Establishment method of grain size prediction model and evaluation method of grain refinement degree

Abstract

The application discloses a kind of grain size prediction model establishing method and the evaluation method of grain refinement degree, it is related to SLM additive evaluation and process adjustment field, including the following steps: manufacturing sample;By forging forming way manufacturing forged sample, by SLM additive way manufacturing additive sample;Parameter acquisition;Grain size d and cutting performance parameters are obtained to forged sample and additive sample;Prediction model training;With cutting performance parameters as input, with grain size d as output, using physical information neural network as the prediction model is trained, until loss function converges, complete the training of the prediction model.Evaluation method applies the prediction model established.The application realizes the nondestructive, online, multi-scale characterization of grain refinement, solves the technical bottleneck that traditional metallographic detection cycle is long, sample destructiveness is big, cannot comprehensively reflect processing service performance, provides theoretical basis and data support for additive manufacturing process optimization and component performance prediction.

Description

Technical Field

[0001] This invention relates to the field of SLM additive manufacturing evaluation and process adjustment, and in particular to a method for establishing a grain size prediction model and a method for evaluating the degree of grain refinement. Background Technology

[0002] Laser melting (SLM) additive manufacturing technology has shown transformative potential in the manufacturing of critical hot-end components such as turbine blades and casings for aero-engines due to its unique manufacturing advantages. In order to ensure that the components (SLM additively manufactured components) have the target performance, it is usually necessary to inspect and evaluate the component performance, and then adjust the SLM additive manufacturing process parameters based on the inspection results so that the newly manufactured components have the target performance.

[0003] Research has revealed a direct relationship between the yield strength and fatigue performance of components and grain size; for example, the grain size of traditionally forged nickel-based superalloy GH4169 is typically 50-200 μm, while SLM additive manufacturing utilizes ultrafast solidification (cooling rate 10... 6 K / s-10 8 Laser power (K / s) can achieve grain refinement to 5-50 μm, or even nanocrystals, significantly improving yield strength and fatigue performance compared to traditionally forged GH4169. However, the grain refinement of SLM additive manufacturing components is influenced by a combination of process parameters such as laser power, scanning strategy, and layer thickness. Since grain refinement directly affects material properties, a grain refinement assessment is necessary during SLM additive manufacturing to determine the appropriate level of grain refinement and provide crucial information for component performance evaluation.

[0004] Currently, the methods for evaluating grain refinement can be broadly categorized into the following types: 1. Metallographic microscopy analysis; This method is destructive to the sample and cannot be used for online testing; it also only characterizes two-dimensional cross-sectional grains, has limited resolution for ultrafine and nanocrystals, and cannot reflect the overall microstructure of the component. 2. Evaluation methods using electron backscatter diffraction (EBSD), scanning electron microscopy (SEM), and transmission electron microscopy (TEM); For EBSD, the electron beam penetration depth is limited, only surface characterization is obtained, and the inspection equipment is expensive and sample preparation is complex; For SEM, it can only observe two-dimensional surfaces, not three-dimensional real grains, and cannot automatically identify grain boundaries and grain orientations; For TEM, sample preparation is complex, the testing range is small, and it is suitable for local microscopic analysis. 3. Indirect mechanical property evaluation methods, such as nanoindentation or micropillar compression; however, these methods have sparse measurement points and a single stress state for the component. 4. Inspection method using ultrasound / eddy current / X-ray; this method has poor adaptability to rough surfaces in additive manufacturing and insufficient resolution, especially when the surface roughness Ra>10μm, the signal attenuation is severe.

[0005] In summary, current methods for inspecting and evaluating component performance have some shortcomings. Therefore, this industry needs a new technical solution for evaluating grain refinement. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned problems by providing a method for establishing a grain size prediction model and an evaluation method for grain refinement. This method achieves non-destructive, online, and multi-scale characterization of grain refinement, solving the technical bottlenecks of traditional metallographic testing, such as long testing cycles, high sample destructiveness, and inability to fully reflect processing and service performance. It provides a theoretical basis and data support for additive manufacturing process optimization and component performance prediction.

[0007] The technical solution adopted in this invention is as follows: A method for establishing a grain size prediction model, wherein the prediction model is for a target component manufactured by SLM additive manufacturing and is used to predict the grain size of the target component, comprising the following steps: A1: Prototype manufacturing; forging prototypes are manufactured by forging forming, and additive prototypes are manufactured by SLM additive manufacturing. A2: Parameter acquisition; Grain size d and cutting performance parameters were acquired for both forged and additive samples; A3: Predictive model training; using cutting performance parameters as input and grain size d as output, the encoder-decoder architecture in the physical information neural network is used as the predictive model for training until the loss function converges, thus completing the training of the predictive model.

