High-entropy alloy crystal structure clustering method fusing language large model knowledge
By cleaning and standardizing atomic-scale data of high-entropy alloys, and generating professional natural language text by combining large language models, numerical and semantic dual-modal features are constructed. This solves the problem that traditional clustering methods cannot incorporate the knowledge of large language models, and realizes accurate clustering of high-entropy alloy crystal structures and guidance for new material design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANXI UNIV
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional clustering methods cannot effectively incorporate linguistic large-scale model knowledge when dealing with high-entropy alloys, resulting in clustering results that are disconnected from physical mechanisms, limited distinguishing ability, and difficulty in applying them to exploratory research on new materials with unknown composition and structure.
By cleaning and standardizing atomic-scale data of high-entropy alloys, structured tabular data is generated. Professional natural language text is generated by combining large language models, numerical and semantic dual-modal features are constructed, and correlation mapping is established through cross-modal inter-distillation mechanism. The clustering objective function is optimized, and the clustering results of high-entropy alloy crystal structures are output.
It improves the resolution and physical interpretability of high-entropy alloy crystal structure clustering. The resulting clustering results accurately reflect the essential differences in the microstructure of the materials, which can guide the design and performance prediction of new materials and improve the efficiency of high-entropy alloy research and development.
Smart Images
Figure CN122024925A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and in particular to a high-entropy alloy crystal structure clustering method that integrates knowledge from large language models. Background Technology
[0002] High-entropy alloys, a significant breakthrough in materials science in recent years, are characterized by multi-principal-element solid solutions composed of five or more principal elements in near-equiatomic proportions. This unique compositional design endows them with four effects unattainable by traditional alloys: high-entropy effect, lattice distortion effect, hysteresis diffusion effect, and "cocktail" effect, resulting in significant advantages in mechanical properties, corrosion resistance, thermal stability, and radiation resistance. However, this complex multi-principal-element system also leads to extreme complexity in its microstructure, including the interweaving of various structural features such as local chemical order, nanoclusters, and lattice distortion regions. This makes traditional structural characterization methods based on single or a few features insufficient to comprehensively and accurately reveal the essence of their microstructure. Therefore, developing novel analytical methods capable of deeply analyzing the complex microstructure of high-entropy alloys has become a key scientific issue for establishing reliable "composition-structure-performance" correlation models and realizing rational material design.
[0003] In the field of microstructure characterization, clustering analysis based on atomic-scale simulation data has become an important tool for studying the local atomic environment of materials. Simulation methods such as molecular dynamics and first-principles calculations can obtain multi-dimensional physical characteristic data, including atomic types, spatial locations, stress states, and energy distributions. Existing techniques typically employ dimensionality reduction methods such as principal component analysis and t-distributed random neighborhood embedding to process high-dimensional features, and then combine them with unsupervised learning algorithms such as K-means clustering and density clustering to group and classify atoms. While this method can distinguish different types of local atomic environments to some extent, it is essentially still a mathematical classification based on numerical similarity and fails to integrate the physical laws of materials science into the clustering process.
[0004] Traditional numerical clustering methods face three core dilemmas when dealing with high-entropy alloy systems: their purely data-driven paradigm cannot incorporate large-scale linguistic model knowledge as guidance, leading to a disconnect between clustering results and physical mechanisms; the numerical features they rely on have limited ability to distinguish between different phase structures with similar physical states, affecting the accuracy of clustering; and most methods require supervised learning based on known structural labels, making them unsuitable for exploratory research on new materials with unknown composition and structure, thus limiting their potential for application in efficient materials discovery. Summary of the Invention
[0005] In view of the aforementioned existing problems, the present invention is proposed.
[0006] Therefore, this invention provides a high-entropy alloy crystal structure clustering method that integrates large-scale linguistic model knowledge to address the problems of purely data-driven paradigms that cannot incorporate large-scale linguistic model knowledge as guidance, leading to a disconnect between clustering results and physical mechanisms; the limited ability of numerical features to distinguish between different phase structures with similar physical states, affecting the accuracy of clustering; and the fact that most methods require supervised learning based on known structural labels, making them unsuitable for exploratory research on new materials with unknown composition and structure, thus limiting their potential application in efficient materials discovery.
