Reverse design method applied to solid electrolyte material

Abstract

The invention discloses a reverse design method applied to a solid electrolyte material. The method comprises the following steps: firstly, constructing an inorganic solid electrolyte material field knowledge base; generating a chemical formula conforming to physical and chemical rationality through a chemical formula generator based on guidance of a domain knowledge base; based on the generated chemical formula, further utilizing a large language model to generate a candidate crystal structure information file, and performing optimization in combination with a Monte Carlo tree search method; in the optimization process, in order to evaluate the performance of the generated candidate crystal structure, a graph neural network model driven by domain knowledge is cooperated to carry out multi-target performance prediction, a prediction result is used as reward feedback, and then the generation quality of the candidate crystal structure is optimized; and finally, by taking activation energy and convex hull energy as indexes, screening out a potential novel inorganic solid electrolyte material meeting performance requirements. The development efficiency of the inorganic solid electrolyte material is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of generative artificial intelligence, and in particular relates to a reverse design method applied to solid electrolyte materials. Background Art

[0002] Solid-state electrolytes are crucial to promoting the development of electrochemical energy storage technology, but their development has been hindered by the large amount of resources required to explore their complex structures. Therefore, there is an urgent need to develop efficient and accurate research methods to accelerate the research and development of inorganic solid-state electrolytes. Although large language models can achieve efficient material reverse design by clarifying the research system and target performance, the generation results of large language models are very random and the generation process is uncontrollable, which can easily lead to results that are inconsistent with domain knowledge. In addition, how to conduct intelligent and automated reverse design of materials from scratch, and how to comprehensively consider the complex relationship between multiple properties and perform multi-objective optimization that balances multiple performance requirements is still a problem that needs to be solved urgently.

[0003] At present, generative artificial intelligence models have shown significant advantages in material reverse design, but the existing attribute-guided generation paradigm can only guide the generation range to the ideal target area, and the generation results are random. Its physical rationality and stability need to be further verified. Moreover, most of its generation processes are still in a "black box" state, lacking clear multi-step reasoning logic and interpretability. This uninterpretability makes it difficult for the generation model to reveal the causal relationship between the generated samples and the target attributes, and it is also impossible to achieve fine-grained control of the material microstructure and key attributes, thereby limiting the reliability and practicality of the generation results. In addition, the performance of solid electrolytes is affected by complex multi-scale factors, and there may be nonlinear coupling and potential conflict relationships between different target attributes, making it extremely challenging to simultaneously optimize multiple attributes of solid electrolyte materials under a unified generation framework. Therefore, in the design and analysis process of solid electrolyte material data, especially in the complete process of reverse design, substantial material field knowledge should be embedded, starting from actual research problems and specific application needs, and comprehensively considering factors such as the material's microstructure, physical and chemical properties, and multi-objective performance indicators. At different research stages and data characteristics, the adaptation model should be selected under the guidance of domain knowledge to ensure that the model can effectively capture the complex structure of solid electrolytes and their structure-activity relationship with ion transport performance. Therefore, in order to accelerate the discovery of new high-performance solid electrolyte materials with reasonable physicochemical properties, a reverse design method applied to solid electrolyte materials is proposed. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a reverse design method applied to solid electrolyte materials, which improves the development efficiency of inorganic solid electrolyte materials.

[0005] To achieve the above object, the present invention provides a reverse design method for solid electrolyte materials, comprising:

[0006] Build a knowledge base in the field of inorganic solid electrolyte materials;

[0007] Based on the inorganic solid electrolyte material field knowledge base, a chemical formula generator is guided to generate a chemical formula that conforms to physical and chemical rationality;

[0008] Based on the generated chemical formula, a large language model is used to generate a candidate crystal structure information file corresponding to the target chemical formula;

[0009] Based on the generated candidate crystal structure information file, the convex hull energy and activation energy of the candidate crystal structure are accurately predicted through a multi-objective performance prediction model based on graph neural network;

[0010] The predicted convex hull energy and activation energy are fed back to the Monte Carlo tree search as rewards, and the large model is used in collaboration with the Monte Carlo tree search to optimize the generation quality of candidate crystal structures and screen out potential new inorganic solid electrolyte materials that meet the performance requirements.