[0008] Furthermore, in step A2, the grain size d is obtained using an electron backscatter diffraction system.

[0009] Furthermore, in step A2, the cutting performance parameters include at least the bending moment polar coordinate area. First-order area moment fractal dimension Anisotropy index Spectral entropy Wavelet energy coefficients Cutting speed Feed per tooth Axial depth of cut Radial depth of cut Cutting time .

[0010] Furthermore, in step A2, when obtaining cutting performance parameters, the bending moment polar coordinate diagram is used as the basis. The bending moment polar coordinate diagram is constructed with the tool rotation angle θ as the polar angle, and the bending moment amplitude is synthesized. Constructing for the polar radius; where: The magnitude of the bending moment in the X direction; This represents the bending moment amplitude in the Y direction.

[0011] Furthermore, in step A2, the polar coordinate area of ​​the bending moment The formula for obtaining it is: ; First-order area moment The formula for obtaining it is: ; fractal dimension The fractal dimension of the polar coordinate plot is calculated using the box counting method; Anisotropy Index The formula for obtaining it is: ; in: This represents the maximum value of the combined bending moment amplitude M. This represents the minimum value of the combined bending moment amplitude M. Spectral entropy The formula for obtaining it is: ; in: The energy percentage of the i-th frequency band; The transient characteristics of cutting force are analyzed using continuous wavelet transform, and wavelet energy coefficients are extracted from these characteristics. .

[0012] Further, in step A3, the grain size d and cutting performance parameters are normalized to form a dataset; the dataset is randomly allocated into a training set and a test set; the prediction model is trained using the training set, and the trained prediction model is tested using the test set.

[0013] Further, in step A3, when training the prediction model, where: Input layer: Cutting performance parameters; Encoder: 3 fully connected layers, with ReLU activation function; Physical embedding layer: The physical constraints include at least Hall-Petch reinforcement constraints, dislocation reinforcement constraints, and physical residuals; among which, The Hall-Petch strengthening constraint is: ; in: For yield limit, The intracrystalline yield strength of a single-crystal matrix. For coefficients, This represents the average grain size. Cutting force ; in: A coefficient related to tool geometry, This is the axial cutting depth. This refers to the feed per tooth. The dislocation reinforcement constraint is: ; in: Taylor factor It is a constant. It is the elastic shear modulus. It is a constant. Dislocation density; Physical residuals for: ; in: Input the measured cutting force value. To strengthen the constraint cutting force model, For coefficients; Decoder: 2 fully connected layers; Loss function: ; in: For the grain size of the forged sample, For the grain size of the additive sample, R_physics represents the physical residual term in the equation. It is a constant. It is a constant. These are the initial conditions; Output: Grain size d.

[0014] A method for evaluating grain refinement, specifically for a target component manufactured by SLM additive manufacturing, is provided to assess the grain refinement of the target component; the method includes the following steps: B1: Perform a cutting operation on the target component to obtain the cutting performance parameters in the establishment method described above; B2: Using the cutting performance parameters from step B1 as input data, input them into the prediction model established by the method described above; the prediction model outputs the predicted grain size d3; B3: Determine the degree of grain refinement based on whether the value of d3 falls within the grain size range.

[0015] Furthermore, it also includes the following steps: B4: Obtain the geometrically necessary dislocation density ρ3 and nanohardness H3 of the target component; B5: Calculate the overall grain refinement index ,in: ; have: The grain size of the forged sample in the method of any one of claims 1-6; The geometrically necessary dislocation density of the forged sample in the establishment method according to any one of claims 1-6; The nanohardness of the forged sample obtained by any one of the establishment methods described in 1-6; 、 、 All are weighting coefficients, where: Furthermore, weights are assigned using the entropy weighting method; B6: If GFI ≥ 30%, then refine significantly; if 15% ≤ GFI < 30%, then refine moderately; if GFI < 15%, then refine lightly.

