Prediction method and system of tensile properties of high entropy alloys based on molecular dynamics simulation

Abstract

The present invention proposes a method and system for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation, which belongs to the technical field of predicting the tensile properties of high entropy alloys; by combining molecular dynamics data into a data set, and dividing the data set into a training set and an evaluation set; training and evaluating a densely connected neural network based on the training set and the evaluation set; calculating a potential function using the densely connected neural network after evaluation; simulating the motion trajectory of atoms during the tensile fracture of the high entropy alloy based on a LAMMPS simulator; predicting the tensile properties of the high entropy alloy according to the simulation results of the motion trajectory. The present invention can achieve accurate prediction of the tensile properties of high entropy alloys on the basis of enhancing the generalization ability of the model.

Description

Technical Field

[0001] The present invention belongs to the technical field of high entropy alloy tensile property prediction, and in particular relates to a high entropy alloy tensile property prediction method and system based on molecular dynamics simulation. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The chemical composition and microstructure of high entropy alloys are extremely complex, which poses great difficulties for actual tensile fracture simulation. In order to solve this problem, technicians in related fields began to use LAMMPS simulators to simulate the tensile fracture of high entropy alloys, so as to reveal the mechanical behavior and fracture mechanism of high entropy alloy materials at the atomic scale, providing important guidance for basic research and practical applications.

[0004] However, the existing method of predicting the tensile properties of high entropy alloys through molecular dynamics simulation based on the LAMMPS simulator still has some technical problems, such as:

[0005] (1) FitSnap, as a new generation of machine learning potential construction tool, is a good assistant for molecular dynamics simulation. The existing method for predicting the tensile properties of high entropy alloys is to use FitSnap software to establish potential functions and prepare for molecular dynamics simulation; however, the FitSnap software used in the existing method is embedded with an MLP (multi-layer perceptron) network architecture. Although the MLP network is a universal function approximator, in an atomic system, the degree of freedom of the system will increase significantly with each additional atom, causing the feature space to expand exponentially, ultimately causing the approximation efficiency of the MLP network to drop sharply in high-dimensional space. Therefore, the MLP network has inherent limitations in describing high-dimensional nonlinear atomic environments, making it difficult to accurately predict the tensile properties of high-entropy alloys.

[0006] (2) In the process of predicting the tensile properties of high-entropy alloys, on the one hand, due to the hugeness of the configuration space, the number of possible configurations of the atomic system grows exponentially, which may cause the training set to be unable to cover all possible situations; on the other hand, high-entropy alloy materials may mutate under certain conditions, which requires the model to accurately capture these nonlinear behaviors. However, the MLP network embedded in the existing FitSnap software often finds it difficult to achieve a good balance between the model's bias and variance when dealing with high-dimensional nonlinear problems, resulting in overfitting or underfitting, and ultimately leading to insufficient generalization ability.

[0007] (3) Atomic interactions in high entropy alloy material systems usually exhibit significant multi-scale characteristics, including short-range interactions (mainly determined by chemical bonds and electron orbital overlap, and sensitive to local structure), medium-range interactions (involving next-nearest neighbor atoms, affecting the elastic and thermal properties of the material), and long-range interactions (such as Coulomb force and van der Waals force, which have an important influence on the macroscopic properties of the material). Accurately capturing these multi-scale characteristics is crucial for simulating the structural, dynamic, and thermodynamic properties of high entropy alloy materials. However, the MLP network embedded in the existing FitSnap software can theoretically fit any complex function, but in practical applications, its feature extraction ability is limited by its own network structure and training method, and it is difficult to capture long-range correlations and information at different spatial scales; therefore, the feature extraction ability is weak.

[0008] (4) The existing method for predicting the tensile properties of high-entropy alloys is based on the LAMMPS software to simulate the molecular motion inside the high-entropy alloy material. However, the Snap (Spectral Neighbor Analysis Potential) force field, as the core module of LAMMPS, has an extremely high computational complexity, which will greatly limit the speed and scale of the simulation. Summary of the invention

[0009] In order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a method and system for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation, which can achieve accurate prediction of the tensile properties of high entropy alloys on the basis of enhancing the generalization ability of the model.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] The first aspect of the present invention provides a method for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation.

