Intelligent research and development system for new material research and development and material research and development method

Abstract

The present disclosure provides an intelligent research and development system and a material research and development method for new material research and development, and belongs to the field of material research and development. The intelligent research and development system comprises: a material database for storing and structurally representing material science data; a multi-agent collaborative module comprising a material design agent, an experimental optimization agent and a knowledge updating agent, for generating a candidate material scheme, an experimental scheme and updating the material database according to experimental feedback; an experimental platform for automatically completing material sample preparation and testing according to the experimental scheme; and an analysis module for collecting and analyzing experimental data and feeding back the results to the material database and the multi-agent collaborative module. The intelligent research and development system and the material research and development method provided by the present disclosure improve the efficiency and success rate of new material screening and process optimization, reduce trial and error costs, and are conducive to forming a sustainable evolving material knowledge system.

Description

Technical Field

[0001] This invention belongs to the field of materials research and development technology, specifically relating to an intelligent research and development system and a materials research and development method for new materials. Background Technology

[0002] New materials research and development is fundamental to advancing technology in fields such as energy, information, and aerospace. Functional materials, represented by solid-state electrolytes and their supporting electrode materials, are crucial for the performance and safety of next-generation energy devices such as all-solid-state batteries. Traditional materials research and development processes typically rely on researchers proposing material formulations and process conditions based on literature and experience. Sample preparation and testing are then manually scheduled on experimental platforms, followed by the researchers compiling and analyzing the experimental results. This linear process, primarily based on manual design and analysis, suffers from long development cycles, high trial-and-error costs, and strong reliance on individual experience, making it difficult to adapt to the ever-increasing complexity of current material systems and the increasingly stringent requirements for target performance indicators.

[0003] To improve R&D efficiency, existing technologies have begun to explore the construction of materials databases and the introduction of machine learning algorithms. Some materials informatics platforms can aggregate literature data, experimental data, and computational data, and provide basic retrieval and statistical functions. Some work also explores the use of machine learning models to predict material properties to assist in screening candidate materials. However, such databases are mostly stored in single data tables or weakly structured formats, lacking unified feature descriptions and vectorized expressions for multi-dimensional information such as composition, structure, properties, and processing conditions. This makes it difficult to support fine-grained similarity retrieval based on material characteristics and systematic variable correlation analysis. Summary of the Invention

[0004] The purpose of this disclosure is to provide an intelligent R&D system and material R&D method for new material development, which can shorten the material R&D cycle, improve R&D efficiency, and reduce labor costs.

[0005] To achieve the above objectives, the technical solution provided in this disclosure is as follows:

[0006] In a first aspect, this disclosure provides an intelligent research and development system for new material research and development, comprising:

[0007] Materials databases are used to store and structure materials science data.

[0008] A multi-agent collaborative module is communicatively connected to the material database. The multi-agent collaborative module includes at least a material design agent, an experimental optimization agent, and a knowledge update agent. The material design agent is used to generate candidate material schemes, the experimental optimization agent is used to generate experimental schemes for candidate materials, and the knowledge update agent is used to update the data and knowledge in the material database based on experimental feedback.

[0009] The experimental platform is communicatively connected to the multi-agent collaborative module and is used to prepare and test material samples according to the experimental scheme generated by the experimental optimization agent.

[0010] The analysis module is communicatively connected to the experimental platform and the material database. It is used to collect and analyze the experimental data output by the experimental platform and feed the analysis results back to the material database and the multi-agent collaborative module.

[0011] In one or more embodiments, the material database stores material science data including at least one or more of the following: material composition information, structural information, performance indicators, and synthesis and characterization conditions information; when the material database performs structured representation of material samples, it performs vectorized encoding of the material samples based on preset material feature descriptors and is configured to support material similarity retrieval based on the material feature descriptors.

[0012] In one or more embodiments, the material database internally constructs a causal relationship model between composition variables, structural variables, process variables and performance indicators. The causal relationship model is at least one of a causal graph model or a Bayesian network model. The multi-agent collaborative module is configured to invoke the causal relationship model to evaluate the degree of influence of different design variables on the target performance, so as to limit the search space or priority of the material design agent and the experimental optimization agent.

[0013] In one or more embodiments, the material design agent includes a generation module running on a computing node. The generation module is used to generate candidate material schemes and corresponding synthesis paths within a joint design space consisting of composition, structure, and performance. The material design agent further invokes a material physics model and a machine learning model to predict the target performance of the candidate materials, and screens the candidate material schemes based on the prediction results.

[0014] In one or more embodiments, the generation module includes at least one of a graph neural network generation model, a variational autoencoder model, or a sequence generation model based on a large language model; the materials design agent further includes a literature semantic parsing unit, which is used to parse the synthesis experience and mechanism descriptions in the materials-related literature, convert the parsing results into rules or feature vectors, and fuse and score them with the candidate material schemes output by the generation module to obtain a list of target candidate materials for issuing experiments.

[0015] In one or more embodiments, the experimental optimization agent sequentially optimizes the experimental parameter space using at least one of Bayesian optimization, reinforcement learning, or evolutionary algorithms. The experimental parameter space includes at least one or more of heat treatment temperature, heat treatment time, atmosphere type and pressure, raw material ratio, grinding time, and forming pressure. The experimental optimization agent updates the agent model or policy function on the parameter space based on the experimental feedback data of the completed experiments and generates the next batch of experimental schemes using the acquisition function or policy network.

[0016] In one or more embodiments, the experimental platform includes a sample preparation sub-platform and a performance testing sub-platform. The sample preparation sub-platform includes at least a raw material storage unit, a batching and mixing unit, a forming unit, and a sintering unit. The performance testing sub-platform includes at least a structural characterization unit and an electrochemical testing unit. The experimental platform uses a preset process control program to control each unit in a coordinated manner, so as to complete the preparation and testing of multiple material samples in parallel or sequentially.

[0017] In one or more embodiments, the analysis module includes a structural characterization analysis unit and a performance analysis unit; the structural characterization analysis unit is used to perform peak position detection, peak shape fitting, and phase identification on the raw spectra from the structural characterization equipment to obtain the phase composition and structural parameters of the material sample; the performance analysis unit is used to perform feature extraction or equivalent circuit fitting on the impedance spectrum or charge-discharge curve from the electrochemical testing equipment to obtain ionic conductivity, charge transfer impedance, specific capacity, and cycle stability performance indicators; the analysis module uses the analysis results as part of the experimental feedback data, correlates them with the composition and process conditions of the corresponding sample, and writes them into the material database.

[0018] Secondly, this disclosure provides a materials research and development method using the aforementioned intelligent research and development system, comprising:

[0019] S1, through a materials design intelligent agent, generates the composition, target structure and synthesis path of candidate materials based on the materials science data stored in the materials database, and uses materials physics model and machine learning model to predict the target performance of each candidate material to obtain a set of candidate materials;

[0020] S2 optimizes the intelligent agent through experiments, models the experimental parameter space, generates an experimental plan containing candidate materials based on existing experimental feedback data, and sends the experimental plan to the experimental platform.