[0011] Optionally, the inorganic solid electrolyte material field knowledge base includes: a chemical element valence information table, an inorganic solid electrolyte chemical formula table, and chemical formula generation screening rules, wherein the chemical formula generation screening rules include screening constraints based on the number of atoms and screening constraints based on charge conservation.

[0012] Optionally, the screening constraints based on the number of atoms include:

[0013] Assume that the chemical formula contains n elements, and the symbol of each element is E. 1 , E 2 , ..., E n ; The number of atoms corresponding to each element is but:

[0014]

[0015] in, The number of atoms representing each element must be a non-negative integer;

[0016] Assume that the chemical formula contains n elements, and the symbol of each element is E. 1 , E 2 , ..., E n ; The number of atoms corresponding to each element is but:

[0017]

[0018] Where d = gcd(x 1 , x2 , .., x n ) is the greatest common divisor of all the numbers of atoms.

[0019] Optionally, the screening constraint based on charge conservation includes:

[0020] Assume that the chemical formula contains n elements, and the symbol of each element is E. 1 , E 2 , ..., E n ; The number of atoms corresponding to each element is The valence number corresponding to each element is but:

[0021]

[0022] Among them, Q total is the total charge of the chemical formula, when Q total = 0, indicating that the compound is neutral; when Q total When ≠0, it means the compound is an ion.

[0023] Optionally, generate chemical formulas that are physicochemically plausible including:

[0024] By combining the inorganic solid electrolyte material domain knowledge base, after inputting the chemical formula structure type into the chemical formula generator, a candidate chemical formula of the inorganic solid electrolyte compound that conforms to the domain knowledge is generated.

[0025] Optionally, the final crystal structure information file corresponding to the target chemical formula is generated including:

[0026] Use a large language model to generate candidate crystal structure information files corresponding to the target chemical formula;

[0027] Construct a multi-objective performance prediction model based on graph neural network, embed it into the crystal structure generator, and perform multi-objective performance prediction and evaluation on candidate crystal structures;

[0028] In the Monte Carlo tree search optimization process, combined with the prediction results of the multi-objective performance prediction model, the candidate crystal structures are optimized and screened through a multi-objective-oriented reward mechanism that integrates domain knowledge to generate the final crystal structure information file corresponding to the target chemical formula.

[0029] Optionally, the multi-objective-oriented reward mechanism integrating domain knowledge includes:

[0030]

[0031] Among them, w 1 and w 2 are the weight coefficients of activation energy and convex hull energy respectively; σ(·) is the Sigmoid function; and are the activation energy prediction value and convex hull energy prediction value output by the multi-objective performance prediction model based on graph neural network; δ 1 and δ 2 They are adjustment parameters respectively.

[0032] Optionally, optimization and screening of the crystal structure includes iterative optimization and integration of multiple constraints, including:

[0033] The screening process is based on iterative optimization, each round of iteration corresponds to a candidate crystal structure i, and the number of iterations is limited to N iter ≤100;

[0034] In each iteration, the crystal structure is required to satisfy both thermodynamic stability and ion transport performance constraints:

[0035]

[0036] in, is the convex hull energy of candidate crystal structure i, The maximum threshold of the convex hull energy is set; is the activation energy of candidate crystal structure i, The maximum activation energy threshold is set.

[0037] Optionally, optimizing and screening the crystal structure also includes screening the target with the best performance among all eligible candidate crystal structures, specifically including:

[0038] Among all eligible candidate crystal structures, select the one with the convex hull energy and activation energy The ultimate goal is to have a structure with the smallest sum:

[0039]

[0040] The constraints are:

[0041]

[0042] in, is the convex hull energy of candidate crystal structure j, is the maximum threshold of the convex hull energy; is the activation energy of candidate crystal structure j, is the maximum threshold of activation energy set; N iter is the number of iterations.

