Sodium battery positive electrode doping structure agent driving design and screening method and system

Abstract

The application provides a sodium battery positive electrode doping structure intelligent agent driving design and screening method and system, comprising: receiving a natural language doping demand instruction, searching professional knowledge in a sodium electricity material simulation knowledge base through retrieval enhancement generation technology, generating an atomic level tool call sequence and an initial state snapshot based on ReAct iterative reasoning by using a large language model; performing geometric optimization, symmetry analysis and three-level physical and chemical rule screening on the crystal structure; through a multi-fidelity screening funnel and a machine learning interatomic potential model rapid relaxation sorting, combining error calibration and context Bayesian optimization active search, the DFT calculation amount is greatly reduced; automatically generating a VASP file and batch calculating, self-adaptively adjusting parameters, completing formation energy reasoning and stability sorting visualization. The application constructs a self-evolution knowledge graph, continuously accumulates historical calculation success and failure experience as structured knowledge, and realizes autonomous enhancement of task planning capability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and computational electrochemical materials science, specifically involving a method and system for intelligent agent-driven design and screening of sodium battery cathode doped structures. In particular, it relates to a fully automated method and system for performing structural doping, structural screening, density functional theory calculation, and workflow scheduling of sodium battery cathode crystal materials using an intelligent agent driven by a large language model. Background Technology

[0002] Sodium-ion batteries have become a research hotspot in the field of large-scale energy storage due to their advantages such as abundant sodium resources, low cost, and strong complementarity with lithium-ion batteries. The cathode materials for sodium-ion batteries mainly include three systems: layered transition metal oxides, polyanionic compounds, and Prussian blue analogues. Among them, layered transition metal oxides are considered the most commercially promising sodium battery cathode material system due to their simple synthesis process, high theoretical specific capacity, and low cost.

[0003] However, layered oxide cathode materials face serious technical bottlenecks in practical applications: irreversible phase transitions, structural degradation, voltage decay, and poor cycle stability severely hinder the commercial development of sodium-ion batteries. To address these issues, doping modification has proven to be one of the most effective strategies—by introducing foreign elements (such as Mg, Zn, Al, Cu, Fe, Ti, Ni, etc.) into transition metal sites or sodium sites, harmful phase transitions can be significantly suppressed, structural stability improved, ion transport kinetics enhanced, and cycle life improved.

[0004] The most relevant prior art to this invention is a large language model-driven framework for automating density functional theory (DFT) computation. This framework is based on a hierarchical intelligent agent driven by a large language model, which automatically transforms natural language requests into a complete computational workspace (including structure files, input files, output files, log files, and execution records). By maintaining persistent project records (recording key facts, constraints, and file pointers), it supports the checkability and restartability of tasks. Combined with a multi-fidelity toolkit covering fast agent relaxation and high-fidelity DFT computation, it optimizes the DFT pipeline and agent model, reducing manual intervention.

[0005] However, the aforementioned prior art has the following technical problems:

[0006] First, existing technologies are primarily geared towards general computing workflows, leaving significant gaps in doping semantic understanding and structure pre-screening. Systems cannot parse doping instructions at specific locations or concentrations, forcing core design work to still rely on manual preprocessing. Furthermore, the lack of a low-cost pre-screening step based on physicochemical rules (such as atomic indices, interatomic spacing, and Ewald energies of structures) results in a large number of physically unreasonable structures directly entering expensive DFT calculations, leading to severe waste of computational resources and inefficient task iteration.

[0007] Second, the existing scheduling modes are too simplistic, making it difficult to support efficient exploration and dynamic adaptive computation across multi-dimensional parameter spaces. When dealing with multi-element, multi-concentration doping combinations, the system cannot achieve automatic task decomposition, parallel resource scheduling, or unified result aggregation and analysis. Furthermore, the rigid sequential execution mode prevents it from flexibly adjusting subsequent processes based on intermediate results (such as convergence status or errors), and lacks conditional branching and intelligent retry mechanisms, thus limiting the system's evolution towards a multi-scale, intelligent adaptive platform. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for intelligent agent-driven design and screening of sodium battery cathode doping structures. By building a dedicated parameter parsing module for crystal doping tasks, it automatically extracts structured parameters such as doping elements and concentrations from the user's natural language for structural doping. Before DFT calculation, a multi-level physical / chemical rule filter is placed to realize the scheduling of structured batch doping calculations and automatic summary of results. The multi-concentration gradient, multi-element or high-entropy combination space is decomposed into parallel tasks, and input files are generated in batches, workflow jobs are submitted, and calculation results are analyzed.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] This invention provides a method for designing and screening intelligent agents driving a sodium battery cathode doped structure, comprising:

[0011] The system receives natural language commands from users regarding sodium-ion crystal doping requirements, searches a pre-defined sodium-ion material simulation knowledge base using a similarity retrieval algorithm, and uses a large language model to perform logical reasoning based on the retrieved domain-specific knowledge entries to generate an atomic-level tool call sequence and an initial state snapshot.

[0012] Structural optimization is performed on the standardized crystal structure in the initial state snapshot. An original set of candidate doped configurations is generated through a symmetry analysis program. Multi-level screening is performed according to the physicochemical rules in the domain expertise entries to obtain a set of screened candidate structures.

[0013] For the selected candidate structure set, VASP input files are automatically generated based on the atomic-level tool call sequence and the domain knowledge entries, and a multi-step calculation workflow is constructed. The multi-step calculation workflow is submitted to the computing server in batches, the calculation convergence status is monitored in real time, and the calculation parameters are adaptively adjusted based on the large language model inference to obtain the DFT calculation results of each doped configuration.

[0014] Based on the DFT calculation results of each doped configuration, the total energy of each doped configuration is extracted and the formation energy calculation inference is performed to generate energy convergence curves and stability visualization charts, thus obtaining the stability ranking results of each doped configuration.

[0015] Optionally, the natural language command for sodium-ion crystal doping requirements input by the user is received, and a similarity retrieval algorithm is used to search a preset sodium-ion material simulation knowledge base. A large language model is then used to perform logical reasoning based on the retrieved domain-specific knowledge entries to generate an atomic-level tool call sequence and an initial state snapshot, including:

[0016] Upon receiving the natural language instruction, the domain-specific knowledge entries are retrieved from the sodium electrochemical material simulation knowledge base using a cosine similarity algorithm. Furthermore, doping elements, doping sites, doping concentrations, and parent materials are extracted from the natural language instruction through keyword matching and slot filling to obtain structured intent metadata.

[0017] The structured intent metadata is mapped into a structured task blueprint by the task orchestration agent, and decomposed into subtasks of structure analysis, doping configuration generation, electrostatic energy screening, VASP input generation and result analysis, generating a directed acyclic graph execution plan.

[0018] Based on the directed acyclic graph execution plan, the structure analysis program is invoked to transform the unstructured input data into the standardized crystal structure, and it is determined whether the standardized crystal structure reaches the computational mechanical stability threshold. If it does, the pseudopotential file, cutoff energy, and k-point sampling accuracy are determined based on the domain expertise entries, the initial state snapshot is constructed, and an interpretable execution trajectory is formed using an iterative inference execution mechanism with a large language model, determining the atomic-level tool call sequence and the physical constraint logic between each subtask.

[0019] Optionally, the normalized crystal structure in the initial state snapshot is subjected to structure optimization, an original candidate doped configuration set is generated through a symmetry analysis program, and multi-level screening is performed according to the physicochemical rules in the domain expertise entries to obtain a screened candidate structure set, including:

[0020] Based on the atomic-level tool call sequence, the geometry optimization program is invoked to start the initial state structure optimization calculation of the standardized crystal structure, and the energy convergence curve and atomic force vector are captured in real time. After the preset convergence threshold is met, the parallel computing subtask set covering multiple different desodium components is executed with high concurrency through the concurrent scheduling engine to obtain the relaxed optimized structure.

[0021] The relaxed optimized structure is subjected to symmetry analysis using PyMatgen to generate all non-equivalent sodium removal configurations and independent doping configurations, which constitute the original candidate doping configuration set.