[0021] S3, the experimental platform performs the preparation and performance testing of material samples according to the experimental plan, and obtains the corresponding raw experimental data;

[0022] S4, the analysis module analyzes and extracts features from the raw experimental data to obtain experimental feedback data, and writes the experimental feedback data into the material database;

[0023] S5, the knowledge updating agent updates the material physics model and machine learning model based on the updated material database, and drives the material design agent and the experimental optimization agent to generate the next round of candidate materials and experimental schemes in the new database state;

[0024] S6. Repeat steps S2 to S5 until the preset termination condition is met.

[0025] In one or more embodiments, in step S2, the experimental optimization agent constructs the experimental feedback data of the current round and the previous rounds into a state, uses a reinforcement learning algorithm to construct the adjustment of experimental parameters into actions, constructs the improvement of target performance or whether the constraints are met into a reward signal, and generates subsequent experimental schemes through policy updates; or, the experimental optimization agent uses a Bayesian optimization algorithm to establish a surrogate model in the experimental parameter space, and selects experimental sample points on the surrogate model based on the acquisition function.

[0026] The intelligent R&D system and materials R&D method disclosed herein achieve the integration of materials design, experimental execution, data feedback and knowledge updating by constructing a closed-loop data flow between the materials database, multi-agent collaborative module, experimental platform and analysis module. Compared with the traditional decentralized, linear and manually driven R&D model, it significantly improves the efficiency and success rate of new material screening and process optimization, reduces trial and error costs, and is conducive to the formation of a sustainable and evolving materials knowledge system. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of an intelligent R&D system according to one embodiment of the present disclosure;

[0029] Figure 2 This is a flowchart of a material development method in one embodiment of the present disclosure. Detailed Implementation

[0030] To enable those skilled in the art to better understand the technical solutions in this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0031] In complex systems such as solid electrolytes and their related electrode materials, the traditional R&D process often involves researchers proposing several formulations and processes based on experience and literature, manually arranging experiments, and then interpreting the results and making decisions for the next round of testing. As the dimensions of material systems increase and performance indicators and engineering constraints become increasingly complex, this linear, manually driven process reveals significant bottlenecks in terms of time cost, resource utilization, and knowledge accumulation efficiency. Large amounts of data are scattered and piled up into data silos, algorithm models are disconnected from actual experiments, high-throughput equipment is merely an automated actuator, and the overall R&D process lacks a mechanism for continuous learning and self-optimization.

[0032] After repeatedly analyzing these issues, the inventors realized that existing work often only strengthens a single part, such as building a material database separately, training a performance prediction model separately, or setting up an automated experimental platform separately. However, these modules are only loosely connected, making it difficult for the system to achieve true autonomous evolution through multiple rounds of experiments. In other words, the current mainstream approach is more about upgrading tools than redesigning the R&D process itself at the system level.

[0033] Based on the foregoing analysis, the technical approach proposed in this disclosure involves reconstructing the entire new materials research and development process into a closed-loop decision-making system. This system is based on knowledge, uses intelligent agents as the brain, employs experimental platforms as executors, and utilizes analytical feedback as the perception channel. Through unified data and knowledge representation, historical information related to materials and new experimental results are aggregated into a material knowledge space that can be intelligently processed. On this basis, a group of intelligent agents with different responsibilities undertake the design, optimization, and summarization tasks that were originally performed separately by researchers. Then, through automated experiments and analytical feedback, the results of each experiment are automatically injected into the aforementioned knowledge space, allowing the system to continuously adjust its judgment and decision-making strategies in a continuous cycle.

[0034] Under this technological approach, the materials database is no longer merely a passive container for storing data, but rather a memory layer that carries and organizes materials knowledge, responsible for expressing the relationships between materials in a way suitable for intelligent processing. The multi-agent collaborative module is regarded as the system's decision-making layer, where agents with different roles work collaboratively based on shared knowledge, providing dynamic adjustment suggestions for two key questions: "What kind of materials should be designed?" and "Under what conditions should they be tested?" The experimental platform is the execution layer, responsible for repeatedly implementing these suggestions and producing experimental results under certain throughput conditions. The analysis module assumes the role of perception and induction, transforming the raw experimental output into structured information that can be written back to the knowledge space. Through repeated cycles between these four layers, the entire system can continuously revise its understanding of the materials system during operation.

[0035] Please refer to Figure 1 The diagram illustrates an intelligent R&D system for new materials development according to an embodiment of this disclosure. The system includes a materials database, a multi-agent collaborative module, an experimental platform, and an analysis module. The materials database stores and structures materials science data. The multi-agent collaborative module is communicatively connected to the materials database. This module includes at least a materials design agent, an experimental optimization agent, and a knowledge update agent. The materials design agent generates candidate material schemes, the experimental optimization agent generates experimental schemes for the candidate materials, and the knowledge update agent updates the data and knowledge in the materials database based on experimental feedback. The experimental platform is connected to the multi-agent collaborative module and performs the preparation and testing of materials samples according to the experimental schemes generated by the experimental optimization agent. The analysis module is connected to the experimental platform and the materials database and collects and analyzes the experimental data output by the experimental platform, feeding back the analysis results to the materials database and the multi-agent collaborative module.

[0036] Materials databases are used to centrally store and structurally represent materials science data. A materials database can contain multiple data tables and index structures to record the composition information of materials (e.g., element types, stoichiometry), structural information (e.g., crystal structure type, space group, lattice constant, microstructure parameters), performance indicators (e.g., ionic conductivity, electronic conductivity, electrochemical window, mechanical properties, etc.), as well as the corresponding synthesis conditions and characterization conditions (e.g., sintering temperature, sintering time, atmosphere type, pressing pressure, test temperature, frequency range, etc.).

[0037] This data is not stored as scattered text, but rather structured using fields, feature vectors, and metadata tags. This allows multi-agent collaborative modules to directly retrieve, filter, and access data from the materials database. For example, in the development of solid-state electrolyte materials, the database can store the proportions (e.g., the molar ratio of Li, P, S, and Cl) of several batches of different sulfide electrolyte samples, sintering curves, room-temperature ionic conductivity obtained through impedance spectroscopy fitting, and corresponding X-ray diffraction pattern analysis results. Through this structured representation, subsequent materials design agents can use existing data to find neighboring samples with performance close to the target and infer new candidate combinations accordingly.

[0038] The multi-agent collaboration module establishes a two-way communication connection with the materials database and is the core of the intelligent R&D system's decision-making. In a preferred implementation, the multi-agent collaboration module includes at least three types of functional entities: a materials design agent, an experimental optimization agent, and a knowledge update agent.

[0039] Materials design agents can construct new candidate material schemes in the composition space and structure space by reading historical sample data from materials databases and combining them with internal generative or inference models. For example, based on existing Li-PS system electrolyte data, they can deduce new formulation structures that introduce a small amount of halogen or adjust the cation ratio, and provide corresponding recommended synthesis routes.

[0040] Experiment optimization agents focus more on how to test these candidate solutions more efficiently. They can model the experimental parameter space as a multi-dimensional space to be optimized and automatically select the temperature, time, atmosphere or ratio combination worth trying in the next round of experiments based on historical experimental feedback recorded in the materials database, thereby avoiding simple grid search or repeated trial and error based on human experience.

[0041] The knowledge updating agent plays the role of summarizing and generalizing. When the analysis module provides new experimental feedback data, the knowledge updating agent can filter and standardize these results based on data quality and consistency, write valuable information back into the material database, and adjust the parameters or strategies of the internal model according to the new data distribution, so that the decision-making basis of the subsequent material design agent and experimental optimization agent always remains up-to-date.

