A method for predicting deformation and residual stress in high-temperature alloy additive manufacturing processes
By using multi-scale simulation and an improved PointNet++ model, combined with microscopic molecular dynamics and macroscopic thermo-mechanical coupling simulation, high-precision and rapid prediction of deformation and residual stress in the additive manufacturing process of high-temperature alloys was achieved. This solves the shortcomings of accuracy and efficiency in existing technologies and supports additive and subtractive manufacturing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2026-05-19
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies cannot effectively solve the problem of balancing the accuracy, efficiency and generalizability of deformation and residual stress in the additive manufacturing process of high-temperature alloys, resulting in large prediction errors and long time consumption, which cannot meet the needs of additive and subtractive manufacturing.
We employ multi-scale simulation modeling and an improved PointNet++ deep learning model, combining microscopic molecular dynamics simulation and macroscopic thermo-mechanical coupling simulation. Through user-defined subroutines, we embed microscopic feature parameters into the macroscopic model, construct an attention mechanism and a dual-output head deep learning model, and use a hybrid loss function for training to achieve simultaneous prediction of deformation and residual stress layer thickness.
It significantly improves prediction accuracy and efficiency, can quickly and accurately identify stress concentration areas, adapt to different process parameters and size specifications, support manufacturing with added and subtracted materials, reduce material waste and component scrap risk, and improve product quality and service safety.
Smart Images
Figure CN122494083A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of additive manufacturing technology, and in particular to a method for predicting deformation and residual stress in the additive manufacturing process of high-temperature alloys. Background Technology
[0002] High-temperature alloys (such as GH4169) are widely used in the manufacture of key components in critical fields such as aerospace and energy due to their excellent high-temperature strength, oxidation resistance, and hot corrosion resistance. Additive manufacturing technologies such as laser-directed energy deposition (L-DED) provide a new approach to achieve integrated, low-cost forming of complex high-temperature alloy components. However, the drastic temperature gradients and repeated melting and solidification during additive manufacturing can easily generate significant deformation and residual stress within the component. The thickness of the residual stress layer is a core indicator for determining the allowance for subsequent subtractive processing; inaccurate prediction of this layer can directly lead to dimensional deviations, cracking, or even scrapping of the component, severely restricting the development and application of additive and subtractive manufacturing technologies.
[0003] Current techniques for predicting deformation and residual stress in high-temperature alloy additive manufacturing are mainly divided into three categories, but all have significant shortcomings: The first category is macroscopic thermo-mechanical coupling simulation technology. This technology, based on the finite element method, solves for thermal stress and deformation by constructing geometric models of components, defining temperature-dependent material properties, and simulating thermal processes using a moving heat source. However, this technology lacks support from microscopic mechanisms, relying on empirically derived material parameters and failing to capture microscopic processes such as atomic diffusion and phase nucleation, leading to significant prediction errors, especially in areas of stress concentration. Furthermore, simulation of a single set of process parameters can take several hours, failing to meet the need for rapid optimization of multiple parameters.
[0004] The second category is microscopic molecular dynamics simulation technology. This technology focuses on the atomic scale, simulating interatomic interactions through embedded atom method (EAM) equipotential functions to reveal the microscopic mechanisms in additive manufacturing processes. However, it can only simulate nanoscale atomic systems, exhibiting a significant scale gap and cannot be directly extended to part-level predictions. Furthermore, a single simulation of a million-atom-scale model takes several days, limiting its engineering application value.
[0005] The third category is purely data-driven deep learning prediction techniques. This technique is based on point cloud processing models and achieves end-to-end prediction of deformation and stress through training on a large amount of data. However, this method only relies on statistical correlation of data and does not incorporate the physical mechanisms of additive manufacturing, resulting in poor model generalization and failure to enhance features for stress concentration areas, making it difficult to meet high prediction accuracy requirements.
