validating complex, multi-physics behaviors of alloys. Further, parallelization and speed can be obtained, because multiple GAN-PINN pairs operate in parallel, enabling rapid generation and validation of large volumes of synthetic data across diverse regions of the material design space. Robustness and generalization may also be obtained: The modular architecture ensures that rare edge cases or unusual physical phenomena are captured more reliably, enhancing the robustness and generalization capability of the overall Al framework. This design significantly accelerates the exploration of high-dimensional alloy spaces while maintaining the physical integrity and quality of the synthetic datasets, leading to more reliable predictions and faster identification of optimal alloy compositions. The synthesis module 21 is configured to generate the synthetic alloy data using the GAN(s) 210 and the PINN(s) 211.
[0052] Here, the GAN 210 (or each of the GANs 210) comprises two interacting neural networks, a generator and a discriminator. The generator is trained to generate data which is similar to real data (e.g., similar to the known-alloy data acquired by the DAQ 20). The generator may be trained with the known-alloy data which assigns performance parameters to alloy compositions. For example, the synthesis module 21 is configured to read the known-alloy data acquired by the DAQ 20 and to identify gaps in the data where in a range of compositions no performance parameters for the corresponding compositions are publicly known. The synthesis module 21 is configured to determine alloy compositions which so far have unknown performance parameters, and to use the GAN 210 to generate performance parameters for these previously unknown alloy compositions as generated data.
[0053] The discriminator is trained to discriminate the generated data of the generator from the real data. If the discriminator identifies the generated data as artificial, the generator generates new data, e.g., refines the data. If the discriminator identifies the generated data as real (even though it is not), this generated data is provided to the PINN 211.
[0054] The PINN 211 may comprise a set of partial differential equations. The PINN 211 isconfigured to validate the generated data based on physical laws and to provide, if successfully validated, the generated data as synthetic alloy data to a proposal module 22.
[0055] The proposal module 22 receives the acquired known-alloy data from the DAQ 20 as well as the synthetic alloy data from the synthesis module 21.
[0056] The proposal module 22 comprises an Al model 220, in the present example with a reinforcement learning algorithm. The model 220 is trained with the acquired known-alloy data and the synthetic alloy data. By the training, the model 220, e.g., creates internal neural network maps based on the training data.
[0057] The model 220 is configured (e.g., in response to a query or prompt provided as input by a user, which may comprise predefined target performance criteria such as required material properties) to generate a proposed new alloy composition which meets predefined target performance criteria, and to provide the proposed new alloy composition at an output 221. The output 221 is, in this example, connected to a display 24, so that the proposed new alloy composition (and its Al determined properties) can be displayed to a user. The proposed new alloy composition can also be output at an interface 225 to be for transmitting the proposed new alloy composition to an external additive-manufacturing database 5. There, the composition can be read, alloyed, and tested in real life without delay.
[0058] The model 220 is, in this example, further configured to send a request to the GAN 210 (or each of theGANs 210) to provide alloy data which meets the predefined target performance criteria (e.g., within predefined or configurable ranges) or parts thereof. The GAN 210 (or each of the GANs 210) are, as an example, configured to generate and provide corresponding alloy compositions to the PINN(s) 211. The PINN(s) 211 are configured to validate these provided alloy compositions and to provide, if validated, the generated data as the synthetic alloy data to the model 200 of the proposal module 22.
[0059] Notably, the exact composition of an alloy can be very valuable and may be proprietary information which different companies or other institutions are not willing to share. However, this information could significantly improve the training of the model 220.
[0060] Therefore, the proposal module 22 comprises a federated-learning module (FED) 224 which is configured to enable multiple institutions to collaboratively train models without sharing proprietary data. The federated-learning module 224 can be configured to receive one or more model parameters for the model 220 from one or more external databases 4 and to provide one or more model parameters, which are a result of the training of the model 220, to the one or more external databases 4 (e.g., of various external institutions). This additional data further improves the quality of the model 220.
[0061] The system 2 further comprises a simulation module 23. The simulation module 23 is configured to simulate material properties of a given alloy composition. The simulation module 23 comprises a quantum-mechanical simulator and / or a classical material simulator for thispurpose. The simulation module 23 may be configured to perform a multi-scale simulation. The simulation module 23 may be configured to assess alloy properties at atomic, microstructural, and macroscopic levels.
[0062] For finding a new alloy, the model generates a new alloy (which is not yet present in the known-alloy data) based on its training and with target performance parameters (e.g., target material properties) as input. The model 220 is additionally configured to perform an optimization of the new preliminary alloy, i.e. , to perform a fine tuning of the composition. Specifically, the model 220 optimizes a new alloy composition by: generating the preliminary alloy composition, providing the preliminary alloy composition to the simulation module 23, receiving an output of the simulation module 23, the output comprising the material properties of the preliminary alloy composition, comparing the material properties of the preliminary alloy composition with the target material properties, and modifying the preliminary alloy composition based on the result of the comparison. This process is repeated iteratively, wherein the alloy composition is modified in each step. For example, one component is increased by an amount, another component is reduced, or the like. The model 220 may minimize the distance of the simulated material properties to the target material properties (as for a linear or exponential fit function). For example, as soon as the optimization converges at a minimum, the proposed new alloy composition is output.