[0042] The experimental platform communicates with a multi-agent collaborative module, responsible for translating the experimental schemes generated upstream into specific material preparation and testing actions. The experimental platform may include an automatic batching device, a mixing and grinding device, a forming device, a sintering furnace array, and interface modules for interfacing with electrochemical testing equipment and structural characterization equipment. For example, when the experimental optimization agent determines that the next round requires sintering samples with several different ratios at three temperature points: 550K, 575K, and 600K, the experimental platform can automatically complete the weighing, mixing, and tableting of raw materials according to the experimental scheme, and sequentially execute the corresponding sintering procedures in different furnace cavities. Subsequent electrochemical testing and structural characterization can also be performed automatically according to a preset process, minimizing errors and delays caused by manual operation.

[0043] The analysis module establishes communication connections with the experimental platform and materials database to receive the raw data output from the experimental platform and process it into a data format that can directly serve intelligent decision-making. In actual implementation, the analysis module may include a raw data receiving unit, a signal processing and feature extraction unit, a structural and performance index calculation unit, and a data labeling and storage unit.

[0044] Taking solid electrolytes as an example, after completing impedance spectroscopy and X-ray diffraction tests, the experimental platform outputs raw frequency-impedance data and raw diffraction intensity-angle data. The analysis module can automatically perform baseline correction, noise filtering, and equivalent circuit fitting, calculating key indicators such as room temperature ionic conductivity, interfacial resistance, and phase composition ratio. These calculation results are then correlated with the composition and processing conditions of the corresponding samples and stored in the materials database. Simultaneously, the analysis module pushes this structured experimental feedback to the multi-agent collaborative module, enabling the experimental optimization agent to update its parameter space model based on the latest results, and allowing the materials design agent to see which design approaches perform well and which perform poorly in real experiments.

[0045] The intelligent R&D system disclosed herein achieves its overall functionality through a closed-loop data flow between a materials database, a multi-agent collaborative module, an experimental platform, and an analysis module. In this closed loop, the materials database provides historical knowledge, the multi-agent collaborative module generates new candidate material schemes and experimental schemes based on this knowledge, the experimental platform puts these schemes into practice, and the analysis module extracts new knowledge from the practical results and feeds it back to the decision-making level through the materials database and the multi-agent collaborative module. Through multiple iterations, the intelligent R&D system gradually narrows the search space for materials and processes, concentrating more resources on high-potential areas. For example, in a solid-state electrolyte system, the initial stage may attempt a large number of combinations, resulting in a relatively dispersed distribution of ionic conductivity; after several rounds of closed-loop optimization, the system will automatically focus on a certain type of anionic framework and the vicinity of a specific doping strategy, further improving performance through more refined parameter adjustments.

[0046] Through structured management of the materials database and continuous maintenance by the knowledge-updating agent, experimental data can be accumulated and reused over a long period, avoiding the problem of a large number of results being scattered in personal notes and disparate files in the traditional R&D model. Through the collaborative work of the materials design agent and the experimental optimization agent, the intelligent R&D system can find better candidate combinations in a broader space of materials and processes with fewer experimental rounds, thereby reducing trial-and-error costs and shortening the R&D cycle.

[0047] In one exemplary embodiment, the materials database stores materials science data including at least one or more of the following: material composition information, structural information, performance indicators, and synthesis and characterization conditions. When representing material samples in a structured manner, the materials database vectorizes the material samples based on preset material feature descriptors and is configured to support material similarity retrieval based on these descriptors.

[0048] Compositional information can include element types and their stoichiometric ratios; for example, the composition information of a solid electrolyte sample is recorded as the molar ratios of elements such as Li, P, S, and Cl. Structural information can include crystal structure type, space group, lattice constant, the occupancy of atomic sites in the unit cell, or the phase composition ratio obtained from the analysis of powder diffraction patterns. Performance indicators can include room temperature ionic conductivity, electronic conductivity, electrochemical stability window, interfacial resistance, and cycle life. Synthesis and characterization conditions information can cover precursor purity, ball milling time, tableting pressure, sintering temperature and time, protective atmosphere type, and the temperature, frequency range, and electrode configuration used during testing. By uniformly storing the above information in a materials database, it can be ensured that any material sample can be retrieved and accessed from multiple dimensions during subsequent decision-making and analysis.

[0049] To facilitate processing of material samples by intelligent algorithms, the material database introduces an intermediate layer of material feature descriptors when representing material samples in a structured manner. Material feature descriptors can be understood as abstract descriptions of materials in terms of composition, structure, processing, and properties. For example, one set of values ​​describes the mole fraction of the elemental composition, another set describes the lattice parameters and symmetry characteristics, and yet another set describes the key nodes of the sintering curve and the testing environment conditions.

[0050] Material feature descriptors can be either explicitly designed features, such as anionic framework type, number of occupiesable ionic sites, bond length and bond angle statistics, or implicit features automatically extracted from raw structural and process data using graph neural networks or autoencoder models. When inputting material samples, the materials database combines these material feature descriptors into feature vectors according to preset dimensions, achieving vectorized encoding. The result of vectorized encoding is that each material sample corresponds to a point in a high-dimensional space. This allows the similarity between different material samples to be measured using distance or a similarity function.

[0051] Vectorized representations built upon material feature descriptors enable material databases to support material similarity retrieval. In one specific embodiment, researchers propose a target performance requirement through a materials design agent, such as seeking a sulfide solid electrolyte with room-temperature ionic conductivity greater than a certain threshold and stability within a specific voltage window. The materials design agent can use the material feature descriptor of a known high-performance reference material sample as a query vector, calculate the similarity between the feature vectors of other material samples and the query vector in the materials database, and sort them according to the similarity magnitude, thereby quickly filtering out a batch of candidate samples that are similar to the reference material in composition and structure. The system can also combine synthesis and characterization condition information to further restrict the process window based on similarity retrieval, for example, prioritizing the return of samples that can achieve high ionic conductivity at lower sintering temperatures, providing a more valuable candidate set for subsequent experimental optimization.

[0052] In this way, materials databases are no longer simply records, but possess the ability to map scattered and heterogeneous materials science data into a unified, computable feature space. The introduction of material feature descriptors and vectorized encoding allows similarity retrieval to automatically find similar materials globally, without relying on manual, line-by-line comparisons, using mathematical distance metrics. This mechanism improves the efficiency of material screening and comparison, reducing the workload of researchers repeatedly searching through massive amounts of literature and experimental records. Furthermore, it provides a high-quality, programmable prior knowledge base for multi-agent collaborative modules, enabling materials design agents and experimental optimization agents to combine and explore from a more reasonable search starting point.

[0053] In one exemplary embodiment, the material database internally constructs a causal relationship model between composition variables, structural variables, process variables, and performance indicators. This causal relationship model is at least one of a causal graphical model or a Bayesian network model. The multi-agent collaborative module is configured to invoke the causal relationship model to evaluate the impact of different design variables on the target performance, thereby limiting the search space or priority of the material design agent and the experimental optimization agent.