[0043] Technical effect of the invention: The invention discloses a reverse design method applied to solid electrolyte materials. Under the guidance of domain knowledge, the large language model and Monte Carlo tree search are combined to realize crystal structure generation, and activation energy and convex hull energy are used as indicators to predict multi-objective performance and optimize screening, and finally obtain potential inorganic solid electrolytes that meet multi-objective performance requirements. This method provides a feasible intelligent reverse design process for materials science research, which not only significantly improves the development efficiency of inorganic solid electrolyte materials, but also lays a theoretical foundation and provides technical support for accelerating the practical application of high-performance electrolyte materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0045] Figure 1 A schematic diagram of a reverse design method for solid electrolyte materials according to an embodiment of the present invention;

[0046] Figure 2 This is a flow chart of reverse design of solid electrolyte materials according to an embodiment of the present invention;

[0047] Figure 3 This is a flow chart of a chemical formula generator according to an embodiment of the present invention;

[0048] Figure 4 For the embodiment of the present invention LiTa 2 PO 8 Schematic diagram of the comparison of generation results, where (a) is the ILMCG generation result, (b) is the FTCP-based generation result, and (c) is the actual result. DETAILED DESCRIPTION

[0049] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0050] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0051] All-solid-state batteries are crucial in new battery technologies due to their high specific energy density and excellent safety performance, and are widely considered to be the core development direction of electrochemical energy storage technology. However, the development of all-solid-state batteries is currently limited by the low ion transport performance of solid electrolytes, and this bottleneck significantly hinders their widespread promotion in practical applications. Therefore, exploring solid electrolytes with high ionic conductivity has become a key scientific issue in promoting the development of solid-state battery technology.

[0052] Traditional solid electrolyte material design usually adopts a forward design approach, starting from the composition and structure of the material, and exploring the material properties through a large number of trial-and-error experiments and computational simulation methods. Limited by the high dimensionality of the chemical component space and the complexity of the physicochemical mechanism, this method is not only time-consuming and laborious, but also has high R&D costs and is difficult to break through the limitations of the existing material system. With the development of the materials genome project and the rise of data-driven methods, the new reverse design paradigm provides a possibility to break through the limitations of traditional methods. Starting from the desired performance indicators such as high ionic conductivity, the composition and structure of the material are reversely deduced, opening up a new path for the research and development of solid electrolyte materials. The purpose of reverse design is to directly generate material structures that meet the constraints of the target properties, for example, through evolutionary algorithms, reinforcement learning, etc. For example, Allahyari et al. (2020) developed a method for predicting materials with one or more optimal target properties-Mendelevian search (MendS), which is based on appropriately defined chemical space, powerful co-evolutionary algorithms and multi-objective Pareto optimization techniques, and is applied to search for low-energy hard materials and superhard materials. Because evolutionary and coevolutionary algorithms have the property of reinforcing sampling of the most promising regions in the search space, each MendS search can discover a large number of materials with excellent properties at a low computational cost. The method successfully discovered almost all known hard systems in a single run and produced a comprehensive chemical map of hard materials. Due to the power and efficiency of the method, it can be used to search for the best materials with any combination of properties under arbitrary conditions. Law et al. (2022) proposed another method for finding new stable structures in a large space of decorative structures with diverse compositions and structures. This method predicted stable structures of 2003 components out of 14.3 million decorative structures. These candidate structures were verified by DFT, confirming that more than 99% of the structures were thermodynamically stable, that is, the decomposition energy was negative. In addition, expanding the search range to more components and prototypes will exponentially increase the search space. For this purpose, a reinforcement learning (RL) enhanced search strategy is demonstrated, which can find stable structures using a proxy stability function at a fraction of the original computational cost.

