A method and system for directional generation and candidate result evaluation of lithium-containing crystal materials

Abstract

This application relates to the field of crystal structure design technology, and particularly to a method and system for directional generation and candidate result evaluation of lithium-containing crystal materials. The method includes: standardizing the structure of a lithium-containing crystal sample; associating the standardized structure of the lithium-containing crystal sample with material property labels to form joint condition information; fine-tuning the basic crystal generation model based on the joint condition information to enhance the model's representation and generation capabilities for lithium-containing systems; requiring the chemical system corresponding to the generated result to contain lithium when sampling the model or inputting condition labels, thereby generating candidate crystal structures; evaluating the generated candidate crystal structures using MatterSim and performing batch analysis of the structures; and outputting lithium-containing candidate crystal structures based on the evaluation results. This application achieves efficient generation, rapid stability assessment, and candidate material prediction of lithium-containing crystal structures by fine-tuning the basic generation model with lithium constraints and combining it with machine learning force field evaluation.

Description

Technical Field

[0001] This application relates to the field of crystal structure design technology, and in particular to a method and system for directional generation and candidate result evaluation of lithium-containing crystal materials. Background Technology

[0002] Crystal structure design is one of the core issues in new material development. Traditional crystal structure prediction methods mainly include structure search methods based on random search, evolutionary algorithms, particle swarm optimization, and first-principles calculations, such as AIRSS, USPEX, and CALYPSO. These methods can usually achieve high accuracy within a small search space, but when the system composition is complex, the space group is diverse, or the search target has explicit elemental constraints, the computational cost increases dramatically, making it difficult to meet the needs of large-scale new material design.

[0003] With the development of artificial intelligence technology, generative model-based crystal structure design methods are gradually becoming a new direction. Existing works such as iMatGen, CrystalGAN, DiffCSP, DiffCSP++, and MatterGen apply variational autoencoders, generative adversarial networks, or diffusion models to crystal structure generation. Among them, MatterGen is a generative model for inorganic material design, capable of generating crystal structures under given conditions, and can introduce chemical composition, space group, or some material properties as control conditions, thus having significant advantages in inverse material design.

[0004] However, existing crystal material generation methods generally suffer from problems such as slow generation speed, insufficient targeting, weak lithium-containing constraints in the generation results, low stability of candidate structures, high screening costs, and a disconnect between property prediction and stability assessment when developing lithium-containing electrode materials or lithium-containing electrolyte materials. It is difficult to efficiently obtain lithium-containing crystal candidate structures with application potential. Summary of the Invention

[0005] This application provides a method and system for directional generation and candidate result evaluation of lithium-containing crystal materials. By fine-tuning the basic generation model with lithium-containing constraints and combining it with machine learning force field evaluation, it achieves efficient generation of lithium-containing crystal structures, rapid stability judgment, and prediction of candidate materials.

[0006] To address the aforementioned technical problems, in a first aspect, embodiments of this application provide a method for the directional generation and candidate result evaluation of lithium-containing crystal materials, comprising the following steps: First, a structurally standardized expression of a lithium-containing crystal sample is performed; then, the structurally standardized expression of the lithium-containing crystal sample is associated with material property tags to form joint condition information; next, based on the joint condition information, the basic crystal generation model is fine-tuned to enhance the model's representation and generation capabilities for lithium-containing systems; when sampling the model or inputting condition tags, the chemical system corresponding to the generated result is required to contain lithium, generating candidate crystal structures; then, the generated candidate crystal structures are evaluated using MatterSim, and the structures are analyzed in batches to output the evaluation results; finally, lithium-containing candidate crystal structures are output based on the evaluation results.

[0007] In some exemplary embodiments, the lithium-containing crystal sample is structurally normalized, including: representing each crystal structure of the lithium-containing crystal sample as lattice parameter information, atom type information, and fractional coordinate information.

[0008] In some exemplary embodiments, the material property labels include: a band gap calculated based on density functional theory, denoted as dft_band_gap; an energy above the thermodynamic convex hull, denoted as energy_above_hull; a crystal space group, denoted as space_group; and a chemical system label, denoted as chemical_system.