[0054] Compositional variables can be understood as adjustable parameters of a material in dimensions such as element types, stoichiometry, and doping concentration. Structural variables can reflect characteristics such as crystal structure type, space group, lattice parameters, site occupancy, and microporous structure. Process variables cover process conditions such as raw material purity, ball milling time, tableting pressure, sintering temperature and time, and atmosphere type and pressure. Performance indicators include characterization results such as room temperature ionic conductivity, electronic conductivity, electrochemical stability window, interfacial impedance, and cycle retention. Causal relationship models use these variables as nodes and the direct influence of one variable on another as directed edges, forming causal graph models or Bayesian network models. Through this formal representation, the experience in areas such as how composition affects structure, how process affects structural density, and how structure and process jointly determine performance is elevated to a calculable and inferable structure.

[0055] In one alternative implementation, the causal graph model can be initialized by materials experts, who can define it as follows: sintering temperature, a process variable, primarily affects ionic conductivity, a performance indicator, by influencing density and grain growth. This is then combined with a large number of historical samples accumulated in a materials database, and structure learning and parameter learning algorithms are used to refine and quantify the edge relationships and conditional probability distributions of the causal graph model. A Bayesian network model, on the other hand, can establish a conditional probability table or distribution for each node, characterizing the probability distribution of performance indicator values ​​given the values ​​of composition, structure, and process variables.

[0056] For example, in the research and development of sulfide solid electrolytes, Li content, P / S ratio, and halogen doping amount can be used as composition variables, anion stacking mode and the number of Li sites that can be occupied can be used as structural variables, and sintering temperature, holding time, and atmosphere purity can be used as process variables. The combined effects of these variables on room temperature ionic conductivity and electrochemical stability window can be learned through a Bayesian network model.

[0057] The multi-agent collaborative module is configured to invoke causal relationship models between compositional variables, structural variables, process variables, and performance indicators to assess the impact of different design variables on target performance. The materials design agent can query these causal relationship models before generating candidate material schemes, identifying key compositional and structural variables that significantly affect target performance. These variables are then prioritized as design degrees of freedom for exploration; for example, prioritizing adjustments to the anionic framework type and the concentration of key dopant elements, while assigning lower weight to minor changes in doping ratios with weaker causal effects.

[0058] When scheduling process parameters, the experimental optimization agent can also use causal relationship models to determine which process variables contribute most to performance indicators. For example, if it is found that sintering temperature and holding time have a much greater causal impact on ionic conductivity and interfacial impedance than ball milling time, then the optimization of the experimental parameter space will focus on a fine scan of the sintering window, while only using a coarser gradation for ball milling time. Through this sensitivity assessment based on causal relationship models, the multi-agent collaborative module can limit the search space or priority of the material design agent and the experimental optimization agent, avoiding wasting a lot of experimental resources on variables with very small causal effects.

[0059] Taking solid electrolytes as an example, in the absence of a causal model, the system may distribute attention evenly in the high-dimensional space of composition, structure and process, and make near-blind attempts at combination. It is easy to be misled by some accidental correlations, such as mistakenly believing that a certain tiny dopant element is strongly correlated with ionic conductivity, when in fact the correlation comes from the mixture of process conditions.

[0060] After introducing a causal relationship model between component variables, structural variables, process variables and performance indicators, the multi-agent collaborative module can identify and separate this mixture through a causal graph model or a Bayesian network model, more accurately determine what should be adjusted, and thus concentrate the search space on the combination of factors that have a substantial impact on the target performance.

[0061] This design narrows the range of variables that the materials design agent and the experimental optimization agent need to explore, accelerates the convergence speed, and improves the efficiency of experimental resource utilization. It also makes the system's decision-making process more interpretable, which is conducive to the continuous correction and enrichment of the causal relationship model itself in long-term operation. This upgrades the materials database from a simple correlation memory to a knowledge base with causal reasoning ability, thereby improving the overall design quality and optimization capability of the intelligent R&D system.

[0062] In one exemplary embodiment, the material design agent includes a generation module running on a computing node. The generation module is used to generate candidate material schemes and corresponding synthesis paths within a joint design space consisting of composition, structure, and performance. The material design agent further invokes a material physics model and a machine learning model to predict the target performance of the candidate materials, and filters the candidate material schemes based on the prediction results.

[0063] When the generation module works, it first searches within a joint design space comprised of composition, structure, and performance. This joint design space can be understood as a high-dimensional space established with material composition variables, target crystal or microstructure characteristics, and desired performance indicators as coordinate axes. The composition dimension can cover parameters such as element types, stoichiometry, dopants, and their concentrations; the structure dimension can encompass information such as crystal structure type, space group, anion framework topology, and distribution of occupiable ion sites; and the performance dimension corresponds to the expected range of target performance such as ionic conductivity, electrochemical stability window, interfacial impedance, and mechanical strength.

[0064] The generation module samples, searches, or generatively constructs this joint design space, outputting a set of candidate material schemes that meet preset constraints. It also generates feasible synthesis paths for each candidate material scheme, providing details such as precursor types, proportions, mixing methods, forming methods, and process flows including heat treatment temperature, time, and atmosphere. Taking solid-state electrolytes as an example, the generation module in the materials design agent can automatically propose a series of candidate material schemes with different halogen doping combinations based on existing sulfide system data. It also provides corresponding high-energy ball milling time intervals and sintering temperature windows for each candidate material scheme, forming a structured list of candidate material schemes.

[0065] To avoid the generation module relying solely on experience or random mechanisms to generate physically infeasible combinations, the materials design agent further invokes materials physics models and machine learning models to predict the target performance of candidate material schemes. Materials physics models can include energy stability assessment models based on first-principles calculations, ion transport barrier calculation models, and elastic constant calculation models, used to verify the physical rationality of candidate material schemes in terms of thermodynamic stability, ion transport channel continuity, etc., at the microscopic scale.

[0066] Machine learning models can establish statistical mapping relationships between composition, structure, process variables and performance indicators through supervised or semi-supervised learning based on a large amount of experimental and computational data already available in materials databases, thereby enabling rapid prediction of target properties such as room temperature ionic conductivity, electrochemical stability window, and interfacial impedance.

[0067] In the actual operation, the materials design agent can call the materials physics model and machine learning model in parallel for each candidate material scheme to obtain performance prediction results in multiple dimensions. Based on this, it will perform comprehensive scoring and ranking, and only those candidate material schemes that reach the preset threshold in target performance and show good stability and syntheticability in the materials physics model will be retained. The remaining candidate material schemes will be eliminated or downgraded.

[0068] Specifically, the generation module includes at least one of a graph neural network generation model, a variational autoencoder model, or a sequence generation model based on a large language model; the material design agent also includes a literature semantic parsing unit, which is used to parse the synthesis experience and mechanism descriptions in the material-related literature, convert the parsing results into rules or feature vectors, and fuse and score them with the candidate material schemes output by the generation module to obtain a list of target candidate materials for issuing experiments.

[0069] Graph neural network generation models can represent crystal structures or local coordination environments as graph structures composed of nodes and edges. Nodes correspond to atom or ion types, and edges represent bonding relationships or relative positional relationships. By learning the topological distribution of existing materials in graph space, new graph structures can be sampled in the latent space and then decoded into specific composition ratios and structural units.