[0053] With the rise of generative artificial intelligence technology, a new path has been opened up for the reverse design of solid electrolyte materials, which can quickly generate, screen, predict and optimize material structures and properties, greatly accelerating the research and development process of materials. Khajeh et al. (2025) introduced a framework that can effectively design polymers with customized properties. The framework includes three core components: a conditional generation model, a computational evaluation module, and a feedback mechanism, all of which are integrated into an iterative framework for material innovation. The framework successfully designed a polymer solid electrolyte with ionic conductivity superior to polyethylene oxide. Yang et al. (2024) used generative artificial intelligence to design polymer solid electrolytes from scratch, effectively generating a large number of novel, diverse and effective polymers with high synthesizable possibilities. Of the 46 candidate materials tested, 17 showed excellent ionic conductivity, exceeding existing polymers in the database, and some even doubled the conductivity value. However, the generation results of the generative model are very random and the generation process is uncontrollable. Specifically, the generative model may generate parameters or structures that are inconsistent with the actual material science data, resulting in research results that are inconsistent with experimental or theoretical calculations, affecting the scientific nature and reliability of the research; due to the "black box" characteristics of the generative model, its internal mechanism is difficult for scientists and engineers to intuitively understand, resulting in the uncontrollable and lack of explainability of the generation process, which in turn affects the credibility of the research. Therefore, if material domain knowledge can be embedded in the entire process of material reverse design, starting from actual research problems and specific application needs, and comprehensively considering factors such as the material's microstructure, physical and chemical properties, and performance indicators, it will help the generative model effectively capture the complex structure of solid electrolytes and their structure-activity relationship with ion transport performance.

[0054] like Figure 1-Figure 2 As shown, this embodiment provides a reverse design method applied to solid electrolyte materials, including:

[0055] S1. Construction of a knowledge base in the field of materials. In order to ensure the rationality of the results generated by the chemical formula generator at the physical and chemical levels, a knowledge base in the field of inorganic solid electrolyte materials was constructed under the guidance of knowledge in the field of materials science.

[0056] S2. Generation of chemical formulas: Based on the domain knowledge base constructed above, after inputting the chemical formula structure type into the chemical formula generator (C-GEN), a candidate chemical formula of the inorganic solid electrolyte compound that conforms to the domain knowledge is generated.

[0057] S3, generation of crystal structure, inputting the candidate chemical formula generated in S2 into the large language model, can generate a candidate crystal structure information file (CIF) corresponding to the target chemical formula.

[0058] S4, multi-objective performance evaluation of crystal structure, input the candidate crystal structure CIF file generated in S3 into the multi-objective performance prediction model based on graph neural network to accurately predict the convex hull energy and activation energy.

[0059] S5, optimization screening of crystal structures, feeds the convex hull energy and activation energy predicted in S4 as rewards to the Monte Carlo Tree Search (MCTS), optimizes the generation quality of candidate crystal structures by collaborating with the large model and the Monte Carlo Tree Search, and formulates a set of systematic screening rules to ensure that the screened crystal structures are optimized in both the convex hull energy and activation energy, two key performance dimensions.

[0060] Specifically, the definition of the material domain knowledge base is:

[0061] In order to ensure the rationality of the generated results of the chemical formula generator at the physical and chemical levels, a knowledge base in the field of inorganic solid electrolyte materials was constructed under the guidance of knowledge in the field of materials science. This knowledge base includes: chemical element valence information table, inorganic solid electrolyte chemical formula table, and generated chemical formula screening rules, providing solid knowledge support for the discovery of new candidate chemical formulas.

[0062] ① Chemical element valence information table. Table 1 is only a partial table of chemical element valence information. In the process of material performance optimization and design, selecting element valences suitable for target performance is the key to improving the rationality of chemical formula generation. Relying on the knowledge in the field of materials science, a comprehensive chemical element valence information table is constructed, covering the element valence data involved in all compounds in the existing data set. Compared with the traditional random chemical formula generation model, the chemical element valence information table provides a clear element valence range for the chemical formula generator to avoid generating chemical formulas that contradict the knowledge in the material field.

[0063] Table 1

[0064]

[0065]

[0066] ②Inorganic solid electrolyte chemical formula table. Table 2 is only a partial inorganic solid electrolyte chemical formula table. An inorganic solid electrolyte chemical formula table is constructed to fundamentally avoid the common problems of valence mismatch, atom number abnormality, and inconsistency with domain knowledge in chemical formula generation. The table includes a variety of inorganic solid electrolyte structure systems with special properties, such as LISICON, NASICON, Garmet, Argyrodite, and Perovskite. Combined with the chemical element valence information table, it can efficiently provide chemical formulas that conform to domain knowledge for large-scale chemical formula generation.