[0009] In some exemplary embodiments, the basic model for crystal generation is the MatterGen model; the basic model for crystal generation is fine-tuned based on joint conditional information, including: using joint conditional information of lithium-containing crystal samples to conduct targeted training of the basic model for crystal generation; the training hardware uses an A100 80GB graphics processing unit to meet the requirements of the diffusion generation model in terms of parameter updates, batch training and memory usage.

[0010] In some exemplary embodiments, during the targeted training process, the model receives joint conditional information of lithium-containing crystal samples and strengthens the conditional response to the lithium-containing chemical system during the loss function optimization process, so that the generated distribution is focused towards the lithium-containing crystal space.

[0011] In some exemplary embodiments, when sampling the model or inputting conditional labels, it is required that the chemical system corresponding to the generated result contains lithium. This includes: in the structure filtering step after model sampling or when inputting conditional labels, it is required that the chemical system corresponding to the generated result is a binary, ternary, or higher-order crystal system composed of lithium and one or more non-lithium elements. By introducing lithium element constraints, a large number of structural outputs in the model that are irrelevant to the research and development goals of lithium battery materials are avoided, thereby improving the generation efficiency and candidate quality.

[0012] In some exemplary embodiments, the generated candidate crystal structures are evaluated using MatterSim, the structures are analyzed in batches, and the evaluation results are output. This includes: inputting the generated candidate crystal structures into the MatterSim evaluation process, using a self-written program to call the MatterSim machine learning force field for rapid evaluation, quickly filtering out obviously unreasonable structures, and retaining candidate crystals that are more likely to be stable; the evaluation results include whether the structure tends to be stable, whether the energy change is reasonable, and whether there is an obviously non-physical configuration.

[0013] In some exemplary embodiments, after outputting lithium-containing candidate crystal structures based on the evaluation results, the method further includes classifying the output lithium-containing candidate crystal structures.

[0014] In some exemplary embodiments, the output lithium-containing candidate crystal structures are classified, including: dividing the output lithium-containing candidate crystal structures into ordinary effective structures, high-stability structures, and stable-unique-novel structures; wherein, a stable-unique-novel structure refers to a structure that, after structural determination, simultaneously meets the stability requirements, does not repeat the generated structures, and is novel relative to the reference structure set.

[0015] Secondly, this application also provides a system for the directional generation and candidate result evaluation of lithium-containing crystal materials. This system is used to implement the method for the directional generation and candidate result evaluation of lithium-containing crystal materials described in the above embodiments. The system includes: a lithium-containing sample input module, an attribute label construction module, a model fine-tuning module, a lithium-containing constraint generation module, an evaluation module, and a candidate structure output module connected in sequence. The lithium-containing sample input module is used to perform structural standardization on lithium-containing crystal samples. The attribute label construction module is used to associate the standardized structural expression of lithium-containing crystal samples with material attribute labels to form joint condition information. The model fine-tuning module is used to fine-tune the basic crystal generation model based on the joint condition information, enhancing the model's representation and generation capabilities for lithium-containing systems. The lithium-containing constraint generation module is used to require the chemical system corresponding to the generated result to contain lithium elements when the model is sampled or condition labels are input, generating candidate crystal structures. The evaluation module is used to perform MatterSim evaluation on the generated candidate crystal structures, conduct batch analysis of the structures, and output the evaluation results. The candidate structure output module is used to output lithium-containing candidate crystal structures based on the evaluation results.

[0016] The technical solution provided in this application has at least the following advantages: This application provides a method and system for directional generation and candidate result evaluation of lithium-containing crystal materials. The method includes the following steps: First, the structure of the lithium-containing crystal sample is standardized; then, the standardized structure of the lithium-containing crystal sample is associated with the material property label to form joint condition information; next, based on the joint condition information, the basic crystal generation model is fine-tuned to enhance the model's ability to represent and generate lithium-containing systems; when the model is sampled or the condition label is input, the chemical system corresponding to the generated result is required to contain lithium, generating candidate crystal structures; then, the generated candidate crystal structures are evaluated using MatterSim, the structures are analyzed in batches, and the evaluation results are output; finally, the lithium-containing candidate crystal structures are output based on the evaluation results.