[0012] The prediction method of high entropy alloy tensile properties based on molecular dynamics simulation includes:

[0013] Collect the calculation tasks for tensile property prediction of high entropy alloy, and send the calculation tasks to the master core; start the slave core thread library based on the master core, and distribute the calculation tasks to multiple slave cores; execute the tensile property prediction of high entropy alloy based on the slave core, specifically:

[0014] Acquire a data set containing molecular dynamics data of a high entropy alloy, and convert the format of the obtained data set;

[0015] Performing data partitioning and parallel data preprocessing on the data set after format conversion; wherein the data set after data partitioning includes a training set and an evaluation set;

[0016] Constructing a densely connected neural network, training and evaluating the densely connected neural network based on the training set and the evaluation set; and calculating the potential function of the high entropy alloy using the densely connected neural network after the evaluation;

[0017] The obtained potential function is input into a LAMMPS simulator, and the motion trajectory of atoms during the tensile fracture of the high entropy alloy is simulated based on the LAMMPS simulator; and the tensile properties of the high entropy alloy are predicted based on the simulation results of the motion trajectory.

[0018] Furthermore, the data set includes a plurality of different types of data sets, and each type of data set includes one or more elements.

[0019] Furthermore, the densely connected neural network consists of a first fully connected layer, a dense block, a transition layer and a second fully connected layer.

[0020] Furthermore, the densely connected neural network is trained and evaluated based on iterative calculations. Specifically, the densely connected neural network is first trained and then evaluated. Each time a complete training and evaluation process of the densely connected neural network is completed, an iteration is considered to be completed. When the evaluation result of the densely connected neural network meets expectations, the iteration is stopped.

[0021] Furthermore, the densely connected neural network is set in the FitSnap software, and the FitSnap software and the LAMMPS simulator are seamlessly integrated and jointly deployed on the Sunway supercomputer.

[0022] Furthermore, before using the densely connected neural network after the evaluation to calculate the potential function of the high entropy alloy, it is necessary to configure the parameters of the FitSnap software, specifically: design a first functional class for verifying whether the environment of the LAMMPS simulator is available, and the first functional class is recorded as LammpsSnap.

[0023] Furthermore, commands corresponding to the LAMMPS simulator are generated based on the first functional class, and neighbor lists and calculation options are set.

[0024] The second aspect of the present invention provides a high entropy alloy tensile property prediction system based on molecular dynamics simulation.

[0025] The high entropy alloy tensile properties prediction system based on molecular dynamics simulation includes:

[0026] The data acquisition module is configured to: acquire a data set containing molecular dynamics data of a high entropy alloy, and perform format conversion on the obtained data set;

[0027] The data processing module is configured to: perform data partitioning and parallel data preprocessing on the data set after the format conversion; wherein the data set after the data partitioning includes a training set and an evaluation set;

[0028] A potential function calculation module is configured to: construct a densely connected neural network, train and evaluate the densely connected neural network based on the training set and the evaluation set; and calculate the potential function of the high entropy alloy using the densely connected neural network after the evaluation.

[0029] The high entropy alloy tensile performance prediction module is configured to: input the obtained potential function into the LAMMPS simulator, simulate the motion trajectory of atoms during the tensile fracture of the high entropy alloy based on the LAMMPS simulator; and predict the tensile properties of the high entropy alloy based on the simulation results of the motion trajectory.

[0030] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation as described in the first aspect of the present invention.

[0031] The fourth aspect of the present invention provides an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation as described in the first aspect of the present invention are implemented.

[0032] One or more of the above technical solutions have the following beneficial effects:

[0033] (1) The present invention constructs a densely connected neural network (DenseNet) and trains and evaluates it; the evaluated DenseNet is used to replace the MLP network to calculate the potential function of the high entropy alloy. DenseNet adopts a dense connection method, that is, the input of each layer contains the output information of all previous layers. This connection method enables each layer to fully utilize the information of the previous layer, thereby reducing information redundancy and loss; in high-dimensional space, dense connections help to maintain the integrity and continuity of information, so that the network can better approximate complex nonlinear functions without the phenomenon of a sharp drop in approximation efficiency in high-dimensional space. Therefore, the present invention can achieve accurate prediction of the tensile properties of high-entropy alloys.

[0034] (2) The dense connections and feature reuse mechanism in DenseNet help reduce the risk of overfitting of the network; in high-dimensional space, this feature helps the network better generalize to unseen data. In addition, compared with the MLP network, although DenseNet has a more complex network structure, it contains a parameter sharing mechanism, which usually leads to a significant improvement in computational efficiency. Therefore, the present invention has higher computational efficiency and strong generalization ability.

