Prediction method for microstructure morphology of 3D printing deposition layer
By combining deep learning with cellular automata, the problem of low efficiency in microstructure prediction in existing technologies is solved, efficient and explainable microstructure morphology prediction is achieved, and the prediction accuracy and adaptability are improved.
Patent Information
- Application Number
- CN202510774005.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-11
- Publication Date
- 2025-09-23
AI Technical Summary
In existing technologies, microstructure prediction relies on complex physical modeling, which is inefficient and has poor adaptability. Traditional machine learning methods ignore spatial coupling and the laws of microstructure evolution, resulting in distribution distortion and local error accumulation in the prediction results.
Combining deep learning with cellular automata, the cooling rate matrix is predicted through multi-layer deposition physical experiments, finite element simulation model calibration, and machine learning models. The grain evolution process is iteratively simulated, and the microstructural characteristics are analyzed using infrared thermal imaging and EBSD. The parameters are optimized to improve prediction accuracy.
Efficient and explainable prediction of microstructure morphology is achieved, and the grain distribution and directionality are highly consistent with physical experiments, with good predictive ability and physical consistency.
Smart Images
Figure BDA0005443785640000041 
Figure HDA0005443785650000011 
Figure HDA0005443785650000012
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of additive manufacturing and materials science, and in particular to a method for predicting the microstructure morphology of a three-dimensional printed deposited layer based on a combination of machine learning and cellular automata. Background Art
[0002] With the rapid development of additive manufacturing technology, metal 3D printing processes such as laser filament fusing are widely used in the manufacture and repair of high-performance structural components. The rapid cooling of the molten pool and the complex thermal cycles during the deposition process lead to the formation of complex microstructures within the material, such as columnar crystals, equiaxed crystals, and grain boundary orientation. These microstructures have a direct impact on the material's mechanical properties, fatigue life, and service reliability.
[0003] Existing technologies often rely on physical experiments or multiscale thermomechanical coupled numerical simulations, such as the finite element method and phase-field method. While these methods offer clear physical mechanisms, they are computationally intensive and highly parameter-dependent, making them difficult to meet the demands of high-throughput, diverse process optimization and microstructural prediction. Furthermore, while traditional machine learning methods possess some predictive capabilities, they often overlook spatial coupling and the evolution of microstructures, leading to distributional distortion and localized error accumulation in their predictions.
[0004] Using deep learning models for high-precision modeling of cooling rate fields, combined with cellular automata, can achieve efficient and interpretable predictions of microstructural morphology. This has been a hot research topic in recent years, but a systematic and stable framework is currently lacking. Therefore, an integrated prediction method combining data-driven models with grain evolution simulation is urgently needed to efficiently restore the microstructural morphology of the deposited layer based on printing process parameters. Summary of the Invention
[0005] To address the shortcomings of the existing technology, the present invention provides a method for predicting the microstructure morphology of 3D-printed deposits. This method combines deep learning with cellular automata to achieve this prediction. This invention aims to address the existing problems of microstructure prediction, which rely on complex physical modeling, suffer from low efficiency, and lack adaptability.
[0006] The technical solutions adopted in the present invention are as follows:
[0007] A method for predicting the microstructure morphology of a 3D printed deposited layer comprises the following steps:
[0008] Step 1: Conduct multi-layer deposition physical experiments and collect experimental data;
[0009] Step 2: Based on physical experiments, construct a finite element simulation model and calibrate the machine learning model;
[0010] Step 3: Build a machine learning model and use the machine learning model to predict the cooling rate matrix;
[0011] Step 4: The cellular automaton iteratively simulates the grain evolution process based on the cooling rate matrix predicted by the machine learning model to perform microstructure prediction.
[0012] Furthermore, infrared thermal imaging and high-temperature thermocouples are used to collect data on temperature changes over time during the deposition process to obtain the thermal cycle history, and finite element simulation is used to supplement the temperature distribution in locations that are difficult to measure directly.
[0013] Furthermore, electron backscatter diffraction (EBSD) was used to perform cross-sectional metallographic analysis on the sample to obtain microstructural characteristics such as grain orientation, size, aspect ratio, and area distribution.