[0017] The method and system for directional generation and candidate result evaluation of lithium-containing crystal materials provided in this application, while retaining the ability of diffusion generation models to efficiently explore structural space, achieves directional generation of lithium-containing crystal materials through lithium-containing constraint fine-tuning and rapid stability evaluation; at the same time, it enhances the model's learning ability on the target crystal distribution by utilizing material property tags, thereby improving the stability and effectiveness of the generation results, and rapidly evaluates candidate structures by calling the MatterSim machine learning force field through a self-developed program, forming a high-efficiency technical process that can be used for the development of lithium battery materials. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0019] Figure 1 This is a flowchart illustrating a method for directional generation and candidate result evaluation of lithium-containing crystal materials, provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of a lithium-containing crystal material directional generation and candidate result evaluation system provided in an embodiment of this application. Detailed Implementation

[0021] As can be seen from the background technology, existing crystal material generation methods generally suffer from problems such as slow generation speed, insufficient targeting, weak lithium-containing constraints in the generation results, low stability of candidate structures, high screening costs, and a disconnect between property prediction and stability assessment when developing lithium-containing electrode materials or lithium-containing electrolyte materials. It is difficult to efficiently obtain lithium-containing crystal candidate structures with application potential.

[0022] The closest existing technical solution to this application is to directly train an existing large-scale inorganic crystal database based on MatterGen or a similar diffusion model, then generate candidate crystal structures under general conditions, and finally perform posterior screening through first-principles calculations or machine learning potential functions. The basic process of this type of solution usually includes: first, representing the crystal structure as lattice parameters, atom types, and atomic fraction coordinates; then, using a diffusion model to learn the crystal structure distribution; outputting candidate crystals based on given composition or property conditions during the generation stage; and finally, screening the generated structures through structure validity checks, energy assessments, and calculations of band structures or thermodynamic indices.

[0023] For example, MatterGen-related schemes typically introduce attribute labels such as band gap, energy above the convex hull value, space group, and chemical system to enhance the model's ability to learn the distribution of the target material. After generation, the stability of the structure is verified by computational simulation or machine learning potential functions. This type of method has proven that the generated models can be used for crystal material design, but its general models often mainly pursue full element coverage and broad generalization ability, and there is still room for further improvement in the targeted generation effect for specific element systems, especially lithium-containing systems.

[0024] The main drawbacks of existing technologies are as follows: First, while general crystal generation models can cover a wide range of elemental systems, they lack specific enhancements for lithium-containing materials, resulting in insufficient proportions of lithium-ion battery-related structures, inadequate structural stability, and insufficient candidate validity in the generated results. Second, existing solutions typically focus more on "whether a structure can be generated," while the evaluation chain for "whether the generated structure is stable, unique, novel, and has further development value" is not tightly integrated, leading to high subsequent screening costs. Third, some methods rely heavily on costly first-principles calculations for posterior verification, making it difficult to achieve rapid closed-loop processing in large-scale candidate generation scenarios. Fourth, although existing models can incorporate property conditions, they lack an integrated technical solution that considers generation, stability assessment, and candidate prediction when fine-tuning around lithium element constraints.

[0025] To address the aforementioned shortcomings, the purpose of this application is to provide a novel method and system for developing lithium-containing crystal materials. While retaining the efficient ability of the diffusion generation model to explore structural space, it achieves the directional generation of lithium-containing crystal materials through lithium-containing constraint fine-tuning and rapid stability evaluation. At the same time, it utilizes material property tags to enhance the model's learning ability on the target crystal distribution, improving the stability and effectiveness of the generation results. Furthermore, it uses a self-developed program to call the MatterSim machine learning force field to quickly evaluate candidate structures, forming a highly efficient technical process that can be used for lithium battery material development.

[0026] The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0027] See Figure 1 This application provides a method for the directional generation and candidate result evaluation of lithium-containing crystal materials, including the following steps: Step S1: Perform structural normalization on the lithium-containing crystal sample.

[0028] Step S2: Associate the standardized structural representation of the lithium-containing crystal sample with the material property label to form joint conditional information.

[0029] Step S3: Based on the joint condition information, fine-tune the basic model of crystal formation to enhance the model's ability to represent and form lithium-containing systems.

[0030] Step S4: When sampling the model or inputting conditional labels, the chemical system corresponding to the generated result must contain lithium to generate candidate crystal structures.