[0063] Notably, the PINN(s) 211 and the simulation module 23 may be separate modules or combined into one module. Also, the synthesis module 21 may use the simulation module 23 to simulate material properties of the alloy.
[0064] The output of the simulation module 23 is also used to train the model 220 (together with the corresponding generated alloy compositions).
[0065] Further, the proposal module 22 comprises an Explainable Al module, XAI module 222. The XAI module 222 is configured to select alloy compositions from the known-alloy descriptions as the known-alloy data to the model 220 which are then used to train the model 220. The XAI module 222 is also configured to provide reasons for the selection it has made (e.g., for display on the display 24, more specifically on an XAI visualization dashboard). This allows a human to understand why the system 2 has proposed a certain alloy composition in general. For example, the XAI module 222 indicates that for the given target properties of a turbine blade, Nickel has been selected as the base material, because nickel-based alloys can have excellent heat resistance, and that further ingredients have been selected for stiffness, corrosion protection, etc. The XAI module 222 is configured to select and synthesize relevant known-alloy data to train the reinforcement learning model 220 and to provide human-understandable explanations for the data selections and alloy design proposals. The model 220 is configured to optimize and propose novel alloy compositions based on the training with both the selected known-alloy data and synthetic alloy data, aiming to meet predefined targetproperties.
[0066] Here, the XAI module 222 is configured to extract and synthesize information from the text documents 30 analyzed by the LLM module 200 to support the alloy design recommendations.
[0067] The system may have various data Interfaces, e.g., APIs for Data Exchange (allow integration with industrial alloy testing databases), secure communication protocols (ensure confidentiality in federated learning contributions) and / or simulation & Al model interfaces (to connect computational simulations with machine learning pipelines).
[0068] FIG. 5 shows a method. Therein, step S1 corresponds to a computer-implemented method to develop an alloy and comprises the following steps:
[0069] Step S10: Acquiring, by the data-acquisition module 20, known-alloy data of known alloys.
[0070] Step S11: Generating, by the synthesis module 21 comprising the at least one Generative Adversarial Network, GAN 210, and / or the at least one Physics-Informed Neural Network, PINN 211, synthetic alloy data.
[0071] Step S12: Generating, by the proposal module 22 comprising the model 220 trained with the acquired known-alloy data and the synthetic alloy data, a proposed alloy composition and providing the proposed alloy composition, e.g., at the output 221.
[0072] The method of FIG. 5 further comprises:
[0073] Step S2: Producing the proposed new alloy composition. Optionally, the new alloy is produced in the form of a powder.
[0074] Step S3: Manufacturing an object with the machine 1 for additive manufacturing using the powder of the proposed new alloy composition. Alternatively, an object can be forged, casted or the like.
[0075] The disclosure is not limited to the embodiments described above, and various modifications and improvements may be made without departing from the concepts described here. Any of the features may be used separately or in combination with any other features, unless they are mutually exclusive, and the disclosure extends to and includes all combinations and sub-combinations of one or more features that are described herein.
[0076] It is to be understood that the elements and features recited in the appended claims may be combined in different ways to produce new claims that likewise fall within the scope of the present disclosure. Thus, whereas the dependent claims appended below depend on only a single independent or dependent claim, it is to be understood that these dependent claims may, alternatively, be made to depend in the alternative from any preceding or following claim, whether independent or dependent, and that such new combinations are to be understood as forming a part of the present specification.
[0077] While the present disclosure has been described above by reference to variousembodiments, it may be understood that many changes and modifications may be made to the described embodiments. It is therefore intended that the foregoing description be regarded as illustrative rather than limiting, and that it be understood that all equivalents and / or combinations of embodiments are intended to be included in this description.
[0078] LIST OF REFERENCE SIGNS
[0079] 1 machine
[0080] 10 case
[0081] 100 build chamber
[0082] 101 wall
[0083] 11 window
[0084] 12 laser
[0085] 14 control unit
[0086] 16 recoater
[0087] 17 powder supply
[0088] 18 optical assembly
[0089] 19 build plate
[0090] 2 system
[0091] 20 data-acquisition module
[0092] 200 LLM module
[0093] 21 synthesis module
[0094] 210 GANs
[0095] 211 PINNs
[0096] 22 proposal module
[0097] 220 model
[0098] 221 output
[0099] 222 XAI model
[0100] 224 federated-learning module
[0101] 225 interface
[0102] 23 simulation module
[0103] 24 display
[0104] 25 memory assembly
[0105] 26 processor assembly
[0106] 3 internet
[0107] 30 text document
[0108] 4 external database5 additive-manufacturing database H hatch distance
[0109] B laser beam
[0110] L laser power
[0111] O object
[0112] P powder
[0113] S velocity
[0114] T thickness
[0115] V vector
Claims
CLAIMS1. A system (2) for developing an alloy, the system comprising:a data-acquisition module (20) for acquiring known-alloy data of known alloys; a synthesis module (21) comprising at least one Generative Adversarial Network, GAN (210), and / or at least one Physics-Informed Neural Network, PINN (211), and configured to generate synthetic alloy data; anda proposal module (22) comprising a model (220) trained with the acquired known- alloy data and the synthetic alloy data, wherein the model (220) is configured to generate a proposed new alloy composition and to provide the proposed new alloy composition at an output (221).