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a high-entropy alloy crystal structure clustering method that integrates knowledge from large-scale language models, comprising: The original high-entropy alloy atomic-scale data was cleaned to remove management fields that were irrelevant to the crystal structure, while retaining core physical properties such as atom type, number of atoms, total energy, three-dimensional coordinates of atoms, and three-dimensional forces. The fields corresponding to missing atom types were uniformly filled with zero values to form structured tabular data. Based on structured tabular data, the attribute values of each sample are used to construct an original feature vector, and standardization is performed on all feature dimensions to eliminate the influence of dimensional differences and obtain a numerical embedding representation. By using a large language model to parse the semantic relationship between attribute names and attribute values in numerical embedding representations, professional natural language text describing the microstructural characteristics of samples is generated. Professional natural language text is input into a dedicated text embedding model and converted into semantic text embedding vectors corresponding to the samples. Using numerical embedding representation and semantic text embedding vector as bimodal inputs, neighborhood distributions are constructed in numerical space and text space respectively, and the association mapping between the two is established through cross-modal mutual distillation mechanism; Based on the neighborhood distribution, the neighborhood distillation loss, modal consistency loss, and entropy regularization loss are calculated, and the clustering objective function is jointly optimized to output the clustering results of high-entropy alloy crystal structures.
[0008] As a preferred embodiment of the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in this invention, the following steps are taken: The original high-entropy alloy atomic-scale data is cleaned to remove administrative fields unrelated to the crystal structure, retaining core physical attributes such as atom type, number of atoms, total energy, atomic three-dimensional coordinates, and three-dimensional forces. Fields corresponding to missing atom types are uniformly filled with zero values to form structured tabular data. Identify fields in the original data that contain sample number, test batch, and experimenter identifier, and remove these fields from the dataset; For the types of atoms appearing in each sample Extract the three-dimensional coordinates of all atoms in this category. , , ) and three-dimensional forces ( , , ),in Indicates the atom type index, Indicates the first in this category One atom; Calculate the mean values for the coordinate components of the same type of atom: ; The mean values of the force components were also calculated separately: ; in, For the types in the sample The number of atoms; If a certain type of atom If it does not appear in the current sample, set its atom count to 0, and set the coordinate mean. With the average force Set all to 0; Total energy of the sample With the total number of atoms It is directly retained as a global attribute field without aggregation or transformation; All fields corresponding to the types of atoms are arranged in a fixed order to form a single structured record.
[0009] As a preferred embodiment of the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in this invention, the following steps are taken: Based on structured tabular data, the attribute values of each sample are constructed into an original feature vector, and standardization processing is performed on all feature dimensions to eliminate the influence of dimensional differences and obtain a numerical embedding representation. Concatenate all numerical fields from each row of the structured table into the original feature vector in a predefined order. ,in This represents the total number of feature dimensions. For the Dimensional features in all Calculate the mean from each sample. with standard deviation : ; For each sample, the first Perform Z-score normalization on 3D features: ; Obtain the standardized numerical embedding representation This is used for subsequent bimodal modeling.
[0010] As a preferred embodiment of the high-entropy alloy crystal structure clustering method integrating large language model knowledge described in this invention, the specific steps of using a large language model to analyze the semantic relationship between attribute names and attribute values in numerical embedding representation to generate professional natural language text describing the microstructural characteristics of samples are as follows: Sequence of column names in a structured table { } and the standardized values of the corresponding samples { The key value pairs are organized as input prompts; The Llama large language model, fine-tuned with a materials science corpus, is invoked to infer and output a natural language description based on the given prompts. This description covers semantic information related to crystal structure, such as atomic composition ratio, typical interatomic spacing trend, overall force balance state and energy stability. Ensure text is generated It does not introduce hypothetical content not included in the original data, but only performs logical deduction and standardized terminology based on the input fields.