[0022] For the original candidate doped configuration set, a three-level screening process is performed sequentially using the Ewald electrostatic energy reverse screening rule, the interatomic spacing threshold elimination rule, and the equivalent configuration deduplication rule to obtain the screened candidate structure set.

[0023] Optionally, for the screened candidate structure set, a VASP input file is automatically generated based on the atomic-level tool call sequence and the domain expertise entries, and a multi-step computation workflow is constructed. The multi-step computation workflow is submitted to the computation server in batches, the computation convergence status is monitored in real time, and the computation parameters are adaptively adjusted based on large language model inference to obtain the DFT calculation results for each doped configuration, including:

[0024] Based on the selected candidate structure set, the geometric and electronic structure features of each candidate configuration are analyzed by the Agent. The matching physical models and expert experience are retrieved from the sodium electrochemical material simulation knowledge base by the RAG, and INCAR, POSCAR, KPOINTS and POTCAR files are automatically generated to form the VASP input file. A multi-step calculation workflow supporting serial, parallel and conditional branch execution is constructed according to the atomic-level tool call sequence, and the multi-step calculation workflow is submitted to the computing server in batches.

[0025] Based on the computing status of the computing server, the energy convergence curve, atomic force vector and charge density residual are captured in real time, and the convergence risk is determined by reasoning using a large language model. If it is determined that there is a local structural collapse anomaly, the subsequent DFT calculation of the corresponding candidate configuration is terminated. If it is determined that there is convergence difficulty, the calculation parameters in the INCAR file are automatically corrected and the self-consistent calculation request is reissued to obtain the DFT calculation results of each doped configuration.

[0026] Optionally, the DFT calculation results for each doped configuration are used to extract the total energy of each doped configuration and perform formation energy calculation inference to generate energy convergence curves and stability visualization charts, thereby obtaining a stability ranking result for each doped configuration, including:

[0027] Based on the DFT calculation results of each doped configuration, the total energy, atomic force convergence trajectory and charge density residual of each doped configuration are extracted from the DFT calculation output file to generate the energy-ion step convergence curve.

[0028] Based on the energy-ion step convergence curve and the total energy, the formation energy calculation reasoning is performed using a large language model. The thermodynamic spontaneity is determined by comparing the energy difference between the doped system and the pure system, and the stability visualization chart is generated to obtain the stability ranking results of each doped configuration.

[0029] Optionally, it also includes a self-evolving knowledge graph update step, including:

[0030] Based on the ontological concepts and domain relationships in the field of sodium-ion cathode doping, the initial knowledge graph is constructed by automatically extracting entity and relation triplets from the sodium-ion material simulation knowledge base, with parent material, doping element, doping site, phase structure, calculation parameters and physical quantity index as node types, and doping initiation relationship, stabilization effect relationship, recommended calculation parameter relationship and calculation success or failure relationship as edge types.

[0031] Based on the initial knowledge graph and the DFT calculation results of each doped configuration, whenever any DFT calculation subtask is completed, computational feature information is automatically extracted from the DFT calculation output file. The computational feature information is written into the initial knowledge graph in the form of standardized triples, and entity alignment and conflict resolution are performed to obtain the merged knowledge graph. If the calculation fails to converge, the failure reason label and corresponding calculation parameter snapshot are recorded to form a failure experience node and written into the merged knowledge graph to obtain the updated knowledge graph. If the calculation converges successfully, the merged knowledge graph is used as the updated knowledge graph.

[0032] Based on the updated knowledge graph, when a new doping task request is received, the system retrieves the set of completed computation cases and the failed experience nodes that match the current doping task. The distribution of computational parameters of successful cases in the set of completed computation cases is used as a strong prior for initializing the computational parameters of the current task. Computational parameters that overlap with the failure modes recorded in the failed experience nodes are skipped or corrected to obtain task planning results optimized based on historical experience. The task planning results are then written into the atomic-level tool call sequence and the domain knowledge entries are updated.