2. The system (2) of claim 1, further comprising a simulation module (23) configured to simulate material properties of an alloy.
3. The system (2) of claim 2, wherein the model (220) of the proposal module (22) is configured to optimize a new alloy composition by: generating a preliminary alloy composition, providing the preliminary alloy composition to the simulation module (23), receiving an output of the simulation module (23), the output being indicative for material properties of the preliminary alloy composition, comparing the material properties of the preliminary alloy composition with predetermined target material properties, and modifying the preliminary alloy composition based on the result of the comparison.
4. The system (2) of claim 3, wherein the model (220) of the proposal module (22) is configured to optimize the preliminary alloy composition iteratively.
5. The system (2) of claim 3 or 4, wherein the model (220) of the proposal module (22) is configured to provide the optimized alloy composition as the proposed new alloy composition if it meets the predetermined target material properties.
6. The system (2) of any one of claims 2 to 5, wherein the proposal module (22) and / or the synthesis module (21) is configured to provide a synthetic alloy composition of the synthetic alloy data to the simulation module (23), to receive an output of the simulation module (23), the output being indicative for material properties of the synthetic alloy composition, and the proposal module (22) is configured to train the model (220) using the output.
7. The system (2) of any one of claims 2 to 6, wherein the simulation module (23) comprises a quantum-mechanical simulator and a classical material simulator.
8. The system (2) of any preceding claim, wherein the model (220) of the proposal module (22) comprises a reinforcement learning algorithm.
9. The system (2) of any preceding claim, wherein the data-acquisition module (20) comprises an LLM module (200) configured to analyze text documents (30) to extract descriptions of known alloys and to output known-alloy descriptions.
10. The system (2) of claim 9, wherein the proposal module (22) comprises an Explainable Al, XAI, module (222) configured for selecting alloy compositions from the known-alloy descriptions, for providing reasons for the selection for display, and for outputting data of the selected alloy compositions as the known-alloy data to the model (220).
11. The system (2) of claim 10, wherein the XAI module (222) is configured to extract and synthesize information from scientific and patent literature analyzed by the LLM module (200) to support alloy design recommendations.
12. The system (2) of any preceding claim, wherein the proposal module (22) comprises a federated-learning module (224) configured to receive one or more model parameters for the model (220) from one or more external databases (4) and / or to provide one or more model parameters being a result of the training of the model (220) to the one or more external databases (4).
13. The system (2) of claim 12, wherein the federated-learning module (224) is configured to enable multiple institutions to collaboratively train models without sharing proprietary data.
14. The system (2) of any preceding claim, further comprising an interface (225) for transmitting the proposed new alloy composition to an external additive-manufacturing database (5).
15. The system (2) of any preceding claim, wherein the at least one GAN (210) comprises two neural networks, a generator and a discriminator, the generator being trained to generate data similar to real data and the discriminator being trained to discriminate the generated data of the generator from real data.1716. The system (2) of claim 15, wherein the real data is at least a part of the known-alloy data.
17. The system (2) of claim 15 or 16, wherein the at least one GAN (210) provides generated data to the at least one PINN (211), wherein the at least one PINN (211) is configured to validate the generated data based on physical laws and to provide, if validated, the generated data as the synthetic alloy data to the proposal module (22).
18. A computer-implemented method for developing an alloy, the method comprising the steps of:acquiring (S10), by a data-acquisition module (20), known-alloy data of known alloys;generating (S11), by a synthesis module (21) comprising at least one Generative Adversarial Network, GAN (210), and / or at least one Physics-Informed Neural Network, PINN (211), synthetic alloy data; andgenerating (S12), by a proposal module (22) comprising a model (220) trained with the acquired known-alloy data and the synthetic alloy data, a proposed new alloy composition and providing the proposed new alloy composition at an output (221).
19. The method of claim 18, further comprising:alloying (S2) the proposed new alloy composition to obtain powder of the proposed new alloy composition; andmanufacturing (S3) an object with a machine (1) for additive manufacturing using powder of the proposed new alloy composition.
20. A computer-readable storage medium (25) storing instructions that, when executed by a processor assembly (26), cause the processor assembly (26) to perform the method of claim 18 or 19.
