A machine learning based laser additive manufacturing substrate and method of manufacturing the same

Abstract

The application discloses a kind of laser additive manufacturing substrate based on machine learning and its manufacturing method, belong to laser additive manufacturing technical field.The method is first obtained the intrinsic of target component, material system and heat processing characteristic parameter, input pre-training model is imported into the adversarial network model containing generator and discriminator, after preliminary training, generator and discriminator are alternately optimized, candidate structure screening verification and data set enhancement iteration, obtain three-dimensional support structure, and then print the structure on the substrate body using laser powder bed fusion technology, finally using laser directional energy deposition technology is printed on support structure target component.This method abandons traditional trial-and-error design method, and efficiently optimizes through the machine learning model integrated with physical laws, and the designed substrate can synergistically control the thermal field, release stress, inhibit cracking, optimize the forming thermal environment, improve the forming stability and service performance of the component, and greatly shorten the development cycle.

Description

Technical Field

[0001] This invention belongs to the field of laser additive manufacturing technology, specifically relating to a laser additive manufacturing substrate based on machine learning and its manufacturing method. Background Technology

[0002] Laser-directed energy deposition (L-DED) is a high-performance machining technology that directly manufactures dense metal parts through point-by-point, layer-by-layer cladding deposition. This technology offers significant advantages in the one-piece forming of large, complex components, the fabrication of functionally graded materials, and the precise repair of damaged parts. However, the extremely high temperature gradient (up to 10⁻⁶℃) during the L-DED process is a major challenge. 6 Rapid cooling (temperatures of ℃ / m) and temperature fluctuations lead to significant internal stress accumulation. Simultaneously, the strong mechanical constraint exerted on the component by the rigid printed support substrate prevents effective release of these internal stresses. When the stress exceeds the material's tensile strength, severe cracking occurs, causing irreversible damage to the manufacturing process, resulting in material waste and a substantial increase in costs.

[0003] To address the aforementioned issues, current solutions primarily employ two approaches: 1) Controlling thermal stress during the printing process. The most common strategies include substrate preheating and process parameter optimization. While these methods can reduce thermal stress to some extent, their effectiveness is often limited for solidification cracking issues caused by the coupling of material properties and L-DED process characteristics in some difficult-to-weld alloys. Furthermore, improper substrate preheating temperature can easily lead to alloy microstructure coarsening, negatively impacting performance; while process parameter optimization is a lengthy trial-and-error process relying on experience, and finding a robust process window incurs high time costs. 2) Adjusting the mechanical constraints of the printing process. Targeted structural design of the substrate can alleviate cracking and deformation caused by strong constraints. However, existing substrate structure designs are relatively simple and functionally limited, failing to meet the high-quality printing requirements of complex components. Moreover, substrate design requires customized design for different components, a cumbersome process. More importantly, changes to the substrate structure inevitably affect its stiffness and thermal conductivity, thus having a complex impact on component deformation and heat accumulation during the printing process. Improper design not only fails to control stress but may also lead to new quality problems.

[0004] It is evident that existing technologies are insufficient for systematically and collaboratively controlling the thermal-mechanical field during the printing process, and cannot simultaneously meet the dual requirements of stress control and optimization of microstructure performance. There is an urgent need for a new method that can intelligently design substrate structures and achieve proactive collaborative management of the thermal-mechanical field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a laser additive manufacturing substrate based on machine learning and its manufacturing method, so as to solve the problem that it is difficult to control the thermal-mechanical field of the printing process in the laser directional energy deposition process, and the problem that stress control and microstructure optimization cannot be satisfied at the same time.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A method for manufacturing a laser additive manufacturing substrate based on machine learning includes the following steps: S1, obtain the intrinsic parameters, material system, and material thermal processing characteristics of the target component to be formed; S2, input the intrinsic parameters, material system and thermal processing characteristic parameters into the pre-trained model, and output them to the adversarial network machine learning model to obtain the three-dimensional support structure; The adversarial network machine learning model includes a generator and a discriminator, and the training process of the adversarial network is as follows: S201, The adversarial network is initially trained using the initial dataset to obtain the initial adversarial network machine learning model; S202, the generator optimization and discriminator optimization are alternately performed on the initial adversarial network machine learning model. In the discriminator optimization stage, the generator parameters are fixed, and the three-dimensional support structure output by the generator is input into the discriminator. The discriminator predicts the temperature field and stress field, and updates the discriminator network parameters through backpropagation by combining the comprehensive loss value output by the discriminator. In the generator optimization stage, the discriminator network parameters are fixed, and the comprehensive loss value output by the discriminator is used as a feedback signal to backpropagate to the generator to adjust the generator network parameters. The generator optimization and discriminator optimization are repeated alternately until the comprehensive loss value no longer decreases significantly and the discriminator evaluation result tends to stabilize. S203: Use the current generator to generate multiple candidate structures, the discriminator calculates the comprehensive loss value, selects the two optimal candidate structures with the lowest comprehensive loss value and the candidate structure with the highest prediction uncertainty, and obtains the real stress field, temperature field and thermal response data through high-fidelity finite element thermo-mechanical coupling simulation. Add the verification data to the initial dataset to form an enhanced database, and re-execute 202 based on the enhanced database. S204, repeat steps 202 and 203 until the set conditions are met to obtain the final three-dimensional support structure. S3 involves printing a three-dimensional support structure onto a substrate by laser powder bed melting, and then fabricating the target component by laser directional energy deposition onto the three-dimensional support structure.