[0033] Optionally, after obtaining the filtered candidate structure set, and before submitting the multi-step computation workflow to the computation server in batches, the process further includes a multi-fidelity filtering funnel execution step, comprising:

[0034] Based on the aforementioned domain expertise items, a base model is loaded from a pre-trained general inter-atomic potential model library. The root mean square error between the machine learning inter-atomic potential prediction of the base model and the DFT reference value is verified for the target parent material system. If the root mean square error is lower than a preset error threshold, the base model is used as the system-specific machine learning inter-atomic potential model. Otherwise, the base model is transferred and fine-tuned using the pre-acquired DFT reference calculation results of the target parent material system as the fine-tuning dataset to obtain the system-specific machine learning inter-atomic potential model.

[0035] Based on the selected candidate structure set, the system-specific machine learning interatomic potential model is invoked to perform fast structure relaxation and energy sorting on each candidate structure, eliminating structures with abnormal interatomic spacing or energy exceeding a preset energy window threshold, thereby obtaining a machine learning interatomic potential fine-screened candidate structure set. Each candidate structure in the machine learning interatomic potential fine-screened candidate structure set carries a machine learning relaxed structure after machine learning interatomic potential relaxation.

[0036] Based on the machine learning interatomic potential screening candidate structure set, the machine learning relaxation structure corresponding to each candidate structure is used as the starting structure for DFT calculation to submit accurate DFT calculation. The accurate formation energy is calculated back from the total DFT energy to obtain the stability ranking of the doped structure. The stability ranking of the doped structure is then input into the multi-step calculation workflow to prioritize the scheduling of a preset number of candidate structures with the highest stability ranking.

[0037] Optionally, after the step of refining the candidate structure set based on the machine learning inter-atomic potential and submitting the accurate DFT calculation using the machine learning relaxation structure corresponding to each candidate structure as the starting structure for DFT calculation, the method further includes a dynamic calibration step for the prediction error of the machine learning inter-atomic potential, including:

[0038] Based on each candidate structure that simultaneously completes machine learning inter-atomic potential pre-screening and precise formation energy back-calculation, calculate the machine learning inter-atomic potential formation energy prediction error value for each candidate structure, establish an error feature mapping model between the error value and the corresponding structural features, and dynamically adjust the retention ratio of subsequent candidate structures based on the error feature mapping model to obtain the calibrated screening threshold.

[0039] Based on the calibrated screening threshold, when the cumulative error of a certain type of doping combination exceeds a preset deviation threshold, candidate structures that have been eliminated by the system-specific machine learning interatomic potential model are automatically extracted and forcibly included in the DFT calculation queue for cross-validation to obtain cross-validation results. Based on the cross-validation results, the system-specific machine learning interatomic potential model is incrementally fine-tuned to obtain an updated system-specific machine learning interatomic potential model, and the updated system-specific machine learning interatomic potential model replaces the system-specific machine learning interatomic potential model for fast structure relaxation of subsequent candidate structures.

[0040] Optionally, after obtaining the filtered candidate structure set, and before submitting the multi-step computation workflow to the computation server in batches, the process further includes an active sequential search step based on contextual Bayesian optimization of a large language model, including:

[0041] Based on the screened candidate structure set, the parent material type, dopant element list, element occupancy ratio, and doping site are extracted for each candidate doping configuration and concatenated into a standardized natural language description string. A Latin hypercube sampling strategy is used to select a preset number of initial candidate doping configurations from the screened candidate structure set to complete the DFT calculation, obtain the formation energy label corresponding to each initial candidate doping configuration, and construct an initial observation set by combining the standardized natural language description string with the corresponding formation energy label.

[0042] Based on the initial observation set, the initial observation set is formatted into a list of few samples and concatenated into the prompt word template of the large language model. The variance is calculated through multiple inferences using temperature sampling to obtain the predicted distribution of all candidate doped configurations in the selected candidate structure set by the current surrogate model. Based on the predicted distribution, a weighted confidence upper bound strategy is used to select a preset batch of candidate doped configurations with the highest collected values ​​from the selected candidate structure set in each iteration and submit them to the DFT calculation to obtain the new formation energy observation values ​​corresponding to each preset batch of candidate doped configurations.