[0011] As a preferred embodiment of the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in this invention, the specific steps of inputting professional natural language text into a dedicated text embedding model and converting it into a semantic text embedding vector corresponding to the sample are as follows: The generated natural language description Input into the nomic-embed-text embedding model; The encoder of this model outputs a fixed dimension. Dense vectors This vector characterizes the structural chemical properties of high-entropy alloys in semantic space; The vector With numerical embedding representation These form bimodal feature pairs for subsequent cross-modal alignment.
[0012] As a preferred embodiment of the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in this invention, the method involves using numerical embedding representations and semantic text embedding vectors as bimodal inputs to construct neighborhood distributions in the numerical space and text space, respectively, and establishing a correlation mapping between the two through a cross-modal mutual distillation mechanism. The specific steps are as follows: First, for the original data vector of each sample Large language model First, the table header, i.e., the attribute names, is parsed. Then, each attribute name is associated with its corresponding sample attribute value for understanding. Finally, a natural language text describing the characteristic of this sample using professional knowledge is generated, as shown in the following formula:
[0013] the remaining, For the first The original vector of each sample, For large language models, The first generation generated after language inference for large models Text description of each sample; Then, the nomic-embed-text model is specifically designed to convert text into vector embeddings, transforming the text descriptions generated by large language models. As input, through a pre-trained nomic-embed-text model The embedding interface can convert it into the corresponding numerical text embedding vector, as shown in the following formula:
[0014] in, This is the embedding interface for the nomic-embed-text model. For the first The text embedding vector of each sample; As a preferred embodiment of the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in this invention, the specific steps are as follows: The calculation of neighborhood distillation loss, modal consistency loss, and entropy regularization loss based on neighborhood distribution, joint optimization of the clustering objective function, and output of the high-entropy alloy crystal structure clustering results are as follows: For each sample Use FAISS to query it Neighbor set, defining the neighborhood distribution of the table sample:
[0015] in, For the first The original data embedding vector of each sample, For the first The text embedding vector of each sample, For similarity function, For the first The neighbor set of a sample The index of the neighbor sample. For tabular data samples With corresponding text neighbors The similarity probability; Text neighborhood distribution:
[0016] in, Text embedding samples Neighbors of the corresponding table data The similarity probability; Neighborhood distillation loss
[0017] in, The total number of samples, For samples in text space The neighborhood distribution vector, For samples in the tabular data space The neighborhood distribution vector, Kullback-Leibler divergence measures the difference between two distributions. and Differences This is the distillation loss, used to align the neighborhood distribution of the table and text. The smaller the value, the more consistent the neighborhood distribution of the two modalities. Modal consistency loss:
[0018] in, Predict the distribution for the data sample in the i-th table. For the predicted distribution of the text, The total number of samples, Modality consistency loss measures the consistency between table modality and text modality in cluster prediction distribution; the smaller the value, the more consistent the predictions of the two modalities. Entropy regularization (to prevent collapse): Designing the entropy term encourages overall uniformity and suppresses overconfidence.
[0019] Among them, the average predicted distribution of the table data samples Average prediction distribution of text data samples , Entropy regularization loss measures the uniformity of the predicted distribution and prevents all samples from being classified into the same class. The larger the value, the higher the entropy and the more uniform the distribution. The total loss is:
[0020] in As weight, The smaller the value, the better the overall training effect of the model.
[0021] As a preferred embodiment of the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in this invention, the clustering results are used to guide the crystal structure design and performance prediction of new high-entropy alloy materials. The specific steps are as follows: Structural commonality analysis was performed on high-entropy alloy samples belonging to the same category in the clustering output to extract statistical regularities of the category in terms of atomic composition ratio, average interatomic spacing, total energy distribution and force equilibrium state. Based on the statistical regularities, a structure-category mapping rule base is constructed. When a new high-entropy alloy atomic-scale data is input, the above full-process processing is performed to obtain its category label. Based on the structural commonalities associated with category labels, the crystal symmetry, phase stability, and mechanical behavior tendencies of new materials can be inferred. The inference results are fed back as prior knowledge into the high-throughput screening or first-principles calculation process for materials, thereby narrowing the search space for candidate structures and improving the efficiency of new material research and development.