[0006] In summary, existing technologies have failed to effectively address the balance between the accuracy of physical mechanisms, prediction efficiency, and generalizability, and cannot provide high-precision, high-speed, and highly adaptable predictive support for the additive and subtractive manufacturing of high-temperature alloys. Summary of the Invention
[0007] The purpose of this invention is to provide a method for predicting deformation and residual stress in the additive manufacturing process of high-temperature alloys, so as to achieve a deep integration of physical mechanisms and data-driven approaches, and significantly improve prediction efficiency while ensuring prediction accuracy.
[0008] To achieve the above objectives, this invention provides a method for predicting deformation and residual stress in the additive manufacturing process of high-temperature alloys, comprising the following steps: S1. Multi-scale simulation modeling: Conduct microscopic molecular dynamics simulation and macroscopic thermo-mechanical coupling simulation of the high-temperature alloy additive manufacturing process, and realize the embedding of microscopic feature parameters into the macroscopic model and the macroscopic-microscopic bidirectional data feedback of stress concentration areas through user-defined subroutines. S2. Dataset Construction: Based on the multi-scale simulation, collect component point cloud data, deformation data and residual stress layer thickness data under different combinations of process parameters, and form a training sample set that integrates process parameters and point cloud geometric features after preprocessing. S3. Model Construction and Training: Construct an improved PointNet++ multi-task deep learning model that includes an attention mechanism and dual output heads. Train the model using the training sample set and the hybrid loss function. The dual output heads are used to output the deformation prediction result and the residual stress layer thickness prediction result, respectively. S4. Fast prediction: Input the process parameters of the working condition to be predicted and the component point cloud data into the trained improved PointNet++ model to obtain the prediction results of deformation and residual stress layer thickness.
[0009] Preferably, step S1 specifically includes: S1.1 Microscopic molecular dynamics simulation: Construct a multi-component atomic system of the target high-temperature alloy, select the matching multi-component embedded atomic normal potential function, simulate the solidification process of the molten pool, and extract microscopic characteristic parameters including atomic diffusion coefficient, strengthening phase nucleation work, and grain boundary activation energy. S1.2 Macroscopic thermo-mechanical coupling simulation: Construct a geometric model of the target component, define temperature-dependent material parameters, use a moving heat source model, and solve to obtain macroscopic deformation cloud map, residual stress distribution and residual stress layer thickness; S1.3 Multi-scale coupling: The microscopic feature parameters are embedded into the macroscopic simulation material property module through a user-defined material subroutine, and the temperature and stress output by the macroscopic simulation are used as the boundary conditions of the microscopic model. The phase transformation strain output by the microscopic model is fed back to the macroscopic stress calculation to correct the stress distribution in the stress concentration area.
[0010] Preferably, step S2 specifically includes: S2.1 Process parameter design: Within the process parameter space of laser power, scanning speed, and powder feeding rate, multiple sets of laser additive manufacturing process parameter combinations are formed by combining them at certain intervals. S2.2 Data Acquisition: Perform multi-scale simulation of step S1 for each combination of process parameters, collect point cloud data in layers according to the height direction of the component, record the corresponding deformation data and residual stress layer thickness data, and form an initial sample set; S2.3 Data Preprocessing: Normalize the process parameters; adopt a point-level prompting mechanism, based on the neighborhood geometric features of defect points, use the K-nearest neighbor algorithm to fill in missing points and correct noise points; stitch the normalized process parameters with the preprocessed point cloud geometric features to form a fused input feature set; S2.4 Data Augmentation and Partitioning: The fused input feature set is augmented and expanded by at least one of the following methods: random rotation, translation, scaling, and Gaussian noise addition, and then divided into training and testing sets according to the proportion.