[0043] Based on the new formation energy observation, the standardized natural language description string corresponding to each of the preset batch of candidate doping configurations and the corresponding new formation energy observation are added to the initial observation set to trigger the next round of iteration until a stable doped structure with a formation energy lower than a preset target threshold is found, the formation energy change of the optimal candidate doping configuration is less than a preset convergence threshold when a preset number of rounds is reached, or the DFT computing resources reach a preset upper limit. The thermodynamically most stable doping configuration recommendation set is obtained, and the doping configuration recommendation set is input into the multi-step computing workflow to prioritize the scheduling of the candidate doping configuration with the lowest formation energy.

[0044] This invention also provides a system for designing and screening intelligent agents driving sodium battery cathode doped structures, comprising:

[0045] The instruction parsing and knowledge retrieval module is used to receive natural language instructions from users regarding sodium electrolytic crystal doping requirements. It then uses a similarity retrieval algorithm to search the preset sodium electrolytic material simulation knowledge base. Based on the retrieved domain-specific knowledge entries, it uses a large language model to perform logical reasoning and generate an atomic-level tool call sequence and an initial state snapshot.

[0046] The candidate structure generation and screening module is used to perform structure optimization on the standardized crystal structure in the initial state snapshot, generate an original candidate doping configuration set through a symmetry analysis program, and perform multi-level screening according to the physicochemical rules in the domain knowledge entries to obtain a screened candidate structure set.

[0047] The computational workflow construction and execution module is used to automatically generate VASP input files and construct multi-step computational workflows for the screened candidate structure set based on the atomic-level tool call sequence and the domain knowledge entries. The multi-step computational workflows are submitted to the computational server in batches, the computational convergence status is monitored in real time, and the computational parameters are adaptively adjusted based on large language model inference to obtain the DFT calculation results of each doped configuration.

[0048] The stability analysis and ranking module is used to extract the total energy of each doped configuration from the DFT calculation results of each doped configuration and perform formation energy calculation inference to generate energy convergence curves and stability visualization charts, thereby obtaining the stability ranking results of each doped configuration.

[0049] Compared with the prior art, the present invention has the following advantages:

[0050] 1. End-to-end automatic generation of natural language doping schemes: This invention has a built-in dedicated parameter parsing module for crystal doping tasks. Users only need to describe the doping requirements in natural language, and the system can automatically extract structured parameters such as doping elements, sites, and concentrations, and establish atomic index mapping and supercell conversion mechanisms to achieve end-to-end conversion from natural language to executable doping schemes, which significantly reduces the threshold for sodium electrode cathode material doping design.

[0051] 2. The pre-screening mechanism based on physical / chemical rules significantly reduces the cost of DFT calculation: This invention incorporates a multi-level physical / chemical rule filter before DFT calculation, supporting atomic index, atomic spacing, and Ewald electrostatic energy screening. The default is to use Ewald energy reverse order screening. Compared with existing technologies, this can reduce the calculation overhead of invalid structures, significantly shorten the task queue waiting time, and maximize the utilization of computing resources.

[0052] 3. Structured batch doping scheduling and efficient exploration of high-entropy combination space: This invention realizes batch management capability tailored to the characteristics of doping tasks. It can automatically decompose multi-concentration gradient, multi-element or high-entropy combination space into subtasks that can be computed in parallel, generate input files in batches, submit workflow jobs and uniformly monitor the computing status, support one-click batch exploration, and automatically sort the batch results by stability, form energy comparison and visualization summary, which greatly improves the efficiency of combination doping screening.

[0053] 4. Intelligent orchestration and adaptive computing capabilities for complex tasks: This invention supports the dynamic decomposition and reorganization of multi-step workflows based on user input. It can automatically adjust the parameters of subsequent tasks based on the convergence of intermediate results or error thresholds, and call heterogeneous computing resources in parallel to accelerate bottleneck links. It realizes conditional branch judgment, intelligent retry of anomalies, and truly adaptive workflow scheduling, which significantly improves the success rate and efficiency of complex system computing.

