Alloy material thermo-mechanical property multi-scale simulation method based on machine learning potential function

Abstract

The invention relates to the technical field of alloy material design and calculation, in particular to an alloy material thermo-mechanical property multi-scale simulation method based on a machine learning potential function. The method comprises the steps of S1, data preparation, S2, potential function construction, S3, multi-scale simulation and S4, result analysis. A potential function capable of accurately describing interaction between atoms of the alloy material is constructed through a machine learning algorithm, and accurate simulation and prediction of the mechanical and thermal properties of the alloy material are achieved in combination with a multi-scale simulation method. The technology provided by the invention has the advantages of high simulation precision, short research and development period, optimized material performance and the like, and is of great significance to design and optimization of alloy materials.

Description

Technical Field

[0001] The present invention relates to the field of alloy material design and computational technology, and in particular to a multi-scale simulation method for mechanical and thermal properties of alloy materials based on machine learning potential functions. Background Art

[0002] With the deepening of materials science research and the development of artificial intelligence technology, the use of machine learning methods to construct and fit potential functions has become the main technical approach to solve the inefficient trial and error problem in traditional potential function fitting.

[0003] Through machine learning potential functions, researchers can more accurately simulate the mechanical and thermal properties of alloy materials, providing strong support for the design and optimization of alloy materials.

[0004] At present, there is no complete multi-scale simulation method for the mechanical and thermal properties of alloy materials based on machine learning potential functions, and this gap needs to be filled urgently. Summary of the invention

[0005] The technical solution adopted by the present invention is as follows:

[0006] A multi-scale simulation method for mechanical and thermal properties of alloy materials based on machine learning potential function, specifically comprising the following steps:

[0007] S1 Data preparation: Collect structural information and thermal property data of several different alloys. The structural information includes but is not limited to crystal structure, composition ratio, lattice arrangement, and defect distribution; the thermal property data includes thermal conductivity and mechanical properties;

[0008] Preprocessing the collected data, wherein the preprocessing operations include any one or more combinations of cleaning, standardization, normalization, removal of outliers, and filling of missing values ​​to ensure data quality;

[0009] The preprocessed data is constructed as a training set for training the machine learning potential function; the training set data should cover the possible phase space of the alloy to ensure the generalization ability of the potential function;

[0010] S2 potential function construction: adopting machine learning algorithm and using training set for learning to obtain machine learning potential function that can accurately describe the interaction between alloy atoms and reflect the complex dependence of the interaction force between atoms; the machine learning algorithm includes any one of deep neural network and support vector machine;

[0011] When designing the neural network, the hierarchical structure is designed according to the complexity of the alloy, including the number of hidden layers, the number of nodes, and the choice of activation function;

[0012] When establishing a support vector machine model, the support vector machine model design includes the selection of kernel function, bandwidth, and regularization;

[0013] S3 multi-scale simulation: Use the machine learning potential function constructed in step S2 and combine it with multi-scale simulation software for simulation; the simulation considers scales from atomic scale to mesoscopic scale; simulate force and heat transport actions, including interatomic force, diffusion on crystal interface, dislocation slip, lattice deformation, heat transport, mesoscopic heat flow, and phase change;

[0014] S4 Result Analysis: After the simulation is completed, analyze the simulation results and extract the force and thermal properties, including force properties, thermal conductivity, mechanical response, phase change, thermal expansion, and thermal conductivity.

[0015] Preferably, after step S4, the simulation value is compared with the experimental value to verify the accuracy of the simulation; if the difference between the simulation value and the experimental value exceeds a set threshold, the data in step S1 is supplemented and then step S2 is performed to optimize the simulation method.

[0016] Preferably, after step S4, the relationship between the simulated mechanical and thermal properties and temperature and pressure is analyzed, and sensitivity analysis is carried out to determine the influence of factors on performance, so as to guide alloy design.

[0017] Preferably, in step S1, structural information of the alloy is collected through experiments or databases.