[0011] Preferably, the improved PointNet++ model described in S3 includes: Input layer: Used to receive fused features that combine normalized process parameters and point cloud geometric features. The number of point cloud samples is preset to 1024. Multi-scale sampling grouping module: adopts a three-level sampling strategy, with each group extracting local features through PointNet and outputting features of different dimensions; Attention mechanism module: Calculates the attention weight of each point, strengthens the features of high-weight points based on the correlation between point cloud curvature and residual stress value, and outputs global features; Fully connected layer: Composed of fully connected layers with progressively decreasing dimensions, using ReLU as the activation function, and introducing Dropout layers between layers to prevent overfitting, performing non-linear transformation and dimensionality compression on the weighted global features; Output layer: Contains two parallel fully connected layers as dual output heads, which are used to synchronously output the deformation vector and the predicted value of the residual stress layer thickness, respectively. The output activation function adopts Linear.
[0012] Preferably, the hybrid loss function in step S3, which integrates the mean absolute error and the root mean square error, specifically includes:
[0013] Where MAE is the mean absolute error, RMSE is the root mean square error, and α and β are preset weighting coefficients used to balance the accuracy and robustness of deformation prediction and residual stress layer thickness prediction.
[0014] Preferably, the model training in step S3 includes: using a deep learning framework, employing multi-GPU parallel training, setting the Adam optimizer, initial learning rate and decay strategy, batch size, number of training rounds, adopting an early stopping strategy to avoid overfitting, training until the loss function converges, and ensuring that the prediction accuracy of the test set meets the requirements.
[0015] Preferably, the thickness of the residual stress layer is the maximum dimension along the thickness direction of the region where the residual stress value reaches or exceeds 300 MPa after high-temperature alloy additive manufacturing; the stress concentration region is the region in the component where the stress value exceeds 1.5 times the average stress.
[0016] Preferably, the hardware preparation work before step S1 is as follows: configuring a hardware system including a simulation workstation, a GPU training server, a laser directional energy deposition additive manufacturing equipment, a 3D scanner and an X-ray diffraction residual stress meter, and software tools including finite element simulation software, molecular dynamics simulation software, deep learning framework and 3D modeling software.
[0017] Preferably, the target high-temperature alloy is GH4169 nickel-based high-temperature alloy, and GH4169 powder with a specific particle size range is used as the experimental material, and the substrate material is the same as that of GH4169 high-temperature alloy.
[0018] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for predicting deformation and residual stress in a high-temperature alloy additive manufacturing process as described in any one of claims 1 to 9.
[0019] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: First, the deep integration of physical mechanisms and data-driven approaches significantly improves prediction accuracy. This invention uses a user-defined material subroutine to precisely embed microscopic characteristic parameters such as atomic diffusion coefficients and phase-nuclear work extracted from molecular dynamics simulations into macroscopic thermo-mechanical coupling simulations. For stress concentration regions, a bidirectional coupling mechanism is designed to input macroscopic temperature and stress to the microscopic level and to feed microscopic phase transition strain back to the macroscopic level. This overcomes the limitations of traditional macroscopic simulations that rely on empirical parameters and cannot capture microscopic mechanisms, enabling prediction results to possess both microscopic physical reliability and macroscopic engineering applicability, significantly improving the accuracy of predicting additive manufacturing deformation and residual stress layer thickness.
[0020] Secondly, the multi-dimensional feature fusion and attention mechanism enhancement improve the model's generalization ability. This invention constructs a multi-dimensional dataset that integrates process parameters and point cloud geometric features. A point-level cueing mechanism is used to complete and correct defect point clouds, and various data augmentation methods are combined to expand the sample space, enabling the model to adapt to high-temperature alloy additive manufacturing scenarios with different sizes, specifications, and combinations of process parameters. Simultaneously, an attention mechanism module based on curvature-stress correlation enhances feature extraction from stress concentration areas, allowing the model to accurately identify high-risk regions and maintain stable predictive performance even when the component's geometry changes. This effectively solves the common problem of poor generalization in purely data-driven models.