[0031] Step S5: Evaluate the generated candidate crystal structures using MatterSim, perform batch analysis on the structures, and output the evaluation results.

[0032] Step S6: Output the lithium-containing candidate crystal structure based on the evaluation results.

[0033] This application provides a method for the directional generation and candidate result evaluation of lithium-containing crystal materials. The method includes six steps: lithium-containing sample input, attribute label construction, MatterGen model fine-tuning, lithium-constrained generation, MatterSim rapid evaluation, and candidate output. The method uses a basic crystal generation model as its core. During the training phase, four types of material attribute labels—band gap, energy above the convex hull, space group, and chemical system—are introduced to fine-tune the model for lithium-containing systems. In the generation phase, lithium constraints are added to the generation conditions, causing the model to preferentially generate crystal structures containing lithium. In the evaluation phase, a self-developed program calls the MatterSim machine learning force field to quickly determine the stability of the generated structures, analyze structural optimization trends, and screen candidates. Finally, the method outputs candidate crystal structures that meet the lithium-containing requirements and have a high probability of stability, and identifies and predicts stable, unique, and novel structures among them.

[0034] In some embodiments, step S1 involves structurally normalizing the lithium-containing crystal sample, including representing each crystal structure of the lithium-containing crystal sample as lattice parameter information, atom type information, and fractional coordinate information. Specifically, the method of this application first performs structurally normalized representation on the input sample. Each crystal structure is represented as lattice parameter information, atom type information, and fractional coordinate information, and further associated with material property labels.

[0035] In some embodiments, the material property labels in step S2 include: the band gap calculated based on density functional theory, denoted as dft_band_gap; the energy above the thermodynamic convex hull, denoted as energy_above_hull; the crystal space group, denoted as space_group; and the chemical system label, denoted as chemical_system. These four types of labels together constitute conditional information, enabling the model to learn not only "what the structure looks like" during training, but also "what kind of structure is closer to the distribution of the target material."

[0036] It should be noted that, in terms of attribute tags, in addition to dft_band_gap, energy_above_hull, space_group, and chemical_system, tags such as bulk modulus, formation energy, electronic conductivity, ion migration barrier, volume, and density can be further introduced to enhance the model's ability to generate targeted data for specific material properties.

[0037] In some embodiments, the basic model for crystal generation in step S3 is the MatterGen model; the basic model for crystal generation in step S3 is fine-tuned based on joint conditional information, including: using joint conditional information of lithium-containing crystal samples to conduct targeted training of the basic model for crystal generation; the training hardware uses an A100 80GB graphics processing unit to meet the requirements of the diffusion generation model in terms of parameter updates, batch training and memory usage.

[0038] This application builds upon the MatterGen crystal generation framework. MatterGen is essentially a generation model for inorganic crystal materials, which can generate new crystal structures by learning the joint distribution between lattice, atom types, and fractional coordinates. This application does not reinvent a basic generation model, but rather builds a dedicated technical process for lithium-containing material development on top of it, enabling the model to be further transformed from a general crystal generation capability into a targeted design capability for lithium-containing crystal materials.

[0039] In some embodiments, during targeted training, the model receives joint conditional information from lithium-containing crystal samples and strengthens the conditional response to the lithium-containing chemical system during loss function optimization, thereby focusing the generated distribution toward the lithium-containing crystal space.

[0040] Specifically, during the model training phase, this application uses lithium-containing crystal samples to fine-tune the basic MatterGen model. The goal of fine-tuning is not to expand the coverage of all elements, but to enhance the model's ability to represent and generate lithium-containing systems. Specifically, during training, the model receives structural representations and attribute labels, and strengthens the conditional response to lithium-containing chemical systems during loss function optimization, focusing the generated distribution towards the lithium-containing crystal space. The training hardware uses... The A100 80GB graphics processing unit is designed to meet the needs of the diffusion generation model in terms of parameter updates, batch training, and memory usage. This training method enables the model to acquire specialized generation capabilities for lithium-containing crystal systems.

[0041] It should be noted that the technical solution of this application can be modified in various ways. Regarding the basic generation model, it is not limited to MatterGen; it can also be replaced with other three-dimensional isotropic generation models that possess crystal structure diffusion generation capabilities, as long as they can achieve the joint generation of lattice, atom type, and fractional coordinates, and support attribute condition control.

