Data-driven complex structure die casting impact failure full-process prediction method

Abstract

The invention relates to the field of impact failure full-process prediction of complex-structure die castings, and discloses a data-driven impact failure full-process prediction method for complex-structure die castings, which comprises the following steps: step 1, performing mold flow simulation on a target part to obtain the solidification time and porosity of any position of the target part; 2, the solidification time and porosity of any position of the target part are input into a multi-layer perceptron model, and the multi-layer perceptron model outputs mechanical parameters of any position of the target part; the multi-layer perceptron model is trained based on data of actually measured mechanical properties; 3, acquiring a collision simulation grid model of the part, mapping the mechanical parameters to each grid unit of the collision simulation grid model, and generating an input file for collision simulation; and 4, performing collision simulation based on the input file in the step 3.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of complex structure die casting impact failure full process prediction, and particularly relates to a data-driven complex structure die casting impact failure full process prediction method. BACKGROUND

[0002] In the engineering design and research and development of key safety structural parts of automobiles, the traditional "design-manufacture-test" iterative mode seriously depends on the destructive crash test of physical samples. This method not only has a long cycle and high cost, but also its effectiveness is largely limited by the experience accumulation and previous data of engineers, and it is difficult to meet the modern industrial demand of product rapid iteration and design complication. Under this background, the crash simulation technology can accurately reproduce the crash working condition in the virtual environment, and provide key basis for evaluating the mechanical properties of parts under new design or new process change, thereby greatly reducing the dependence on physical test.

[0003] However, the accuracy of high-fidelity crash simulation fundamentally depends on the authenticity of the input material parameters. For thin-walled, multi-jointed or internally featured aluminum alloy die castings, the mechanical properties thereof present significant spatial non-uniformity, and it is difficult to obtain a sufficient number of representative samples through conventional uniaxial tensile test due to the limitation of complex body features and sampling feasibility, resulting in inaccurate simulation input data, which becomes the main bottleneck restricting the prediction accuracy.

[0004] Therefore, a data-driven complex structure die casting impact failure full process prediction method is needed, which integrates high-precision mold flow simulation, machine learning driven parameterized process optimization, and crash failure prediction model, to efficiently and accurately predict die casting impact failure, thereby realizing efficient and accurate prediction of the mechanical properties of die castings. SUMMARY

[0005] The present application aims to provide a data-driven complex structure die casting impact failure full process prediction method, which efficiently and accurately predicts die casting impact failure.

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical solution: a data-driven complex structure die casting impact failure full process prediction method, comprising the following steps: Step one, mold flow simulation is performed on the target part to obtain the solidification time and porosity of the target part at any position; Step two, the solidification time and porosity of the target part at any position are input into a multilayer perception machine model, and the multilayer perception machine model outputs the mechanical parameters of the target part at any position; the multilayer perception machine model is trained based on the data of measured mechanical properties; Step three, obtaining the collision simulation grid model of the part, mapping the mechanical parameters to each grid unit of the collision simulation grid model to generate an input file for collision simulation; Step four, performing collision simulation based on the input file in step three.

[0007] The beneficial effects of the present scheme are: 1. Integrating high-precision mold flow simulation, machine learning driven parameterized process optimization, and collision failure prediction model, the acquisition of part mechanical properties will be transformed from "physical experiment post-validation" to "virtual simulation prior optimization" paradigm.

[0008] 2. The multi-layer perception model is the core content of the present application, including input layer, hidden layer and output layer. Its core advantage is that it can learn the non-linear relationship in complex data through multi-layer structure.

[0009] The multi-layer perception model is trained based on the measured mechanical properties data, which can predict the mechanical properties of aluminum alloy structural parts under the coupling action of multiple defect factors (solidification time, porosity). Without directly finding the specific physical model, the relationship between independent variables and dependent variable labels can be found, the dependent variable can be directly predicted through the independent variable, and the influence of solidification time and porosity on the mechanical properties of different positions of the part can also be evaluated. Thus, the mechanical parameters of any position of the target part are output, and the input file for collision simulation is generated. Further, the die casting impact failure is efficiently and accurately predicted.

[0010] Further, in step one, the method for obtaining the solidification time and porosity of any position of the target part comprises the following steps: Step 1.1, data extraction is performed on the mold flow simulation results of the die casting, and the simulation results including point cloud data of solidification time and porosity are output; Step 1.2, the die casting is meshed, the geometry of the die casting is discretized into a plurality of tetrahedrons with specific geometric characteristics and topological relationships, thereby forming a mesh model, the point cloud data obtained by data extraction is aligned with the mesh model in the coordinate system, a spatial mapping relationship is established, and a die casting model is formed; the tetrahedrons without point cloud data inside are removed Step 1.3, the geometry is divided into a plurality of topologically related sub-regions based on the secondary axis-aligned bounding box; Step 1.4, the neighborhood aggregation method based on virtual sphere scanning is used to simplify the data of the geometry, and all grid nodes inside the virtual sphere are compressed into a new node.

[0011] So set up, the data result of the data volume of the mold flow simulation is huge, the scheme processes these data, divides the data into each tetrahedron, and then rejects the redundant data, so as to improve the efficiency of subsequent data processing. Taking the mold flow simulation result of the software FLOW-3D as an example, the existence of the auxiliary structure such as the pouring system, the ladle and the cooling system in the die casting model leads to the increase of the difficulty of subsequent analysis and data extraction. In the scheme, the tetrahedron not containing point cloud data is removed, that is, the auxiliary structure such as the pouring system, the ladle and the cooling system is removed, so as to improve the data extraction accuracy and realize the data reduction of the numerical simulation result of the die casting process.