[0014] Furthermore, after comparing the difference between the temperature measurement points in the physical experiment and the temperature measurement points in the simulation, the machine learning calibration model was used to fine-tune the boundary conditions in the finite element simulation.
[0015] Further, the process of step 2 is as follows:
[0016] Step 2.1: Statistically analyze the parameter types and ranges that need to be calibrated. After determining the relevant parameters that need to be optimized, use Latin hypercube to design experimental samples, and then perform batch simulation to obtain training data set samples;
[0017] Step 2.2: Build a machine learning calibration model for training, construct a mapping between optimization parameters and output temperature values, and use a multi-objective genetic algorithm to optimize parameters simultaneously for different power conditions. The objective function is to minimize the temperature error. From the Pareto front solution set, the solution closest to the Utopia point is selected as the optimal parameter set.
[0018] Step 2.3: Bring the optimal parameter set into the simulation model to complete the calibration of the model.
[0019] Furthermore, the training method of the neural network model is:
[0020] Step 3.1: Based on the calibrated model, calculate the cooling rates of different spatial points in the sedimentary layer and organize them into a cooling rate matrix of a fixed size, which serves as supervised labels for machine learning.
[0021] Step 3.2, take eight process parameters including laser power, scanning speed, wire feeding speed, interlayer time, and ambient temperature as the model input feature X;
[0022] The cooling rate matrix calculated at different spatial points is used as the prediction target. Based on the established automated batch simulation-data collection framework, the output cooling rate matrix is flattened into a vector form Y, and input and output sample pairs (Xi, Yi) are constructed to form a model training dataset.
[0023] Step 3.3: Build a CNN-Transformer hybrid neural network model, where the CNN module extracts the local interaction features between process parameters, and the Transformer module models their long-range influence on the overall distribution of the cooling rate matrix;
[0024] Step 3.4: Train the model based on the training dataset.
[0025] Step 3.5: Introduce the Bayesian optimization algorithm to automatically iterate the key hyperparameters of the model.
[0026] Furthermore, in step 3.4, the Huber loss function (Huber Loss) is used as the regression loss.
[0027] Furthermore, the key hyperparameters of the model include convolution kernel size, number of Transformer attention heads, hidden layer dimension, and learning rate, thereby improving prediction accuracy and generalization ability.
[0028] Furthermore, the process of microstructure prediction in step 4 is as follows:
[0029] Step 4.1, construct a two-dimensional cellular automaton grid that is consistent with the size of the predicted cooling rate matrix;
[0030] Step 4.2: Set up the grain nucleation, orientation, growth and competition mechanisms. The grain growth rate is controlled by the local cooling rate, reflecting the coupling mechanism of thermal influence and microstructure.
[0031] Step 4.3: Post-process the cooling rate matrix predicted by the machine learning model, convert it into time-temperature data that can be read by the cellular automaton, input it into the prediction model, iteratively simulate the grain evolution process, and output the predicted grain morphology image.
[0032] Beneficial effects of the present invention:
[0033] The predicted microstructure images were quantitatively compared with EBSD experimental results, including indicators such as average grain size, orientation distribution histogram, and area ratio. Verification results show that the proposed method is highly consistent with physical experiments in terms of grain distribution, orientation, and morphology, demonstrating good physical consistency and predictive power. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 Schematic diagram of the cooling rate matrix of the 3D printed deposited layer profile in the present invention.
[0035] Figure 2 The present invention compares the simulation data and the predicted results of each element of the temperature matrix of the prediction model in the form of a heat map. Figure 2 (a) is the simulation model extraction result, and (b) is the prediction model output result.
[0036] Figure 3 It is a scatter density diagram between the true value and the predicted value of the prediction model prediction result in the present invention.
[0037] Figure 4 This is a comparison between the physical experiment scanning results and the cellular automation prediction results in the present invention. Figure 4 (a) is the physical experiment scanning result, (b) is the physical scanning image of EBSD, and (c) is the cellular automation prediction result.