[0007] A further improvement of the present invention is that: Preferably, in S1, the intrinsic parameters include geometric dimensions, cross-sectional features, height, wall thickness, and spatial orientation information; the material thermal processing characteristic parameters include physical property parameters, mechanical property parameters, phase transformation temperature, and strengthening phase precipitation temperature range; and based on the service performance requirements and manufacturing process limitations of the target component, multi-objective constraint conditions are established, including residual stress loss terms, thermal history constraint terms, and manufacturability constraint terms.

[0008] Preferably, in S201, the generator adopts a multilayer perceptron neural network architecture, with random noise vector as input and three-dimensional support structure parameters as output. The three-dimensional support structure parameters include topology type, feature size, and spatial gradient distribution. The discriminator integrates a physical information neural network and a performance evaluation module. The physical information neural network takes the three-dimensional support structure parameters as input and outputs predicted temperature field and stress field. The performance evaluation module is used to calculate the difference between the multi-objective constraints and the comprehensive performance value.

[0009] Preferably, in S202, the comprehensive loss value is composed of a weighted average of residual stress loss, thermal history loss, and manufacturability loss; the heat conduction control equation and the linear elasticity equation are embedded in the physical information neural network as physical constraints.

[0010] Preferably, in S203, the step of selecting the two optimal candidate structures with the lowest comprehensive loss value and the candidate structure with the highest prediction uncertainty specifically includes: the discriminator calculates and sorts the comprehensive loss value of all candidate structures, and determines the two optimal candidate structures with the lowest comprehensive loss value and the candidate structure with the highest prediction uncertainty by the discriminator based on the sorting of comprehensive loss values.

[0011] Preferably, in S204, the set conditions are: the comprehensive loss value of the generated structure does not fall below the historical best record value for P consecutive iterations, and the error between the discriminator prediction result and the high-fidelity simulation verification result is less than a preset threshold, and the adversarial network machine learning model is mature and converged.

[0012] Preferably, in S3, the three-dimensional parametric model of the three-dimensional support structure and the three-dimensional model of the substrate body are integrated and assembled to form a substrate CAD model, which is then imported into a laser powder bed melting device for layer slicing and scanning path planning, and the three-dimensional support structure is printed layer by layer on the substrate body.

[0013] Preferably, a dense structural layer with a thickness of 0.1 mm to 0.8 mm is printed on the upper surface of the three-dimensional support structure, and then the target component is printed on the dense structural layer.

[0014] Preferably, before S2, the step of constructing an initial dataset is also included: based on finite element thermo-mechanical coupling simulation software, various basic topology types are selected as support structures, different structural parameters are set to form multiple sets of support structure parameter samples, thermo-mechanical coupling simulation calculations are performed, and the thermo-mechanical response data corresponding to each set of support structures is obtained to construct an initial dataset with support structure parameters as input features and thermo-mechanical response data as output labels.

[0015] A laser additive manufacturing substrate based on machine learning, manufactured by any of the above manufacturing methods, includes a substrate body and a support structure disposed on the substrate substrate.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a machine learning-based method for manufacturing a laser additive manufacturing substrate. This method designs a support structure on the substrate, abandoning traditional trial-and-error methods or experience-based structural design methods during the manufacturing process. Through a machine learning model integrating physical laws, it can automatically and efficiently search for the optimal structure that satisfies multiple objectives, greatly shortening the R&D cycle. This invention combines machine learning intelligent design with additive manufacturing to design a substrate capable of actively controlling the thermo-mechanical field. This substrate can not only effectively release stress and suppress cracking through fine internal structural design, but also achieve the optimal forming thermal environment through customized thermal management functions, promoting the diffusion and precipitation of strengthening phase-forming elements, thereby simultaneously improving the overall stability of component forming and its final service performance.

[0017] The second aspect of this invention discloses a laser additive manufacturing substrate based on machine learning. This substrate incorporates a support structure on a traditional substrate body, integrating the substrate body and the support structure as a standardized component that can be directly installed into existing laser directional energy deposition equipment without requiring modification to the host machine. The prepared substrate, in addition to its traditional support function, integrates stress suppression and heat flow management. Simultaneously, the support structure can be customized to meet the specific geometric, material, and performance requirements of different components, achieving precise control of thermo-mechanical-structural aspects. This substrate improves manufacturing results by regulating the physical field, making it suitable for all laser additive manufacturing material systems sensitive to thermal stress or requiring specific thermal history control, particularly for easily cracked, difficult-to-weld alloys and precipitation-strengthened alloys, exhibiting broad technical applicability. This substrate can synergistically regulate the stress field and thermal history during the printing process, thereby suppressing stress generation and optimizing the microstructure in situ during printing. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a laser-directed energy deposition substrate and its manufacturing method based on machine learning-driven design, according to the present invention.

[0019] Figure 2 A schematic diagram illustrating the process of designing support structures on a machine learning-driven substrate.

[0020] Figure 3 Schematic diagrams of two substrate structures designed for machine learning.

[0021] Figure 4 Finite element models are provided for a standard substrate and two different substrates.

[0022] Figure 5 This is a comparison of the stress field results for a standard substrate and a substrate with a support structure.

[0023] Figure 6 This is a comparison chart of the temperature history of a standard substrate and a substrate with a support structure. Detailed Implementation

[0024] Hereinafter, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature.