[0012] Further, step 1.3 specifically comprises the following steps: Step 1.3.1, the AABB space comprises primary AABB space division and secondary AABB space division; The primary AABB space division divides the geometry of the die casting into a plurality of topologically related sub-regions; The secondary AABB space division divides the sub-region into a plurality of tetrahedral units; The KDTree is used to construct a binary tree structure reorganization in the sub-region; Step 1.3.2, the target coordinates are regarded as a special bounding box with a volume of 0, the special bounding box is sequentially subjected to bounding box intersection detection with the primary bounding box, so as to screen out the sub-region where the special bounding box is located, then the tetrahedral units in the sub-region are traversed, the special bounding box is sequentially subjected to bounding box intersection detection with the secondary bounding box, so as to screen out the tetrahedral unit where the special bounding box is located, and the target coordinates in the tetrahedral unit are located through the binary tree structure.

[0013] So set up, the following effects are obtained: 1. In order to efficiently process the large-scale three-dimensional space data of the die casting, the space index structure is introduced to replace the inefficient 'traversal search'. The core lies in that the unnecessary intersection judgment times are reduced by constructing a multi-level space index structure, so that the time complexity is reduced from O(NXM) to close to O(NlogM), and efficient space query and unit positioning are realized.

[0014] First, the primary and secondary AABB space partitioning of the geometric body is carried out, and the AABB is a secondary axis-aligned bounding box, which is an enclosing box constructed based on the projection range of the object in the three-dimensional space coordinate axis, and the boundary surface thereof is strictly parallel to the X, Y and Z axes of the Cartesian coordinate system. The core feature thereof is to quickly determine a cuboid region through the minimum value and the maximum value of the vertex coordinates of the object, which can be mathematically expressed as two extreme points: the minimum point (Xmin, Ymin, Zmin) and the maximum point (Xmax, Ymax, Zmax). This design makes it only necessary to compare and record the coordinate extreme values in the calculation process, without the need for complex geometric operations, thereby greatly reducing the calculation cost.

[0015] The secondary AABB space of tens of thousands of tetrahedral units is taken as a data point to construct a 16-layer KD-Tree. When it is necessary to query which tetrahedron a point belongs to, it is only necessary to traverse a few leaf nodes from the tree root, without the need to traverse all the tetrahedral units, thereby realizing "quick pruning".

[0016] Then, it is only necessary to judge whether the target coordinates are within these secondary AABB spaces. Since the AABB intersection detection is very fast, the efficiency is much higher than that of directly judging whether the point is within the tetrahedron.

[0017] 2. The present scheme reorganizes all the secondary AABB spaces in the sub-region using a KD-Tree. If only secondary AABB is used, the second screening still needs to traverse all the units in the sub-region, and the complexity is still O(n). However, after embedding the KD-Tree into each sub-region, the efficiency of the second screening is also improved from O(n) to O(log n). Thus, the "cooperation of module-level coarse screening (O(1)) and unit-level accurate positioning (O(log n))" is realized, and the total complexity is finally reduced to O(log n). Furthermore, by introducing a spatial index structure and an algorithm optimization strategy, the calculation complexity is significantly reduced, and the spatial query efficiency problem of large-scale grid models is solved.

[0018] Further, in step two, the training method of the multi-layer perception model is as follows: Step 2.1, the mold flow simulation of the training part, the mold flow simulation result includes the solidification time and the porosity; Step 2.2, prepare and train a flat plate mold with the same material as the part, and perform basic mechanical property testing on the flat plate mold to obtain reference mechanical properties; Step 2.3, sample three-point bending test samples at different positions on the training part, and perform three-point bending tests on the three-point bending test samples to obtain measured mechanical properties; Step 2.4, input the reference mechanical properties into the material card, simulate the three-point bending sample based on the material card, the material card parameters include the hardening curve scaling factor and the fracture surface scaling factor, adjust the hardening curve scaling factor and the fracture surface scaling factor, so as to calibrate the measured mechanical properties and the reference mechanical properties, and obtain the adjusted hardening curve scaling factor and the fracture surface scaling factor; Step 2.5, train the multi-layer perception model, take the mold flow simulation result as the input feature, and take the mechanical parameters at different positions as the target variable, the mechanical parameters include the hardening curve scaling factor and the fracture surface scaling factor.

[0019] In this way, the following effects are achieved: Based on the multi-layer perception model, the function relationship between the solidification time, the porosity and the mechanical parameters at different positions is constructed, since the data of the mold flow simulation has continuity, the mechanical parameters at any position can be output through the multi-layer perception model; in the present application, the mechanical parameters are the hardening curve scaling factor and the fracture surface scaling factor, specifically, the constitutive property at any position is obtained by multiplying the constitutive curve and the hardening curve scaling factor, and the fracture property at any position is obtained by multiplying the fracture surface and the fracture surface scaling factor. In order to obtain the material constitutive relationship and fracture parameters at any position of the die casting, the accuracy and continuity of the data are ensured. Further, in step 2.5, before training the multi-layer perception model, the hardening curve scaling factor and the fracture surface scaling factor are normalized. Specifically, the constitutive scaling factor and the fracture scaling factor are processed according to certain rules, so that they fall into the decimal interval (0, 1), thereby eliminating the influence of different dimensions on the final result.