[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This invention establishes a grain size prediction model using an encoder-decoder architecture in a physical information neural network, constructing a relationship between cutting performance parameters and grain size. By cutting the target component to obtain cutting data, the prediction model can effectively predict the grain size, thereby enabling a quantitative assessment of the grain refinement degree and providing data for subsequent process parameter adjustments. Compared to existing solutions described in the background art, this invention reduces the cutting time of the target component to an average of 8-10 minutes per piece, far lower than the 4-6 hours required for metallographic sample preparation and observation, thus improving the time for grain refinement degree assessment and enabling online evaluation. Furthermore, this invention integrates multiple scales such as metallographic structure and cutting performance, resulting in more accurate grain size prediction and performance characterization. Moreover, the cutting path can achieve full coverage, allowing for more comprehensive grain characterization in all directions. Detailed Implementation

[0017] Example 1 A method for establishing a grain size prediction model, taking the construction of a grain size prediction model for nickel-based superalloy GH4169 (hereinafter referred to as GH4169 for the sake of brevity) as an example, this prediction model is used to predict the grain size of a target component made of nickel-based superalloy GH4169 produced by SLM additive manufacturing, and includes the following steps: A1: Prototype manufacturing; forging prototypes are manufactured by forging forming, and additive prototypes are manufactured by SLM additive manufacturing. Among them: 800 forged samples with grain size d1 of 80μm-150μm were selected; 120 additive samples were manufactured with laser power P=200W-400W, scanning speed v=800mm / s-2000mm / s, layer thickness t=20μm-50μm, and spacing h=80μm-120μm, and grain size d2 of 5μm-40μm were selected.

[0018] A2: Parameter acquisition; Grain size d and cutting performance parameters were acquired for both forged and additive samples; the details are as follows.

[0019] Grain size d was acquired using an electron backscatter diffraction system with parameters set to a step size u = 0.5 μm–2 μm, an accelerating voltage of 20 kV, and a sampling area ≥ 500 μm × 500 μm, including the grain size d1 of forged samples and the grain size d2 of additive samples. Simultaneously, the geometrically necessary dislocation density could also be obtained. And the calculation of nanohardness using the Oliver-Pharr method For example, obtaining the geometrically necessary dislocation density of the forged sample. To obtain the nano-hardness of the forged sample This provides data support for subsequent calculations of GFI.

[0020] Cutting performance parameters should include at least bending moment and polar area. First-order area moment fractal dimension Anisotropy index Spectral entropy Wavelet energy coefficients Cutting speed Feed per tooth Axial depth of cut Radial depth of cut Cutting time These parameters are obtained based on a bending moment polar coordinate diagram, which is constructed using the tool rotation angle θ as the polar angle to synthesize the bending moment amplitude. Constructing for the polar radius; where: The specific methods for obtaining cutting performance parameters are as follows.

[0021] Employs a Kistler 9170A or equivalent precision wireless force-measuring toolholder, integrating three-dimensional cutting force. With double bending moment Sensor, measuring range Natural frequency ≥ 5kHz, sampling frequency Synchronous acquisition of spindle encoder signals achieves precise synchronization of cutting angles (resolution 0.1°). A brand-new carbide end mill (D10R4, TiAlN-based nano-composite PVD coating, coating thickness 3-5μm) is used. Tool pre-wear testing ensures consistent tool wear conditions (flank wear) during cutting for both samples. Tool wear was monitored using an online microscope or tactile scanner, and a forced tool change was performed when wear exceeded tolerance. A five-factor, four-level orthogonal cutting test was conducted, and the data from the cutting test are shown in Table 1 below.

[0022] Table 1: Five-Factor, Four-Level Orthogonal Cutting Test Scheme

[0023] Each specimen underwent at least 64 sets of parameter combinations to ensure statistical significance.

[0024] In step A2, the polar coordinate area of ​​the bending moment The formula for obtaining it is: ; First-order area moment The formula for obtaining it is: ; fractal dimension The fractal dimension of the polar coordinate plot is calculated using the box counting method; Anisotropy Index The formula for obtaining it is: ; in: This represents the maximum value of the combined bending moment amplitude M. This represents the minimum value of the combined bending moment amplitude M. Spectral entropy The formula for obtaining it is: ; in: The energy percentage of the i-th frequency band; The transient characteristics of cutting force are analyzed using continuous wavelet transform, and wavelet energy coefficients are extracted from these characteristics. .

[0025] A3: Predictive Model Training; Using cutting performance parameters as input and grain size d as output, the predictive model is trained using an encoder-decoder architecture in a physical information neural network until the loss function converges, thus completing the training of the predictive model. Specific steps are as follows.

[0026] The grain size d and cutting performance parameters are normalized to form a dataset. The dataset is then randomly divided into a training set and a test set. The training set contains 800 data sets, of which 700 are forged samples and 100 are additive samples. The remainder is the test set. The prediction model is trained using the training set and tested using the test set.