[0021] Third, the improved PointNet++ model, designed in this invention, employs a dual-output head structure to simultaneously output deformation vectors and residual stress layer thickness prediction results, enabling multi-task parallel prediction. The hybrid loss function integrates mean absolute error and root mean square error, balancing the accuracy and robustness of multi-task prediction. After model training, the response time from input process parameters and point cloud data to output prediction results can be achieved in milliseconds under a single operating condition. This represents a significant improvement in prediction efficiency compared to the hours-long processing time of traditional macroscopic thermo-mechanical coupling simulations, meeting the needs of rapid optimization and online decision-making for process parameters in additive manufacturing.
[0022] Fourth, it has significant engineering application value, supporting the integrated manufacturing of additive and subtractive manufacturing processes. The residual stress layer thickness prediction results output by this invention can be directly used to guide the accurate planning of subtractive processing allowances, significantly improving the efficiency of residual stress removal and avoiding material waste, increased trial-and-error costs, and component scrapping risks caused by inaccurate predictions. By providing predictive support for integrated manufacturing of additive and subtractive manufacturing processes with high precision, speed, and strong adaptability, it effectively reduces manufacturing costs, improves product quality and service safety, and has important engineering significance for the intelligent manufacturing of key high-temperature alloy components in aerospace, energy, and power fields.
[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a schematic flowchart of an embodiment of the method for predicting deformation and residual stress in a high-temperature alloy additive manufacturing process according to the present invention. Figure 2 This is a schematic diagram of the cross-scale coupled simulation modeling and high-quality sample library construction process in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction and training process of a deep learning network that integrates model point clouds and process parameters in an embodiment of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0028] Example This embodiment uses GH4169 nickel-based superalloy as an example and takes a cylindrical part as an example to describe the method of the present invention in detail.
[0029] Before starting the prediction process, hardware preparation is required. This includes a simulation workstation, a GPU training server, an L-DED additive manufacturing system, a 3D scanner, and an XRD residual stress analyzer. The simulation workstation is equipped with a high-performance CPU, ample memory, and a professional GPU. The GPU training server supports multi-GPU parallel training to ensure efficient data processing and model training. Software tools encompass finite element simulation software, molecular dynamics simulation software, deep learning frameworks, programming languages, and 3D modeling software, meeting the full-process requirements of multi-scale simulation, dataset construction, model training, and validation. The experimental material uses GH4169 powder with a specific particle size range, and the substrate material is consistent with the high-temperature alloy to ensure material compatibility and forming quality during additive manufacturing.
[0030] like Figure 1 As shown, a method for predicting deformation and residual stress in a high-temperature alloy additive manufacturing process is described, with the following implementation steps: S1. Conduct multi-scale simulation modeling for high-temperature alloy additive manufacturing. For example... Figure 1 As shown, it specifically includes: S1.1, Microscopic Molecular Dynamics Simulation Stage: A multi-component atomic system of GH4169, containing elements such as nickel, chromium, iron, niobium, and titanium, was constructed using specialized software. The system comprised 500,000 atoms, with a simulation region size of 100×100×100 nm and periodic boundary conditions. Multi-component embedded atomic normal functions were selected to ensure that interatomic interactions matched the thermodynamic properties of GH4169. The solidification stage of the additive manufacturing process was simulated, controlling temperature and time steps. A specific ensemble was used to control temperature and the natural atomic motion. Microscopic parameters such as atomic diffusion coefficient, γ'' phase nuclear work, and grain boundary activation energy were extracted using post-processing tools.
[0031] S1.2 Macroscopic Thermo-Mechanical Coupling Simulation: A GH4169 cylindrical component with dimensions of φ100×200mm and a wall thickness of 3mm was constructed using 3D modeling software. The substrate dimensions were 200×200×20mm. After modeling, the component was imported into macroscopic simulation software. Temperature-dependent material parameters were defined, covering key properties such as elastic modulus, thermal conductivity, and coefficient of thermal expansion from room temperature to 900℃, while Poisson's ratio remained constant. A moving Gaussian heat source model was adopted, with initial laser power and spot radius set. The heat source characteristics were described using a specific power density distribution function. Boundary conditions were set, with fixed constraints at the bottom of the substrate and convective and radiative heat dissipation on the component surface. Argon was used as the protective gas, and the ambient temperature was controlled at room temperature. The mesh generation adopted a strategy combining global coarsening and local refinement. The global mesh size was 1mm, while the molten pool region and stress concentration region were locally refined to 0.05mm. Eight-node linear thermo-mechanical coupling elements were selected, and reasonable analysis step time steps and convergence criteria were set to obtain macroscopic deformation contour maps, residual stress distribution, and residual stress layer thickness.