[0020] Further, in step 2.2, the reference mechanical properties include the reference hardening curve , the dimensionless strain rate ε * , the stress triaxiality η, the material strain rate sensitivity coefficient D4 and the material failure parameters D1, D2 and D3; In step 2.3, the measured mechanical properties include the sampling point constitutive curve and the sampling point fracture surface ; Step 2.4 specifically includes the following steps: Step 2.4.1, solid modeling is performed on all samples, the reference mechanical properties are input into the material card, the solid model is simulated based on the material card, and the hardening curve scaling factor and the fracture surface scaling factor are adjusted, so as to calibrate the measured mechanical properties and the reference mechanical properties; The hardening curve scaling factor satisfies: ; The fracture surface scaling factor Satisfy: ; 2.4.2 Shell modeling is performed on the sample smaller than the design thickness, and the shell model is simulated based on the hardening curve scaling factor and the fracture surface scaling factor. The material card parameters also include the hardening curve scaling factor and the fracture surface scaling factor of the shell model. The material card parameters are adjusted twice based on the following formula: ; ; λ1 and λ2 are both second-order correction coefficients; The adjusted hardening curve scaling factor and fracture surface scaling factor are used as the mechanical parameters of the training part at the position smaller than the design thickness.

[0021] With such settings, the following effects are achieved: 1. The process parameters of the simulation part at each position are continuous and complete, but due to the calculation by the machine throughout the process, there is a lack of actual measurement data, so the accuracy is not high, and there are defects. Although the mechanical parameters of different positions of the part can be obtained by actual measurement, for thin-walled, multi-jointed or internally featured die castings, the number of samples is insufficient, and the data obtained is not continuous.

[0022] Therefore, the multi-layer perception machine model constructs the functional relationship between the defect data and the mechanical parameters at different positions, thereby realizing the prediction of the constitutive relation and fracture at any position.

[0023] 2. Due to the difficulties in obtaining standard samples on the part by conventional means, the present scheme processes the samples on the flat plate mold of the same material as the part to obtain the samples required for basic mechanical property testing, and obtains the reference material card through simulation calculation; to preliminarily solve the sampling difficulty problem. The simulation results include the mechanical property information at each position, but the accuracy of the data is not high; the test results of the samples are accurate, but the number of samples is not enough; therefore, the two groups of data are adjusted, the adjusted material card parameters are used for machine learning, and the mechanical properties at all positions on the part are predicted by the multi-layer perception machine model, to solve the problem of detection blind area of complex parts.

[0024] 3. Since the shell element cannot accurately represent the evolution of strain and cumulative damage along the thickness direction, the shell modeling in the three-point bending simulation of thick plates has a large difference between the theoretical assumption and the implementation, so the prediction accuracy of the solid model is higher than that of the shell model. But part of the position thickness is thin, the shell model is more efficient, so for the whole part, mixed modeling is adopted, and for these thin positions, solid modeling is needed first, and the first parameter adjustment is needed, and then shell modeling is needed, and based on the first adjustment result, the second parameter adjustment is needed, so as to ensure the efficiency of subsequent mechanical learning.

[0025] Further, in step 2.3, different thickness samples are selected for testing by sampling method; in step 2.4.2, after obtaining the material card parameters of the samples with different thickness, the material card parameters of the shell model of the samples with intermediate thickness are obtained by interpolation method. By this setting, the mechanical property information of more thickness samples is obtained by difference method, so as to improve the training effect; the "intermediate thickness" in the application refers to the thickness between any two thickness values of the samples in step two.

[0026] Further, step three specifically includes the following steps: Step 3.1, obtaining the collision simulation grid of the part; Step 3.2, mapping the mechanical parameter data to each grid unit, and the mechanical parameter also includes yield strength; Step 3.3, merging the grid units that are adjacent and have a yield strength difference within a preset threshold to form multiple sub-regions; Step 3.4, assigning each sub-region a dedicated material card based on the mechanical property data for generating the input file of the collision simulation.

[0027] The beneficial effects of the present scheme are: 1. Yield strength is the critical stress threshold for the transition of material from elastic deformation to plastic deformation. When the local stress of the die casting exceeds the yield strength, irreversible plastic deformation will occur, resulting in irreversible plastic deformation. The fracture strain reflects the deformation capacity of the material before failure. In actual working conditions, the safety factor is usually set based on the yield strength, and the local stress field matches the yield stress, which can accurately distribute the material properties and avoid using high-strength materials in low-stress areas. 2. In the collision simulation, the constitutive model describes the stress-strain relationship of the material from loading to failure, providing global history data such as stress, strain, and damage variable for each integration point, which considers global physical mechanisms. The yield strength and the key parameters in the constitutive model have significant correlation, and higher yield strength often has lower fracture strain, so the strength is selected as the basis for dividing parts to improve the accuracy of simulation analysis. To achieve intelligent division of regions and accurate distribution of material properties, improve the accuracy and efficiency of collision simulation.

[0028] 3. It allows for the allocation of high-strength materials in high-stress areas and low-strength materials in low-stress areas, avoiding resource waste and aligning with the lightweight design concept. Attached Figure Description

[0029] Figure 1 This is the overall flowchart of the embodiment; Figure 2 This is a flowchart of step one of the embodiments; Figure 3 This is a flowchart of step two in the embodiment; Figure 4 This is a flowchart of step three in the embodiment; Figure 5 Porosity distribution diagram after data compression in the example Figure 6 Solidification time distribution diagram after data compression in the example Figure 7 This is a diagram showing the division of eight sub-regions in an embodiment; Figure 8 This is a sub-region filtering diagram for an example. Figure 9 for Figure 8 Enlarged view of the sub-region; Figure 10 This is a schematic diagram illustrating the range of influence of the material card parameters on the hardening curve in an embodiment. Detailed Implementation

[0030] The following detailed description illustrates the specific implementation method: Example A data-driven method for predicting the entire impact failure process of complex die-cast parts, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps: Step 1: Perform mold flow simulation on the target part to obtain the solidification time and porosity at any location of the target part; Corresponding process Figure 1 The detailed flowchart is shown in the middle left column. Figure 2 As shown, the specific steps include: Step 1.1, Data Extraction: Through research, aluminum alloys are characterized by their lightweight, high plasticity, and corrosion resistance. However, during die casting, defects such as porosity, shrinkage, and hot cracking are easily generated. Based on the priority of factors affecting the quality of die castings, the main factors affecting the quality of die castings are solidification time, porosity, and flow length. However, the flow length varies depending on the pouring position, so its impact on die casting is usually not considered. Only when die castings are formed using high-pressure die casting, where the flow rate and cooling rate of the molten metal are relatively fast, must the impact of flow length on the quality of die castings be considered. This embodiment is used for aluminum subframes, which employ low-pressure casting. Therefore, the main considerations are the impact of solidification time and porosity on die castings. The mold flow simulation software used is Magmasoft or Flow3D.