[0018] Preferably, in step S1, the data requirements for the training set include:

[0019] a. Composition and structure: The training set data includes the chemical composition ratios of several different alloys, including different element compositions, concentration ranges, and element ratios; the structural information includes ordered structures, disordered structures, and disordered structures include non-lattice arrangements and defective structures, and the defective structures include vacancies, dislocations, and interfaces;

[0020] b. State: The training set data includes state data under several different temperature and pressure conditions; the state data includes thermodynamic response, thermal conductivity, phase change, pressure-induced phase change, and lattice structure change data under high temperature and high pressure; the training set is set to simulate different temperature and pressure ranges, from room temperature to extreme conditions, to ensure that the model can predict the dynamic changes of alloys in various thermodynamic and thermal properties;

[0021] c. Dynamic and defect effects: The training set data introduces dynamic factor data, including defects and dynamic processes; specific dynamic and defect effect data include dislocation movement, diffusion, lattice hopping, interface migration, and defect dynamic behavior data; the simulation of the dynamic process can enhance the performance of the model in non-equilibrium state.

[0022] Preferably, in step S1, the mechanical property parameters include elastic modulus, hardness, yield point, and thermal expansion coefficient.

[0023] Preferably, in step S2, during the training process, a gradient descent algorithm is used to minimize the loss function; during the training, a cross-validation set is monitored to prevent overfitting.

[0024] Preferably, in step S3, simulation is performed using multi-scale simulation software, including VASP and LAMMPS.

[0025] Furthermore, during the simulation process of step S3, a parallel evolution algorithm is used to optimize the simulation efficiency, and the parallel evolution algorithm includes a time step algorithm.

[0026] Preferably, in step S3, the simulation includes the following contents:

[0027] a. Forces: simulate direct interactions between atoms, including bonding and repulsion, and determine the stability of matter;

[0028] b. Diffusion on the interface: Diffusion of atoms on the crystal interface; this is the key to the material growth and nucleation process, such as an important link in solid-state reactions;

[0029] c. Lattice deformation: the structural response of a material when it is subjected to load, reflecting plasticity, elasticity, and toughness;

[0030] d. Heat transport: including heat conduction, heat flow, study of thermal properties of materials, influence of heat dissipation and thermal management;

[0031] e. Phase change: The phase transition of materials under different conditions, including changes in solid and liquid morphology, affects the properties of the material.

[0032] This technology provides efficient tools for alloy material design and optimization through precision and short cycle, which is of great significance to the optimization of alloy performance. The machine learning potential function combined with multi-scale simulation provides accurate simulation for alloy science and provides strong support for new material discovery and performance design. In the entire simulation, the time step algorithm is the key. Through intelligent adjustment, the simulation efficiency and accuracy are ensured so that each time step can adapt to dynamic events at different scales, such as atomic motion, lattice diffusion, heat conduction, etc., and can be accurately captured. In this way, through algorithm optimization, the simulation finds the best balance in simulation time and spatial resolution, ensuring high efficiency and accuracy from microscopic to mesoscopic scales.

[0033] Beneficial effects of the present invention:

[0034] 1. Improve simulation accuracy: Through machine learning potential functions, the interaction between atoms in alloy materials can be more accurately described, thereby improving the accuracy and reliability of simulation results.

[0035] 2. Shorten the R&D cycle: The multi-scale simulation method can complete a large number of simulation experiments in a shorter period of time, thereby shortening the R&D cycle of alloy materials.

[0036] 3. Optimize material properties: Providing suggestions for the design and optimization of alloy materials based on simulation results can help researchers better understand and control the mechanical and thermal properties of alloy materials and achieve optimization of material properties. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 : Construct a machine learning potential flow chart;

[0038] Figure 2 : Multi-scale simulation of mechanical and thermal properties of alloy materials;

[0039] Figure 3 : Training set construction flow chart;

[0040] Figure 4 : Sensitivity analysis chart.

[0041] Figure 5 : (a) Overall planning flow chart (b) Dataset construction (c) Performance of machine learning training, including the evolution of the loss function and the differences between the force, energy, and virial calculated by DTF, respectively.