[0022] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the high-entropy alloy crystal structure clustering method that integrates large model knowledge of language as described in the first aspect of the present invention.
[0023] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in the first aspect of the present invention.
[0024] The beneficial effects of this invention are as follows: By integrating professional semantic knowledge generated by a large language model with original atomic-scale numerical features, a dual-modal representation system is constructed, and a cross-modal distillation mechanism is introduced to achieve structural alignment between the numerical space and the textual semantic space, thereby improving the resolution and physical interpretability of high-entropy alloy crystal structure clustering. At the same time, through the joint optimization of neighborhood distillation loss, modal consistency loss, and entropy regularization term, cluster collapse is effectively avoided and the category distribution is ensured to be balanced. The final clustering results not only accurately reflect the essential differences in the microstructure of the materials, but can also be directly used to guide the design of new materials and the screening of high-performance candidate structures, effectively improving the efficiency and intelligence level of high-entropy alloy R&D. Attached Figure Description
[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 A flowchart for a high-entropy alloy crystal structure clustering method that integrates knowledge from a large language model.
[0027] Figure 2 A model diagram for a high-entropy alloy crystal structure clustering method that integrates knowledge from large language models.
[0028] Figure 3 A flowchart for language large model text generation using a high-entropy alloy crystal structure clustering method that integrates knowledge from large language models. Detailed Implementation
[0029] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0030] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0031] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0032] Reference Figures 1-3 This is the first embodiment of the present invention, which provides a high-entropy alloy crystal structure clustering method that integrates knowledge from a large language model, including the following steps: S1. Perform cleaning processing on the original high-entropy alloy atomic-scale data, remove management fields that are not related to the crystal structure, retain core physical properties such as atom type, number of atoms, total energy, three-dimensional coordinates of atoms and three-dimensional forces, and uniformly fill the fields corresponding to missing atom types with zero values to form structured tabular data.
[0033] Furthermore, identify fields in the original data that contain sample number, test batch, and experimenter identifier, and remove these fields from the dataset; For the types of atoms appearing in each sample Extract the three-dimensional coordinates of all atoms in this category. , , ) and three-dimensional forces ( , , ),in Indicates the atom type index, Indicates the first in this category One atom; Calculate the mean values for the coordinate components of the same type of atom: ; The mean values of the force components were also calculated separately: ; in, For the types in the sample The number of atoms; If a certain type of atom If it does not appear in the current sample, set its atom count to 0, and set the coordinate mean. With the average force Set all to 0; Total energy of the sample With the total number of atoms It is directly retained as a global attribute field without aggregation or transformation; All fields corresponding to the types of atoms are arranged in a fixed order to form a single structured record.
[0034] It should be noted that by aggregating the mean values of the relevant attributes of the atom types and uniformly filling the missing fields with zero, not only is the noise interference caused by the random fluctuations of single atoms eliminated, but the feature dimensions of different samples are also strictly aligned. This provides a structurally consistent and physically meaningful input basis for subsequent numerical modeling, effectively improving the stability and repeatability of the clustering algorithm.
[0035] S2. Based on structured tabular data, the attribute values of each sample are used to construct the original feature vector, and standardization is performed on all feature dimensions to eliminate the influence of dimensional differences and obtain numerical embedding representation.
[0036] Furthermore, all numerical fields in each row of the structured table are concatenated into the original feature vector in a predefined order. ,in This represents the total number of feature dimensions. For the Dimensional features in all Calculate the mean from each sample. with standard deviation : ; For each sample, the first Perform Z-score normalization on 3D features: ; Obtain the standardized numerical embedding representation This is used for subsequent bimodal modeling.
[0037] It should be noted that the Z-score standardization process effectively eliminates the feature weight bias caused by the differences in the dimensions and orders of magnitude of physical quantities such as energy, coordinates, and force, so that each dimension contributes equally in the clustering process, avoiding high-order features from dominating the model learning direction, thereby ensuring that the clustering results truly reflect the essential differences in the microstructure of the material rather than the numerical scale effect.