[0042] In some embodiments, step S4 requires that the chemical system corresponding to the generated result contains lithium elements when sampling the model or inputting conditional labels. This includes requiring that the chemical system corresponding to the generated result is a binary, ternary, or higher-order crystal system composed of lithium elements and one or more non-lithium elements in the structure filtering step after model sampling or when inputting conditional labels. By introducing lithium element constraints, a large number of structural outputs in the model that are unrelated to the research and development goals of lithium battery materials are avoided, thereby improving the generation efficiency and candidate quality.

[0043] Specifically, in the generation stage, this application introduces a lithium element constraint. This lithium element constraint means that, during model sampling or condition input, the chemical system corresponding to the generated result is explicitly required to contain lithium, preferably a binary, ternary, or higher-order crystal system composed of Li and one or more non-lithium elements. This constraint can be reflected in the condition label or in the post-sampling structure filtering step.

[0044] It should be noted that, in terms of elemental constraints, this application can also be extended from lithium elemental constraints to energy storage-related elements such as sodium, potassium, magnesium, and aluminum, to achieve the directional generation of sodium battery, potassium battery, or multivalent ion battery materials.

[0045] In some embodiments, the generated candidate crystal structures are evaluated using MatterSim, and the structures are analyzed in batches to output evaluation results. This includes: inputting the generated candidate crystal structures into the MatterSim evaluation process, using a self-written program to call the MatterSim machine learning force field for rapid evaluation, quickly filtering out obviously unreasonable structures, and retaining candidate crystals that are more likely to be stable; the evaluation results include whether the structure tends to be stable, whether the energy change is reasonable, and whether there are obvious non-physical configurations.

[0046] Specifically, in the structural evaluation stage, this application does not directly rely on full first-principles calculations for posterior screening. Instead, it uses a self-developed program to call MatterSim machine learning force fields for rapid evaluation. MatterSim is a deep learning atomic simulation model that can estimate the energy, stability trends, and structural rationality of crystal structures at a relatively low computational cost. This application uses a self-developed program to automatically feed the generated structures into the MatterSim evaluation process for batch analysis, outputting results including whether the structure tends to be stable, whether the energy changes are reasonable, and whether there are obvious non-physical configurations. Based on this, obviously unreasonable structures can be quickly screened out, while more likely stable candidate crystals are retained.

[0047] It should be noted that, in the stability assessment stage, in addition to the MatterSim machine learning force field, other inter-atom machine learning potential function models can also be used, or a method of cascading verification of machine learning force field and first-principles calculation can be adopted.

[0048] In some embodiments, after outputting lithium-containing candidate crystal structures based on the evaluation results, the method further includes classifying the output lithium-containing candidate crystal structures.

[0049] In some embodiments, the output lithium-containing candidate crystal structures are classified, including: dividing the output lithium-containing candidate crystal structures into ordinary effective structures, high-stability structures, and stable-unique-novel structures; wherein, stable-unique-novel structures refer to structures that, after structural determination, simultaneously meet the stability requirements, do not repeat existing structures, and are novel relative to the reference structure set. Such structures can serve as key candidate targets for subsequent lithium battery material development.

[0050] An implementation result of this application shows that the lithium-containing confined generation model obtained using the above scheme can improve the stability of the generated crystal structure to 70% to 80%. Simultaneously, approximately 20% of the generated results are stable, unique, and novel structures, indicating that this application can not only generate lithium-containing crystals but also output a high proportion of new candidate structures with potential for further development. This result was obtained through evaluation using a self-developed program calling the MatterSim machine learning force field.

[0051] It should be noted that, in addition to being used for the development of lithium battery cathode, anode and electrolyte materials, this application can also be used for high-throughput prediction and screening of lithium-containing superionic conductors, lithium-containing functional ceramics and other lithium-containing inorganic crystals.