[0031] 1) In the mold flow simulation, the liquidus temperature, solidus temperature and latent heat of fusion are input, and the temperature field distribution at the end of filling is used as the initial value to output solidification data, which includes solidus velocity, thermal gradient and cooling rate. 2) Porosity is extracted from data based on a simplified shrinkage model. The simplified shrinkage model is used to evaluate the solidification shrinkage volume of each isolated liquid region in each time step of the die casting. Then, the volume is subtracted from the top of the liquid region based on the amount of liquid metal available in the cells from which the liquid is removed. The “top” of the liquid region is defined by the direction of gravity. The shrinkage model can quickly simulate the shrinkage of complex die castings. The volume does not include the evolution of gas. The porosity volume is determined by the cooling rate and the solid / liquid density difference. 3) Extract data from the mold flow simulation results of the die casting and output the simulation results, which include point cloud data of porosity and solidification data; 4) Calculate the solidification time based on solidification data. The solidification time will be calculated based on the metal temperature difference at the beginning and end of the time cycle. If the metal temperature difference is less than the design temperature difference, the solidification time will be set to the current time. When the metal temperature difference is greater than the design temperature difference, it is considered to have solidified. Remove redundant information, including duplicate data, molds, and gating systems. Step 1.2, Data Preprocessing: HyperMesh software is used to mesh the die casting, discretizing the geometry of the die casting into several tetrahedrons with specific geometric features and topological relationships to form a mesh model. The point cloud data obtained from data extraction is then compared with the mesh model to establish a coordinate system and a spatial mapping relationship, thus forming a die casting model. HyperMesh software was used to mesh the parts, forming a mesh model that accurately describes the spatial characteristics of the geometry. This transformed the parts from shell structures into "pseudo-solid" models, i.e., solid models. Three mesh sizes were used during the part meshing process, as shown in the table below. In the Python programming environment, the key data after part meshing in the *.k file was parsed to accurately obtain detailed information about each tetrahedron in the model, including its constituent nodes, elements, and topological relationships. This laid the foundation for subsequent in-depth comparison and accurate calculation of the point set and the geometric features (such as shape, size, curvature, etc.) and topological relationships (such as adjacency, containment, intersection, etc.) of the part geometry.

[0032] Remove tetrahedrons that do not contain point cloud data; Specifically, calculate the directed volume of the cloud data coordinates to each face of the tetrahedron and the total volume of the tetrahedron. Calculate the directed volume of the small tetrahedron formed by the target coordinates and each face of the tetrahedron. If the sum of the directed volumes of all the small tetrahedrons equals the volume of the original tetrahedron, and the signs of the directed volumes of all the small tetrahedrons are the same as the signs of the volume of the original tetrahedron, then it indicates that the cloud data coordinates are within the tetrahedron.

[0033] When using the volume method for determination, the method is simple and not limited to whether the tetrahedron's geometric shape is regular. It is robust and suitable for calculating complex geometric shapes. However, when there are many tetrahedrons, the computational load is large and the efficiency is low. Step 1.3, Time Complexity Optimization: After extracting and preprocessing the simulation data, which covers 900,000 coordinate points, each point requires traversing all tetrahedrons. Without optimization, the total time complexity is at least O(900,000 * 76,000). Therefore, spatial indexing and algorithm optimization are needed to reduce the time complexity. This embodiment proposes a composite spatial partitioning strategy based on quadratic axis-aligned bounding boxes (AABB). The core algorithm flow includes: Step 1.3.1: Divide the geometry into AABB space, which includes primary AABB space partitioning and secondary AABB space partitioning; The specific steps for performing AABB space partitioning include: b1. Use a common axis-aligned bounding box to surround the entire space; b2, such as Figure 7 As shown, the interior of the total axis-aligned bounding box is divided into several sub-regions, and each sub-region is enclosed by several primary bounding boxes. b3. Use several secondary bounding boxes to enclose the tetrahedrons in the sub-region.

[0034] The primary AABB space partitioning divides the geometry of the die casting into several topologically related sub-regions; the primary bounding box is the smallest cuboid parallel to the XYZ coordinate axes, which can just enclose the sub-regions. The secondary AABB spatial partitioning divides the sub-region into several tetrahedral units; the secondary bounding box is the smallest cuboid parallel to the XYZ coordinate axes, which can just enclose the tetrahedral units.

[0035] The secondary AABB spatial partitioning method divides the subregion into several tetrahedral units as follows: The geometry is meshed to form a network model, and the geometry is discretized into several tetrahedrons with geometric features and topological relationships. The coordinate system of the point set of the simulation data and the mesh model is unified, and the Boolean relationship between the point set and the tetrahedral elements is established through spatial geometric algorithms.