[0042] Figure 6 :(a) Uniaxial tensile stress-strain stretching along

[100] ,

[110] and

[111] directions for random model and short-range order model at 300K (b) Local shear strain distribution of random model and short-range order model when the tensile strain is 40% (c) Spatial distribution of atoms with shear strain exceeding 50%. (d) Generalized superposition failure curves of random samples and short-range order samples (e) Schematic diagram of Peierls stress calculation for random model and short-range order model and their corresponding shear stress-strain curves. (f) Stress-strain curve of shear process. DETAILED DESCRIPTION

[0043] The design concept and formation mechanism of the present invention are further described in detail below in conjunction with the accompanying drawings and exemplary embodiments to make the technical solution of the present invention clearer.

[0044] Example 1

[0045] Taking the simulation design of ternary W-Mo-V alloy as an example, the specific operations include the following:

[0046] 1. Data collection and collation: The specific process of collecting W-Mo-V alloy information and thermal property data is achieved through a series of precise experiments and calculation methods. First, X-ray diffraction technology (XRD) and electron microscopy (TEM) are used to observe and analyze the crystal structure of the ternary W-Mo-V alloy, including lattice arrangement and defect distribution; the composition information is determined by energy spectrum analysis. Thermal property data is measured by a thermal conductivity meter, and mechanical properties are tested by a universal testing machine to test the yield point. Subsequently, all collected data are integrated and archived to ensure data integrity.

[0047] 2. Preprocessing: The details of data preprocessing involve four key steps: data cleaning, standardization, missing value processing, and outlier processing. In the data cleaning stage, statistical analysis is used to identify and remove outliers within the abnormal range, including data points outside the cluster, to ensure data accuracy; missing value processing uses average value filling (mean replacement), interpolation, or model prediction filling; standardization makes the data set centralized, using minimum-maximum value normalization and mean-standardization to ensure that the data is comparable and aligned on the same scale.

[0048] 3. Training set construction: To ensure the generalization of the potential function, the key is to construct a training set that covers a wide range of phase spaces of alloy materials. This process involves the following detailed steps:

[0049] a. Composition and structural diversity: First, ensure that the training set should include a wide range of chemical composition ratios of ternary W-Mo-V alloys, representing different composition ranges, and determine data points within WxMoyVz (x+y+z=100). Structural information is not limited to common ordered structures, but also includes non-lattice arrangements, including disorder, defects, such as vacancies, dislocations, interfaces, etc., as well as complex characteristics at phase interfaces, which helps the model learn the behavior of materials under different conditions.

[0050] b. State coverage: Temperature and pressure conditions: The training set also needs to simulate room temperature -2000℃ and pressure 0-20GPa, and determine the temperature and pressure condition data points within the above range to ensure that the model can predict the dynamic changes of various thermodynamic and thermal properties of the alloy. This includes thermodynamic response, thermal conductivity, phase change, pressure-induced phase change, and lattice structure changes under high temperature and high pressure.

[0051] c. Dynamic and defect effects: The introduction of dynamic factors includes defects and dynamic processes, including dislocation movement, diffusion, lattice hopping, interface migration, and defect dynamic behavior. The simulation of these dynamic processes can enhance the performance of the model in non-equilibrium states.

[0052] d. Diversified data points: Through the above steps, the diversity of the data set is ensured, and each training point represents the alloy in the possible state space, so that the model generalizes rather than being limited to a specific configuration. This requires that the data points in the training set be as varied as possible, but at the same time remain efficient and avoid redundancy, and achieve model generalization by carefully selecting key representative data points.

[0053] e. Verification and iteration: Verify the training results through model testing and validation set data, and continuously iterate and adjust the training set to ensure generalization ability. If the model has poor prediction performance in the uncovered phase space, it may be necessary to go back and add more data points to further refine the training to ensure the accuracy and generalization of the model in unknown areas.