[0052] Furthermore, this application also provides a system for the directional generation and candidate result evaluation of lithium-containing crystal materials. This system is used to implement the method for the directional generation and candidate result evaluation of lithium-containing crystal materials described in the above embodiments. The structural diagram of the system is shown below. Figure 2 As shown, the modules work collaboratively to develop lithium-containing crystalline materials. The system includes: a lithium-containing sample input module, an attribute label construction module, a model fine-tuning module, a lithium-containing constraint generation module, an evaluation module, and a candidate structure output module, connected in sequence. Specifically, the lithium-containing sample input module is used to standardize the structure of lithium-containing crystal samples; the attribute label construction module is used to associate the standardized structure of lithium-containing crystal samples with material attribute labels to form joint condition information; the model fine-tuning module is used to fine-tune the basic crystal generation model based on the joint condition information, enhancing the model's ability to represent and generate lithium-containing systems; the lithium-containing constraint generation module requires the chemical system corresponding to the generated result to contain lithium when the model is sampled or condition labels are input, generating candidate crystal structures; the evaluation module is used to evaluate the generated candidate crystal structures using MatterSim, perform batch analysis of the structures, and output the evaluation results; and the candidate structure output module is used to output lithium-containing candidate crystal structures based on the evaluation results. These modules can be deployed on the same computing server or in a distributed computing platform.

[0053] The key features of the method and system for directional generation and candidate result evaluation of lithium-containing crystal materials provided in this application are: 1. Based on the MatterGen crystal generation model, construct a directional fine-tuning scheme for the development of lithium-containing crystal materials.

[0054] 2. Introduce lithium element constraints during model training and generation to limit or preferentially limit the generated results to lithium-containing crystal structures.

[0055] 3. Four attribute labels, namely dft_band_gap, energy_above_hull, space_group, and chemical_system, are used as joint training conditions to improve the model's ability to generate lithium-containing systems and its structural stability.

[0056] 4. A self-developed program was used to call MatterSim machine learning force field to perform rapid stability evaluation and candidate screening of the generated lithium-containing crystal structure.

[0057] 5. Establish an integrated technical process for “lithium-containing directional generation – rapid stability assessment – ​​stable / unique / novel candidate output”.

[0058] 6. Based on the above process, lithium-containing crystals with a stability of 70% to 80% were generated, and approximately 20% of the stable, unique, and novel structures were obtained.

[0059] In a preferred embodiment, fine-tuning training is performed using 4,634 lithium-containing crystal sample data points, with training hardware employing an NVIDIA® A100 80GB graphics processing unit. The resulting model can limit the generated results to lithium-containing crystal structures, achieving a stability of 70% to 80% for the generated crystals, while identifying approximately 20% of stable, unique, and novel structures.

[0060] Compared with existing schemes based on general crystal generation models, the method and system for directional generation and candidate result evaluation of lithium-containing crystal materials provided in this application have the following advantages: First, this application does not generalize to generate arbitrary inorganic crystals, but rather focuses on model-oriented enhancement around the development goals of lithium-containing crystal materials. Therefore, it is more in line with the research and development scenarios of lithium battery materials and can significantly improve the correlation between the generated results and the application goals.

[0061] Secondly, this application combines four types of attribute labels—band gap, energy above the convex hull, space group, and chemical system—for model training, enabling the model to learn not only structural forms but also stability and property-related distribution characteristics, thus resulting in more stable generated results.

[0062] Furthermore, this application uses lithium element confinement to suppress the output of irrelevant chemical systems, which can reduce invalid sampling and improve the efficiency of candidate structure generation.

[0063] Finally, this application uses a self-developed program to call MatterSim machine learning force field for rapid evaluation, which is more efficient than the traditional screening method that relies on large-scale first-principles calculations and is suitable for rapid initial screening of large-scale candidate structures.

[0064] From a technical perspective, the model finally trained in this application can improve the stability of generating lithium-containing crystal structures to 70% to 80%, while being able to identify about 20% of stable, unique, and novel structures. This shows that this application is superior to simple general generation schemes in terms of candidate structure quality and novel candidate discovery capability.

[0065] This application has been validated through model training and simulation evaluation, proving its feasibility. During implementation, lithium-containing crystal samples were used to fine-tune and train the MatterGen base model. The training hardware was... The A100 80GB graphics processing unit was used. During training, four attribute labels—dft_band_gap, energy_above_hull, space_group, and chemical_system—were used in combination to improve the model's learning performance on lithium-containing crystal systems. After training, the model's generated results were evaluated in batches, and stability analysis was performed using a self-developed program calling the MatterSim machine learning force field.

