Convolutional neural network-based additive manufacturing zirconium alloy magnetic susceptibility prediction method

Abstract

The invention discloses an additive manufacturing zirconium alloy magnetic susceptibility prediction method based on a convolutional neural network, and belongs to the technical field of material performance prediction and intelligent manufacturing. The method comprises the following steps: firstly, preparing a plurality of groups of zirconium alloy samples with different process parameters through an electron beam powder bed melting process; secondly, performing X-ray diffraction analysis on the sample to obtain normalized diffraction intensity data of each crystal face, and measuring the magnetic susceptibility of each crystal face; thirdly, constructing a convolutional neural network model, and performing model training by taking the normalized diffraction intensity data and the phase composition data as input features and the magnetic susceptibility as an output label; and finally, predicting the zirconium alloy magnetic susceptibility under unknown process parameters by using the trained model. According to the method, the CNN model reveals that the crystal orientation is a leading factor influencing the magnetic susceptibility under the condition that the material components are the same, efficient and accurate prediction of the magnetic susceptibility of the zirconium alloy is realized, and a new way is provided for design and optimization of an MRI compatible implant material.

Description

Technical Field

[0001] Zirconium alloys, due to their excellent biocompatibility, corrosion resistance, and low magnetic susceptibility, have broad application prospects in the field of biomedical implantation. Magnetic susceptibility is a key factor affecting the size of artifacts in magnetic resonance imaging (MRI) of implanted materials; materials with low magnetic susceptibility can significantly reduce MRI artifacts and improve image quality. Currently, the magnetic susceptibility of additively manufactured zirconium alloys is significantly affected by process parameters, and traditional trial-and-error optimization methods are inefficient and difficult to achieve synergistic control of multiple properties. Convolutional neural networks (CNNs), as a deep learning model, have shown powerful capabilities in predicting the correlation between material structure and properties. Existing research mainly focuses on the influence of phase composition on magnetic susceptibility, lacking systematic analysis of crystal orientation and texture intensity. Therefore, developing a CNN-based magnetic susceptibility prediction method is of great significance for promoting the intelligent design and manufacturing of MRI-compatible zirconium alloys. Background Technology

[0002] Zirconium alloys, especially zirconium alloys, are considered ideal candidates for next-generation MRI-compatible orthopedic implants due to their excellent biocompatibility, corrosion resistance, and inherently low magnetic susceptibility. In clinical MRI examinations, the difference in magnetic susceptibility between implants and human tissue can lead to local magnetic field distortion, producing artifacts and severely interfering with the diagnosis of tissues surrounding the implant. Therefore, developing implant materials with ultra-low and controllable magnetic susceptibility is crucial. However, material magnetic susceptibility is a performance indicator influenced by a complex combination of factors. For zirconium alloys prepared using additive manufacturing techniques such as electron beam powder bed melting, their final magnetic susceptibility is significantly affected by a complex chain: process parameters (such as energy density and scan speed) → microstructure (phase composition, grain size and morphology) → crystallographic texture. Traditional trial-and-error methods, involving numerous experiments to explore the process-performance relationship, are not only time-consuming and costly but also struggle to analyze the nonlinear interactions between multiple variables, making precise directional control of magnetic susceptibility impossible.

[0003] In recent years, artificial intelligence, especially deep learning technology, has provided revolutionary tools for revealing the complex relationships between material processing, microstructure, and properties. In the field of additive manufacturing, deep learning has been successfully applied to predict microstructure, optimize process parameters, and reduce defects. In predicting magnetic properties, artificial intelligence methods can mine massive amounts of data to establish hidden relationships between material composition, structure, and magnetic properties, overcoming some limitations of traditional theoretical models. Existing research largely focuses on adjusting phase composition by changing alloy composition, thereby affecting magnetic susceptibility, often neglecting crystal texture—a factor that is easily formed during additive manufacturing and has a potentially significant impact on magnetic susceptibility. Developing a method that can efficiently and accurately predict the magnetic susceptibility of additively manufactured zirconium alloys and clearly reveal key influencing factors (especially texture) has significant theoretical and engineering value for accelerating the development of high-performance MRI-compatible implant materials. Summary of the Invention

[0004] This invention aims to provide a method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks, so as to achieve rapid and accurate prediction of the magnetic susceptibility of zirconium alloys and provide theoretical support and technical means for their application in biomedical implant materials.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks includes the following steps:

[0007] 1) Prepare multiple sets of zirconium alloy samples with different process parameters through metal additive manufacturing processes such as electron beam powder bed melting, laser powder bed melting, laser direct deposition, and selective laser sintering;

[0008] 2) Perform X-ray diffraction tests on the alloy sample prepared in step 1) to obtain the normalized diffraction intensity data of each crystal plane and measure the magnetic susceptibility value of the alloy sample.

[0009] 3) Construct a convolutional neural network model, using the normalized diffraction intensity data and phase composition data obtained in step 2) as input features and the magnetic susceptibility value as the output label, and train the model.