[0032] S1.3 Multi-scale Coupling: Parameters extracted from microscopic molecular dynamics are embedded into the material property module of macroscopic simulation through user-defined material subroutines, correcting the γ'' phase precipitation kinetic model and replacing the original empirical parameters. For stress concentration regions, local bidirectional coupling is employed, using the temperature and stress output from the macroscopic simulation as boundary conditions for the microscopic model. The γ'' phase transformation strain output from the microscopic model is fed back to the macroscopic stress calculation, correcting the stress distribution in that region.
[0033] S2. Dataset Construction: Based on the multi-scale simulation, point cloud data, deformation data, and residual stress layer thickness data of components under different combinations of process parameters are collected. After preprocessing, a training sample set integrating process parameters and point cloud geometric features is formed; for example... Figure 3 As shown, it specifically includes: S2.1 Process parameter design: Within the process parameter space of laser power, scanning speed, and powder feeding rate, multiple sets of laser additive manufacturing process parameter combinations are formed by combining them at certain intervals. S2.2 Data Acquisition: Perform multi-scale simulation in step S1 for each combination of process parameters, and collect point cloud data in layers according to the height direction of the component. Each sample includes process parameter vector, point cloud geometric features, deformation amount and residual stress layer thickness, and a total of 100 valid samples are collected. S2.3 Data Preprocessing: Normalize the process parameters; adopt a point-level prompting mechanism, based on the neighborhood geometric features of defect points, use the K-nearest neighbor algorithm to fill in missing points and correct noise points; stitch the normalized process parameters with the preprocessed point cloud geometric features to form a fused input feature set; S2.4 Data Augmentation and Partitioning: The fused input feature set is augmented and expanded by at least one of the following methods: random rotation, translation, scaling, and Gaussian noise addition, to expand the dataset to 200 sets and divide it into training and test sets proportionally.
[0034] S3. Construct an improved PointNet++ multi-task deep learning model that includes an attention mechanism and dual output heads. Train the model using the aforementioned training sample set and a hybrid loss function. The dual output heads are used to output the deformation prediction result and the residual stress layer thickness prediction result, respectively. The improved PointNet++ model includes: Input layer: Used to receive fused features that combine normalized process parameters and point cloud geometric features. The number of point cloud samples is preset to 1024. Multi-scale sampling grouping module: adopts a three-level sampling strategy, with each group extracting local features through PointNet and outputting features of different dimensions; Attention mechanism module: Calculates the attention weight of each point, enhances the features of high-weight points based on the correlation between point cloud curvature and residual stress value, and outputs global features.
[0035] Fully connected layer: Composed of fully connected layers with progressively decreasing dimensions, using ReLU as the activation function, and introducing Dropout layers between layers to prevent overfitting, performing non-linear transformation and dimensionality compression on the weighted global features; Output layer: Contains two parallel fully connected layers as dual output heads, which are used to synchronously output the deformation vector and the predicted value of the residual stress layer thickness, respectively. The output activation function adopts Linear.
[0036] In step S3, the hybrid loss function integrates the mean absolute error and the root mean square error, specifically as follows:
[0037] Where MAE is the mean absolute error, RMSE is the root mean square error, and α and β are preset weighting coefficients used to balance the accuracy and robustness of deformation prediction and residual stress layer thickness prediction.