[0066] The results show that the model, finely tuned by the method described in this application, can stably generate Li-containing crystal structures with a stability of 70% to 80%. Among the generated samples, approximately 20% exhibit stable, unique, and novel structures, indicating that this application not only achieves the directional generation of lithium-containing crystals but also possesses good novel structure discovery capabilities and subsequent screening value. These results fully demonstrate the feasibility of the technical solution presented in this application, making it suitable for the early-stage development and prediction of lithium-containing electrode materials, lithium-containing solid electrolyte materials, and other lithium-containing functional crystal materials.

[0067] It should be noted that the technical solution of this application can be modified in various ways. First, regarding the basic generation model, it is not limited to MatterGen; it can be replaced with other three-dimensional isotropic generation models with crystal structure diffusion generation capabilities, as long as they can achieve joint generation of lattice, atom type, and fractional coordinates, and support attribute condition control. Second, regarding attribute tags, in addition to dft_band_gap, energy_above_hull, space_group, and chemical_system, tags such as bulk modulus, formation energy, electronic conductivity, ion migration barrier, volume, and density can be further introduced to enhance the model's ability to generate specific material properties in a targeted manner. Third, in the stability evaluation stage, in addition to the MatterSim machine learning force field, it can be replaced with other interatomic machine learning potential function models or a method of cascading verification using machine learning force fields and first-principles calculations can be adopted. Furthermore, regarding elemental constraints, this application can be extended from lithium element constraints to energy storage-related elements such as sodium, potassium, magnesium, and aluminum to achieve the targeted generation of sodium-ion batteries, potassium-ion batteries, or multivalent ion battery materials. Finally, in addition to its application in the development of lithium-ion battery cathode, anode, and electrolyte materials, this application can also be used for high-throughput prediction and screening of lithium-containing superionic conductors, lithium-containing functional ceramics, and other lithium-containing inorganic crystals.

[0068] Based on the above technical solutions, this application provides a method and system for directional generation and candidate result evaluation of lithium-containing crystal materials. The method includes the following steps: First, the structure of the lithium-containing crystal sample is standardized; then, the standardized structure of the lithium-containing crystal sample is associated with the material property label to form joint condition information; next, based on the joint condition information, the basic model for crystal generation is fine-tuned to enhance the model's ability to represent and generate lithium-containing systems; when sampling the model or inputting condition labels, the chemical system corresponding to the generated result is required to contain lithium, generating candidate crystal structures; then, the generated candidate crystal structures are evaluated using MatterSim, the structures are analyzed in batches, and the evaluation results are output; finally, the lithium-containing candidate crystal structures are output based on the evaluation results.

[0069] The method and system for directional generation and candidate result evaluation of lithium-containing crystal materials provided in this application, while retaining the ability of diffusion generation models to efficiently explore structural space, achieves directional generation of lithium-containing crystal materials through lithium-containing constraint fine-tuning and rapid stability evaluation; at the same time, it enhances the model's learning ability on the target crystal distribution by utilizing material property tags, thereby improving the stability and effectiveness of the generation results, and rapidly evaluates candidate structures by calling the MatterSim machine learning force field through a self-developed program, forming a high-efficiency technical process that can be used for the development of lithium battery materials.

[0070] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A method for directional generation and candidate result evaluation of a lithium-containing crystal material, characterized in that, Includes the following steps: Structural normalization representation of lithium-containing crystal samples; The standardized structural representation of lithium-containing crystal samples is associated with material property labels to form joint conditional information; Based on joint conditional information, the basic model for crystal formation is fine-tuned to enhance the model's ability to represent and generate lithium-containing systems. When sampling the model or inputting conditional labels, it is required that the chemical system corresponding to the generated result contains lithium, and candidate crystal structures are generated. The generated candidate crystal structures are evaluated using MatterSim, and the structures are analyzed in batches to output the evaluation results. Based on the evaluation results, output the candidate crystal structure containing lithium.

2. The method for directional generation and candidate result evaluation of lithium-containing crystal materials according to claim 1, characterized in that, Structural normalization representation of lithium-containing crystal samples, including: Each crystal structure of the lithium-containing crystal sample is represented by lattice parameter information, atom type information, and fractional coordinate information.