[0025] The method provided in this application can be applied to mobile phones, tablets, wearable devices, in-vehicle devices, augmented reality (AR) / virtual reality (VR) devices, laptops, and ultra-mobile personal computers. In this application, the specific type of terminal device is not limited to terminal devices such as mobile personal computers (UMPCs), netbooks, and personal digital assistants (PDAs).

[0026] It should be noted that the terms "first," "second," etc., used in the specification and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] As the background section points out, laser-directed energy deposition (LDED) of difficult-to-weld alloys faces severe bottlenecks of stress cracking and suboptimal microstructure. The root cause lies in the combined effect of the inherent high temperature gradient and the extremely strong mechanical constraints of the substrate. On the one hand, the extremely high cooling rate significantly inhibits the diffusion and precipitation of strengthening phase-forming elements; on the other hand, the coupling of enormous thermal stress and the strong mechanical constraints of the substrate easily leads to component cracking. Traditional methods struggle to address this contradiction in a coordinated manner.

[0028] To address the aforementioned problems, this invention provides a machine learning-driven laser-directed energy deposition (LDED) substrate and its manufacturing method. This method, based on machine learning technology, designs a multi-scale complex support structure on the substrate body and manufactures it using laser powder bed melting technology. Finally, the support structure and the substrate body are used as the substrate in laser-directed energy deposition. The core idea and technical problem solved by this invention is as follows: Traditionally, the substrate in laser-directed energy deposition only serves as a support platform, and its strong mechanical constraints are a significant factor inducing residual stress in the component. Adding a complex support structure to the substrate, using both the substrate and the support structure as active thermo-mechanical management components, not only reduces mechanical constraints but also actively regulates heat flow. This achieves low-stress forming while creating a more favorable thermal environment for printing, meeting the requirements for tissue growth. Based on this principle, machine learning methods are used to finely design the support structure, which is then manufactured using laser powder bed melting technology and applied to the laser-directed energy deposition process, ultimately achieving thermo-mechanical synergistic control during laser-directed energy deposition.

[0029] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] S1. Obtain the 3D CAD model of the target component to be formed, and extract the intrinsic parameters of the component, including geometric dimensions, cross-sectional features, height, wall thickness, and spatial orientation information. Determine the material information of the target component, including the material system and material thermal processing characteristics. The material system includes, but is not limited to, commonly used materials for laser-directed energy deposition such as nickel-based superalloys, titanium alloys, and aluminum alloys. Define the thermal processing characteristics of the selected material, specifically including physical property parameters, mechanical property parameters, phase transformation temperature, and strengthening phase precipitation temperature range. Physical property parameters include thermal conductivity, coefficient of thermal expansion, and elastic modulus, while mechanical property parameters include tensile strength and yield strength. Based on the component's service performance requirements and manufacturing process limitations, establish three multi-objective constraints to construct the subsequent comprehensive loss function.

[0031] 1) Residual stress loss item: Based on the service safety requirements of the component, the specific upper limit of the maximum residual stress after the component is formed is quantitatively defined; or a relative constraint standard is set, requiring the residual stress of the substrate forming component of the present invention to be reduced by a preset ratio (such as 50%, 70%, etc.) compared with the traditional solid substrate forming component. This ratio needs to be determined in combination with the cracking sensitivity of the component material. 2) Thermal history constraint: To meet the microstructure control requirements of component materials, the residence time of the deposited layer in specific temperature ranges such as the precipitation temperature range of the strengthening phase and the phase transformation temperature range is limited, or a relative constraint standard is set, requiring that the residence time be extended by a preset multiple (such as 2 times, 3 times, etc.) compared with the traditional substrate forming, so as to ensure that the strengthening phase diffuses and precipitates sufficiently and optimizes the microstructure of the component. 3) Manufacturability constraints: These constraints limit the dimensions of independent structures within the designed support structure. Specifically, each independent structure is the smallest geometric functional forming unit within the support structure that cannot be further divided, can be individually controlled in terms of size and process constraints, and directly determines the laser powder bed melting forming effect. This includes, but is not limited to, parameters such as rod diameter, wall thickness, hole diameter, and gap width. At the same time, process constraints such as suspension angle and structural continuity formability are specified to ensure that the designed support structure can be stably and precisely formed using laser powder bed melting equipment without the risk of processing failure.

[0032] S2, Construct the initial dataset, which includes the target constraints; and the dataset supporting the structural parameters and thermal response.

[0033] Based on the three types of constraints defined in S1, samples are combined within their reasonable value ranges to form multiple sets of differentiated target constraints, such as (50% reduction in residual stress, 2-fold increase in thermal history), (60% reduction in residual stress, 3-fold increase in thermal history), and (70% reduction in residual stress, 2.5-fold increase in thermal history). These target constraints will serve as "target condition labels" for subsequent data samples to guide the diversified design of the support structure.

[0034] S201, based on finite element thermal analysis Force coupling simulation software, such as ABAQUS and ANSYS, is used to establish an integrated finite element simulation model of the substrate body, intelligent support structure and target component. Mesh generation is performed according to the geometric characteristics of the component and the design requirements of the support structure. Mesh refinement is performed on the root of the component and the key stress / heat transfer area of ​​the support structure to ensure the accuracy of simulation calculation. S202. Based on the target setting, select structures such as lattice structure, TPMS structure (porous medium topology optimization), and origami structure as basic topology types, and set different structural parameters such as cell size, wall thickness, porosity, and gradient distribution mode to ensure that each set of samples meets the manufacturability constraints in its corresponding target constraints, forming multiple sets of support structure parameter samples. Through parameter combination, multiple sets of support structure parameter samples with differentiated characteristics are formed. The number of samples needs to cover different topology types and parameter ranges to ensure the diversity of the dataset.