[0036] KDTree is used to construct binary tree structure reorganization within subregions; Constructing binary tree structure reorganization within a subregion using KDTree includes the following steps: Step a1: Calculate the variance of the point set on each coordinate axis of the simulation data in the sub-region, and select the coordinate axis with the largest variance; Step a2: Use the median of all data points on the selected coordinate axis as the split point; Step a3: Create a plane perpendicular to the selected coordinate axis at the dividing point to divide the current space into two parts. Step a4: Recursively repeat steps 1.1-1.3 for the two parts of the space until the number of data points in the subspace is less than the set threshold or the depth of the tree reaches the preset value, then stop the partitioning and form leaf nodes.

[0037] Step 1.3.2: Determine if the target coordinate point is within the sub-region. Traverse the tetrahedral cells of the sub-region and locate the target coordinates using a binary tree structure. The overall algorithm diagram is shown below. Figure 2 As shown, by constructing a multi-level spatial index structure, the number of unnecessary intersection checks can be reduced.

[0038] Specifically, the target coordinates are considered as a special bounding box with a volume of 0, such as Figure 8 , Figure 9 As shown, the special bounding box is sequentially subjected to bounding box intersection detection with the primary bounding box to filter out the sub-regions where the special bounding box is located. Then, the tetrahedral elements of the sub-regions are traversed, and the special bounding box is sequentially subjected to bounding box intersection detection with the secondary bounding box to filter out the tetrahedral elements where the special bounding box is located.

[0039] After traversing the tetrahedral cells in the sub-region, the selected tetrahedral cells are used as the calculation objects. The directed volume of the small tetrahedron formed by the target coordinates and each face of the tetrahedral cell is calculated. If the sum of the directed volumes of all the small tetrahedrons is equal to the volume of the original tetrahedron, and the signs of the directed volumes of all the small tetrahedrons are the same as the signs of the volume of the original tetrahedron, then it means that the target coordinates are within the tetrahedral cells.

[0040] Step 1.4, Data Compression: Data simplification is performed using a neighborhood aggregation method based on virtual sphere scanning. Step 1.4.1: Using the virtual sphere as a tool, scan all the grid nodes within the virtual sphere; In point cloud data processing, the selection of the virtual sphere radius and scanning step size has a significant impact on the accuracy of the results. The sphere radius can be selected based on the density of the point cloud to ensure that the sphere can encompass an appropriate number of points. When considering actual data compression requirements, if a 3D geometric model of the part is provided, it can be converted into coordinate point cloud data, and the sphere radius can be dynamically adjusted according to the information content of the converted data. If no 3D model information is available, the part thickness can be used as the default sphere radius. Both methods take into account applications in different scenarios, have flexibility and applicability, and help improve the efficiency and accuracy of point cloud processing. The formula for calculating the radius γ of the virtual sphere is: , 5; d avg- - The average distance from each point in the point cloud data to its nearest neighbor; α - Scale factor; The ratio of scanned points to the total number of point clouds is the coverage rate C. After data reduction, the coverage rate should meet the following requirement: 95% ≤ C ≤ 99%. Step 1.4.2: Perform geometric mean calculation on the grid nodes within the virtual sphere, and compress these grid nodes into a new node; Step 1.4.3: The virtual sphere moves in steps not exceeding its radius, gradually covering the surface of the part until surface sampling is completed; The material of the parts in this project is AlSi7Mg. 0.3 The part weighs 17kg, has dimensions of 704.4mm × 1036mm × 225mm, and a minimum thickness of approximately 2mm. The minimum thickness is used as the radius and scan step size of the virtual sphere. After data simplification, the data points remain within the expected range while ensuring geometric accuracy. The retained nodes fully represent the geometric characteristics, and areas with high curvature are not oversimplified. The porosity distribution map after data compression is shown below. Figure 5 As shown in the figure, the solidification time distribution after data compression is as follows: Figure 6 As shown; The error assessment after data compression and dimensionality reduction adopts a multidimensional statistical assessment framework. The reduced data is reconstructed in three dimensions through visualization. While meeting the expected range, the geometric accuracy of the results must be guaranteed. It is necessary to check whether the retained nodes can fully express the geometric characteristics and whether the regions with high curvature have been oversimplified. Step 1.5, Data Assembly: The KDTree structure retrieval method is used to merge the "porosity" and "solidification time" data into a standard format document. Training Data Location: The positions of the training data points in the simulation results are learned; Data Output: All training elements at all locations are output.

[0041] Step 2: Input the solidification time and porosity at any location of the target part into the multilayer perceptron model. The multilayer perceptron model outputs the mechanical parameters at any location of the target part. The multilayer perceptron model is trained based on the measured mechanical property data. (Flowchart follows.) Figure 1 The middle column, the specific flowchart is as follows: Figure 3 As shown; The training method for the multilayer perceptron model is as follows: Step 2.1: Perform mold flow simulation on the training parts. The mold flow simulation results include solidification time and porosity. Step 2.2: Prepare a flat mold made of the same material as the training parts. The method for making the flat mold is as follows: when producing the training parts, use the same process to die-cast a homogeneous flat mold. Using a flat plate mold as a sampling carrier, standard specimens for uniaxial tensile, notched, shear, and cupping tests are obtained. Basic mechanical property tests are then performed on these standard specimens to obtain reference mechanical properties, including reference constitutive curves and reference fracture surfaces. The reference constitutive curve includes a reference hardening curve. Dimensionless strain rate ε * Stress triaxiality η, material strain rate sensitivity coefficient D4, and material failure parameters D1, D2, and D3; The hardening curve is a curve that describes the relationship between flow stress and deformation during plastic deformation of a material; Dimensionless strain rate is usually defined as the ratio of the current strain rate to the reference strain rate; Stress triaxiality is defined as the ratio of mean stress to equivalent stress, and is a key indicator for measuring whether the stress state is "tension" or "compression". Material failure parameters and material strain rate sensitivity coefficients are constants that need to be determined through a series of mechanical experiments (such as tensile tests and shear tests with different notch radii).