[0054] 5. Streaming data-driven closed-loop feedback and dynamic decision-making: This invention adopts the ReAct paradigm to capture streaming physical indicators such as energy convergence curves, atomic force vectors, and charge density residuals in real time during the calculation process. It also uses a large language model for dynamic reasoning. When anomalies such as local structural collapse or SCF oscillation are detected, the system can automatically terminate invalid calculations or correct parameters and reissue them, enabling the system to have the autonomous decision-making ability of a materials calculation expert.

[0055] 6. Automatic result summarization and multimodal visualization output: This invention automatically collects key physical quantities generated by batch DFT calculation tasks, including total energy, formation energy, lattice distortion rate, charge density convergence curves and energy-ion step curves of structural relaxation for each doped / desodium configuration, and generates graphic analysis reports with the assistance of a large language model, which greatly improves the decision-making efficiency of material screening. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart illustrating the intelligent agent driving design and screening method for sodium battery cathode doped structure provided in an embodiment of the present invention.

[0058] Figure 2 A schematic diagram of a multi-stage screening funnel process based on physical / chemical rules provided in an embodiment of the present invention;

[0059] Figure 3This is a schematic diagram of the intelligent agent-driven design and screening system for the positive electrode doping structure of a sodium battery provided in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] Example 1

[0062] like Figure 1 As shown, this invention provides a method for designing and screening intelligent agents driving a sodium battery cathode doped structure, comprising:

[0063] Step S1: Receive the user's input of the sodium electrochemical crystal doping requirement in natural language, search in the preset sodium electrochemical material simulation knowledge base through a similarity retrieval algorithm, and use a large language model to perform logical reasoning based on the retrieved domain professional knowledge entries to generate an atomic-level tool call sequence and an initial state snapshot.

[0064] Specifically, the system's input layer is oriented towards the vertical scientific task of sodium-ion crystal doping, constructing a Retrieval-Enhanced Generation (RAG) architecture. Upon receiving the user's natural language doping request, relevant professional knowledge is retrieved from the sodium-ion material simulation knowledge base to assist the large language model in accurately extracting structured parameters such as doping elements, sites, and concentrations. The similarity retrieval algorithm includes, but is not limited to, the cosine similarity algorithm, which extracts the core semantic feature vector of the user's command and matches it with knowledge entries in the knowledge base to obtain domain-specific professional knowledge entries related to the current request. Based on these domain-specific professional knowledge entries, the large language model performs logical reasoning to determine the atomic-level tool call sequence required to complete the request and the physical constraint logic between each step. Simultaneously, it constructs an initial state snapshot, providing a definite physical benchmark for subsequent calculation processes.

[0065] Step S2: Perform structure optimization on the standardized crystal structure in the initial state snapshot, generate an original candidate doped configuration set through a symmetry analysis program, and perform multi-level screening according to the physicochemical rules in the domain expertise entries to obtain a screened candidate structure set.

[0066] Specifically, based on the atomic-level tool call sequence, a geometry optimization program is invoked to initiate initial-state structure optimization calculations for the standardized crystal structure, capturing energy convergence curves and atomic force vectors in real time. After satisfying a preset convergence threshold, the relaxed optimized structure is obtained. The relaxed optimized structure is then subjected to symmetry analysis using PyMatgen to generate all non-equivalent desodium configurations and independent doping configurations, forming the original candidate doping configuration set. For this original candidate doping configuration set, a three-level screening process is performed sequentially using the Ewald electrostatic energy reverse order screening rule, the interatomic spacing threshold elimination rule, and the equivalent configuration deduplication rule to obtain the screened candidate structure set. Through this multi-level physical / chemical rule screening, a large number of physically unreasonable structures can be eliminated before DFT calculations, significantly reducing the resource consumption of subsequent expensive calculations.

[0067] Step S3: For the selected candidate structure set, automatically generate VASP input files based on the atomic-level tool call sequence and the domain knowledge entries, and construct a multi-step calculation workflow. Submit the multi-step calculation workflow to the calculation server in batches, monitor the calculation convergence status in real time, and adaptively adjust the calculation parameters based on large language model inference to obtain the DFT calculation results of each doped configuration.