[0038] S3. Utilize a large language model to parse the semantic relationship between attribute names and attribute values in numerical embedding representations, and generate professional natural language text describing the microstructural characteristics of samples.
[0039] Furthermore, the column name sequence of the structured table { } and the standardized values of the corresponding samples {
[0040] The key value pairs are organized as input prompts; The Llama large language model, fine-tuned with a materials science corpus, is invoked to infer and output a natural language description based on the given prompts. This description covers semantic information related to crystal structure, such as atomic composition ratio, typical interatomic spacing trend, overall force balance state and energy stability. Ensure text is generated It does not introduce hypothetical content not included in the original data, but only performs logical deduction and standardized terminology based on the input fields.
[0041] It should be noted that by using a large language model fine-tuned from materials science corpus to transform structured numerical values into professional natural language descriptions, a semantic leap from raw data to domain knowledge is achieved. The generated text contains structure-performance correlation logic that can be understood by human experts, providing a high-quality and interpretable semantic bridge for subsequent clustering enhancement using prior knowledge of language models.
[0042] S4. Input professional natural language text into a dedicated text embedding model and convert it into semantic text embedding vectors corresponding to the samples.
[0043] Furthermore, the generated natural language description Input into the nomic-embed-text embedding model; The encoder of this model outputs a fixed dimension. Dense vectors This vector characterizes the structural chemical properties of high-entropy alloys in semantic space; vector With numerical embedding representation These form bimodal feature pairs for subsequent cross-modal alignment.
[0044] It should be noted that by using a dedicated scientific text embedding model to transform professional descriptions into dense semantic vectors, implicit knowledge such as the chemical composition and mechanical state of high-entropy alloys is encoded into a continuous vector space. This semantic text embedding and numerical embedding form complementary representations, together constituting a dual-modal input that combines physical accuracy and knowledge richness, effectively enhancing feature expression capabilities.
[0045] S5. Using numerical embedding representation and semantic text embedding vector as bimodal inputs, neighborhood distributions are constructed in the numerical space and text space respectively, and the correlation mapping between the two is established through a cross-modal mutual distillation mechanism.
[0046] Furthermore, firstly, for each sample's original data vector... Large language model First, the table header, i.e., the attribute names, is parsed. Then, each attribute name is associated with its corresponding sample attribute value for understanding. Finally, a natural language text describing the characteristic of this sample using professional knowledge is generated, as shown in the following formula:
[0047] the remaining, For the first The original vector of each sample, For large language models, The first generation generated after language inference for large models Textual descriptions of each sample.
[0048] Then, the nomic-embed-text model is specifically designed to convert text into vector embeddings, transforming the text descriptions generated by large language models. As input, through a pre-trained nomic-embed-text model The embedding interface can convert it into the corresponding numerical text embedding vector, as shown in the following formula:
[0049] in, This is the embedding interface for the nomic-embed-text model. For the first The text embedding vector of each sample.
[0050] It should be noted that by constructing a bimodal neighborhood distribution based on FAISS and forcing numerical values to align with the local structure of the text through a mutual distillation mechanism, the clustering model can simultaneously absorb data-driven patterns and language model reasoning knowledge during the learning process. This effectively alleviates the semantic ambiguity caused by missing labels in unsupervised scenarios and improves the clarity of cluster boundaries and the cohesion of categories.
[0051] S6. Calculate neighborhood distillation loss, modal consistency loss and entropy regularization loss based on neighborhood distribution, jointly optimize the clustering objective function, and output the clustering results of high-entropy alloy crystal structures.
[0052] Furthermore, for each sample Use FAISS to query it Neighbor set, defining the neighborhood distribution of the table sample:
[0053] in, For the first The original data embedding vector of each sample, For the first The text embedding vector of each sample, For similarity function, For the first The neighbor set of a sample The index of the neighbor sample. For tabular data samples With corresponding text neighbors The similarity probability.
[0054] Text neighborhood distribution:
[0055] in, Text embedding samples Neighbors of the corresponding table data The similarity probability.