[0067] Table 2

[0068]

[0069]

[0070] ③Generate chemical formula screening rules:

[0071] Although there are clear restrictions on the valence states and types of elements, problems such as abnormal number of atoms, redundant multiple chemical formulas, or charge non-conservation may still occur during the chemical formula generation process. To this end, combined with relevant knowledge in the field of inorganic solid electrolytes, chemical formula screening rules are proposed. These rules, as key constraints in the chemical formula generation process, not only ensure the rationality of the generated results at the physical and chemical levels, but also improve their practical application value in material design, further enhancing the scientificity and effectiveness of the generated results.

[0072] Screening rule 1: Screening constraint based on the number of atoms

[0073] Upper limit constraint on the total number of atoms: Assume that the chemical formula contains n elements, and the symbol of each element is E 1 , E 2 , ..., E n ; The number of atoms corresponding to each element is but:

[0074]

[0075] in, The number of atoms representing each element must be a non-negative integer. If the total number of atoms exceeds this upper limit, the chemical formula will be eliminated.

[0076] Chemical formula normalization: Assume that the chemical formula contains n elements, and the symbol of each element is E 1 , E 2 , ..., E n ; The number of atoms corresponding to each element is but:

[0077]

[0078] Where: d = gcd(x 1 , x 2 , ..., x n ) is the greatest common divisor of all atomic numbers; if The index is omitted ( Indicated as E i ). Reduce the generated multiple chemical formulas to the simplest chemical formula to avoid redundant generation results and ensure the uniqueness of the generated chemical formula.

[0079] Screening Rule 2: Screening Constraints Based on Charge Conservation

[0080] To ensure the stability of the generated compounds, the electrical neutrality condition must be strictly met, which is the basic condition for the stable existence of chemical substances. In view of this situation, combined with the knowledge in the field of materials, the following screening rules are introduced:

[0081] Assume that the chemical formula contains n elements, and the symbol of each element is E. 1 , E 2 , ..., E n ; The number of atoms corresponding to each element is The valence number corresponding to each element is but:

[0082]

[0083] Among them, Q total is the total charge of the chemical formula. total = 0, indicating that the compound is neutral; when Q total When ≠0, it means that the compound is an ion (charged), that is, it does not meet the constraints and should be eliminated.

[0084] The chemical formula is generated as Figure 3 As shown, in the process of material structure generation, the crystal structure generation based on the large language model depends on the input of a specific chemical formula. However, how to quickly generate a large number of chemical formulas that conform to the knowledge of the material field is a key problem that needs to be solved urgently. In response to this challenge, the present invention deeply explores the composition rules of material chemical formulas, and proposes and designs a chemical formula generator C-GEN guided by the knowledge of the material science field. The chemical formula generator (C-GEN) combines the inorganic solid electrolyte chemical formula table and the chemical element valence information table, and after inputting the chemical formula structure type, generates candidate chemical formulas of inorganic solid electrolyte compounds that conform to the field knowledge. The above-mentioned screening rules based on the number of atoms and charge conservation are further applied to eliminate the generation results that do not conform to physical and chemical rationality. Finally, the screened chemical formula is not only consistent with the knowledge base of the inorganic solid electrolyte material field, but also meets the requirements of element valence and charge conservation, ensuring the scientificity and practicality of the generation results, and providing solid data and knowledge support for subsequent material crystal structure generation and performance optimization.

[0085] Generation of crystal structure:

[0086] In order to generate the target chemical formula text that meets the material domain knowledge into an inorganic solid electrolyte crystal structure with thermodynamic stability and excellent ion transport performance, under the guidance of domain knowledge, the present invention develops a multi-performance crystal structure generator (ILMCG) based on large model collaborative Monte Carlo tree search. The generator generates the candidate crystal structure information file (CIF) corresponding to the target chemical formula through a large language model, and combines the multi-objective performance prediction model based on graph neural network to perform performance evaluation and optimization screening, and quickly identify inorganic solid electrolyte materials that meet multiple performance requirements.