[0042] Step 2.3: Take three-point bend samples at different locations on the training part. The cross-sectional dimensions of the three-point bend samples are uniformly 40mm × 30mm, and the thickness is adjusted according to the sampling location. The span during testing is set to 30mm. Perform three-point bend tests on the samples to obtain the measured mechanical properties, including the constitutive curves at the sampling points. and the fracture surface at the sampling point ; Step 2.4: Input the reference mechanical properties into the material card, simulate the three-point bend specimen based on the material card, adjust the parameters of the material card, and thus align the measured mechanical properties with the reference mechanical properties. Specifically, the following steps are included: 2.4.1 Solid modeling was performed on all three-point bend specimens. The solid model describes the object as a completely solid closed volume filled with material. The reference mechanical properties were input into the material card, and the solid model was simulated based on the material card. The parameters of the material card were adjusted to align the measured mechanical properties with the reference mechanical properties. The material card parameters include Young's modulus, yield strength, hardening curve shrinkage factor, and fracture surface shrinkage factor. Young's modulus is a physical quantity that describes the ability of a solid material to resist deformation, and is defined as the ratio of stress to strain.

[0043] like Figure 10 As shown, the constitutive curves of the sampling points obtained from the three-point bending force test can be divided into four stages: elastic deformation stage, yielding stage, uniform plastic deformation stage, and necking stage. To improve simulation benchmarking efficiency, three parameters were selected for material card parameter benchmarking: Young's modulus (controlling the elastic deformation stage), hardening curve scaling factor (regulating plastic deformation behavior), and fracture surface scaling factor (determining fracture characteristics), corresponding to... Figure 10 The three regions are divided by the dashed line.

[0044] Hardening curve scaling factor Satisfies (based on MAT24 material card): ; Fracture surface scaling factor Satisfies (based on JC principles): ; The Johnson-Cook model is an empirical failure model widely used for metallic materials and certain polymers under large deformation and high strain rate scenarios such as high-speed impact, explosion, and cutting. It describes how the failure strain of a material (i.e., the strain at which the material begins to fracture) is affected by stress state, strain rate, and temperature.

[0045] These three parameters, D1, D2, and D3, are the core failure parameters of this model. Together, they determine the dependence of the failure strain of the material on the stress triaxiality under quasi-static, room temperature conditions.

[0046] The specific definitions of the three parameters are as follows: D1: An approximate failure strain under pure shear (zero-stress triaxiality); D1 can be regarded as the toughness benchmark of a material under shear-dominated failure modes (such as perforation and shearing). It is a fitting parameter, but its value is close to the failure strain measured in a pure shear test.

[0047] D2: The attenuation coefficient of failure strain under high stress triaxiality; the larger the D2, the greater the attenuation of failure strain relative to the shear state (D1) under high stress triaxiality.

[0048] D3: Controls the sensitivity of failure strain to stress triaxiality; D3 determines the "steepness" of the failure strain curve as stress triaxiality changes. The larger the absolute value of D3, the steeper the curve drops, meaning that the material is more "sensitive" to stress triaxiality.

[0049] D2 and D3 work together to describe the change in failure strain with triaxiality. D2 and D3 are strongly coupled and cannot be understood separately. Together, they fit the shape of the curve of failure strain as a function of triaxiality.

[0050] 2.4.2 Shell modeling is performed on the three-point bend specimen with a thickness less than the design thickness. The shell model simplifies the object into a curved or flat surface without thickness; it is essentially a two-dimensional geometric representation. Simulation is performed on the shell model based on the material card parameters obtained in step 2.4.1. The material card parameters also include the scaling factor of the shell model's hardening curve and the scaling factor of the shell model's fracture surface. The material card parameters are then adjusted a second time based on the following formula: ; ; λ1 and λ2 are both second-order correction coefficients; The adjusted hardening curve scaling factor and fracture surface scaling factor are used as mechanical parameters for the training part at locations where the thickness is less than the design thickness.

[0051] After obtaining the material card parameters of samples with different thicknesses, the material card parameters of the shell model of the sample with intermediate thickness are derived by interpolation.

[0052] Step 2.5: Perform data cleaning and analysis on the mechanical parameters, solidification time and porosity obtained in steps 2.1-2.4. Data cleaning involves analyzing the expected value and standard deviation of the original data, and then removing data outside the distribution band formed by the expected value and standard deviation. The data analysis assumes that the data follows a normal distribution, meaning that the values ​​of the ordinates under the same x-axis follow a normal distribution. A smoothing window normal distribution probability weighting method is proposed. Specifically, the data is divided into several equal intervals according to the x-axis range. At the midpoint of each interval, the expected value and variance of all data points within that interval are calculated. Using the normal distribution assumption, the probability of each data point is calculated. The probabilities of all data points within the interval are normalized into weights, and the final weighted value is output. This analytical method, through probability weighting, eliminates the interference of data randomness and provides an intuitive predictive trend for the influence of solidification time and porosity on the constitutive shrinkage factor.

[0053] The scaling factors of the hardening curve and the fracture surface were normalized and trained using a multilayer perceptron (MLP) model. The input features of the MLP model were the solidification time and porosity from step one, and the target variables were Young's modulus, the scaling factor of the hardening curve, and the scaling factor of the fracture surface.

[0054] The multilayer perceptron model includes a preprocessing module, a model training module, and a data prediction module. The preprocessing module uses Pandas functions to read data, handle missing values, and remove outliers from the solidification time and porosity datasets of the training parts.