[0035] S203 couples the laser directional energy deposition process parameters, including laser power, scanning speed, powder feeding rate, layer thickness, etc., into the simulation model to carry out multi-cycle, full-process thermo-mechanical coupling simulation calculations and obtain thermo-mechanical response data such as component temperature field evolution, stress field distribution, deformation, and residence time in key temperature ranges for each support structure. S204, for each set of supporting structure parameter samples, accurately extracts the core thermodynamic response data of the components from the simulation calculation results, including but not limited to the dynamic evolution curve of the temperature field, the spatial distribution cloud map of the stress field, the overall deformation of the component, the deformation trend of key parts, the residence time of the deposited layer in a specific temperature range, and the peak value and distribution location of residual stress. Finally, using multi-objective constraints, the topology type of the supporting structure, the intrinsic parameters of the components, and the material system as input features, and using the core thermodynamic response data such as the residual stress, temperature history, and deformation of the components as output labels, a standardized initial dataset of supporting structure parameters and component thermodynamic responses is constructed. The dataset is then cleaned and normalized to remove invalid data and ensure data quality. The thermodynamic response data specifically includes the temperature field, temperature history, and stress field.

[0036] S3, Construct and train a generative adversarial network machine learning model that incorporates physical information. This process constructs a Generative Adversarial Network (GAN) architecture that integrates Physical Information Neural Network (PINN) as the core machine learning model. The model is trained and iteratively optimized using an initial dataset, enabling it to accurately predict the thermal performance of support structures and autonomously generate optimal support structure design schemes. The specific implementation is as follows: The S301 adversarial network machine learning model consists of two core modules: a generator and a discriminator. The generator uses a multilayer perceptron neural network (MLP) architecture to generate three-dimensional parametric design schemes for the support structure. The discriminator is composed of a physical information neural network (PINN) and a performance evaluation module, which is used to predict the thermal performance and evaluate the multi-objective performance of the support structure schemes output by the generator.

[0037] During the model training phase, the input of the multilayer perceptron neural network is a random noise vector. During the model application phase, the input is the encoding vector of the intrinsic parameters, material information, and multi-objective constraints of the target component. The output is the three-dimensional support structure parameters, which include key geometric features such as the topology, feature dimensions, and spatial gradient distribution of the structure. The input of PINN is the three-dimensional support structure parameters output by the generator, and the output is the approximate temperature and stress field prediction results generated during the laser-directed energy deposition printing process. The performance evaluation module consists of a stress gap calculation unit, a thermal history gap calculation unit, and a loss value integration unit. Its input is the temperature and stress field data predicted by PINN and the three multi-objective constraints set in step one. The output is the comprehensive loss value. The comprehensive loss value is calculated through a comprehensive loss function. The comprehensive loss value uses the calculated stress, thermal history, and dimensions of the three-dimensional support structure as predicted values. The performance evaluation module quantitatively compares the predicted values ​​with the target values ​​of the three multi-objective constraints to calculate the residual stress loss value and the thermal history loss value, and integrates them into a comprehensive performance gap score, which serves as the basis for the discriminator to judge the feasibility of the structure.

[0038] Specifically, the comprehensive loss function encodes the three multi-objective constraints set in S1 into the model's comprehensive loss function. This function consists of three weighted components: a residual stress loss term, a thermal history loss term, and a manufacturability loss term. The residual stress loss term measures the quantitative difference between the model's predicted residual stress and the preset residual stress target; the thermal history loss term measures the quantitative difference between the model's predicted residence time in the key temperature range and the preset thermal history target; and the manufacturability loss term applies a penalty weight to design schemes that violate laser powder bed melting process constraints (such as feature size less than a threshold, overhang angle exceeding process limits, etc.). The training objective of the comprehensive loss function is to drive the model to autonomously generate support structure design schemes that simultaneously achieve low residual stress and optimal thermal history, while satisfying the hard manufacturability constraints.

[0039] S303, Embedded Physical Constraints: In the Physical Information Neural Network (PINN) of the discriminator, the heat conduction control equation and the linear elasticity equation are embedded as physical constraints in the training process. By constructing a physical constraint loss function, the prediction results of the model strictly follow the basic physical laws of heat conduction and mechanical deformation, avoiding the generation of support structure schemes that do not conform to physical principles. At the same time, the Physical Information Neural Network has the ability of a lightweight surrogate solver, which can quickly complete the approximate solution of the thermal performance of the support structure.

[0040] S304, Model Training and Iteration: First, the model is initially trained using the initial dataset constructed in step S2, enabling the generator and discriminator to establish basic generation and evaluation capabilities, respectively.

[0041] The generator's initial learning: Using the supporting structure parameters in the initial dataset as the training target, the generator is trained to learn the mapping from random noise vectors to real structural parameters. The training objective at this stage is to enable the generator to initially grasp the geometric characteristics, topological types, and parameter distribution patterns of various supporting structures. The discriminator's initial learning: Using the supporting structure parameters in the initial dataset as input and their corresponding thermal response data (stress field, temperature field, etc.) as output labels, supervised learning training is performed on the Physical Information Neural Network (PINN) part of the discriminator. The purpose of this step is to enable the discriminator to initially possess the ability to quickly predict thermal responses from structural parameters. Simultaneously, the performance evaluation module learns how to calculate the comprehensive loss value based on these prediction results. When the generator can generate reasonable structures and the discriminator can make preliminary and accurate predictions and evaluations of the thermal performance of these structures, the initial training of the model is considered complete. Based on this, the generator and discriminator undergo alternating optimization adversarial training to obtain the initial adversarial network machine learning model. This training phase consists of two layers: an inner layer of generative adversarial network (GAN) adversarial training and an outer layer of active learning iterative optimization.