[0010] 4) Using a trained convolutional neural network model, the magnetic susceptibility of zirconium alloys prepared under unknown process parameters is predicted.

[0011] The method for predicting the magnetic susceptibility of additive manufacturing zirconium alloys based on convolutional neural networks, wherein the alloy mentioned in step 1) is an additive manufacturing zirconium alloy including alpha-type zirconium alloy, near-alpha-type zirconium alloy, alpha+beta-type zirconium alloy, metastable beta-type zirconium alloy, or beta-type zirconium alloy.

[0012] The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks, wherein the process parameters in step 1) include scanning speed, energy density, chamber spacing, layer thickness, and energy density.

[0013] The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks, wherein the normalized diffraction intensity data in step 2) includes the normalized intensity values ​​of each crystal plane of α-Zr.

[0014] The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks, wherein step 3) involves training the convolutional neural network model using a loss function for a regression task. This loss function includes at least one or more of the following: mean square error (MSE) loss function, root mean square error (RMSE) loss function, mean absolute error (MAE) loss function, Huber loss function / Smooth L1 loss function, LogCosh loss function, and / or quantile / Pinball Loss. The network weights are updated using a backpropagation algorithm combined with gradient descent optimization methods until the model converges.

[0015] The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks, wherein the input features in step 3) also include the phase composition factor of the sample, which is calculated by XRD spectrum.

[0016] The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks describes the application of the trained convolutional neural network model in predicting the magnetic susceptibility of zirconium alloys in additive manufacturing.

[0017] The method for predicting the magnetic susceptibility of additively manufactured zirconium alloys based on convolutional neural networks is applied in the design and optimization of MRI-compatible zirconium alloy implant materials.

[0018] Advantages and beneficial effects of the present invention:

[0019] 1. The CNN model constructed in this invention achieves accurate prediction of the magnetic susceptibility of zirconium alloys by directly learning the intrinsic mapping relationship between X-ray diffraction texture data obtained from experimental samples and magnetic susceptibility. Verification shows that the deviation between the model's predicted values ​​and experimental measurements can be controlled within 8%, and the predicted parameter range is not limited to the range used in training, demonstrating high engineering reliability.

[0020] 2. The findings of this invention differ from the traditional view that "phase composition" is the dominant factor affecting the magnetic susceptibility of zirconium alloys. Through interpretability analysis using a CNN model, this invention, for the first time, clearly reveals that for zirconium alloys with a fixed composition, the intensity of their crystallographic texture, particularly the orientation distribution of crystal planes such as (100), (101), and (201), is the most critical factor determining their magnetic susceptibility. This discovery provides a new and more direct direction for precisely controlling magnetic susceptibility through process control of texture.

[0021] 3. This method establishes a rapid prediction channel from readily available XRD data to final magnetization performance. Compared to time-consuming and labor-intensive traditional trial-and-error experiments, this invention can effectively evaluate and screen magnetization based solely on the texture characteristics that may be caused by process parameters before preparing physical samples, thereby greatly shortening the material development cycle and reducing experimental costs.

[0022] 4. The CNN model used in this invention is not only a high-performance prediction tool, but its kernel weight analysis also endows the model with good interpretability, enabling its conclusions to serve a deeper understanding of physical mechanisms. Furthermore, the model maintains good predictive ability for samples outside the training parameter range, demonstrating its strong generalization ability and suitability for exploring complex process windows in practical engineering.

[0023] 5. This method provides an efficient technical platform for the intelligent design and performance optimization of MRI-compatible implant materials. By combining this predictive model with a process simulation model, it is expected to enable the pre-assessment and design of magnetic susceptibility and MRI artifacts in complex-shaped implants, thus promoting the development of personalized, high-performance medical implants. Attached Figure Description

[0024] Figure 1 A schematic diagram illustrating the construction of a method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks.

[0025] Figure 2 This is a flowchart illustrating the properties of zirconium powder and electron beam enhancement in Example 1.

[0026] Figure 3 These are normalized crystal orientation intensity data maps used to train the CNN model in Example 1: (a) (0 0 2), (b) (1 0 0), (c) (1 0 1), (d) (1 0 2), (e) (1 0 3), (f) (1 1 0), (g) (1 1 2), (h) (2 0 1)).

[0027] Figure 4 The following are examples from Example 1: (a) Evolution of results differences during CNN model training, (b) Weights of each influencing factor, (c) Evolution of output results during training, (d) Comparison of predicted and actual magnetic susceptibility curves, and (e) Deviation measurement results of predicted and actual measurement data for four non-training samples.

[0028] Figure 5 These are MRI artifact test images and quantitative analysis results of the low magnetic susceptibility samples screened in Example 2, with biomedical Ti-6Al-4V as the reference sample. Detailed Implementation

[0029] The present invention will be further described below with reference to the embodiments, but the scope of protection of the present invention is not limited thereto.