[0035] (3) In high-dimensional space, the internal information in DenseNet can be freely transmitted between different layers of the network, so that long-distance correlation information can be effectively captured; at the same time, dense connections enable each layer to reuse the features of the previous layer and integrate these features. This feature reuse and integration mechanism helps the network extract richer and more diverse feature representations from the input data, thereby improving the network's ability to capture information at different spatial scales. Therefore, the present invention can better capture long-distance correlations and information at different spatial scales, and has strong feature extraction capabilities.

[0036] (4) The densely connected neural network constructed by the present invention is set in the FitSnap software, and the FitSnap software and the LAMMPS simulator are seamlessly integrated and deployed together on the Sunway supercomputer. Under the heterogeneous architecture of the Sunway supercomputer, the parallel optimization calculation of the Snap force field is realized through the reasonable allocation of tasks between the master core and the slave core, the efficient design of data communication and the optimization of computing logic; the simulation efficiency is significantly improved, and the performance acceleration is achieved by dozens of times.

[0037] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0039] Figure 1 This is a flow chart of a method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation in Example 1 of the present invention.

[0040] Figure 2 Schematic diagram of the structure of the DenseNet network in Example 1 of the present invention.

[0041] Figure 3 This is a task execution flow chart based on the Sunway supercomputer in Example 1 of the present invention.

[0042] Figure 4Schematic diagram of the initial atomic state of the high entropy alloy before tensile fracture simulation in Example 1 of the present invention.

[0043] Figure 5 Schematic diagram of the atomic state of the high entropy alloy after tensile fracture simulation in Example 1 of the present invention.

[0044] Figure 6 Schematic diagram of the atomic state during the high entropy alloy tensile fracture simulation process in Example 1 of the present invention. DETAILED DESCRIPTION

[0045] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0046] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0047] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0048] Embodiment 1

[0049] This embodiment discloses a method for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation.

[0050] like Figure 1 As shown, the method for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation includes:

[0051] Collect the calculation tasks for tensile property prediction of high entropy alloy, and send the calculation tasks to the master core; start the slave core thread library based on the master core, and distribute the calculation tasks to multiple slave cores; execute the tensile property prediction of high entropy alloy based on the slave core, specifically:

[0052] Step S1, obtaining a data set containing molecular dynamics data of a high entropy alloy, and performing format conversion on the obtained data set;

[0053] Step S2, performing data partitioning and parallel data preprocessing on the data set after the format conversion; wherein the data set after the data partitioning includes a training set and an evaluation set;

[0054] Step S3, constructing a densely connected neural network, training and evaluating the densely connected neural network based on the training set and the evaluation set; using the densely connected neural network after the evaluation to calculate the potential function of the high entropy alloy;

[0055] Step S4, inputting the obtained potential function into a LAMMPS simulator, simulating the motion trajectory of atoms during the tensile fracture of the high entropy alloy based on the LAMMPS simulator; and predicting the tensile properties of the high entropy alloy according to the simulation results of the motion trajectory.

[0056] Based on the above process, the present invention can achieve accurate prediction of the tensile properties of high entropy alloys on the basis of enhancing the generalization ability of the model. To facilitate the understanding of the technical solution of the present invention, the specific implementation steps in the technical solution of the present invention are further explained and illustrated below.

[0057] This embodiment provides an efficient parallel implementation method for the structural characteristics of the new generation of Shenwei supercomputers and the needs of tensile fracture performance prediction of high entropy alloys. The Shenwei supercomputer adopts a master-slave heterogeneous structure, which consists of a general computing master core (MPE) and a streamlined computing slave core (CPE). The two work together to achieve efficient computing performance; among them, the master core is mainly responsible for the initialization of applications, data communication, file I / O and other modules, as well as the startup and management of slave core tasks. It is similar to a traditional processor core, can run the operating system, and support various standard application program interfaces; the slave core is mainly responsible for the calculation acceleration of the computing module of the application. They are organized in the form of an array, and each slave core has a high-speed local local data storage space (LDM) for storing temporary data and calculation results; the slave cores are interconnected through the network within the array to achieve efficient parallel computing. The working architecture of the Shenwei supercomputer adopted in this embodiment is a single-core group working mode, that is, 1 master core-64 slave cores, and the specific implementation process can be implemented by steps S1-S4.

[0058] Step S1, obtaining a data set containing molecular dynamics data of a high entropy alloy, and performing format conversion on the obtained data set.