[0038] Figure 5 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0039] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0040] refer to Figure 1-5 This embodiment is a method for predicting the microstructure morphology of 3D printed deposits based on a combination of machine learning and physical modeling. The following is a detailed description of the method of the present invention with reference to the following examples:
[0041] Step 1: Laser additive manufacturing experiments were conducted using 10CrNi3MoV steel, a typical additive manufacturing alloy. Using industrial-grade laser deposition equipment, key process parameters such as laser power, scanning speed, wire feed speed, and interlayer cooling time were adjusted to produce single-pass, multi-layer deposition samples under different operating conditions. During the experiments, high-temperature thermocouples were used to measure the temperature at fixed points on the samples.
[0042] After the experiment is completed, the prepared samples are sampled, ground, polished and corroded, and the electron backscatter diffraction (EBSD) technology is used to obtain the grain morphology image of the deposited layer cross section, and the data such as grain orientation distribution, size, and area ratio are extracted as a microstructure reference.
[0043] Step 2: Build a finite element simulation model consistent with the physical experiment. After comparing the temperature measurement points in the physical experiment with those in the simulation, use a machine learning calibration model to fine-tune the boundary conditions in the finite element simulation. The optimal boundary parameter combination that minimizes error is then incorporated into the model, achieving simulation model calibration and improving accuracy.
[0044] The finite element method was used to simulate heat conduction during the laser deposition process. The temperature curves of each grid cell on the central cross-section of the deposited layer were obtained over time. From these curves, the cooling rate (i.e., the temperature gradient with respect to time) at each point was calculated. This was then constructed into a 10×20 two-dimensional matrix, which served as the prediction target for the machine learning model.
[0045] In step 3, each set of experimental process parameters (laser power, scanning speed, wire feed speed, interlayer time, ambient temperature, etc.) is used as the model input feature, and the cooling rate matrix at the corresponding position is expanded into a one-dimensional vector of length 200 as the output. The input and output data are normalized through standardization to ensure the numerical stability of model training. A CNN-Transformer hybrid neural network structure is also established, using the Huber loss function as the training objective. A Bayesian optimization algorithm is used to automatically adjust the model hyperparameters to obtain a model with strong generalization capabilities and high prediction accuracy.
[0046] In order to verify the effectiveness of the hybrid neural network proposed in this paper in the task of predicting the cooling rate of deposited layers in laser additive manufacturing, a long short-term memory network (LSTM), a pure CNN model, a Transformer model, and a graph convolutional neural network (GCN) were used as comparison methods. Comparative experiments were conducted under the same dataset and loss function. The comparison results are shown in Table 1. 2 The hybrid neural network proposed in the present invention performs best in terms of the three evaluation indicators of mean square error and root mean square error. The model of the present invention is significantly better than the above-mentioned comparison models in terms of overall prediction accuracy, spatial consistency and prediction stability in the high cooling rate area. Figure 3 As shown in the figure, in the cooling rate matrix prediction composed of 200 spatial nodes, almost all the prediction points of the proposed model fall within the relative error range of ±5%.
[0047] Step 4: Use the trained model to predict the input of the new process parameter combination and obtain the cooling rate distribution of each position in the deposited layer. The cooling rate matrix is input into the predefined grain growth model based on cellular automation to simulate the grain nucleation and growth process in different regions and output the grain orientation map, grain size distribution map, etc. The results are as follows: Figure 4 As shown in the figure, the image taken by optical microscope (a), the physical scanning image of EBSD (b) and the cellular automaton prediction image (c) are consistent in the grain orientation map and grain size distribution.
[0048] Table 1 Comparison of prediction accuracy evaluation indicators of different models
[0049]
[0050]
[0051] The microstructure images predicted by the present invention were quantitatively compared with EBSD experimental results, including indicators such as average grain size, orientation distribution histogram, and area ratio. Verification results show that the proposed method is highly consistent with physical experiments in terms of grain distribution, orientation, and morphology, demonstrating good physical consistency and predictive power.
[0052] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.