[0030] Example 1:

[0031] The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks in this embodiment includes the following steps:

[0032] 1. An electron beam powder bed melting system was adopted, with 40 different process parameters set (energy density 25-230 J·mm). -3 Scanning speed 1000-8000 mm·s -1 Zirconium alloy cylindrical samples were prepared. Specific process parameters are shown in Table 1.

[0033] Table 1

[0034]

[0035] 2. Obtain the XRD patterns of each sample using an X-ray diffractometer, and calculate the normalized diffraction intensity (I) of each crystal plane (including (100), (101), (201), etc.) using the following formula (1). std (hkl)), where I0 is the derivatization intensity of the annealed zirconium powder, and I(hkl) is the measured diffraction intensity of the (hkl) crystal plane. The phase composition factor was obtained after normalizing the volume ratio of α-Zr to β-Zr. The magnetic susceptibility of each sample in a magnetic field from -1.5 T to 1.5 T was measured using a vibrating sample magnetometer.

[0036] (1)

[0037] 3. A convolutional neural network (CNN) model is constructed to obtain the prediction formula for additive manufacturing of zirconium alloys. The input features are normalized intensity data of eight crystal planes and phase composition factors. The model can linearly combine nine input matrices to generate an output matrix corresponding to the magnetic susceptibility. The CNN uses the backpropagation algorithm to adjust the weights, achieving efficient and accurate fitting of the data. Backpropagation calculates the gradient of the loss function with respect to each weight using the chain rule, enabling the model to learn from errors and optimize the prediction results. The model training uses the mean squared error loss function, which improves prediction accuracy by penalizing larger errors and is particularly suitable for regression tasks. The trained model generates a set of weight coefficients, representing the contribution of each input factor to the magnetic susceptibility. These weight coefficients converge after 2000 iterations, are extracted from the convolution kernel, and after normalization, finally obtain the prediction equation for the magnetic susceptibility of additively manufactured zirconium alloys.

[0038] 4. Four sets of process parameters that were not included in the training were selected (energy densities of 40, 75, 200, and 230 J·mm⁻). 3Verification samples were prepared, and the measured XRD data were input into a trained CNN model to predict their magnetic susceptibility. The results showed that the deviation between the predicted and experimental values ​​was between 0.2% and 8.0%, indicating that the model has high accuracy and generalization ability.

[0039] Example 2:

[0040] The method of this invention was applied to the MRI compatibility design of zirconium alloy implant materials. A CNN model was used to predict the magnetic susceptibility under different orientation characteristics, and materials with magnetic susceptibility below 1.0 × 10⁻⁶ were selected. -6 cm 3 ·g -1 By utilizing the process window, biomedical alloy samples with artifact volume reduced by approximately 58% compared to Ti-6Al-4V alloy were successfully prepared.

Claims

1. A method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks, characterized in that, Includes the following steps: 1) Prepare multiple sets of zirconium alloy samples with different process parameters through metal additive manufacturing processes such as electron beam powder bed melting, laser powder bed melting, laser direct deposition, and selective laser sintering; 2) Perform X-ray diffraction tests on the alloy sample prepared in step 1) to obtain the normalized diffraction intensity data of each crystal plane and measure the magnetic susceptibility value of the alloy sample. 3) Construct a convolutional neural network model, using the normalized diffraction intensity data and phase composition data obtained in step 2) as input features and the magnetic susceptibility value as the output label, and train the model. 4) Using a trained convolutional neural network model, the magnetic susceptibility of zirconium alloys prepared under unknown process parameters is predicted.

2. The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks according to claim 1, characterized in that: The alloy mentioned in step 1) is an additive manufacturing zirconium alloy, including alpha-type zirconium alloy, near-alpha-type zirconium alloy, alpha+beta-type zirconium alloy, metastable beta-type zirconium alloy, or beta-type zirconium alloy.

3. The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks according to claim 1, characterized in that: The process parameters mentioned in step 1) include scanning speed, energy density, chamber spacing, layer thickness, and energy density.

4. The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks according to claim 1, characterized in that: The normalized diffraction intensity data mentioned in step 2) includes the normalized intensity values ​​of each crystal plane of a-Zr.

5. The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks according to claim 1, characterized in that: The convolutional neural network model described in step 3) is trained using a loss function for regression tasks. The loss function includes at least one or more of the following: mean squared error (MSE) loss function, root mean square error (RMSE) loss function, mean absolute error (MAE) loss function, Huber loss function / smooth L1 loss function, log-hyperbolic cosine loss function, and / or quantile / pinball loss function. The network weights are updated using a backpropagation algorithm combined with gradient descent optimization methods until the model converges.

6. The method for predicting the magnetic susceptibility of zirconium alloys in additive manufacturing based on convolutional neural networks according to claim 1, characterized in that: The input features mentioned in step 3) also include the phase composition factor of the sample, which is calculated by XRD spectrum.

7. An application of a convolutional neural network model trained according to any one of claims 1 to 6 in predicting the magnetic susceptibility of zirconium alloys in additive manufacturing.

8. The application of the method according to any one of claims 1 to 6 in the design and optimization of MRI-compatible zirconium alloy implant materials.

Images