[0087] Generation of candidate crystal structures:

[0088] The large language model can generate a target crystal structure that conforms to the laws of physical chemistry according to the input chemical formula by performing autoregressive training on the CIF file format. During the generation process, the candidate chemical formula generated by the chemical formula generator that conforms to the domain knowledge is input into the large language model. The large language model can gradually generate a complete CIF file including space group information and atomic coordinates from the chemical formula start mark based on the learned crystallographic rules, chemical bond characteristics and space group information and other complex crystallographic information. The generated structure conforms to the crystallographic rules and has physical feasibility. Compared with the traditional generation method, the large language model significantly improves the efficiency and rationality of crystal structure prediction, has stronger flexibility and adaptability, and can handle more diverse material systems. Combined with other prediction tools (such as the formation energy prediction model based on the graph neural network), the large language model can further screen out structures with higher stability, laying a solid foundation for the automation and efficiency of material design. In the present invention, the CrystaLLM large language model is used as a case model for candidate crystal structure generation in the full process of reverse design, and other similar large language models are also applicable to the process.

[0089] In order to accurately evaluate the performance of the generated candidate crystal structures, a multi-objective performance prediction model based on graph neural network was constructed and embedded in the crystal structure generator ILMCG. Based on the previously proposed "divide and conquer" graph neural network prediction model guided by material domain knowledge, SDCGNN, the convex hull energy data corresponding to 17,730 CIFs were used to train the SDCGNN model, and a multi-objective performance prediction model based on graph neural network was obtained. This model synergistically integrates the two complementary perspectives of crystal structure and gap network in SDCGNN, and comprehensively characterizes the crystal structure from three levels: geometry, chemistry, and topology. The domain knowledge is symbolized and embedded in the convolution operation of the graph neural network. By designing an adaptive weight fusion mechanism, the embedding vectors of the crystal structure graph and the gap network graph are unified into a shared representation space. Through this in-depth and multifaceted representation, the model can accurately predict the activation energy (E a, unit: eV) and convex hull energy (E hull , unit: meV / atom).

[0090] Crystal structure optimization and screening:

[0091] In the Monte Carlo tree search optimization process, combined with the prediction results of the multi-objective performance prediction model based on graph neural network, the optimization of comprehensive performance relies on a multi-objective oriented reward mechanism that integrates domain knowledge. Its reward is defined as:

[0092]

[0093] Among them, w 1 and w 2 are the weight coefficients of activation energy and convex hull energy, respectively, reflecting the importance of the two performance objectives; σ(·) is the Sigmoid function, which is used to normalize the performance value to the [0,1] interval to improve the sensitivity of the algorithm to different target scales; and They represent the activation energy prediction value and convex hull energy prediction value output by the multi-objective performance prediction model based on graph neural network, δ 1 and δ 2 To adjust the parameters.

[0094] This reward mechanism guides the path selection of the generated structure during the Monte Carlo Tree Search (MCTS) optimization process. By introducing knowledge in the field of materials science (such as ion transport performance and thermodynamic stability), the optimization process is not only data-driven, but also incorporates domain-specific knowledge, thereby promoting goal-oriented precision optimization. This mechanism dynamically adjusts the reward value to gradually approach the target performance requirements, achieves a performance balance between activation energy and convex hull energy, and ultimately ensures that the generated crystal structure meets multiple performance goals at the same time. It significantly improves the efficiency of the model in exploring the material space, avoids the inefficiency and redundancy of traditional random search, and provides an efficient multi-objective material design method.

[0095] On this basis, combined with the knowledge in the field of inorganic solid electrolyte materials, a set of systematic screening rules is proposed to efficiently screen material structures through iterative optimization and integration of multiple constraints. This rule system focuses on thermodynamic stability and ion transport performance to ensure that the screened crystal structure is optimal in both of these two key performance dimensions. Specifically, it includes the following constraints:

[0096] Screening rules: iterative optimization and integration of multiple constraints

[0097] The screening process is based on iterative optimization, each round of iteration corresponds to a candidate crystal structure i, and the number of iterations is limited to N iter≤100. In each round of iteration, the crystal structure is required to satisfy both thermodynamic stability and ion transport performance constraints:

[0098]

[0099] in:

[0100]

[0101] in, is the convex hull energy of candidate crystal structure i, is the maximum threshold of the convex hull energy; is the activation energy of candidate crystal structure i, The maximum activation energy threshold is set.