[0070] Variational autoencoder models can compress the composition, structure, and some process conditions of known materials into a continuous latent space. By interpolating, perturbing, or moving along the target performance gradient direction within this latent space, candidate material schemes that are statistically similar to high-performance samples but also novel can be generated.

[0071] Sequence generation models based on large language models can encode chemical formulas, crystal structure markers, synthesis steps, and key process parameters into sequence markers. Through conditional generation, given target performance, material type, or constraints, they can output candidate material schemes and corresponding synthesis paths that meet the constraints. For example, they can generate structured sequences like "Li10.5Si1.5P1.5S12Cl0.5, first high-energy ball milling for 6 h, then heat treatment at 510K for 8 h".

[0072] The materials design intelligent agent also incorporates a literature semantic parsing unit to automatically extract synthesis experience and mechanism descriptions from a large amount of materials-related literature that are difficult to obtain directly from numerical databases. This unit can access the full text of papers, patent specifications, and technical reports in electronic format. Through a domain-adaptive natural language processing model, it identifies entities such as material names, component ratios, process conditions, testing methods, and performance results, and mines causal or empirical descriptions, such as "appropriate halogen doping in sulfide solid electrolytes is beneficial for reducing grain boundary impedance" or "sintering temperatures exceeding a certain threshold will lead to the formation of secondary phases." This semantic information is then converted into rule sets or feature vectors through rule templates or embedded encoding. Rules can be expressed as "recommended sintering temperature range given a combination of framework type and doping elements" or "a certain type of element should not be introduced simultaneously at high concentrations," while feature vectors can serve as a quantitative characterization of the similarity of a candidate material scheme in the empirical knowledge space.

[0073] After the generation module outputs a batch of candidate material schemes, the materials design intelligence will fuse and score these candidate material schemes with the rules or feature vectors generated by the literature semantic analysis unit. The fusion scoring process can include two parts: one part is the scoring of the rationality and diversity of the candidate material schemes in the latent space by the generation module, and the other part is the scoring of the consistency between the candidate material schemes and existing experience and mechanism descriptions by the literature semantic analysis unit.

[0074] For example, a graph neural network generative model provides several sulfide electrolyte schemes with novel anionic frameworks. A variational autoencoder model assigns higher expected ionic conductivity scores to some of these combinations in the performance potential space. Meanwhile, a literature semantic parsing unit penalizes candidate material schemes with excessively high halogen content based on the empirical rule that "excessive halogen doping easily leads to secondary phases." The final fused scoring results can be weighted across multiple dimensions, selecting candidate material schemes with higher scores that do not violate known mechanistic principles to form a target candidate material list. This list is then distributed to the experimental optimization agent and experimental platform for subsequent experimental verification.

[0075] Taking the development of solid electrolytes for all-solid-state batteries as an example, without the participation of a literature semantic analysis unit, the generation module may propose a large number of combinations that seem mathematically reasonable but have already been proven unstable or difficult to synthesize in experimental practice, resulting in a significant waste of experimental resources on repeatedly encountering problems. After introducing a literature semantic analysis unit, the system can not only proactively avoid failed formulations already reported in the literature among candidate material schemes, but also prioritize recommending combinations that share commonalities with successful cases in terms of synthesis paths and structural characteristics. For example, it can prioritize halogen doping ranges and sintering temperature ranges similar to a certain type of high-performance sulfide. As a result, the list of target candidate materials is far smaller than the initial generated set, but has a higher hit rate and feasibility.

[0076] In one exemplary embodiment, the experimental optimization agent sequentially optimizes the experimental parameter space using at least one of Bayesian optimization, reinforcement learning, or evolutionary algorithms. The experimental parameter space includes at least one or more of heat treatment temperature, heat treatment time, atmosphere type and pressure, raw material ratio, grinding time, and forming pressure. The experimental optimization agent updates the agent model or policy function on the parameter space based on the experimental feedback data of completed experiments and generates the next batch of experimental schemes using the acquisition function or policy network.

[0077] The experimental parameter space encompasses at least one or more of the following: heat treatment temperature, heat treatment time, atmosphere type and pressure, raw material ratio, grinding time, and forming pressure. Each parameter can be considered as a continuous or discrete adjustable dimension, and multiple dimensions are combined to form a high-dimensional experimental parameter space. The experimental optimization agent employs at least one of Bayesian optimization, reinforcement learning, or evolutionary algorithms to sequentially optimize this parameter space. Sequential optimization refers to re-evaluating, based on the currently available information, which regions in the parameter space are more likely to produce high-performance samples after each batch of experiments is completed and feedback data is obtained, and prioritizing parameter combinations from these regions in the next batch of experiments, rather than enumerating all possible combinations in advance.

[0078] In one implementation, a surrogate model in the parameter space can be constructed using a Bayesian optimization method. Variables such as heat treatment temperature, heat treatment time, atmosphere type and pressure, raw material ratio, grinding time, and forming pressure are used as inputs, while the target performance index (e.g., room temperature ionic conductivity or interfacial impedance) is used as the output. The model is then fitted to the currently completed experimental data using Gaussian process regression or other regression models. In each iteration, the experimental optimization agent uses the surrogate model to estimate the expected value and uncertainty of performance under different parameter combinations. Then, it selects the next batch of parameter points to be actually verified in the parameter space using acquisition functions such as expected improvement and upper confidence bound.

[0079] In another implementation, the experimental optimization agent can employ reinforcement learning algorithms. It constructs states from current parameter selections and historical experimental results, treats adjustments to parameters such as heat treatment temperature, raw material ratio, and grinding time as actions, and considers improvements in target performance, achievement of targets, or failure penalties as reward signals. Through updates to the policy function or value function, it gradually learns how to choose the next experimental plan in different states. In an evolutionary algorithm implementation, a set of parameter combinations can be considered as a population of individuals. Through crossover, mutation, and selection operations, high-performing individuals are retained in multiple iterations, and a new generation of parameter combinations is generated based on these, until the population performance converges or a preset termination condition is reached.

[0080] Through the aforementioned mechanism, the introduction of the experimental optimization agent transforms the allocation of experimental resources from a static, manually designed experimental matrix into a dynamic, self-learning decision-making process. Most low-value regions in the experimental parameter space can be quickly identified and their focus reduced during iterations, concentrating more experimental opportunities on regions deemed high-potential by the proxy model or policy function, significantly reducing ineffective experiments and repetitive explorations. Furthermore, feedback from each round of experiments flows back in the form of structured data, driving continuous updates to the proxy model or policy function, making the system's understanding of the relationship between process variables and performance indicators increasingly accurate.

[0081] In one exemplary embodiment, the experimental platform includes a sample preparation sub-platform and a performance testing sub-platform. The sample preparation sub-platform includes at least a raw material storage unit, a batching and mixing unit, a forming unit, and a sintering unit. The performance testing sub-platform includes at least a structural characterization unit and an electrochemical testing unit. The experimental platform uses a preset process control program to control each unit in a coordinated manner, so as to complete the preparation and testing of multiple material samples in parallel or sequentially.