[0059] In this embodiment, the data sets used are all open source data sets specifically used for materials science research; wherein, the acquired data sets contain multiple different types of data sets, each type of data set contains one or more elements, specifically: the data set contains the WBe_PRB2019 data set, the Ta_Linear_JCP2014 data set, the Fe_Linear_NPJ2021 data set, and a data set containing multiple elements such as NbMoTaW; wherein, the WBe_PRB2019 data set contains tungsten (W) and beryllium (Be) elements, the Ta_Linear_JCP2014 data set contains tantalum (Ta) elements, the Fe_Linear_NPJ2021 data set contains (Fe) elements, and the data set containing multiple elements such as NbMoTaW is an open source data set that has been converted to a format. These data sets were originally used to generate multi-component linear Snap potential functions, and to analyze and predict the performance of materials under different conditions. Preferably, the data set is provided in a JSON file format, containing information such as atomic coordinates, energy, and forces, to support computational simulation and model development of materials.

[0060] Furthermore, these data sets need to be converted into a data format supported by the FitSnap software before use.

[0061] Step S2: performing data partitioning and parallel data preprocessing on the data set after format conversion; wherein the data set after data partitioning includes a training set and an evaluation set.

[0062] Step S2-1, divide the data set into data parts. Specifically, in this embodiment, the data is divided in a ratio of training set: evaluation set = 7:3, that is, 70% of the data in the obtained data set is used as the training set, and 30% of the data in the obtained data set is used as the evaluation set.

[0063] Step S2-2: perform parallel data preprocessing on the data set.

[0064] The neural network used in the subsequent work of this embodiment is set in the FitSnap software. At the same time, the FitSnap software and the LAMMPS simulator are seamlessly integrated and deployed together on the Shenwei supercomputer; Among them, the FitSnap software is a software tool related to machine learning interatomic potentials, which is mainly used to generate machine learning interatomic potentials, i.e., potential functions, for the LAMMPS simulator; the LAMMPS simulator is a large-scale atomic / molecular parallel simulator, which is a parallel computing tool for large-scale atomic and molecular simulations. Based on the above design, the data set is preprocessed in parallel.

[0065] First, a class is built to parse and process molecular dynamics-related JSON files. The data is read by traversing the file list, and the key information is extracted after structural verification. The data is then standardized using predefined conversion rules. Data standardization includes field conversion, coordinate operations (rotation and translation), energy correction, and redundant field removal. The ultimate goal of data standardization is to store the cleaned data in a unified format for subsequent analysis or model training, ensuring data integrity, consistency, and efficient processing.

[0066] Next, a first functional class is designed to verify whether the LAMMPS simulator environment is available, and the first functional class is recorded as LammpsSnap. Specifically: a LammpsSnap class is designed as the first functional class in the FitSnap software for Snap force field calculation based on LAMMPS. The LammpsSnap class defines key properties in the initialization phase; among them, the key properties include data for storing calculation data, shared array indexes i and row_index, and lmp instances that can interact with LAMMPS. During the initialization process, the LammpsSnap class can verify whether the LAMMPS environment is available and prepare for subsequent calculation processes.

[0067] Subsequently, the LAMMPS environment is dynamically configured based on the LammpsSnap class, and the corresponding LAMMPS commands are generated according to the Snap parameters provided by the user. After the configuration is completed, as long as the LAMMPS calculation is completed, the LammpsSnap class can directly extract the Snap descriptor matrix, reference potential energy, force, and stress data calculated by LAMMPS. In addition, by assigning weights to each physical quantity, its contribution in the subsequent model training process is adjusted; all data are ultimately stored in a defined shared array and managed through distributed indexing to ensure the consistency and integrity of the data in a parallel computing environment.

[0068] Finally, configure the Snap force field calculation environment for LAMMPS, that is, prepare the required configuration and parameters for the Snap force field calculation of LAMMPS. Specifically: first, load the molecule or crystal structure on LAMMPS to ensure that the atomic position and unit cell parameters are accurate; then, define the specific values ​​of key variables according to the user configuration; among them, key variables include the upper limit of angular momentum, truncation radius factor and scaling factor; configure element-related parameters for each atom type, such as weight parameters and atomic radius parameters. Then, by declaring the reference force field, ensure that LAMMPS can correctly calculate the interaction between atoms; at the same time, set the calculation logic of the Snap descriptor based on the built-in function of LAMMPS, and initialize the neighbor list to optimize the neighbor search method.

[0069] Based on the above process, the comprehensive configuration of LAMMPS and the parallel data preprocessing of the dataset were completed, on this basis, the efficiency and accuracy of the Snap force field calculation can be ensured.