[0102] If there is a solution that meets the screening rules, all candidate crystal structures that meet the requirements are retained; if there is no solution, the current generation is terminated.

[0103] Optimal performance goals:

[0104] Among all eligible candidate crystal structures, select the one with the convex hull energy and activation energy The ultimate goal is to have a structure with the smallest sum:

[0105]

[0106] The constraints are:

[0107]

[0108] in, is the convex hull energy of candidate crystal structure j, is the maximum threshold of the convex hull energy; is the activation energy of candidate crystal structure j, is the maximum threshold of activation energy set; N iter The maximum number of iterations is 100.

[0109] This method provides an efficient screening path and can be widely used in the accelerated discovery and development of materials, ensuring efficiency and rationality in the materials design process.

[0110] In view of the shortcomings of the prior art, a reverse design method for solid electrolytes is proposed. This method aims to perform intelligent reverse design of solid electrolyte materials under the guidance of domain knowledge to accelerate the discovery of new candidate high-performance solid electrolyte materials. In order to solve the problems of large randomness of the generation results, uncontrollable generation process, and easy occurrence of results that contradict domain knowledge in the current reverse design method using generative models, this method embeds the domain knowledge of materials science into the whole process of material reverse design, avoids the results that contradict domain knowledge from the source, and improves the controllability of the generation process. In order to improve the efficiency of the material screening and optimization process, the reverse design process is automated, and the multi-objective performance evaluation prediction and screening of the crystal structure are embedded in the optimization process of the crystal structure generation quality, so as to realize the systematic closed-loop operation of "iterative optimization-performance evaluation-condition screening". In addition, the existing methods mostly focus on single-objective performance. Although this method can achieve initial success in some specific fields, it often ignores the complex relationship between multiple performances. In response to this problem, this application proposes a multi-objective performance optimization reward mechanism of convex hull energy and activation energy, and develops a multi-objective optimization method that can balance multiple performance requirements.

[0111] The generation results of other generation models and the crystal structure generator (ILMCG) proposed in this application are compared, such as Figure 4As shown, (a) is the ILMCG generation result, (b) is the FTCP generation result, and (c) is the real result. In order to fully capture the diversity of material crystals and meet the reversibility of sample representation, Zekun et al. proposed a crystallographic representation method (Fourier transformed crystal properties / FTCP framework) that combines the CIF features of real space and the Fourier transform features of reciprocal space. The variational autoencoder limits the type and number of elements to achieve mass generation of material crystal structures. Through experiments and result comparisons, it is found that the CIF representation of the FTCP framework is poor and the reversibility is low when generating ternary or higher compounds. Existing work has demonstrated this. However, ILMCG guided by domain knowledge still maintains a good generation effect, and the atomic distribution is closer to the real structure than the FTCP framework. In addition, FTCP cannot effectively grasp the element valence state of the generated results, and often the results calculated based on the normal valence state cannot meet the charge conservation phenomenon. In addition, ILMCG's training data is more complex and contains information such as atomic site occupancy and space groups. These structural information not only provides constraints for material structure generation, but also makes the model show stronger generation potential. In terms of generation speed, ILMCG is comparable to other generation models. At the same time, compared with generation models such as variational autoencoders, ILMCG can generate material structures based on research objectives, reducing the randomness of the generation results. In addition, with the introduction of the gap network, the ion channel information of the material is supplemented, and the structural information of the material crystal space is improved from two structural perspectives, which improves the reliability of the generated material structure information.

[0112] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A reverse design method applied to solid electrolyte materials, characterized in that: include: Build a knowledge base in the field of inorganic solid electrolyte materials; Based on the inorganic solid electrolyte material field knowledge base, a chemical formula generator is guided to generate a chemical formula that conforms to physical and chemical rationality; Based on the generated chemical formula, a large language model is used to generate a candidate crystal structure information file corresponding to the target chemical formula; Based on the generated candidate crystal structure information file, the convex hull energy and activation energy of the candidate crystal structure are accurately predicted through a multi-objective performance prediction model based on graph neural network; The predicted convex hull energy and activation energy are fed back to the Monte Carlo tree search as rewards, and the large model is used in collaboration with the Monte Carlo tree search to optimize the generation quality of candidate crystal structures and screen out potential new inorganic solid electrolyte materials that meet the performance requirements.