[0082] The sample preparation sub-platform in the experimental platform includes at least a raw material storage unit, a batching and mixing unit, a forming unit, and a sintering unit. These units are spatially connected via transport mechanisms, transfer trays, or robotic arms, and are controlled by a single process control program. The raw material storage unit can employ a multi-compartment, sealed container structure to store different types of precursor powders or solutions, and uses a weighing module to quantitatively output single or multiple raw materials. The batching and mixing unit receives raw materials from the storage unit and automatically weighs, dispenses, and mixes them to prepare a mixture with the target proportions. The forming unit can be equipped with an automatic tablet press or molding device to press the mixed powder into sheet or block-shaped green bodies under a certain pressure, ensuring the densification process and the consistency of the finished product's geometric dimensions during subsequent sintering. The sintering unit can include multiple independent temperature-controlled chambers, each with programmable heating, holding, and cooling curves, enabling heat treatment of different samples in a controlled atmosphere.

[0083] The performance testing sub-platform is designed for the structural and performance characterization of material samples, and includes at least a structural characterization unit and an electrochemical testing unit. The structural characterization unit can be connected to characterization equipment such as powder diffractometers, Raman spectrometers, and scanning electron microscopes through standardized interfaces. With the help of sample tray coding and barcode recognition mechanisms, it enables automatic sample loading and unloading and test program invocation.

[0084] The electrochemical testing unit can be integrated with an electrochemical workstation, an electrochemical impedance spectroscopy (EIS) module, and a constant current charge-discharge (NCD) module. It performs electrochemical performance tests on pressed samples or assembled half-cell and full-cell structures under preset temperature, pressure, and voltage windows, recording the impedance spectroscopy data, charge-discharge curve data, sample identification, and test conditions. Through this integrated structure and performance characterization configuration, the performance testing sub-platform can continuously output standardized and traceable experimental data with minimal human intervention.

[0085] The experimental platform uses a pre-set process control program to coordinate and control the various units within the sample preparation and performance testing sub-platforms. This program can be deployed in a host computer or distributed control system, parsing the raw material proportioning and process parameter schemes into a series of fixed steps and condition parameters. These parameters are then distributed via communication interfaces to the raw material storage unit, batching and mixing unit, forming unit, sintering unit, structural characterization unit, and electrochemical testing unit. For multiple material samples within the same batch, the process control program can either sequentially complete the preparation and testing, or it can execute in parallel across different chambers and testing channels according to equipment throughput and time windows, thereby maximizing output with limited resources.

[0086] Since all steps are driven by process control programs, the experimental conditions, execution sequence, and data acquisition process have good repeatability and traceability, which greatly reduces the errors and uncertainties caused by human operation differences, and provides a reliable data foundation for subsequent model training, causal analysis, and optimization decision-making of intelligent R&D systems.

[0087] In an exemplary embodiment, the analysis module includes a structural characterization analysis unit and a performance analysis unit. The structural characterization analysis unit is used to perform peak position detection, peak shape fitting, and phase identification on the raw spectra from the structural characterization equipment to obtain the phase composition and structural parameters of the material sample. The performance analysis unit is used to perform feature extraction or equivalent circuit fitting on the impedance spectrum or charge-discharge curve from the electrochemical testing equipment to obtain ionic conductivity, charge transfer impedance, specific capacity, and cycle stability performance indicators. The analysis module uses the analysis results as part of the experimental feedback data, correlates them with the composition and process conditions of the corresponding sample, and writes them into the material database.

[0088] The structural characterization and analysis unit primarily handles raw spectral data from structural characterization equipment, such as intensity-angle spectra output from powder X-ray diffractometers and intensity-displacement curves output from Raman spectrometers. The unit first performs baseline correction and smoothing on the raw spectra to eliminate background noise and instrument drift. Then, it performs peak detection, automatically identifying the angle or wavenumber positions of diffraction peaks or characteristic peaks using algorithms. Based on this, it performs peak shape fitting to obtain precise peak position, full width at half maximum (FWHM), and peak intensity for each peak.

[0089] By comparing the detected peak positions with diffraction cards or spectral characteristics of different phases in a standard database, the structural characterization analysis unit can identify phases, determine the types and approximate proportions of various crystalline phases present in the material sample, and extract structural parameters such as lattice constants, cell parameters, or grain sizes. For example, in the development of solid electrolyte samples, the structural characterization analysis unit can identify the coexistence of the target conductive phase and secondary phases, quantitatively assess the content of the target phase and the influence of sintering conditions on the crystal structure. These results, expressed in the form of phase composition and structural parameters, can serve as the basis for subsequent performance interpretation and process optimization.

[0090] The performance analysis unit targets raw time-domain or frequency-domain data from electrochemical testing equipment, such as frequency-impedance data obtained from electrochemical impedance spectroscopy and voltage-capacity curves obtained from constant current charge-discharge tests. The unit can extract features from impedance spectra or perform equivalent circuit fitting, decomposing complex Nyquist or Bode plots into several physically meaningful component parameters, thereby calculating indicators such as ionic conductivity and charge transfer impedance of the sample under specific temperature and frequency conditions. For charge-discharge curves, the unit can extract characteristic parameters such as first-cycle discharge specific capacity, coulombic efficiency, and capacity retention during cycling to characterize the specific capacity and cycle stability of the material sample.

[0091] After completing the structural characterization and performance analysis units, the analysis module does not save these results in isolation. Instead, it incorporates the analysis results, such as phase composition, structural parameters, ionic conductivity, charge transfer impedance, specific capacity, and cycle stability, as part of the experimental feedback data. These results are then linked one-to-one with the composition and processing conditions of the corresponding material samples and written into the material database. This linking is achieved through unique sample identifiers, constructing a consistent record entry for the entire set of information—composition, structure, process, and performance. This allows subsequent material design and experimental optimization agents to directly obtain complete, multi-dimensional sample characterization results when searching the database, eliminating the need for repeated manual analysis of the original spectra and test curves.

[0092] Through the above design, the analysis module plays a crucial role in transforming experimental results into knowledge within the entire intelligent R&D system. The structural characterization analysis unit and the performance analysis unit automatically convert raw spectra and raw electrochemical test data into high-value structural parameters and performance indicators, significantly reducing the workload of manually analyzing spectra and curves one by one, and improving the consistency and objectivity of data processing. By closely linking the analysis results with sample composition and process conditions and writing them into the materials database, experimental feedback data can be stored in a structured form for a long time and repeatedly accessed, providing a solid data foundation for causal relationship modeling, machine learning modeling, and multi-agent decision-making.

[0093] This disclosure also provides a materials development method using the aforementioned intelligent R&D system, comprising:

[0094] S1, through a materials design intelligent agent, generates the composition, target structure and synthesis path of candidate materials based on the materials science data stored in the materials database, and uses materials physics models and machine learning models to predict the target performance of each candidate material to obtain a set of candidate materials.

[0095] In step S1, the materials design agent retrieves existing materials science data from the materials database, including compositional information, structural information, performance indicators, and corresponding synthesis and characterization conditions for different material samples. Based on this data, the materials design agent generates the composition, target structure, and synthesis path of candidate materials. Specifically, it can explore new combinations in the composition and structure spaces through its internal generation module. Simultaneously, it combines materials physics models and machine learning models to predict the target performance of each candidate material, thereby obtaining a set of candidate materials labeled with predicted performance.

[0096] For example, in the research and development of solid electrolytes for all-solid-state batteries, the material design intelligence can propose several new halogen-doped formulations based on existing data of sulfide systems, provide corresponding sintering process routes, and predict the ionic conductivity, electrochemical stability window, and other indicators of these new formulations at room temperature, providing preliminary screening basis for subsequent experiments.