[0056] Neighborhood distillation losses:
[0057] in, The total number of samples, For samples in text space The neighborhood distribution vector, For samples in the tabular data space The neighborhood distribution vector, Kullback-Leibler divergence measures the difference between two distributions. and Differences This is the distillation loss, used to align the neighborhood distribution of the table and text. The smaller the value, the more consistent the neighborhood distribution of the two modalities. Modal consistency loss:
[0058] in, Predict the distribution for the data sample in the i-th table. For the predicted distribution of the text, The total number of samples, Modality consistency loss measures the consistency between the table modality and the text modality in the clustering prediction distribution. The smaller the value, the more consistent the predictions of the two modalities.
[0059] Entropy regularization (to prevent collapse): Designing the entropy term encourages overall uniformity and suppresses overconfidence.
[0060] Among them, the average predicted distribution of the table data samples Average prediction distribution of text data samples , Entropy regularization loss measures the uniformity of the predicted distribution, preventing all samples from being classified into the same class. A larger value indicates higher entropy and a more uniform distribution.
[0061] The total loss is:
[0062] in As weight, This represents the total loss; a smaller value indicates a better overall training effect of the model. Clustering results are used to guide the crystal structure design and performance prediction of new high-entropy alloy materials. The specific steps are as follows: Structural commonality analysis was performed on high-entropy alloy samples belonging to the same category in the clustering output to extract statistical regularities of the category in terms of atomic composition ratio, average interatomic spacing, total energy distribution and force equilibrium state. Based on statistical regularities, a structure-category mapping rule base is constructed. When a new high-entropy alloy atomic-scale data is input, the above full-process processing is performed to obtain its category label. Based on the structural commonalities associated with category labels, the crystal symmetry, phase stability, and mechanical behavior tendencies of new materials can be inferred. The inference results are fed back as prior knowledge into the high-throughput screening or first-principles calculation process for materials, thereby narrowing the search space for candidate structures and improving the efficiency of new material research and development.
[0063] It should be noted that jointly optimizing the neighborhood distillation loss, modal consistency loss, and entropy regularization term not only promotes the coordinated convergence of the bimodal prediction results, but also actively suppresses cluster collapse and promotes the balanced distribution of categories. The final output cluster division has both physical rationality and structural distinguishability, which can be directly used for structural classification and performance prediction in new material design, effectively accelerating the closed loop of high-entropy alloy R&D.
[0064] This embodiment also provides a computer device applicable to the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as proposed in the above embodiment.
[0065] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0066] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the high-entropy alloy crystal structure clustering method for fusing large model knowledge of language as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0067] In summary, this invention constructs a dual-modal representation system by integrating professional semantic knowledge generated from a large language model with original atomic-scale numerical features. It also introduces a cross-modal distillation mechanism to achieve structural alignment between the numerical space and the textual semantic space, thereby improving the resolution and physical interpretability of high-entropy alloy crystal structure clustering. Simultaneously, through joint optimization of neighborhood distillation loss, modal consistency loss, and entropy regularization, it effectively avoids cluster collapse and ensures balanced category distribution. The resulting clustering not only accurately reflects the essential differences in the microstructure of the materials but can also be directly used to guide the design of new materials and the screening of high-performance candidate structures, effectively improving the efficiency and intelligence level of high-entropy alloy R&D.
[0068] Example 2, please refer to the table, is the second embodiment of the present invention.
[0069] This embodiment is experimentally validated on the HEA25S dataset. This high-entropy alloy dataset includes parameters such as atom types, coordinates of each atom, force on each atom, total number of atoms, cell size, lattice scaling factor, total energy of the system, virial stress tensor of the system, and free energy of the system. It is divided into training, validation, and test sets. The model is trained using the training and validation sets, and tested using the test set. Experiments are conducted under 20 different seeds, and the results are evaluated using three metrics: ACC, NMI, and ARI. The average value is used as the experimental result. The experiment is compared with the k-means method, and the results show that the proposed clustering method outperforms the k-means method.
[0070] Table 1. Results of a high-entropy alloy crystal structure clustering method incorporating knowledge from a large language model. ; Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.