2. The reverse design method for solid electrolyte materials according to claim 1, characterized in that: The inorganic solid electrolyte material field knowledge base includes: a chemical element valence information table, an inorganic solid electrolyte chemical formula table, and a generated chemical formula screening rule, wherein the generated chemical formula screening rule includes a screening constraint based on the number of atoms and a screening constraint based on charge conservation.

3. The reverse design method for solid electrolyte materials according to claim 2, characterized in that: The screening constraints based on the number of atoms include: Assume that the chemical formula contains n elements, and the symbol of each element is E1, E2, ..., E n ; The number of atoms corresponding to each element is but: in, The number of atoms representing each element must be a non-negative integer; Assume that the chemical formula contains n elements, and the symbol of each element is E1, E2, ..., E n ; The number of atoms corresponding to each element is but: Where d = gcd(x1, x2, ..., x n ) is the greatest common divisor of all the numbers of atoms.

4. The reverse design method for solid electrolyte materials according to claim 2, characterized in that: The screening constraints based on charge conservation include: Assume that the chemical formula contains n elements, and the symbol of each element is E1, E2, ..., E n ; The number of atoms corresponding to each element is The valence number corresponding to each element is but: Among them, Q total is the total charge of the chemical formula, when Q total = 0, indicating that the compound is neutral; when Q total When ≠0, it means the compound is an ion.

5. The reverse design method for solid electrolyte materials according to claim 1, characterized in that: Generating chemical formulas that are consistent with physicochemical plausibility includes: By combining the domain knowledge base of inorganic solid electrolyte materials, after inputting the chemical formula structure type into the chemical formula generator, candidate chemical formulas of inorganic solid electrolyte compounds that conform to the domain knowledge are generated.

6. The reverse design method for solid electrolyte materials according to claim 1, characterized in that: The final crystal structure information file corresponding to the target chemical formula includes: Use a large language model to generate candidate crystal structure information files corresponding to the target chemical formula; Construct a multi-objective performance prediction model based on graph neural network, embed it into the crystal structure generator, and perform multi-objective performance prediction and evaluation on candidate crystal structures; In the Monte Carlo tree search optimization process, combined with the prediction results of the multi-objective performance prediction model, the candidate crystal structures are optimized and screened through a multi-objective-oriented reward mechanism that integrates domain knowledge to generate the final crystal structure information file corresponding to the target chemical formula.

7. The reverse design method for solid electrolyte materials according to claim 6, characterized in that: The multi-objective-oriented reward mechanism integrating domain knowledge includes: Where w1 and w2 are the weight coefficients of activation energy and convex hull energy respectively; σ(·) is the Sigmoid function; and are the activation energy prediction value and convex hull energy prediction value output by the multi-objective performance prediction model based on graph neural network, respectively; δ1 and δ2 are the adjustment parameters, respectively.

8. The reverse design method for solid electrolyte materials according to claim 6, characterized in that: Optimization and screening of crystal structures include iterative optimization and integration of multiple constraints, including: The screening process is based on iterative optimization, each round of iteration corresponds to a candidate crystal structure i, and the number of iterations is limited to N iter ≤100; In each iteration, the crystal structure is required to satisfy both thermodynamic stability and ion transport performance constraints: in, is the convex hull energy of candidate crystal structure i, is the maximum threshold of the convex hull energy; is the activation energy of candidate crystal structure i, The maximum activation energy threshold is set.

9. The reverse design method for solid electrolyte materials according to claim 6, characterized in that: Optimizing and screening the crystal structure also includes screening the target with the best performance among all eligible candidate crystal structures, including: Among all eligible candidate crystal structures, select the one with the convex hull energy and activation energy The ultimate goal is to have a structure with the smallest sum: The constraints are: in, is the convex hull energy of candidate crystal structure j, The maximum threshold of the convex hull energy is set; is the activation energy of candidate crystal structure j, is the maximum threshold of activation energy set; N iter is the number of iterations.

Images