[0097] S2 optimizes the agent through experiments, models the experimental parameter space, generates an experimental scheme containing candidate materials based on existing experimental feedback data, and sends the experimental scheme to the experimental platform.

[0098] In step S2, the experimental optimization agent models the experimental parameter space, treating process variables such as heat treatment temperature, heat treatment time, atmosphere type and pressure, raw material ratio, grinding time, and forming pressure as adjustable dimensions. A surrogate model or strategy function is constructed in the parameter space based on existing experimental feedback data. Based on this modeling result, the experimental optimization agent generates an experimental plan containing candidate materials and sends the plan to the experimental platform. The experimental plan includes not only the specific ratio information of each candidate material but also the specific combination of process parameters to be used for each sample.

[0099] For example, for a candidate sulfide formulation, sintering can be performed at two or three different temperature windows. By comparing the performance differences corresponding to different temperature curves, sensitive process conditions can be identified. Compared to traditional experimental matrices designed based on human experience, this step automatically plans the distribution of experimental resources in the parameter space through a data-driven approach, making the experiments more targeted and informative.

[0100] Specifically, in step S2, the experimental optimization agent constructs the experimental feedback data of the current round and the previous rounds into a state, uses a reinforcement learning algorithm to construct the adjustment of experimental parameters into actions, constructs the improvement of target performance or whether the constraints are met into a reward signal, and generates subsequent experimental schemes through policy updates; or, the experimental optimization agent uses a Bayesian optimization algorithm to establish a surrogate model in the experimental parameter space, and selects experimental sample points on the surrogate model based on the acquisition function.

[0101] For reinforcement learning paths, the experimental optimization agent first constructs a state by unifying the experimental feedback data from the current round and historical rounds. This state can include information such as the composition of candidate materials, the current combination of process parameters, the target performance distribution of recent experimental samples, and the proportion of unexplored areas. This information is then used to form a vector representation that can be processed by the policy network through embedding encoding or feature concatenation. The adjustment of experimental parameters is characterized as an action, that is, a specific adjustment scheme made in a given state for dimensions such as heat treatment temperature, heat treatment time, atmosphere parameters, raw material ratio, or grinding time.

[0102] For example, the sintering temperature is adjusted from 520K to 540K, and the halogen doping amount is adjusted from 5% to 7%. The improvement of the target performance or whether the preset constraints are met is constructed as a reward signal. When the new round of experiments is better than the historical samples in terms of ionic conductivity, charge transfer impedance, etc., or meets the constraint range for the first time, a positive reward is given. Otherwise, a zero or negative reward is given.

[0103] By continuously collecting state-action-reward-new state sequences during multiple rounds of experiments, the experimental optimization agent continuously updates its internal policy function using policy gradient, Q-learning, or other reinforcement learning algorithms. This enables the policy network to gradually learn to prioritize generating experimental schemes that are more likely to bring performance improvements or information gains under different states, thereby achieving self-learning optimization of process parameters.

[0104] For the Bayesian optimization path, the experimental optimization agent constructs a surrogate model in the experimental parameter space, taking process variables as input and target performance as output. It uses Gaussian process regression or other probabilistic models to fit the parameter-performance mapping relationship based on the completed experimental samples, providing both the performance prediction value and the uncertainty estimate for each point. Based on this, it defines a collection function, such as expected improvement, upper confidence bound, or information gain, and comprehensively considers the current prediction performance and the uncertainty of the model in this region, thereby selecting the next batch of experimental sample points on the surrogate model.

[0105] Taking the optimization of the sintering window of solid electrolytes as an example, reinforcement learning can learn through repeated experiments how to adjust the sintering temperature and time to maximize ionic conductivity under different ratios and temperature distributions, while Bayesian optimization can gradually narrow the exploration area on the temperature-time plane and focus on sampling in areas with high performance and large uncertainty. Both essentially improve the experimental scheme through feedback-driven decision-making.

[0106] By introducing sequential optimization methods such as reinforcement learning or Bayesian optimization, the experimental optimization agent no longer relies on a fixed trial matrix or human experience. Instead, it transforms the results of each round of experiments into an update of the understanding of the parameter space, enabling subsequent experimental schemes to achieve a better balance between information content and success rate. This significantly reduces the proportion of invalid experiments, accelerates the convergence speed towards optimal process conditions and high-performance material combinations, and improves the overall efficiency of the entire intelligent R&D system in terms of experimental resource utilization and performance exploration.

[0107] S3, the experimental platform performs material sample preparation and performance testing according to the experimental plan, and obtains the corresponding raw experimental data.

[0108] In step S3, the experimental platform executes the preparation and performance testing of material samples according to the experimental plan issued by the experimental optimization agent. Sample preparation is completed through processes such as raw material storage, batching and mixing, forming, and sintering. Then, structural characterization equipment and electrochemical testing equipment are used to complete structural and performance measurements. The experimental platform outputs raw experimental data for each sample, such as powder diffraction intensity-angle patterns, impedance spectral frequency-impedance data, and charge-discharge voltage-capacity curves. This raw data directly reflects the material's true performance under the given process conditions.

[0109] S4, the analysis module analyzes and extracts features from the raw experimental data to obtain experimental feedback data, and writes the experimental feedback data into the material database.

[0110] In step S4, the analysis module analyzes and extracts features from the raw experimental data. The structural characterization analysis unit performs peak position detection, peak shape fitting, and phase identification to obtain phase composition and structural parameters. The performance analysis unit performs impedance spectrum equivalent circuit fitting or charge-discharge curve feature extraction to obtain performance indicators such as ionic conductivity, charge transfer impedance, specific capacity, and cycle stability. These analysis results are then compiled into experimental feedback data and written into the material database, so that each sample corresponds to a complete record in the database containing composition, structure, process, and performance.

[0111] S5, based on the updated materials database, updates the materials physics model and machine learning model, and drives the materials design agent and the experimental optimization agent to generate the next round of candidate materials and experimental schemes in the new database state.

[0112] In step S5, after the materials database is updated with a new round of experimental feedback data, the knowledge-updating agent updates the materials physics model and the machine learning model accordingly. On the one hand, new sample data is incorporated into the training set, and the machine learning model can adjust its parameters through incremental training or periodic retraining, so that its predictive ability continuously improves with the accumulation of data. On the other hand, by comparing the deviation between the model prediction and the experimental feedback, the approximate assumptions in the materials physics model can be corrected or expanded to better reflect the actual behavior of the current materials system.

[0113] Based on this, the knowledge updating agent drives the material design agent and the experimental optimization agent to generate the next round of candidate materials and experimental schemes in the new database state, so that the design and optimization strategy of the entire system is always based on the latest knowledge.

[0114] S6. Repeat steps S2 to S5 until the preset termination condition is met.

[0115] In step S6, by repeatedly executing steps S2 to S5, a continuously running closed-loop iterative process is formed until a preset termination condition is met, such as finding several material formulations and process windows that meet the target performance threshold, or achieving performance convergence within a specified experimental budget and time range.