[0055] The model training module uses the random forest regression algorithm for learning and a nested 10-fold cross-validation method for training data.

[0056] The data prediction module combines the mean absolute error (MAE) and the mean absolute percentage error (MAPE) to evaluate the accuracy and generalization ability of the model predictions. The calculation formula is as follows: ; ; Represents experimental values. m represents the predicted value, and m represents the number of experimental values. The lower the MAPE value, the better the prediction performance; the closer the MAE value is to 0, the more accurate the model prediction.

[0057] Iterate through all data points to obtain the constitutive curves and fracture surfaces for all points.

[0058] Step 3: Obtain the collision simulation mesh model of the part, map the mechanical parameters to each mesh cell of the collision simulation mesh model, and generate the input file for collision simulation; the flowchart corresponds to...Figure 1 The right column details the process as follows: Figure 4 As shown, the specific steps include: Step 3.1: Based on global mechanical parameters, construct three-dimensional anisotropic constitutive cloud map and fracture toughness cloud map to realize the visual characterization of the spatial distribution of material properties.

[0059] The collision simulation mesh model uses a hybrid modeling approach combining shell and solid methods. Step 3.2: Map the mechanical property data to each mesh cell; Specifically, a sphere scanning algorithm accelerated by a KD tree structure is used for information mapping. For the dataset D={x_i ϵR^3,i=1,...,N}, the segmentation dimension d and segmentation value υ are set as follows, and efficient retrieval is achieved by constructing a multi-level spatial index structure.

[0060] ; ; ; The process involves recursively splitting the nodes until the number of nodes in the leaf nodes is less than or equal to K_max, constructing a KD tree with levels L of 1, 2, 3, etc., and r_L as the search radius, forming a pyramid structure. Higher levels provide a wide range of coarse search coverage, while lower levels provide fine search.

[0061] Information mapping based on the sphere scanning method includes the following steps: 3.2.1. Based on the geometric properties of the elements, an equivalent sphere is established, with the centroid of the mesh element as the center of the equivalent sphere. For the shell model, the radius of the equivalent sphere is obtained by the equivalent area method; for the solid model, the radius of the sphere is obtained by the equivalent volume method.

[0062] 3.2.2 Scan the input mechanical property data and determine that any data point whose coordinates fall within the sphere is a match for that grid cell, and the mechanical property data carried by the data point is associated with that grid cell; 3.2.3. The mechanical property data of each equivalent sphere is arithmetically averaged and used as the mechanical property data of that grid element.

[0063] Data mapping is performed using Gaussian radial basis function (RBF) interpolation. RBF interpolation is a distance-weighted multidimensional data interpolation method. Its core idea is to construct an interpolation surface using a radially symmetric Gaussian function. It is suitable for irregularly distributed high-dimensional data, and its mathematical form is as follows.

[0064] ; Indicate the points to be interpolated x and x iThe Eulerian distance is given by ϵ, where ϵ is the bandwidth parameter (shape parameter), controlling the local sensitivity and decay rate of the function. The interpolation function is a linear combination of Gaussian functions for all data points.

[0065] By using the Gaussian radial basis function (RBF) interpolation method, discrete mechanical property data are accurately mapped to the collision simulation mesh. Through iterative calculation, it is ensured that each mesh element obtains a set of mechanical property parameters that match its spatial location, including key indicators such as elastic modulus, yield strength, and fracture strain. This establishes a high-precision material-structure coupled simulation model, providing accurate material property inputs for collision dynamics analysis.

[0066] In this step, each unit obtains independent mechanical property data.

[0067] Step 3.3: Merge adjacent mesh cells whose yield strength difference is within a preset threshold to form multiple sub-regions; specifically, this includes the following steps: Step 3.3.1: Extract the upper and lower boundaries of the yield strength of the global element of the part, and divide the range formed by the upper and lower boundaries of the yield strength into several continuous gradient intervals. Step 3.3.2: Traverse all mesh elements and determine the specific performance gradient range to which they belong based on their respective yield strength values; Step 3.3.3: Reassemble the grid cells that fall into the same gradient interval and assign them a unified number to form multiple sub-regions with serial numbers. Merge the sub-regions whose volume (or area) is less than the minimum threshold into their adjacent sub-regions with the closest average yield strength. Step 3.4: Assign a unique material card based on mechanical property data to each sub-region to generate the input file for collision simulation; assign the same material properties to the solid model part of the same region based on mechanical property data, and assign the same material properties to the shell model part of the same region based on the corrected mechanical parameters (including the scaling factor of the hardening curve of the shell model and the scaling factor of the fracture surface of the shell model); The method for obtaining material cards based on mechanical property data includes the following steps: Step 3.4.1: Calculate the average value of all elements within the sub-region and output the fracture surface and constitutive curve; Step 3.4.2: Combine each sub-region and constitutive curve, and output the file in the format of the simulation collision to form a material card. In this embodiment, the *.k file is output in the lsdyna simulation format.

[0068] Step 4: Based on the *.k file from Step 3, perform LSDYNA collision service simulation, output the comparison and verification results with the traditional solution, and form the software and solution.

[0069] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A data-driven method for predicting the entire process of impact failure in complex die-cast parts, characterized in that, Includes the following steps: Step 1: Perform mold flow simulation on the target part to obtain the solidification time and porosity at any location of the target part; Step 2: Input the solidification time and porosity of any location on the target part into the multilayer perceptron model. The multilayer perceptron model outputs the mechanical parameters of any location on the target part. The multilayer perceptron model is trained based on the measured mechanical property data. Step 3: Obtain the collision simulation mesh model of the part, map the mechanical parameters to each mesh cell of the collision simulation mesh model, and generate the input file for collision simulation; Step 4: Perform collision simulation based on the input file from Step 3.