[0042] GAN Adversarial Training: Given a training dataset (initially the initial dataset constructed using S2, later augmented datasets), the model is trained through alternating optimization of the generator and discriminator. First, the generator is fixed, and the discriminator is trained. The candidate structures currently generated by the generator are input into the discriminator. The discriminator updates its network parameters using backpropagation based on the temperature and stress fields predicted by PINN and the comprehensive loss value calculated by the performance evaluation module using a comprehensive loss function, thereby improving the accuracy of its structural performance evaluation. This step is repeated multiple times until the discriminator's evaluation capability stabilizes under the current generator. Then, the discriminator is fixed again, and the generator is trained. The comprehensive loss value output by the discriminator is used as a feedback signal backpropagated to the generator. The generator adjusts its network parameters accordingly, enabling its next generated candidate structures to achieve lower comprehensive loss values. This step is also repeated multiple times. After repeating the above two steps a sufficient number of times, until the comprehensive loss value of the candidate structures generated by the generator does not fall below the historical best value for P consecutive iterations (e.g., P=20), it is determined that the loss value has stopped decreasing, training stops, and the discriminator's evaluation results tend to stabilize, meaning the variance of the comprehensive loss value output by the discriminator is less than a preset threshold (e.g., 1%). This completes one round of inner-layer GAN training. At this point, the generator has learned to generate better support structures that satisfy multi-objective constraints on the current dataset.

[0043] To further improve the model's generalization ability and generation quality, active learning and iterative optimization are initiated based on the above. The specific process is as follows: (1) Generate candidate structures: Use the generator that has completed the inner layer training to generate multiple new candidate design schemes for support structures.

[0044] (2) Rapid evaluation and screening: The currently trained discriminator performs rapid thermodynamic performance evaluation and manufacturability verification on all candidate structures. Based on the comprehensive loss value, the two optimal candidate structures with the lowest comprehensive loss value and the candidate structure with the highest discriminator prediction uncertainty are selected. The prediction uncertainty is calculated by enabling Monte Carlo Dropout during the discriminator prediction stage to calculate the variance of the predicted temperature field and stress field. The larger the variance, the higher the prediction uncertainty.

[0045] (3) High-fidelity simulation verification: The selected candidate structures are imported into a high-fidelity finite element thermo-mechanical coupling simulation model, and the entire process is accurately calculated to obtain their real stress field, temperature field and thermo-mechanical response data.

[0046] (4) Model retraining: Add the real data obtained from high-fidelity simulation to the current dataset to form an augmented database. Then, based on this augmented database, restart a new round of inner GAN adversarial training (i.e., repeat the above inner loop) to enable the model to learn a more accurate structure-performance mapping relationship on the new data.

[0047] (5) Cyclic Convergence: Repeat steps (1)-(4) multiple times (e.g., three times) until the following two convergence conditions are met, at which point the model training is considered mature and converged: ① The comprehensive loss value (i.e., the difference between the candidate structure generated by the model and the multi-objective constraint conditions set in S1) obtained by the discriminator tends to be stable and no longer decreases significantly with iteration, indicating that the model can stably generate near-optimal structures that meet the target requirements; ② The error between the discriminator's evaluation results of the candidate structure (comprehensive loss value, temperature field, stress field, etc.) and the high-fidelity simulation verification results tends to be stable and less than the preset threshold (e.g., error <5%), indicating that the discriminator can accurately predict the actual thermal response of the structure, and its accuracy as a "surrogate solver" is reliable enough. When the above two conditions are met simultaneously, the model training is considered mature and converged.

[0048] The model employs a generative adversarial network (GAN) architecture, the core of which lies in the co-evolution of the generator and discriminator through alternating optimization and mutual promotion. In each training iteration, the generator parameters are first fixed, and the candidate structures currently generated by the generator are input into the discriminator. The discriminator updates its own network parameters through backpropagation based on the prediction results of the physical information neural network and the comprehensive loss value calculated by the performance evaluation module, improving the accuracy of the evaluation of structural performance. Subsequently, the discriminator parameters are fixed, and the comprehensive loss value output by the discriminator is used as a feedback signal backpropagated to the generator. The generator adjusts its own network parameters accordingly, so that the structures it generates next can obtain lower comprehensive loss values. The generator and discriminator continuously evolve in this process of alternating optimization and mutual driving, ultimately enabling the generator to learn to achieve multi-objective optimization of low residual stress and specific high-temperature dwell time while meeting manufacturability constraints. The optimization direction of the model is defined by a comprehensive loss function, which consists of three terms: a residual stress loss term, which measures the difference between the predicted peak residual stress and the target residual stress value; a thermal history loss term, which measures the difference between the predicted high-temperature dwell time and the target time; and a manufacturability loss term, which penalizes designs that violate laser powder bed melting process constraints (such as feature size less than 0.3 mm, overhang angle exceeding 45°, etc.). The final comprehensive loss value is obtained by weighted summing of the three losses.