3. The method for directional generation and candidate result evaluation of lithium-containing crystal materials according to claim 1, characterized in that, The material property labels include: The band gap calculated based on density functional theory is denoted as dft_band_gap; The energy excess relative to the thermodynamic convex hull is represented as energy_above_hull; Crystal space group, denoted as space_group; The chemical system label is represented as chemical_system.

4. The method of claim 1, wherein the method further comprises: determining a plurality of candidate results; and determining a plurality of candidate result scores for the plurality of candidate results. The fundamental model for crystal formation is the MatterGen model; based on joint conditional information, the fundamental model for crystal formation is fine-tuned, including: The basic model for crystal formation is trained in a targeted manner using the joint conditional information of lithium-containing crystal samples; The training hardware uses an A100 80GB graphics processing unit to meet the needs of the diffusion generative model in terms of parameter updates, batch training, and memory usage.

5. The method for directional generation and candidate result evaluation of lithium-containing crystal materials according to claim 4, characterized in that, During the targeted training process, the model receives joint conditional information from lithium-containing crystal samples and strengthens the conditional response to the lithium-containing chemical system during the loss function optimization process, so that the generation distribution focuses on the lithium-containing crystal space.

6. The method of claim 1, wherein the method further comprises: When sampling the model or inputting conditional labels, the chemical system corresponding to the generated result must contain lithium, including: In the structure filtering step after model sampling or when inputting condition labels, it is required that the chemical system corresponding to the generated result is a binary, ternary or higher-order crystal system composed of lithium and one or more non-lithium elements. By introducing lithium element constraints, a large number of structural outputs in the model that are irrelevant to the research and development goals of lithium battery materials are avoided, thereby improving the generation efficiency and candidate quality.

7. The method of claim 1, wherein the method further comprises: determining a plurality of candidate results; and evaluating the plurality of candidate results. The generated candidate crystal structures are evaluated using MatterSim, and batch analysis of the structures is performed. The evaluation results are output, including: The generated candidate crystal structures are input into the MatterSim evaluation process. A self-written program is used to call the MatterSim machine learning force field for rapid evaluation, quickly eliminating obviously unreasonable structures and retaining more likely stable candidate crystals. The evaluation results include whether the structure tends to be stable, whether the energy changes are reasonable, and whether there are obvious non-physical configurations.

8. The method for directional generation and candidate result evaluation of lithium-containing crystal materials according to claim 1, characterized in that, After outputting lithium-containing candidate crystal structures based on the evaluation results, the process also includes classifying the output lithium-containing candidate crystal structures.

9. The method for directional generation and candidate result evaluation of lithium-containing crystal materials according to claim 8, characterized in that, The output lithium-containing candidate crystal structures are classified, including: The output lithium-containing candidate crystal structures are divided into ordinary effective structures, high-stability structures, and stable-unique-novel structures. Among them, stable-unique-novel structures refer to structures that simultaneously meet the stability requirements after structure determination, do not repeat the generated structures, and are novel relative to the reference structure set.

10. A system for directional generation and candidate result evaluation of lithium-containing crystal materials, the system being used to implement the method for directional generation and candidate result evaluation of lithium-containing crystal materials as described in any one of claims 1 to 9, characterized in that, The system includes: a lithium-containing sample input module, an attribute label construction module, a model fine-tuning module, a lithium-containing constraint generation module, an evaluation module, and a candidate structure output module, connected in sequence; among them, The lithium-containing sample input module is used to perform structural standardization expression of lithium-containing crystal samples; The attribute tag construction module is used to associate the standardized structural expression of lithium-containing crystal samples with material attribute tags to form joint conditional information; The model fine-tuning module is used to fine-tune the basic model of crystal formation based on joint condition information, thereby enhancing the model's ability to represent and generate lithium-containing systems. The lithium-containing constraint generation module is used to require the chemical system corresponding to the generated result to contain lithium element when the model is sampled or the condition label is input, and to generate candidate crystal structures. The evaluation module is used to evaluate the generated candidate crystal structures using MatterSim, perform batch analysis of the structures, and output the evaluation results. The candidate structure output module is used to output lithium-containing candidate crystal structures based on the evaluation results.

Images