[0071] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A high-entropy alloy crystal structure clustering method integrating knowledge from a large language model, characterized in that: include: The original high-entropy alloy atomic-scale data was cleaned to remove management fields that were irrelevant to the crystal structure, while retaining core physical properties such as atom type, number of atoms, total energy, three-dimensional coordinates of atoms, and three-dimensional forces. The fields corresponding to missing atom types were uniformly filled with zero values to form structured tabular data. Based on structured tabular data, the attribute values of each sample are used to construct an original feature vector, and standardization is performed on all feature dimensions to eliminate the influence of dimensional differences and obtain a numerical embedding representation. By using a large language model to parse the semantic relationship between attribute names and attribute values in numerical embedding representations, professional natural language text describing the microstructural characteristics of samples is generated. Professional natural language text is input into a dedicated text embedding model and converted into semantic text embedding vectors corresponding to the samples. Using numerical embedding representation and semantic text embedding vector as bimodal inputs, neighborhood distributions are constructed in numerical space and text space respectively, and the association mapping between the two is established through cross-modal mutual distillation mechanism; Based on the neighborhood distribution, the neighborhood distillation loss, modal consistency loss, and entropy regularization loss are calculated, and the clustering objective function is jointly optimized to output the clustering results of high-entropy alloy crystal structures.
2. The high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in claim 1, characterized in that: The process involves cleaning the original high-entropy alloy atomic-scale data, removing management fields irrelevant to the crystal structure, and retaining core physical properties such as atom type, number of atoms, total energy, atomic three-dimensional coordinates, and three-dimensional forces. Missing atom type fields are uniformly filled with zero values to form structured tabular data. The specific steps are as follows: Identify fields in the original data that contain sample number, test batch, and experimenter identifier, and remove these fields from the dataset; For the types of atoms appearing in each sample Extract the three-dimensional coordinates of all atoms in this category. , , ) and three-dimensional forces ( , , ),in Indicates the atom type index, Indicates the first in this category One atom; Calculate the mean values for the coordinate components of the same type of atom: ; The mean values of the force components were also calculated separately: ; in, For the types in the sample The number of atoms; If a certain type of atom If it does not appear in the current sample, set its atom count to 0, and set the coordinate mean. With the average force Set all to 0; Total energy of the sample With the total number of atoms It is directly retained as a global attribute field without aggregation or transformation; All fields corresponding to the types of atoms are arranged in a fixed order to form a single structured record.
3. The high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in claim 2, characterized in that: The method based on structured tabular data constructs an original feature vector from the attribute values of each sample, and performs standardization on all feature dimensions to eliminate the influence of dimensional differences, thereby obtaining a numerical embedding representation. The specific steps are as follows: Concatenate all numerical fields from each row of the structured table into the original feature vector in a predefined order. ,in This represents the total number of feature dimensions. For the first Dimensional features in all Calculate the mean from each sample. with standard deviation : ; For each sample, the first Perform Z-score normalization on 3D features: ； Obtain the standardized numerical embedding representation This is used for subsequent bimodal modeling.
4. The high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in claim 3, characterized in that: The method of using a large language model to analyze the semantic relationship between attribute names and attribute values in numerical embedding representations to generate professional natural language text describing the microstructural characteristics of samples involves the following steps: Sequencing the column names of the structured table { } and the standardized values of the corresponding samples { The key-value pair format is used to organize the input prompts; the Llama large language model, fine-tuned with a materials science corpus, is invoked to infer and output a natural language description based on the prompts. This description encompasses semantic information related to crystal structure, such as atomic composition ratios, typical interatomic spacing trends, overall force equilibrium state, and energy stability; ensuring the generated text... It does not introduce hypothetical content not included in the original data, but only performs logical deduction and standardized terminology based on the input fields.
5. The high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in claim 4, characterized in that: The specific steps for inputting professional natural language text into a dedicated text embedding model and converting it into a semantic text embedding vector corresponding to the sample are as follows: The generated natural language description... The input is fed into the nomic-embedded-text embedding model; the encoder of this model outputs a fixed-dimensional... Dense vectors This vector characterizes the structural chemical properties of high-entropy alloys in semantic space; the vector With numerical embedding representation These form bimodal feature pairs for subsequent cross-modal alignment.