[0049] S4, input target parameters, including intrinsic parameters of the target component to be formed, material system and material thermal processing characteristics parameters, and generate the optimal substrate CAD model from the well-trained model.

[0050] S401 inputs the 3D CAD model of the target component, key material processing characteristics, and quantified multi-objective constraint requirements extracted from S1 into a well-trained machine learning model in a standardized format.

[0051] S402: After receiving the input parameters, the model performs rapid inference and optimization based on the learned structure-performance mapping relationship. The generator produces multiple candidate 3D support structure design schemes according to the input conditions. The discriminator performs rapid performance evaluation and manufacturability verification on these candidate schemes, and determines the optimal 3D support structure design scheme with the lowest comprehensive loss value based on the overall loss value. Specifically, it outputs the structural, geometric parameters and material distribution information of the 3D support structure.

[0052] S403, the model finally outputs a complete three-dimensional parametric model of the optimal support structure, which is then integrated with the three-dimensional model of the substrate body to form a complete substrate CAD model that can be directly used for manufacturing. At the same time, it outputs the key design parameters and process hints of the support structure.

[0053] S5 uses laser powder bed melting technology to manufacture the substrate. The substrate CAD model obtained in S4 is imported into the dedicated control software of the laser powder bed melting equipment for subsequent manufacturing preparation and precision printing. Due to the complex dimensions of the three-dimensional support structure, the following steps are adopted, the specific implementation steps of which are as follows: S501, based on the processing precision of the laser powder bed melting equipment and the topological characteristics of the support structure, performs layer slicing and scanning path planning, optimizes the path for the small geometric units and dense areas of the support structure, and ensures forming precision and structural density; S502. Based on the functional requirements of the support structure (such as heat transfer performance, mechanical support performance, and metallurgical bonding performance with the substrate body / component), select metal powder that matches the substrate body material and meets the requirements of laser powder bed melting and forming, and clarify the key indicators such as particle size distribution, sphericity, and loose packing density of the powder. S503: Fix the substrate body onto the forming worktable of the laser powder bed melting equipment, start the equipment according to the slicing and path planning parameters, and complete the precision printing of the support structure; S504. To ensure a good metallurgical bond between the deposited layer and the substrate during subsequent laser directional energy deposition (DED) and to avoid defects such as poor bonding and delamination, a thin, dense structural layer is designed and printed on the upper surface of the support structure corresponding to the deposition area of ​​the component during laser powder bed fusion printing of the support structure. The thickness of this dense structural layer is in the range of 0.1mm to 0.8mm and needs to be determined in combination with the printing material and process parameters, taking into account both the metallurgical bonding effect and the thermal / mechanical control function.

[0054] S505 After printing, the substrate undergoes preliminary processing such as powder removal, cleaning, and surface pretreatment to remove residual powder and forming burrs. The forming quality of the support structure is checked to ensure that there are no manufacturing defects such as cracks, pores, or deformation. The support structure is then prepared on the substrate body.

[0055] S601, applying a substrate to a laser-directed energy deposition modeling target component. The substrate with supporting structure, manufactured and qualified in S5, is installed onto the working platform of the laser-directed energy deposition equipment. After equipment debugging and process parameter setting, laser-directed energy deposition of the target component is carried out. The specific implementation is as follows: S601 precisely positions and securely fixes the substrate on the worktable of the laser directional energy deposition equipment, ensuring the substrate's levelness and coaxiality, and adjusting core components such as the laser head and powder feeder to ensure normal equipment operation; S602 selects appropriate laser-directed energy deposition core process parameters, including laser power, scanning speed, powder feeding rate, layer thickness, and scanning path, based on the material system and geometric characteristics of the target component, without the need for additional energy field auxiliary control devices; S603, the laser directional energy deposition equipment is started to deposit the material of the target component layer by layer on the surface of the dense layer of the support structure of the substrate. Since the substrate designed in this invention has a customized variable structure and thermal conductivity, it can autonomously achieve coordinated control of thermal and mechanical fields during the deposition process. Without the need for additional energy field (such as ultrasound, electromagnetic) assistance, it can effectively reduce the residual stress of the component and optimize the temperature history. S604: During the deposition process, the online monitoring system of the equipment can be used to observe the state of the molten pool and the quality of the deposited layer in real time, and to deal with any sudden process problems in a timely manner, so as to ensure the stability of the component forming process.

[0056] S7, Component Separation and Substrate Post-Processing After the target component is formed by laser-directed energy deposition, the substrate and component are cooled to room temperature before the component is separated and the substrate is further processed. The specific implementation is as follows: S701 uses precision machining methods such as wire cutting and electrical discharge machining to separate the formed target component from the support structure of the substrate. During the separation process, protective measures must be taken to avoid damage to the surface quality and dimensional accuracy of the component. S702 involves subsequent surface processing and heat treatment of the separated target components to further optimize the dimensional accuracy and microstructure properties of the components and meet actual service requirements. S703. Based on the forming quality, material properties, and actual production needs of the support structure, the substrate is subjected to differentiated post-processing: if the support structure has no obvious deformation or damage, and its design is applicable to the printing of similar components, the substrate can be cleaned, repaired, and reused; if the support structure is a customized design for specific irregular components, or if there is irreparable deformation or damage after forming, the substrate is treated as a single-use substrate and the substrate body is recycled and reused.