[0070] Step S3, constructing a densely connected neural network, training and evaluating the densely connected neural network based on the training set and the evaluation set; and calculating the potential function of the high entropy alloy using the densely connected neural network after the evaluation.

[0071] Step S3-1, construct a densely connected neural network.

[0072] Define basic network modules such as DenseLayer, DenseBlock and Transition, and combine these basic network modules into Densenet (densely connected neural network); the combined Densenet can support the stacking of DenseBlock and the feature dimension reduction of the Transition layer, and can also support the dynamic construction of network architectures that adapt to different task requirements based on input parameters. Specifically, the dense layer is the basic component unit of DenseNet. Each dense layer receives features from all previous layers as input and passes its output to all subsequent layers; the dense block is composed of multiple dense layers stacked together, and the input of each dense layer is the concatenation of the outputs of all previous dense layers; the transition layer is used to connect two dense blocks and control the dimension of the features so that they are not too complicated.

[0073] like Figure 2 As shown, the Densenet densely connected neural network consists of a first fully connected layer, a dense block, a transition layer, and a second fully connected layer. Specifically, the first fully connected layer and the second fully connected layer are connection layers with the same structure; each dense layer establishes a connection with all previous layers, thereby achieving full transfer and reuse of features.

[0074] Step S3-2: parameter configuration.

[0075] The Densenet densely connected neural network is set up in the FitSnap software. The FitSnap software and the LAMMPS simulator are seamlessly integrated and deployed together on the Sunway supercomputer.

[0076] In terms of distributed processing, Fitsnap software can use shared arrays to store and access data, and supports cross-process data interaction. By managing the data blocks of each process, the correct distribution and access of data is ensured; at the same time, by storing metadata such as atom types and row types, the traceability and integration of data are ensured. Since the configuration module in Fitsnap supports a variety of Snap configuration options (including nonlinear descriptors, atom-by-atom descriptors, zero-value feature elimination, etc.), it can adapt to diverse computing needs.

[0077] Step S3-3, initializing the constructed densely connected neural network model.

[0078] Since the Densenet densely connected neural network is built on the open source deep learning framework Pytorch, after building the Densenet densely connected neural network, a FitTorch class in the Pytorch framework is used to encapsulate the network model. By receiving multiple network instances and parameters such as the number of descriptors related to molecular dynamics, element type, and data type, it supports single-element or multi-element type configuration; at the same time, by loading the existing model weights, initializing the optimizer (such as AdamW) and the learning rate scheduler (such as the cosine annealing restart scheduler), the model is initialized.

[0079] Furthermore, multiple network instances and molecular dynamics-related parameters (such as the number of descriptors, element types, and data types) are obtained through configuration files or dynamic generation to build network models that adapt to task requirements; among them, the number of descriptors determines the size and data dimension of the model input layer, the element type supports the processing of multi-element systems, and captures its characteristics by creating independent network instances for each element type, and the data type controls the accuracy and stability of the calculation. These parameters are closely related to the model input, network instance selection, and numerical calculation during training, evaluation, and prediction, providing core support for the efficient completion of molecular dynamics tasks.

[0080] Furthermore, loading model weights is achieved by reading the pre-trained model file. Specifically, first, check whether the weight path is specified in the configuration file. If it exists, the weight file is directly loaded and the weights are applied to the model; at the same time, the state of the optimizer is restored. If the weight file is not available, the optimizer and model parameters are reinitialized. By loading weights, training can be continued or transfer learning can be performed after training is interrupted, and when weights are not provided, the model will be initialized from scratch to ensure the integrity and flexibility of training.

[0081] Step S3-4: Train and evaluate the densely connected neural network based on the training set and the evaluation set.

[0082] The densely connected neural network is trained and evaluated based on iterative calculation. Specifically: the densely connected neural network is trained first, and then evaluated. Each time a complete training and evaluation process of the densely connected neural network is completed, one iteration is considered to be completed. When the evaluation result of the densely connected neural network meets the expectation, the iteration is stopped. Specifically: the model parameters are optimized through multiple rounds of iterations, and each round of training is divided into two stages: training set calculation and validation set evaluation. The MSE loss function is used to optimize the predicted values ​​of energy and force, and the learning rate is dynamically adjusted through the scheduler.