[0054] 4. Machine learning algorithm: In machine learning, neural network design, especially the core of deep learning, is built through multi-layer neural networks (DNNs) and convolutional networks (CNNs), and each layer learns complex patterns. In the hidden layer, the activation function ReLU (Rectified Linear Unit) is used to promote nonlinear activation and increase the learning ability of the network, and the Sigmoid function is used for the probability output layer mapping transformation to handle the ternary classification problem.

[0055] 5. Constructing potential function: Through the training process, the gradient optimization algorithm (Gradient Descent Optimization) is used to gradually adjust the weights, aiming to minimize the loss function (Mean Squared Error). In this way, the model gradually converges and finally obtains a potential function that can accurately describe the atomic interactions between alloys. Constructing the potential function is a machine learning model obtained through the training process, specifically embodied in the neural network potential (NNPs), which can accurately simulate the interaction force between atoms and lay a solid foundation for multi-scale simulation.

[0056] Specific to this embodiment:

[0057] The W-Mo-V dataset consists of 4247 structures, which are divided into training and test sets in a ratio of 9:1. The dataset covers a wide range of compositions from unary systems to ternary systems, effectively covering the composition and configuration space of WMoV alloys.

[0058] Figure 5 (a) is the overall scheme of this study. First, we constructed the training and test sets required for machine learning. Then, we trained the dataset and evaluated the difference between the predicted values ​​of the potential function and DFT. Finally, we performed parameter adjustments. Figure 5 (b) illustrates that the database includes different crystal structures of single and multi-components. Figure 5(c) illustrates the evolution of the loss function with respect to generations during training. The machine learning model for the WMoV alloy shows a high level of accuracy with a root mean square error (RMSE) of 15.6 meV / atom for energy, force, and virial, respectively. and 92.3 meV / atom. In addition, the lattice constants predicted by the machine learning model match well with those calculated by DFT.

[0059] 6. Multi-scale simulation: The multi-scale simulation process uses the micro-scale first-principles calculation software VASP and the meso-scale molecular dynamics simulation software LAMMPS to achieve a full range of simulations from the atomic scale to the meso-scale. The simulation covers a wide range of physical phenomena, including:

[0060] a. Forces: simulate direct interactions between atoms, including bonding and repulsion, which determine the stability of matter;

[0061] b. Diffusion on the interface: Diffusion of atoms on the crystal interface, which is the key to the material growth and nucleation process, such as an important link in solid-state reactions;

[0062] c. Lattice deformation: the structural response of a material when it is subjected to load, reflecting plasticity, elasticity, and toughness;

[0063] d. Heat transport: including heat conduction and heat flow, to study the thermal properties of materials and influence heat dissipation and thermal management;

[0064] e. Phase change: examine the transformation of materials under different conditions, including solid and liquid changes, which affect the properties of the materials.

[0065] To evaluate the origin of the increase in yield strength due to short-range order, the resistance to dislocation nucleation and slip was calculated. Figure 6 (d) shows the generalized stacking fault energy curves for the models with and without short-range order. The model with short-range order shows an increase in unstable stacking fault energy of 161 mJ / m compared to the random sample. 2 In the short-range order model, clusters with stronger bonding ability, including WW and Mo-V pairs, are shown to increase the anti-slip property, thereby suppressing dislocation nucleation. Peierls stress is also calculated to evaluate the pinning effect of lattice friction on dislocations. Figure 6 As shown in (e), the Burgers vector along the screw dislocation

[111] direction is placed in the model with and without short-range order, and shear strain is applied along the dislocation line. The stress-strain curve of the shear process is shown in Figure 6 (f) Short-range order leads to an increase in Peierl stress of 0.4 GPa. Therefore, the increase in yield strength is attributed to the increase in slip barriers and the enhancement of lattice friction.

[0066] 7. Sensitivity analysis: After simulation, extract the thermal and mechanical properties, conductivity, response, and mechanical properties, compare with the experiment, analyze the temperature, pressure, sensitivity, and identify the key factors that cause the impact.