6. The high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in claim 5, characterized in that: The method uses numerical embedding representations and semantic text embedding vectors as bimodal inputs, constructs neighborhood distributions in the numerical space and text space respectively, and establishes a correlation mapping between the two through a cross-modal mutual distillation mechanism. The specific steps are as follows: First, for the original data vector of each sample... Large language model First, the table header, i.e., the attribute names, is parsed. Then, each attribute name is associated with its corresponding sample attribute value for understanding. Finally, a natural language text describing the characteristic of this sample using professional knowledge is generated, as shown in the following formula: the remaining, For the first The original vector of each sample, For large language models, The first generation generated after language inference for large models The text descriptions of each sample; then, the nomic-embed-text model is specifically designed to convert the text into vector embeddings, transforming the text descriptions generated by the large language model. As input, through a pre-trained nomic-embed-text model The embedding interface can convert it into the corresponding numerical text embedding vector, as shown in the following formula: in, This is the embedding interface for the nomic-embed-text model. For the first The text embedding vector of each sample.
7. The high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in claim 6, characterized in that: The method involves calculating neighborhood distillation loss, modal consistency loss, and entropy regularization loss based on neighborhood distribution, jointly optimizing the clustering objective function, and outputting the clustering results of high-entropy alloy crystal structures. The specific steps are as follows: For each sample... Use FAISS to query it Neighbor set, defining the neighborhood distribution of the table sample: in, For the first The original data embedding vector of each sample, For the first The text embedding vector of each sample, For similarity function, For the first The neighbor set of a sample The index of the neighbor sample. For tabular data samples With corresponding text neighbors Similarity probability; text neighborhood distribution: in, Text embedding samples Neighbors of the corresponding table data The similarity probability; Neighborhood distillation losses: in, The total number of samples, For samples in text space The neighborhood distribution vector, For samples in the tabular data space The neighborhood distribution vector, Kullback-Leibler divergence measures the difference between two distributions. and Differences This is the distillation loss, used to align the neighborhood distribution of the table and text. The smaller the value, the more consistent the neighborhood distribution of the two modalities. Modal consistency loss: in, Predict the distribution for the data sample in the i-th table. For the predicted distribution of the text, The total number of samples, Modality consistency loss measures the consistency between table modality and text modality in cluster prediction distribution; the smaller the value, the more consistent the predictions of the two modalities. Entropy regularization (to prevent collapse): Designing the entropy term encourages overall uniformity and suppresses overconfidence. Among them, the average predicted distribution of the table data samples Average prediction distribution of text data samples , Entropy regularization loss measures the uniformity of the predicted distribution and prevents all samples from being classified into the same class. The larger the value, the higher the entropy and the more uniform the distribution. The total loss is: in As weight, The smaller the value, the better the overall training effect of the model.
8. The high-entropy alloy crystal structure clustering method that integrates large-scale language model knowledge as described in claim 7, characterized in that: The clustering results are used to guide the crystal structure design and performance prediction of new high-entropy alloy materials. The specific steps are as follows: perform structural commonality analysis on high-entropy alloy samples belonging to the same category in the clustering output, and extract the statistical regularities of the category in terms of atomic composition ratio, average interatomic spacing, total energy distribution and force equilibrium state; construct a structure-category mapping rule base based on the statistical regularities; when a new high-entropy alloy atomic-scale data is input, perform the above full-process processing to obtain its category label. Based on the structural commonalities associated with category labels, the crystal symmetry, phase stability, and mechanical behavior tendencies of new materials are inferred. The inferred results are then fed back as prior knowledge into high-throughput screening or first-principles calculation processes to narrow the search space for candidate structures and improve the efficiency of new material research and development.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the high-entropy alloy crystal structure clustering method according to any one of claims 1 to 8, which incorporates large-scale language model knowledge.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the high-entropy alloy crystal structure clustering method according to any one of claims 1 to 8, which incorporates large model knowledge of fused language.