[0116] In summary, the intelligent R&D system and materials R&D method provided in this disclosure achieve the integration of materials design, experimental execution, data feedback, and knowledge updating by constructing a closed-loop data flow among the materials database, multi-agent collaborative module, experimental platform, and analysis module. Compared with the traditional decentralized, linear, and manually driven R&D model, it significantly improves the efficiency and success rate of new material screening and process optimization, reduces trial and error costs, and is conducive to the formation of a sustainable and evolving materials knowledge system.

[0117] It will be apparent to those skilled in the art that this disclosure is not limited to the details of the exemplary embodiments described above, and that this disclosure can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of this disclosure is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this disclosure. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0118] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An intelligent R&D system for new material research and development, characterized in that, include: A materials database is used to store and structure materials science data, including the proportions of sulfide solid electrolyte samples, sintering curves, room temperature ionic conductivity obtained by impedance spectroscopy fitting, and X-ray diffraction pattern analysis results. A multi-agent collaborative module is communicatively connected to the material database. The multi-agent collaborative module includes at least a material design agent, an experimental optimization agent, and a knowledge update agent. The material database internally constructs a causal relationship model between composition variables, structural variables, process variables, and performance indicators. The causal relationship model is at least one of a causal graph model or a Bayesian network model. The composition variables include Li content, P / S ratio, and halogen doping amount. The structural variables include anion stacking mode and the number of Li sites that can be occupied. The process variables include sintering temperature, holding time, and atmosphere purity. The performance indicators include room temperature ionic conductivity and electrochemical stability window. The material design agent is used to read historical sample data from the material database and call the causal relationship model to identify key component variables and structural variables that have a significant impact on the target performance. The key component variables and structural variables are used as the design freedoms to be explored first, so as to generate candidate material schemes and corresponding synthesis paths. The experimental optimization agent is used to call the causal relationship model, determine the contribution of different process variables to the performance index, and limit the experimental parameter space or experimental parameter priority based on the experimental feedback data of the completed experiments, so as to generate the experimental scheme of the candidate material. The experimental platform is communicatively connected to the multi-agent collaborative module and is used to prepare and test material samples according to the experimental scheme generated by the experimental optimization agent, and output the original frequency-impedance data and the original diffraction intensity-angle data. The analysis module is communicatively connected to the experimental platform and the material database. It is used to analyze the raw frequency-impedance data and the raw diffraction intensity-angle data to obtain room temperature ionic conductivity, interfacial resistance and phase composition ratio. The room temperature ionic conductivity, interfacial resistance and phase composition ratio are correlated with the composition and process conditions of the corresponding sample and written into the material database. At the same time, the structured experimental feedback is pushed to the multi-agent collaborative module. The knowledge updating agent is used to incorporate new sample data into the training set after the material database is updated by a new round of experimental feedback data, and to update the material physics model and machine learning model by comparing the deviation between the model prediction and the experimental feedback, so as to drive the material design agent and the experimental optimization agent to generate the next round of candidate material schemes and experimental schemes in the new database state.

2. The intelligent R&D system according to claim 1, characterized in that, The materials database stores materials science data including at least one or more of the following: material composition information, structural information, performance indicators, and synthesis and characterization conditions. When the material database performs a structured representation of material samples, it performs vectorized encoding of the material samples based on preset material feature descriptors and is configured to support material similarity retrieval based on the material feature descriptors.

3. The intelligent R&D system according to claim 1, characterized in that, The material design agent includes a generation module running on a computing node. The generation module is used to generate candidate material schemes and corresponding synthesis paths in a joint design space consisting of composition, structure and performance. The material design agent further invokes material physics models and machine learning models to predict the target performance of candidate materials, and screens candidate material schemes based on the prediction results.

4. The intelligent R&D system according to claim 3, characterized in that, The generation module includes at least one of a graph neural network generation model, a variational autoencoder model, or a sequence generation model based on a large language model; The materials design agent also includes a literature semantic parsing unit, which is used to parse the synthesis experience and mechanism descriptions in materials-related literature, convert the parsing results into rules or feature vectors, and fuse and score them with the candidate material schemes output by the generation module to obtain a list of target candidate materials for issuing experiments.

5. The intelligent R&D system according to claim 1, characterized in that, The experimental optimization agent uses at least one of Bayesian optimization, reinforcement learning or evolutionary algorithm to sequentially optimize the experimental parameter space. The experimental parameter space includes at least one or more of the following: heat treatment temperature, heat treatment time, atmosphere type and pressure, raw material ratio, grinding time and forming pressure. The experimental optimization agent updates the agent model or policy function in the parameter space based on the experimental feedback data of the completed experiments, and generates the next batch of experimental schemes using the acquisition function or policy network.

6. The intelligent R&D system according to claim 1, characterized in that, The experimental platform includes a sample preparation sub-platform and a performance testing sub-platform. The sample preparation sub-platform includes at least a raw material storage unit, a batching and mixing unit, a forming unit, and a sintering unit. The performance testing sub-platform includes at least a structural characterization unit and an electrochemical testing unit. The experimental platform uses a preset process control program to control each unit in a coordinated manner, so as to complete the preparation and testing of multiple material samples in parallel or sequentially.

7. The intelligent R&D system according to claim 1, characterized in that, The analysis module includes a structural characterization analysis unit and a performance analysis unit; The structural characterization analysis unit is used to perform peak position detection, peak shape fitting, and phase identification on the raw spectra from the structural characterization equipment to obtain the phase composition and structural parameters of the material sample. The performance analysis unit is used to extract features or fit equivalent circuits to impedance spectra or charge-discharge curves from electrochemical testing equipment to obtain performance indicators such as ionic conductivity, charge transfer impedance, specific capacity and cycle stability. The analysis module incorporates the analysis results as part of the experimental feedback data, correlates them with the composition and process conditions of the corresponding samples, and then writes them into the material database.

8. A material development method using the intelligent R&D system according to any one of claims 1 to 7, characterized in that, include: S1, through a materials design intelligent agent, generates the composition, target structure and synthesis path of candidate materials based on the materials science data stored in the materials database, and uses materials physics model and machine learning model to predict the target performance of each candidate material to obtain a set of candidate materials; S2 optimizes the intelligent agent through experiments, models the experimental parameter space, generates an experimental plan containing candidate materials based on existing experimental feedback data, and sends the experimental plan to the experimental platform. S3, the experimental platform performs the preparation and performance testing of material samples according to the experimental plan, and obtains the corresponding raw experimental data; S4, the analysis module analyzes and extracts features from the raw experimental data to obtain experimental feedback data, and writes the experimental feedback data into the material database; S5, the knowledge updating agent updates the material physics model and machine learning model based on the updated material database, and drives the material design agent and the experimental optimization agent to generate the next round of candidate materials and experimental schemes in the new database state. S6. Repeat steps S2 to S5 until the preset termination condition is met.

9. The material development method according to claim 8, characterized in that, In step S2, the experimental optimization agent constructs the state from the experimental feedback data of the current round and the previous rounds, constructs the action from the adjustment of experimental parameters using the reinforcement learning algorithm, constructs the reward signal from the improvement of target performance or whether the constraint is met, and generates the subsequent experimental plan through policy update. Alternatively, the experimental optimization agent may employ a Bayesian optimization algorithm to establish a surrogate model in the experimental parameter space, and select experimental sample points on the surrogate model based on the acquisition function.

Images