[0038] Based on a deep learning framework, multi-GPU parallel training was employed. The Adam optimizer was used with an initial learning rate of 0.001, and cosine annealing was used for learning rate decay. The batch size was set to 32, and the number of training epochs was 500. An early stopping strategy was used: training was stopped when the validation set loss no longer decreased within 20 epochs to ensure model convergence and avoid overfitting. S4. Rapid prediction and verification of deformation and residual stress layer.
[0039] Input the process parameters and point cloud data of the cylindrical part under the new working condition. After preprocessing, input the data into the trained model. Output the deformation and residual stress layer thickness, and record the prediction time. Print the GH4169 cylindrical part under this working condition. Measure the actual deformation and residual stress layer thickness using a 3D scanner and an XRD residual stress meter. Compare the prediction results with the experimental data to verify the prediction accuracy and speed, ensuring that the deformation prediction error does not exceed 10%, the residual stress layer thickness prediction error does not exceed 10%, and the prediction time does not exceed 20 seconds.
[0040] In summary, this embodiment successfully achieved rapid prediction of additive deformation and residual stress layer in high-temperature alloys based on molecular dynamics and deep learning. By employing a multi-scale coupling mechanism, parameters such as atomic diffusion coefficient and γ'' phase nuclear work extracted from microscopic molecular dynamics are accurately embedded into macroscopic thermo-mechanical coupling simulation, significantly improving the physical reliability of the prediction results. By improving the PointNet++ model and introducing a curvature-stress correlation attention mechanism, feature extraction of stress concentration regions is enhanced, achieving simultaneous and accurate prediction of deformation vectors and residual stress layer thickness. Dataset construction and hybrid loss function design ensure the model's generalization ability and prediction robustness. Experimental verification shows that the prediction results of this method agree well with the measured data, and the prediction efficiency is significantly improved compared to traditional macroscopic simulation, providing rapid and accurate prediction support for additive and subtractive manufacturing of high-temperature alloys.
[0041] The remaining technical features in the above embodiments can be flexibly selected by those skilled in the art to meet different specific practical needs according to actual circumstances. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims. In the above description, numerous specific details have been set forth to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to implement the present invention. In other instances, to avoid obscuring the present invention, well-known techniques, such as specific construction details, operating conditions, and other technical conditions, have not been specifically described.
Claims
1. A method of deformation and residual stress prediction for a high-temperature alloy additive manufacturing process, characterized by, The steps are as follows: S1. Multi-scale simulation modeling: Conduct microscopic molecular dynamics simulation and macroscopic thermo-mechanical coupling simulation of the high-temperature alloy additive manufacturing process, and realize the embedding of microscopic feature parameters into the macroscopic model and the macroscopic-microscopic bidirectional data feedback of stress concentration areas through user-defined subroutines. S2. Dataset Construction: Based on the multi-scale simulation, collect component point cloud data, deformation data and residual stress layer thickness data under different combinations of process parameters, and form a training sample set that integrates process parameters and point cloud geometric features after preprocessing. S3. Model Construction and Training: Construct an improved PointNet++ multi-task deep learning model that includes an attention mechanism and dual output heads. Train the model using the training sample set and the hybrid loss function. The dual output heads are used to output the deformation prediction result and the residual stress layer thickness prediction result, respectively. S4. Fast prediction: Input the process parameters of the working condition to be predicted and the component point cloud data into the trained improved PointNet++ model to obtain the prediction results of deformation and residual stress layer thickness.
2. A method of deformation and residual stress prediction for a high-temperature alloy additive manufacturing process according to claim 1, characterized in that, Step S1 specifically includes: S1.1 Microscopic molecular dynamics simulation: Construct a multi-component atomic system of the target high-temperature alloy, select the matching multi-component embedded atomic normal potential function, simulate the solidification process of the molten pool, and extract microscopic characteristic parameters including atomic diffusion coefficient, strengthening phase nucleation work, and grain boundary activation energy. S1.2 Macroscopic thermo-mechanical coupling simulation: Construct a geometric model of the target component, define temperature-dependent material parameters, use a moving heat source model, and solve to obtain macroscopic deformation cloud map, residual stress distribution and residual stress layer thickness; S1.3 Multi-scale coupling: The microscopic feature parameters are embedded into the macroscopic simulation material property module through a user-defined material subroutine, and the temperature and stress output by the macroscopic simulation are used as the boundary conditions of the microscopic model. The phase transformation strain output by the microscopic model is fed back to the macroscopic stress calculation to correct the stress distribution in the stress concentration area.