[0057] The above-described implementation steps of the present invention can be flexibly adapted to target components with different materials and geometric features. Through intelligent design of machine learning models and precise implementation of laser additive manufacturing technology, a substrate with active thermal-mechanical field collaborative management function can be created, effectively solving the technical bottlenecks of stress cracking and unsatisfactory microstructure in traditional laser directional energy deposition processes, and significantly improving the forming quality and service performance of components.

[0058] The following is a further explanation with reference to specific embodiments. This case uses IN738LC, a nickel-based superalloy that is highly prone to cracking in additive manufacturing, as the printing material. The target component is a representative rectangular frame structure to clearly verify the universality and effectiveness of this method. The specific steps are as follows: S1. First, define the component model and expected goals. Target component dimensions: A rectangular frame with a length and width of 50mm, a thickness of 5mm, and a height of 30mm. Multiple objective requirements: 1) The designed substrate should reduce the maximum residual stress value of the component by 70% compared to a standard substrate. This objective is verified through finite element simulation; 2) The residence time of any region of the component on the designed substrate within the temperature range of 800℃-1100℃ should be at least 3 times longer than that on a standard substrate; 3) All quantifiable smallest geometric unit dimensions (aperture width, wall thickness, etc.) in this support structure must be greater than 0.3mm.

[0059] S2. Eight different support structures were designed, including three body-centered cubic lattice structures with different cell sizes of 3mm, 4mm, and 5mm, and all with a member diameter of 0.5mm; one face-centered cubic lattice structure with a cell size of 4mm and a member diameter of 0.5mm; one octagonal truss lattice structure with a cell size of 5mm and a member diameter of 0.6mm; one Gyroid TPMS structure with a period of 4mm and a wall thickness of 0.4mm; one Diamond TPMS structure with a period of 5mm and a wall thickness of 0.5mm; and one Miura-ori origami structure with a unit size of 6mm and a thickness of 0.5mm. The printing of these different support structures was simulated using ABAQUS finite element simulation software, with rectangular frames used for printing. After printing, the residual stress distribution at different locations of the components and the thermal history data of typical areas were recorded, forming a "Support Structure Parameters-Thermal Response" database containing eight sets of data.

[0060] S3, the machine learning model, adopts a generative adversarial network-based architecture, and its complete workflow is as follows: Figure 2As shown, the generator of this model employs a multilayer perceptron neural network (MLP). During the training phase, the input is a set of random noise vectors, and the output is a complete set of three-dimensional support structure parameters, specifically including key geometric features such as the structure's topology, feature dimensions, and spatial gradient distribution. The model's discriminator consists of a Physical Information Neural Network (PINN) and a performance evaluation module. The PINN is pre-trained, and during its training, the heat conduction control equation and the linear elasticity equation are embedded in its loss function as physical constraints, enabling it to act as a lightweight surrogate solver to quickly predict the approximate temperature and stress fields generated by the support structure during the printing process. The performance evaluation module receives the prediction results from the PINN and, based on the set multi-objective requirements, calculates a comprehensive performance gap score using a weighted comprehensive loss function composed of residual stress loss, thermal history loss, and manufacturability loss terms. This score serves as the basis for the discriminator to judge the feasibility of the structure.

[0061] The model training process is as follows Figure 2 As shown, the generator and discriminator are first trained using the aforementioned eight sets of initial data to establish a preliminary structure-performance mapping relationship. Then, an active learning iterative optimization loop is initiated. In the first round, the generator produces 30 new candidate support structures; the discriminator quickly evaluates these 30 structures, sorts them according to their comprehensive loss values, and selects the two structures with the lowest comprehensive loss and the structure with the highest discriminator prediction uncertainty. The three selected structures are then imported into a high-fidelity finite element simulation model for full-process accurate calculations to obtain their true stress and temperature field data. These three sets of new data are added to the initial database, forming an enhanced database containing 11 sets of data, which is used to train the model in the second round. This process is repeated three times. After three iterations, the candidate structures generated by the model show consistent performance in discriminator evaluation and high-fidelity simulation verification, and the variance of the comprehensive loss value is less than the preset stability threshold (set to 0.005 in this embodiment), indicating that the model has converged and matured.

[0062] The dimensions of the target component's rectangular frame, the material system of IN738LC, and the corresponding target requirements are input into the mature model. The model automatically calculates and optimizes, ultimately outputting two support structures that meet the requirements. Figure 3 To verify the practical feasibility of the output structure, two substrate structure models were imported into finite element simulation software for simulation calculations. Figure 4 Mesh models of three different substrates are provided. The simulation results for the three substrates are compared. Figure 5The stress field simulation results show that the residual stress of the ordinary substrate is extremely high, mainly concentrated at the root of the component, with a maximum residual stress value of 1995.8 MPa. The residual stress values ​​of the other two types of substrates are smaller, also concentrated at the root of the component. Among them, the maximum residual stress value of the component on the columnar support structure substrate is 440.12 MPa, and that on the lattice support structure is 474.37 MPa, with a reduction of more than 70% in both cases.