[0027] In step A3, when training the prediction model, physical constraints and loss functions need to be pre-input into the physical information neural network; where: Input layer: Cutting performance parameters; Encoder: 3 fully connected layers, with ReLU activation function; Physical embedding layer: Physical constraints include at least Hall-Patch reinforcement constraints, dislocation reinforcement constraints, and physical residuals; among which, The Hall-Petch strengthening constraint is: ; in: For yield limit, The intracrystalline yield strength of a single-crystal matrix. For coefficients, Given the average grain size, we have: It can be calibrated by combining microscopic characterization and mechanical data; Cutting force ; in: A coefficient related to tool geometry, This is the axial cutting depth. This refers to the feed per tooth. The dislocation reinforcement constraint is: ; in: Taylor factor It is a constant. It is the elastic shear modulus. It is a constant. For dislocation density; it is usually determined by the elastic modulus E and Poisson's ratio ν, through It can be obtained by calculation or by direct measurement through dynamic elastic modulus testing (such as the resonant frequency method); It needs to be quantified by direct observation with TEM or XRD linear analysis (such as the Williamson-Hall method), and it is often strengthened in conjunction with substructures (cells, grain boundaries). The overall strength cannot be predicted in isolation using the classical dislocation strengthening formula. Physical residuals For equation 7: ; in: Input the measured cutting force value. To strengthen the constraint cutting force model, Let be the coefficient; then: , For lever arm; Decoder: 2 fully connected layers; Loss function is given by equation 8: ; in: For the grain size of the forged sample, For the grain size of the additive sample, R_physics represents the physical residual term in the equation. It is a constant. It is a constant. These are the initial conditions; Output: Grain size d.

[0028] Example 2 A method for evaluating grain refinement, specifically for a target component manufactured by SLM additive manufacturing, is provided to assess the grain refinement of the target component; the method includes the following steps: B1: Perform a cutting operation on the target component to obtain the cutting performance parameters in the establishment method described in Example 1; B2: Using the cutting performance parameters from step B1 as input data, input them into the prediction model established using the method described in the embodiment; the prediction model outputs the predicted grain size d3. B3: Determine the degree of grain refinement based on whether the value of d3 falls within the grain size range.

[0029] It also includes the following steps: B4: Obtain the geometrically necessary dislocation density ρ3 and nanohardness H3 of the target component; B5: Calculate the overall grain refinement index ,in: ; have: The grain size of the forged sample in the method of any one of claims 1-6; The geometrically necessary dislocation density of the forged sample in the establishment method according to any one of claims 1-6; The nanohardness of the forged sample obtained by any one of the establishment methods described in 1-6; , , All are weighting coefficients, where: Furthermore, weights are assigned using the entropy weighting method; B6: If GFI ≥ 30%, then refine significantly; if 15% ≤ GFI < 30%, then refine moderately; if GFI < 15%, then refine lightly.

[0030] Using the evaluation method disclosed in this embodiment, two batches of target components made of nickel-based superalloy GH4169 manufactured by SLM additive manufacturing were evaluated, and the cutting performance parameters are shown in Table 2.

[0031] Table 2: Machining performance parameters of target components in two batches

[0032] Table 3 compares the predicted and measured data of the target components in the two batches.

[0033] Table 3: Comparison of predicted and measured data of target components in two batches

[0034] Table 3 shows that the prediction model established by this method has an error rate of less than 1.8% in predicting grain size. If the predicted error rate is less than 4.65%, it fully demonstrates the feasibility of the scheme, and thus the scheme has the following advantages: A grain size prediction model is established using an encoder-decoder architecture in a physical information neural network. This model constructs the relationship between cutting performance parameters and grain size. By cutting the target component to obtain cutting data, the grain size can be effectively predicted, allowing for a quantitative assessment of grain refinement and providing data for subsequent process parameter adjustments. Compared to existing methods described in the background art, this method shortens the cutting time of the target component, averaging 8-10 minutes per piece, far lower than the 4-6 hours required for metallographic sample preparation and observation. This improves the time for assessing grain refinement and enables online evaluation. Furthermore, this invention integrates multiple scales, including metallographic structure and cutting performance, resulting in more accurate grain size prediction and performance characterization. The cutting path can achieve full coverage, providing a more comprehensive anisotropic characterization of the grains.