3. The method of claim 1, wherein, Step S2 specifically includes: S2.1 Process parameter design: Within the process parameter space of laser power, scanning speed, and powder feeding rate, multiple sets of laser additive manufacturing process parameter combinations are formed by combining them at certain intervals. S2.2 Data Acquisition: Perform multi-scale simulation of step S1 for each combination of process parameters, collect point cloud data in layers according to the height direction of the component, record the corresponding deformation data and residual stress layer thickness data, and form an initial sample set; S2.3 Data Preprocessing: Normalize the process parameters; adopt a point-level prompting mechanism, based on the neighborhood geometric features of defect points, use the K-nearest neighbor algorithm to fill in missing points and correct noise points; stitch the normalized process parameters with the preprocessed point cloud geometric features to form a fused input feature set; S2.4 Data Augmentation and Partitioning: The fused input feature set is augmented and expanded by at least one of the following methods: random rotation, translation, scaling, and Gaussian noise addition, and then divided into training and testing sets according to the proportion.
4. The method of claim 1, wherein: The improved PointNet++ model described in S3 includes: Input layer: Used to receive fused features that combine normalized process parameters and point cloud geometric features. The number of point cloud samples is preset to 1024. Multi-scale sampling grouping module: adopts a three-level sampling strategy, with each group extracting local features through PointNet and outputting features of different dimensions; Attention mechanism module: Calculates the attention weight of each point, strengthens the features of high-weight points based on the correlation between point cloud curvature and residual stress value, and outputs global features; Fully connected layer: Composed of fully connected layers with progressively decreasing dimensions, using ReLU as the activation function, and introducing Dropout layers between layers to prevent overfitting, performing non-linear transformation and dimensionality compression on the weighted global features; Output layer: Contains two parallel fully connected layers as dual output heads, which are used to synchronously output the deformation vector and the predicted value of the residual stress layer thickness, respectively. The output activation function adopts Linear.
5. The method of claim 1, wherein: The hybrid loss function described in step S3, which fuses the mean absolute error and the root mean square error, specifically involves: where MAE is the mean absolute error, RMSE is the root mean square error, and a and β are preset weight coefficients for balancing the accuracy and robustness of the deformation prediction and the residual stress layer thickness prediction.
6. The method of claim 1, wherein: The model training described in step S3 includes: training in parallel using multiple GPUs based on a deep learning framework, setting the Adam optimizer, initial learning rate and decay strategy, batch size, number of training rounds, adopting an early stopping strategy to avoid overfitting, training until the loss function converges, and ensuring that the prediction accuracy of the test set meets the requirements.
7. The method of claim 1, wherein: The residual stress layer thickness is the maximum dimension along the thickness direction of the region where the residual stress value reaches or exceeds 300 MPa after high-temperature alloy additive manufacturing; the stress concentration region is the region in the component where the stress value exceeds 1.5 times the average stress.
8. The method of claim 1, wherein: The hardware preparation work before step S1 is as follows: configure a hardware system including a simulation workstation, a GPU training server, a laser directional energy deposition additive manufacturing equipment, a 3D scanner and an X-ray diffraction residual stress meter, and software tools including finite element simulation software, molecular dynamics simulation software, deep learning framework and 3D modeling software.
9. The method of claim 1, wherein: The target high-temperature alloy is GH4169 nickel-based high-temperature alloy. GH4169 powder with a specific particle size range is used as the experimental material, and the substrate material is the same as that of GH4169 high-temperature alloy.
10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for predicting deformation and residual stress in a high-temperature alloy additive manufacturing process as described in any one of claims 1 to 9.