Claims
1. A method for predicting the microstructure morphology of a 3D printed deposited layer, characterized in that: The following steps are involved: Step 1: Conduct multi-layer deposition physical experiments and collect experimental data; Step 2: Based on physical experiments, construct a finite element simulation model and calibrate the machine learning model; Step 3: Build a machine learning model and use the machine learning model to predict the cooling rate matrix; Step 4: The cellular automaton iteratively simulates the grain evolution process based on the cooling rate matrix predicted by the machine learning model to perform microstructure prediction.
2. A method for predicting the microstructure morphology of a 3D printed deposited layer according to claim 1, characterized in that: Infrared thermal imaging and high-temperature thermocouples are used to collect data on temperature changes over time during the deposition process to obtain the thermal cycle history, and finite element simulation is used to supplement the temperature distribution in locations that are difficult to measure directly.
3. The method for predicting the microstructure morphology of a 3D printing deposited layer according to claim 1, wherein: Electron backscatter diffraction is used to perform cross-sectional metallographic analysis on the sample to obtain microstructural characteristics such as grain orientation, size, aspect ratio, and area distribution.
4. The method for predicting the microstructure morphology of a 3D printed deposited layer according to claim 1, wherein: After comparing the differences between the temperature measurement points in the physical experiment and the temperature measurement points in the simulation, the boundary conditions in the finite element simulation were fine-tuned using the machine learning calibration model.
5. The method for predicting the microstructure morphology of a 3D printing deposited layer according to claim 1, wherein: The process for step 2 is as follows: Step 2.1: Statistically analyze the parameter types and ranges that need to be calibrated. After determining the relevant parameters that need to be optimized, use Latin hypercube to design experimental samples, and then perform batch simulation to obtain training data set samples; Step 2.2: Build a machine learning calibration model for training, construct a mapping between optimization parameters and output temperature values, and use a multi-objective genetic algorithm to optimize parameters simultaneously for different power conditions. The objective function is to minimize the temperature error. From the Pareto front solution set, the solution closest to the Utopia point is selected as the optimal parameter set. Step 2.3: Bring the optimal parameter set into the simulation model to complete the calibration of the model.
6. The method for predicting the microstructure morphology of a 3D printing deposited layer according to claim 1, characterized in that: The training method of the neural network model is: Step 3.1: Based on the calibrated model, calculate the cooling rates of different spatial points in the sedimentary layer and organize them into a cooling rate matrix of a fixed size as supervised labels for machine learning; Step 3.2, take eight process parameters including laser power, scanning speed, wire feeding speed, interlayer time, and ambient temperature as the model input feature X; The cooling rate matrix calculated at different spatial points is used as the prediction target. Based on the established automated batch simulation-data collection framework, the output cooling rate matrix is flattened into a vector form Y, and input and output sample pairs (Xi, Yi) are constructed to form a model training dataset. Step 3.3: Build a CNN-Transformer hybrid neural network model, where the CNN module extracts the local interaction features between process parameters, and the Transformer module models their long-range influence on the overall distribution of the cooling rate matrix; Step 3.4: Train the model based on the training dataset. Step 3.5: Introduce the Bayesian optimization algorithm to automatically iterate the key hyperparameters of the model.
7. A method for predicting the microstructure morphology of a 3D printing deposited layer according to claim 6, characterized in that: In step 3.4, the Huber loss function is used as the regression loss.
8. A method for predicting the microstructure morphology of a 3D printed deposited layer according to claim 6, characterized in that: The key hyperparameters of the model include convolution kernel size, number of Transformer attention heads, hidden layer dimension, and learning rate, which can improve prediction accuracy and generalization ability.
9. The method for predicting the microstructure morphology of a 3D printing deposited layer according to claim 1, wherein: The process of microstructure prediction in step 4 is as follows: Step 4.1, construct a two-dimensional cellular automaton grid that is consistent with the size of the predicted cooling rate matrix; Step 4.2: Set up the grain nucleation, orientation, growth and competition mechanisms. The grain growth rate is controlled by the local cooling rate, reflecting the coupling mechanism of thermal influence and microstructure. Step 4.3: Post-process the cooling rate matrix predicted by the machine learning model, convert it into time-temperature data that can be read by the cellular automaton, input it into the prediction model, iteratively simulate the grain evolution process, and output the predicted grain morphology image.