[0083] During the training process, record the target values, predicted values, and losses of energy and force in each iteration, and finally save the best model. First, save the Desenet densely connected neural network model in a format compatible with LAMMPS; then, use the IgnoreElems function to encapsulate the single-element model and the ElemwiseModels function to encapsulate the multi-element model; both the IgnoreElems function and the ElemwiseModels function are built-in method functions of LAMMPS. At the same time, load the saved Desenet densely connected neural network model for evaluation or further training.

[0084] In the actual evaluation stage, the performance of the Desenet densely connected neural network model is tested on a single or multiple configurations. The predicted values ​​are generated by standardizing the input descriptors, and the model performance is evaluated by comparing them with the target values. The results can be saved as files for analysis.

[0085] Step S3-5: Calculate the potential function of the high entropy alloy using the densely connected neural network after the evaluation.

[0086] Obtain the weight parameters of the densely connected neural network, and compose the obtained weight parameters and related data content into a potential function. Specifically, the machine learning potential function combines elements such as neural network parameters (weights), input data (such as descriptors), target values ​​(such as potential energy or force), and training data.

[0087] Step S4, input the obtained potential function into the LAMMPS simulator, and simulate the motion trajectory of atoms during the tensile fracture of the high entropy alloy based on the LAMMPS simulator; predict the tensile properties of the high entropy alloy according to the simulation results of the motion trajectory, and optimize the simulation process in parallel on the Sunway computer. After configuration, the task execution implementation process based on the Sunway supercomputer is as follows: Figure 3As shown in the figure, specifically: with the transmission and processing of data between local storage (LDM) and main memory (MEM) as the core, from initialization to final cumulative calculation, an efficient task execution process is demonstrated. First, based on the Sunway supercomputer, the computing task of predicting the tensile properties of high-entropy alloys is collected and sent to the main core; based on the main core, the slave core thread library is started, the computing task is distributed to multiple slave cores and LDM space and cache are dynamically allocated; based on the slave core, the tensile properties of high-entropy alloys are predicted; combined with the cutoff radius, neighbor data is collected from the main memory and transmitted to the LDM for preliminary processing to calculate intermediate variables (such as ui, zi, bi) and feature quantities (such as bispectrum, that is, the first i The spectral vector of atoms is then calculated based on the intermediate results, and the high-level coefficients (such as beta and deidrj) are further calculated, and the results are transferred to the main memory through DMA (direct access memory) for accumulation processing; after each slave core completes the calculation task, it returns the calculation result to the master core; where ui represents the atom i displacement or velocity, zi represents the atomic i The charge or atomic number of the atom. i Some characteristic quantities of can be spectral vectors or local environment descriptors; beta represents the coefficient used to describe the material response, such as the elastic coefficient, deidrj represents the atomic i With Atom j The rate of change of the interaction between them is usually the gradient of potential energy. During the task execution, the main memory not only serves as the core of data storage and interaction, but also transmits the results back to the LDM to support iterative calculations. Finally, by accumulating all atomic forces and sub-energies, the calculation of the overall physical quantity of the system is realized. The entire process ensures the efficient completion of computing tasks and the optimal use of resources through efficient data transmission, distributed task allocation and dynamic memory management.

[0088] First, set the basic parameters and conditions of the simulation. Specifically, use the metal unit system (units metal) and simple atom type (atom_style atomic) to meet the simulation requirements of metal materials; at the same time, use periodic boundary conditions (boundary ppp) to ensure that the boundary effect will not affect the simulation results. Next, load the initial atomic structure file and introduce the Snap potential function file based on machine learning; the Snap potential function can describe the complex atomic interactions in high entropy alloys with high precision and is one of the key force fields for simulation. Subsequently, the system is optimized based on two stages; the first stage is energy minimization, that is, using the steepest descent method for optimization, gradually reducing the total energy of the system, and eliminating unreasonable forces and stresses between atoms; this stage is to ensure that the system can adjust from a high energy state to a relatively reasonable low energy state. The second stage is a dynamic simulation. Under constant temperature conditions, the system is further optimized through 50,000 steps of molecular dynamics simulation so that it can reach thermodynamic equilibrium under a temperature-controlled environment of 300K; the simulation time step is set to 0.001 picoseconds to take into account both the accuracy and stability of the calculation.

[0089] Furthermore, the thermodynamic information of the system is output every 1000 steps, including key parameters such as the number of steps, temperature, and pressure. The atomic trajectory file is saved to record the motion trajectory of atoms during the simulation process, providing data support for subsequent structural visualization and dynamic analysis. The final system state is saved to a file to provide optimized initial conditions for subsequent simulations.