[0035] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for establishing a grain size prediction model, wherein the prediction model is for a target component manufactured by SLM additive manufacturing and is used to predict the grain size of the target component, characterized in that: Includes the following steps: A1: Manufacturing a sample; Forged samples are manufactured by forging, and additive samples are manufactured by SLM additive manufacturing. A2: Parameter acquisition; Grain size d and cutting performance parameters were acquired for both forged and additive samples; A3: Predictive model training; using cutting performance parameters as input and grain size d as output, the encoder-decoder architecture in the physical information neural network is used as the predictive model for training until the loss function converges, thus completing the training of the predictive model.

2. The method for establishing according to claim 1, characterized in that: In step A2, the grain size d is obtained using an electron backscatter diffraction system.

3. The method for establishing according to claim 1, characterized in that: In step A2, the cutting performance parameters include at least the bending moment and polar area. First-order area moment fractal dimension Anisotropy index Spectral entropy Wavelet energy coefficients Cutting speed Feed per tooth Axial depth of cut Radial depth of cut Cutting time .

4. The method for establishing according to claim 3, characterized in that: In step A2, when obtaining cutting performance parameters, the bending moment polar coordinate diagram is used as the basis. The bending moment polar coordinate diagram is constructed with the tool rotation angle θ as the polar angle, and the bending moment amplitude is synthesized. Constructing for the polar radius; where: The magnitude of the bending moment in the X direction; This represents the bending moment amplitude in the Y direction.

5. The method for establishing according to claim 3, characterized in that: In step A2, the polar coordinate area of ​​the bending moment The formula for obtaining it is: ； First-order area moment The formula for obtaining it is: ； fractal dimension The fractal dimension of the polar coordinate plot is calculated using the box counting method; Anisotropy Index The formula for obtaining it is: ； in: This represents the maximum value of the combined bending moment amplitude M. This represents the minimum value of the combined bending moment amplitude M. Spectral entropy The formula for obtaining it is: ； in: The energy percentage of the i-th frequency band; The transient characteristics of cutting force are analyzed using continuous wavelet transform, and wavelet energy coefficients are extracted from these characteristics. .

6. The method for establishing according to claim 1, characterized in that: In step A3, the grain size d and cutting performance parameters are normalized to form a dataset; the dataset is randomly allocated into a training set and a test set; the prediction model is trained using the training set, and the trained prediction model is tested using the test set.

7. The method for establishing according to claim 6, characterized in that: In step A3, when training the prediction model, where: Input layer: Cutting performance parameters; Encoder: 3 fully connected layers, with ReLU activation function; Physical embedding layer: The physical constraints include at least Hall-Petch reinforcement constraints, dislocation reinforcement constraints, and physical residuals; among which, The Hall-Petch strengthening constraint is: ； in: For yield limit, The intracrystalline yield strength of a single-crystal matrix. For coefficients, This represents the average grain size. Cutting force ; in: A coefficient related to tool geometry, The axial cutting depth, This refers to the feed per tooth. The dislocation reinforcement constraint is: ； in: Taylor factor It is a constant. It is the elastic shear modulus. It is a constant. Dislocation density; Physical residuals for: ； in: Input the measured cutting force value. To strengthen the constraint cutting force model, For coefficients; Decoder: 2 fully connected layers; Loss function: ； in: For the grain size of the forged sample, For the grain size of the additive sample, R_physics represents the physical residual term in the equation. It is a constant. It is a constant. These are the initial conditions; Output: Grain size d.

8. A method for evaluating grain refinement, the method being used to evaluate the grain refinement of a target component manufactured by SLM additive manufacturing; characterized in that: Includes the following steps: B1: Perform a cutting operation on the target component to obtain the cutting performance parameters in the establishment method described in any one of claims 1-4; B2: Using the cutting performance parameters from step B1 as input data, input them into the prediction model established using the method described in any one of claims 1-4; the prediction model outputs the predicted grain size d3; B3: Determine the degree of grain refinement based on whether the value of d3 falls within the grain size range.

9. The evaluation method according to claim 8, characterized in that: It also includes the following steps: B4: Obtain the geometrically necessary dislocation density of the target component. and nano hardness ; B5: Calculate the overall grain refinement index ,in: ； have: The grain size of the forged sample in the method of any one of claims 1-6; The geometrically necessary dislocation density of the forged sample in the establishment method according to any one of claims 1-6; The nanohardness of the forged sample obtained by any one of the establishment methods described in 1-6; , , All are weighting coefficients, where: Furthermore, weights are assigned using the entropy weighting method; B6: If GFI ≥ 30%, then refine significantly; if 15% ≤ GFI < 30%, then refine moderately; if GFI < 15%, then refine lightly.

Images