[0068] Specifically, the system analyzes the geometric and electronic structure features of each candidate configuration using an agent, retrieves matching physical models and expert experience from the sodium electrochemical material simulation knowledge base using RAG, and automatically generates INCAR, POSCAR, KPOINTS, and POTCAR files to form the VASP input file. A multi-step computational workflow supporting serial, parallel, and conditional branch execution is constructed based on the atomic-level tool call sequence, and this workflow is submitted to the computation server in batches. During computation, the system captures energy convergence curves, atomic force vectors, and charge density residuals in real time, uses a large language model to infer and determine convergence risks, and adaptively adjusts computational parameters based on the determination results to ensure successful computation.

[0069] Step S4: For the DFT calculation results of each doped configuration, extract the total energy of each doped configuration and perform formation energy calculation inference to generate energy convergence curves and stability visualization charts, and obtain the stability ranking results of each doped configuration.

[0070] Specifically, the total energy, atomic force convergence trajectories, and charge density residuals for each doped configuration are extracted from the DFT calculation output file to generate energy-ion step convergence curves. Based on these energy-ion step convergence curves and the total energy, a formation energy calculation inference is performed using a large language model. The thermodynamic spontaneity is determined by comparing the energy difference between the doped system and the pure system, generating the stability visualization chart and obtaining the stability ranking results for each doped configuration. If the formation energy is negative and low, it is inferred that the doped structure has high thermodynamic stability, ultimately providing users with an intuitive ranking of doped structure stability and thermodynamic trend analysis.

[0071] Example 2

[0072] In this embodiment, the natural language command for sodium-ion crystal doping requirements input by the user is retrieved from a preset sodium-ion material simulation knowledge base using a similarity retrieval algorithm. A large language model is then used to perform logical reasoning based on the retrieved domain-specific knowledge entries to generate an atomic-level tool call sequence and an initial state snapshot, including:

[0073] Upon receiving the natural language instruction, the system retrieves the domain-specific knowledge entries from the sodium electrochemical material simulation knowledge base using a cosine similarity algorithm. Then, it extracts doping elements, doping sites, doping concentrations, and parent materials from the natural language instruction using keyword matching and slot filling to obtain structured intent metadata.

[0074] The structured intent metadata is mapped into a structured task blueprint by the task orchestration agent, and decomposed into subtasks of structure analysis, doping configuration generation, electrostatic energy screening, VASP input generation, and result analysis, generating a directed acyclic graph execution plan.

[0075] Based on the directed acyclic graph execution plan, the structure analysis program is invoked to transform the unstructured input data into the standardized crystal structure, and it is determined whether the standardized crystal structure reaches the computational mechanical stability threshold. If it does, the pseudopotential file, cutoff energy, and k-point sampling accuracy are determined based on the domain expertise entries, the initial state snapshot is constructed, and an interpretable execution trajectory is formed using an iterative inference execution mechanism with a large language model, determining the atomic-level tool call sequence and the physical constraint logic between each subtask.

[0076] Specifically, firstly, the sodium electrode material simulation knowledge base is a pre-constructed specialized domain knowledge base for sodium electrode cathode doping calculation tasks. Its content sources include academic literature in the field of sodium electrode cathode doping, expert experience rules for materials calculation, VASP software parameter manuals, and historical calculation case data, stored in vector form. Each knowledge entry corresponds to a piece of professional text with clear physical meaning (such as recommended calculation parameters for a specific material system, physicochemical property constraints for a specific doping element, etc.). Upon receiving a user's natural language command, the command is first semantically encoded using an encoder of a pre-trained language model, extracting its core semantic feature vector. Then, a cosine similarity algorithm is used to calculate the similarity between the semantic feature vector and the vectors of all knowledge entries in the sodium electrode material simulation knowledge base, selecting the entries with the highest similarity as the domain-specific professional knowledge entries K most relevant to the current doping requirements. rag Returning data is used to assist in subsequent large language model inference.