[0090] In the simulation, the z direction of the system is divided into two regions; the atoms in the fix_gu1 region remain fixed and do not participate in the displacement; while the atoms in the fix_gu2 region gradually apply a tensile force that increases with the time step, that is, the fix_gu1 region is the fixed end, serving as a reference point; the fix_gu2 region is the end that is subjected to force, used to apply external force to simulate the tensile behavior of the material. This design controls the external force through variables, allowing the system to gradually accumulate stress until the tensile force exceeds the critical strength of the material, resulting in the breaking of the interatomic bonds. The simulation process truly reproduces the tensile process of the material from force to fracture, similar to the tensile test in the experiment.

[0091] To ensure the accuracy of the simulation, this embodiment ran 3 million steps at a constant temperature of 300 K, with a total simulation time of 300 picoseconds. During the simulation, the stress-strain relationship, thermodynamic parameters, and atomic trajectory data of the system were recorded. Through these data, the entire process of the material from elastic deformation to plastic deformation and ultimately fracture can be clearly seen.

[0092] like Figure 4 , Figure 5As shown in the figure, this simulation reveals the tensile fracture process of NbMoTaW high entropy alloy under tensile conditions, especially the effects of internal defects, dislocations and interatomic forces on the fracture mode. The simulation results not only provide an important reference for the evaluation of the mechanical properties of materials, but also provide a theoretical basis for optimizing the design and application of high entropy alloys, laying the foundation for improving their reliability in engineering. Figure 4 , Figure 5 The red spheres are Nb atoms, the blue spheres are Mo atoms, the yellow spheres are Ta atoms, and the purple spheres are W atoms.

[0093] like Figure 6 As shown, the tensile fracture process of the tested high entropy alloy shows a tensile fracture strength of 1337MPa. Specifically: Figure 6 The cross-sectional area of ​​the rectangular box is , at step 75, the force on a single atom is , convert this force into SI units ;in, Electron volt is a unit of energy that represents the energy obtained by an electron under a voltage of 1 volt; Angstrom is a unit of length, common sense. In this example, there are 500 atoms in the system, so the total force is ; According to the pressure calculation formula, the applied pressure is 1337MPa. In general, for NbMoTaW high entropy alloys, if the tensile fracture strength is greater than 1200MPa, it can be considered to have good tensile properties. Since the NbMoTaW high entropy alloy in this experiment has a high tensile fracture strength of 1337 MPa, it shows that it has superior mechanical properties under high stress environments, and is particularly suitable for high-load application scenarios such as aerospace and energy fields. Therefore, overall, the tensile properties of the alloy are still good and suitable for use in high-strength environments.

[0094] Furthermore, in order to verify the superiority of the present invention over other prior arts in predicting the tensile properties of high entropy alloys, this embodiment uses MSE (mean square error) and RMSE (root mean square error) as evaluation indicators to conduct a model performance comparison experiment. The experimental results are shown in Tables 1 to 4, namely:

[0095] Table 1 Performance of each model under the MSE indicator under the training data set (strength)

[0096]

[0097] Table 2 Performance of each model under the RMSE indicator under the training data set (strength)

[0098]

[0099] Table 3 Performance of each model under the MSE indicator under the test data set (strength)

[0100]

[0101] Table 4 Performance of each model under the RMSE indicator under the test data set (strength)

[0102]

[0103] For the two indicators of MSE (mean square error) and RMSE (root mean square error), the lower the value, the better the performance. From the results shown in Tables 1 to 4, it can be seen that compared with other existing models, the model of the present invention has achieved the best performance on all data sets, which proves that the method of the present invention has a unique advantage in processing this type of data.

[0104] The present invention obtains a machine learning potential function by training a densely connected neural network for subsequent dynamic simulation, and simulates the tensile fracture process of a high entropy alloy based on dynamic simulation. Among them, the core process of molecular dynamics simulation is to calculate the interaction force through the position information of the atomic coordinate points, and then derive the acceleration through the mechanical relationship, obtain the speed from the acceleration, and then calculate the displacement at the next moment, update the position of the atom, and realize the position-force-speed cycle iteration. After the above experiments, the superiority of the model built by the present invention is verified.

[0105] Embodiment 2

[0106] This embodiment discloses a high entropy alloy tensile property prediction system based on molecular dynamics simulation.