[0063] Figure 6 By comparing the temperature history of the midpoint of the rectangular frame sidewall on the three substrates, it can be seen that the average temperature of this point on the two substrate components designed by machine learning is significantly higher than that on the ordinary substrate. The dwell time of this point at 800℃-1100℃ on the ordinary substrate is 23s, on the columnar support structure substrate it is 162s, and on the lattice support structure substrate it is 151s, all of which meet the preset temperature history requirements.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for manufacturing a laser additive manufacturing substrate based on machine learning, characterized in that, Includes the following steps: S1, obtain the intrinsic parameters, material system, and material thermal processing characteristics of the target component to be formed; S2, input the intrinsic parameters, material system and thermal processing characteristic parameters into the pre-trained model, and output them to the adversarial network machine learning model to obtain the three-dimensional support structure; The adversarial network machine learning model includes a generator and a discriminator, and the training process of the adversarial network is as follows: S201, The adversarial network is initially trained using the initial dataset to obtain the initial adversarial network machine learning model; S202, the generator optimization and discriminator optimization are alternately performed on the initial adversarial network machine learning model. In the discriminator optimization stage, the generator parameters are fixed, and the three-dimensional support structure output by the generator is input into the discriminator. The discriminator predicts the temperature field and stress field, and the discriminator network parameters are updated through backpropagation by combining the comprehensive loss value output by the discriminator. During the generator optimization phase, the discriminator network parameters are fixed, and the overall loss value output by the discriminator is used as a feedback signal to propagate back to the generator to adjust the generator network parameters. The generator optimization and discriminator optimization are repeated alternately until the overall loss value no longer decreases significantly and the discriminator evaluation result tends to stabilize. S203: Use the current generator to generate multiple candidate structures, the discriminator calculates the comprehensive loss value, selects the two optimal candidate structures with the lowest comprehensive loss value and the candidate structure with the highest prediction uncertainty, and obtains the real stress field, temperature field and thermal response data through high-fidelity finite element thermo-mechanical coupling simulation. Add the verification data to the initial dataset to form an enhanced database, and re-execute 202 based on the enhanced database. S204, repeat steps 202 and 203 until the set conditions are met; S3 involves printing a three-dimensional support structure onto a substrate by laser powder bed melting, and then fabricating the target component by laser directional energy deposition onto the three-dimensional support structure.

2. The method for manufacturing a laser additive manufacturing substrate based on machine learning according to claim 1, characterized in that, In S1, the intrinsic parameters include geometric dimensions, cross-sectional features, height, wall thickness, and spatial orientation information; the material thermal processing characteristic parameters include physical property parameters, mechanical property parameters, phase transformation temperature, and strengthening phase precipitation temperature range; and based on the service performance requirements and manufacturing process limitations of the target component, multi-objective constraint conditions are established, including residual stress loss terms, thermal history constraint terms, and manufacturability constraint terms.

3. The method for manufacturing a laser additive manufacturing substrate based on machine learning according to claim 1, characterized in that, In S201, the generator adopts a multilayer perceptron neural network architecture, with random noise vector as input and three-dimensional support structure parameters as output. The three-dimensional support structure parameters include topology type, feature size, and spatial gradient distribution. The discriminator integrates a physical information neural network and a performance evaluation module. The physical information neural network takes the three-dimensional support structure parameters as input and outputs predicted temperature and stress fields. The performance evaluation module is used to calculate the difference between the multi-objective constraints and the comprehensive performance value.

4. The method for manufacturing a laser additive manufacturing substrate based on machine learning according to claim 1, characterized in that, In S202, the comprehensive loss value is composed of a weighted average of residual stress loss, thermal history loss, and manufacturability loss; the heat conduction control equation and the linear elasticity equation are embedded in the physical information neural network as physical constraints.

5. The method for manufacturing a laser additive manufacturing substrate based on machine learning according to claim 1, characterized in that, In S203, the step of selecting the two optimal candidate structures with the lowest comprehensive loss value and the candidate structure with the highest prediction uncertainty specifically includes: the discriminator calculates and sorts the comprehensive loss value of all candidate structures, and determines the two optimal candidate structures with the lowest comprehensive loss value and the candidate structure with the highest prediction uncertainty of the discriminator based on the sorting of comprehensive loss values.

6. The method for manufacturing a laser additive manufacturing substrate based on machine learning according to claim 1, characterized in that, In S204, the set conditions are: the comprehensive loss value of the generated structure does not fall below the historical best record value for P consecutive iterations, and the error between the discriminator prediction result and the high-fidelity simulation verification result is less than a preset threshold, and the adversarial network machine learning model is mature and converged.

7. The method for manufacturing a laser additive manufacturing substrate based on machine learning according to claim 1, characterized in that, In S3, the three-dimensional parametric model of the three-dimensional support structure and the three-dimensional model of the substrate body are integrated and assembled to form a substrate CAD model. The model is then imported into a laser powder bed melting device for layer slicing and scanning path planning, and the three-dimensional support structure is printed layer by layer on the substrate body.

8. The method for manufacturing a laser additive manufacturing substrate based on machine learning according to claim 7, characterized in that, A dense structural layer with a thickness of 0.1 mm to 0.8 mm is printed on the upper surface of the three-dimensional support structure, and then the target component is printed on the dense structural layer.

9. The method for manufacturing a laser additive manufacturing substrate based on machine learning according to claim 1, characterized in that, Before S2, the process also includes the step of constructing an initial dataset: based on finite element thermo-mechanical coupling simulation software, various basic topology types are selected as supporting structures, different structural parameters are set to form multiple sets of supporting structure parameter samples, thermo-mechanical coupling simulation calculations are performed, and the thermo-mechanical response data corresponding to each set of supporting structures is obtained to construct an initial dataset with supporting structure parameters as input features and thermo-mechanical response data as output labels.

10. A laser additive manufacturing substrate based on machine learning, manufactured by any one of claims 1-9, characterized in that, It includes the substrate body and the support structure disposed on the substrate.

Images