2. The data-driven method for predicting the entire process of impact failure in complex die-cast parts according to claim 1, characterized in that: Step one, the method for obtaining the solidification time and porosity at any location of the target part includes the following steps: Step 1.1: Extract data from the mold flow simulation results of the die casting and output the simulation results, which include point cloud data of solidification time and porosity. Step 1.2: Mesh the die casting, discretize the geometry of the die casting into several tetrahedrons with specific geometric features and topological relationships to form a mesh model. Unify the coordinate system of the point cloud data obtained from data extraction with the mesh model to establish a spatial mapping relationship and form a die casting model; remove tetrahedrons that do not contain point cloud data. Step 1.3: Spatial partitioning of the geometry based on a quadratic axis-aligned bounding box, dividing the geometry into several topologically related sub-regions; Step 1.4: Use the neighborhood aggregation method based on virtual sphere scanning to simplify the geometry data, compressing all the mesh nodes within the virtual sphere into a new node.

3. The data-driven method for predicting the entire process of impact failure in complex die-cast parts according to claim 2, characterized in that, Step 1.3 specifically includes the following steps: Step 1.3.1: AABB space includes primary AABB space partitioning and secondary AABB space partitioning; The primary AABB spatial partitioning divides the geometry of the die-casting part into several topologically related sub-regions; The secondary AABB spatial partitioning divides the subregion into several tetrahedral units; KDTree is used to construct binary tree structure reorganization within subregions; Step 1.3.2: Treat the target coordinates as a special bounding box with a volume of 0. Perform bounding box intersection checks with the primary bounding box in turn to filter out the sub-regions where the special bounding box is located. Then, traverse the tetrahedral cells of the sub-regions and perform bounding box intersection checks with the secondary bounding boxes in turn to filter out the tetrahedral cells where the special bounding box is located. Locate the target coordinates within the tetrahedral cells using a binary tree structure.

4. The data-driven method for predicting the entire process of impact failure in complex die-cast parts according to claim 3, characterized in that: In step two, the training method for the multilayer perceptron model is as follows: Step 2.1: Perform mold flow simulation on the training parts. The mold flow simulation results include solidification time and porosity. Step 2.2: Prepare and train a flat mold with the same material as the parts, and conduct basic mechanical property tests on the flat mold to obtain reference mechanical properties; Step 2.3: Take samples of three-point bending specimens at different positions on the training part, perform three-point bending tests on the three-point bending specimens, and obtain the measured mechanical properties. Step 2.4: Input the reference mechanical properties into the material card, and simulate the three-point bend specimen based on the material card. The material card parameters include the hardening curve shrinkage factor and the fracture surface shrinkage factor. Adjust the hardening curve shrinkage factor and the fracture surface shrinkage factor to align the measured mechanical properties with the reference mechanical properties and obtain the adjusted hardening curve shrinkage factor and fracture surface shrinkage factor. Step 2.5: Train the multilayer perceptron model, using the simulation results of the model flow as input features and the mechanical parameters at different locations as target variables. The mechanical parameters include the scaling factor of the hardening curve and the scaling factor of the fracture surface.

5. The data-driven method for predicting the entire process of impact failure in complex die-cast parts according to claim 4, characterized in that: In step 2.5, before training the multilayer perceptron model, the scaling factor of the hardening curve and the scaling factor of the fracture surface are normalized.

6. The data-driven method for predicting the entire process of impact failure in complex die-cast parts according to claim 5, characterized in that: In step 2.2, the reference mechanical properties include the reference hardening curve. Dimensionless strain rate ε * Stress triaxiality η, material strain rate sensitivity coefficient D4, and material failure parameters D1, D2, and D3; In step 2.3, the measured mechanical properties include the constitutive curves at the sampling points. and the fracture surface at the sampling point ; Step 2.4 specifically includes the following steps: Step 2.4.1: Perform solid modeling on all samples, input the reference mechanical properties into the material card, simulate the solid model based on the material card, and adjust the scaling factor of the hardening curve and the scaling factor of the fracture surface to align the measured mechanical properties with the reference mechanical properties. Hardening curve scaling factor satisfy: ； Fracture surface scaling factor satisfy: ； 2.4.2 Shell modeling is performed on specimens with a thickness less than the design thickness. The shell model is simulated based on the hardening curve scaling factor and the fracture surface scaling factor. The material card parameters also include the hardening curve scaling factor and the fracture surface scaling factor of the shell model. The material card parameters are then adjusted a second time based on the following formula: ； ； λ1 and λ2 are both second-order correction coefficients; The adjusted hardening curve scaling factor and fracture surface scaling factor are used as mechanical parameters for the training part at locations where the thickness is less than the design thickness.

7. The data-driven method for predicting the entire process of impact failure in complex die-cast parts according to claim 6, characterized in that: In step 2.3, samples of different thicknesses are selected for testing using a sampling method; in step 2.4.2, after obtaining the material card parameters of samples of different thicknesses, the material card parameters of the shell model of the intermediate thickness sample are derived using an interpolation method.

8. The data-driven method for predicting the entire process of impact failure in complex die-cast parts according to claim 7, characterized in that: Step three specifically includes the following steps: Step 3.1: Obtain the collision simulation mesh for the part; Step 3.2: Map the mechanical parameter data to each grid cell. The mechanical parameters also include yield strength. Step 3.3: Merge adjacent mesh cells whose yield strength difference is within a preset threshold to form multiple sub-regions; Step 3.4: Assign a unique material card based on mechanical property data to each sub-region to generate the input file for collision simulation.

Images