[0107] The high entropy alloy tensile properties prediction system based on molecular dynamics simulation includes:

[0108] The data acquisition module is configured to: acquire a data set containing molecular dynamics data of a high entropy alloy, and perform format conversion on the obtained data set;

[0109] The data processing module is configured to: perform data partitioning and parallel data preprocessing on the data set after the format conversion; wherein the data set after the data partitioning includes a training set and an evaluation set;

[0110] A potential function calculation module is configured to: construct a densely connected neural network, train and evaluate the densely connected neural network based on the training set and the evaluation set; and calculate the potential function of the high entropy alloy using the densely connected neural network after the evaluation.

[0111] The high entropy alloy tensile performance prediction module is configured to: input the obtained potential function into the LAMMPS simulator, simulate the motion trajectory of atoms during the tensile fracture of the high entropy alloy based on the LAMMPS simulator; and predict the tensile properties of the high entropy alloy based on the simulation results of the motion trajectory.

[0112] Embodiment 3

[0113] The purpose of this embodiment is to provide a computer-readable storage medium.

[0114] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation as described in the first embodiment of the present disclosure.

[0115] Embodiment 4

[0116] The purpose of this embodiment is to provide an electronic device.

[0117] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation as described in the first embodiment of the present disclosure are implemented.

[0118] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0119] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0120] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation, characterized in that: include: Collecting computational tasks for predicting tensile properties of high entropy alloys, and sending the computational tasks to the main core; Start the slave core thread library based on the master core and distribute the computing tasks to multiple slave cores; Based on the prediction of tensile properties of high entropy alloys from the core, specifically: Acquire a data set containing molecular dynamics data of a high entropy alloy, and convert the format of the obtained data set; Performing data partitioning and parallel data preprocessing on the data set after format conversion; wherein the data set after data partitioning includes a training set and an evaluation set; Constructing a densely connected neural network, training and evaluating the densely connected neural network based on the training set and the evaluation set; and calculating the potential function of the high entropy alloy using the densely connected neural network after the evaluation; The obtained potential function is input into a LAMMPS simulator, and the motion trajectory of atoms during the tensile fracture of the high entropy alloy is simulated based on the LAMMPS simulator; and the tensile properties of the high entropy alloy are predicted based on the simulation results of the motion trajectory.

2. The method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation according to claim 1, characterized in that: The data set includes a plurality of different types of data sets, and each type of data set includes one or more elements.

3. The method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation according to claim 1, characterized in that: The densely connected neural network consists of a first fully connected layer, a dense block, a transition layer and a second fully connected layer.

4. The method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation according to claim 1, characterized in that: The densely connected neural network is trained and evaluated based on iterative calculation. Specifically, the densely connected neural network is trained first and then evaluated. Each time a complete training and evaluation process of the densely connected neural network is completed, an iteration is considered to be completed. When the evaluation result of the densely connected neural network reaches the expectation, the iteration is stopped.

5. The method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation according to claim 1, characterized in that: The densely connected neural network is set in the FitSnap software, and the FitSnap software and the LAMMPS simulator are seamlessly integrated and jointly deployed on the Sunway supercomputer.

6. The method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation according to claim 1, characterized in that: Before using the densely connected neural network after evaluation to calculate the potential function of the high entropy alloy, it is necessary to configure the parameters of the FitSnap software. Specifically: design a first functional class for verifying whether the environment of the LAMMPS simulator is available, and record the first functional class as LammpsSnap.

7. The method for predicting tensile properties of high entropy alloys based on molecular dynamics simulation according to claim 6, characterized in that: Generate commands corresponding to the LAMMPS simulator based on the first functional class and set the neighbor list and calculation options.

8. A high entropy alloy tensile properties prediction system based on molecular dynamics simulation, characterized in that: include: The data acquisition module is configured to: acquire a data set containing molecular dynamics data of a high entropy alloy, and perform format conversion on the obtained data set; The data processing module is configured to: perform data partitioning and parallel data preprocessing on the data set after the format conversion; wherein the data set after the data partitioning includes a training set and an evaluation set; A potential function calculation module is configured to: construct a densely connected neural network, train and evaluate the densely connected neural network based on the training set and the evaluation set; and calculate the potential function of the high entropy alloy using the densely connected neural network after the evaluation. The high entropy alloy tensile performance prediction module is configured to: input the obtained potential function into the LAMMPS simulator, simulate the motion trajectory of atoms during the tensile fracture of the high entropy alloy based on the LAMMPS simulator; and predict the tensile properties of the high entropy alloy based on the simulation results of the motion trajectory.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for predicting the tensile properties of high entropy alloys based on molecular dynamics simulation as described in any one of claims 1 to 7 are implemented.

Images