Claims

1. A multi-scale simulation method for mechanical and thermal properties of alloy materials based on machine learning potential function, characterized in that: The specific steps include: S1 Data preparation: Collect structural information and thermal property data of several different alloys, including crystal structure, composition ratio, lattice arrangement, and defect distribution; The thermal property data include thermal conductivity and mechanical properties; Preprocessing the collected data includes any one or more operations of cleaning, standardization, normalization, removing outliers, and filling missing values; The preprocessed data is constructed as a training set for training the machine learning potential function; S2 potential function construction: adopt machine learning algorithm, use training set to learn, obtain machine learning potential function describing the interaction between alloy atoms, reflecting the complex dependence of the interaction force between atoms; the machine learning algorithm includes deep neural network and support vector machine; S3 multi-scale simulation: Use the machine learning potential function constructed in step S2 and combine it with multi-scale simulation software for simulation; the simulation considers scales from atomic scale to mesoscopic scale; Simulate force and heat transport, including interatomic force, diffusion on crystal interface, dislocation slip, lattice deformation, heat transport, mesoscopic heat flow, and phase change; S4 Result Analysis: After the simulation is completed, analyze the simulation results and extract the force and thermal properties, including force properties, thermal conductivity, mechanical response, phase change, thermal expansion, and thermal conductivity.

2. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 1 is characterized in that: After step S4, the simulation value is compared with the experimental value to verify the accuracy of the simulation; if the difference between the simulation value and the experimental value exceeds a set threshold, the data in step S1 is supplemented and then step S2 is performed to optimize the simulation method.

3. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 1 is characterized in that: After step S4, the relationship between the simulated mechanical and thermal properties and temperature and pressure is analyzed, and sensitivity analysis is carried out to determine the influence of factors on performance and guide alloy design.

4. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 1 is characterized in that: In step S1, the structural information of the alloy is collected through experiments or databases; the data requirements for the training set include: a. Composition and structure: The training set data includes the chemical composition ratios of several different alloys, including different element compositions, concentration ranges, and element ratios; the structural information includes ordered structures, disordered structures, and disordered structures include non-lattice arrangements and defective structures, and the defective structures include vacancies, dislocations, and interfaces; b. State: The training set data includes state data under several different temperature and pressure conditions; the state data includes thermodynamic response, thermal conductivity, phase change, pressure-induced phase change, and lattice structure change data under high temperature and high pressure; the training set is set to simulate different temperature and pressure ranges, from room temperature to extreme temperature and extreme pressure conditions; c. Dynamic and defect effects: The training set data introduces dynamic factor data, including defects and dynamic processes; specific dynamic and defect effect data include dislocation movement, diffusion, lattice jumping, interface migration, and defect dynamic behavior data.

5. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 1 is characterized in that: In the step S2, when designing the neural network, the hierarchical structure is designed according to the complexity of the alloy, including the number of hidden layers, the number of nodes, and the selection of activation functions; when establishing the support vector machine model, the support vector machine model design includes the selection of kernel function, bandwidth, and regularization.

6. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 1 is characterized in that: In the step S1, the mechanical property parameters include elastic modulus, hardness, yield point, and thermal expansion coefficient.

7. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 1 is characterized in that: In step S2, during the training process, a gradient descent algorithm is used to minimize the loss function; during the training, a cross-validation set is used for monitoring to prevent overfitting.

8. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 1 is characterized in that: In the step S3, simulation is performed using multi-scale simulation software, including VASP and LAMMPS.

9. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 8, characterized in that: During the simulation process of step S3, a parallel evolution algorithm is used to optimize the simulation efficiency, and the parallel evolution algorithm includes a time step algorithm.

10. The multi-scale simulation method of mechanical and thermal properties of alloy materials based on machine learning potential function according to claim 1, characterized in that: In step S3, the simulation includes the following contents: a. Force action: simulate direct interactions between atoms, including bonding and repulsion; b. Diffusion on the interface: atomic diffusion on the crystal interface; c. Lattice deformation: the structural response of a material when it is subjected to load; d. Heat transport: including heat conduction and heat flow; e. Phase change: the phase change of materials under different conditions.

Images