[0077] While completing knowledge retrieval, the intent recognition process is initiated, employing a combination of keyword matching and slot filling to extract structured doping parameters from the natural language command. Slot filling refers to pre-defining several parameter slots strongly related to the sodium electrostatic doping task, including parent material slots (for filling crystal types, such as NaMnO2, Na2Mn3O6, etc.), dopant element slots (for filling foreign dopant elements, such as Mg, Al, Ti, Zn, etc.), doping site slots (for filling the lattice location where doping occurs, such as transition metal TM sites or sodium Na sites), and doping concentration slots (for filling the doping mole fraction or atomic percentage, such as 5%). The intent recognition sub-agent first identifies semantic fragments in the user command corresponding to each slot using keyword matching rules, and then normalizes and maps the identified semantic fragments to the corresponding slots through the slot filling mechanism, forming structured intent metadata. For example, when a user inputs "doping 5% Mg at the transition metal Mn site of NaMnO2", the system automatically fills the parent material slot with "NaMnO2", the dopant element slot with "Mg", the dopant site slot with "TM site (Mn site)", and the dopant concentration slot with "5%". Simultaneously, it establishes an atomic index mapping and supercell conversion mechanism. If the number of Mn atoms in the structure provided by the user is insufficient to meet the 5% doping accuracy, the system automatically infers the required supercell multiple, constructs a supercell structure that meets the doping concentration requirements, and reconstructs the atomic index mapping relationship accordingly to ensure that subsequent doping operations are performed at the correct atomic sites.

[0078] Secondly, after receiving the structured intent metadata, the task orchestration agent maps the overall workflow into a structured task blueprint based on the inherent logical dependencies of the mixed tasks. A directed acyclic graph (DAG) execution plan refers to the subtask dependency topology expressed using a DAG data structure. Each node in the graph represents an independent executable subtask, and each directed edge represents the data dependency between two subtasks (i.e., the output of the predecessor subtask is the input of the successor subtask). There are no directed cycles in the graph, thus ensuring the determinism and deadlock-free nature of the task execution order. In this invention, the directed acyclic graph (DAG) execution plan includes the following five types of subtask nodes and their dependencies: the structure analysis subtask is responsible for converting user-provided unstructured crystal data into standardized crystal structure objects, serving as the data entry point for the entire workflow; the doping configuration generation subtask takes the output of the structure analysis subtask as input to generate all candidate doping configurations; the electrostatic energy screening subtask takes the output of the doping configuration generation subtask as input to perform multi-level physicochemical rule screening; the VASP input generation subtask takes the output of the electrostatic energy screening subtask as input to automatically generate DFT calculation input files for each candidate configuration; and the result analysis subtask takes the output file after VASP calculation as input to perform formation energy inference and visualization. The dependencies between the subtasks are explicitly encoded in the DAG, and the scheduling engine automatically determines the set of subtasks that can be executed in parallel at each time step based on this graph, achieving task-level parallel acceleration.

[0079] Finally, according to the directed acyclic graph execution plan, the structure parsing subtask is triggered first, invoking the preset structure parser to process the user-provided structure input data. The structure input data can be in various unstructured formats, including material database indexes (such as mp-id numbers from the Materials Project database), CIF format structure files, POSCAR format files, structure strings (such as SMILES or crystal structure description strings), or user-uploaded file handles. The structure parser converts the above-mentioned input data in various formats into a standardized crystal structure object D that meets the requirements of the downstream computing engine. std The standardized crystal structure object includes lattice parameters (lattice constants a, b, c and lattice angles α, β, γ), atomic coordinates (fractional coordinates or Cartesian coordinates), a list of elemental composition and symmetry space group information, which can be directly called by mainstream computational materials science libraries such as pymatgen and ASE.

[0080] After the structural analysis is completed, the standardized crystal structure object D is... stdPerform a computational mechanical stability assessment. The computational mechanical stability threshold refers to a set of physical parameter constraints used to determine whether the input crystal structure possesses a reasonable physical state and can be directly used for DFT calculations. Specifically, these include: a lower limit threshold for interatomic spacing (the distance between any two atoms must not be less than a preset proportion of the sum of the covalent bond radii of the corresponding elements to prevent atomic overlap from causing DFT calculation divergence), elemental composition rationality constraints (the types of elements must be within the pseudopotential library supported by VASP), and lattice parameter validity constraints (the lattice constant must be positive and the lattice angles must meet geometric validity conditions). The return status of the structure analysis program and the statistical information of the above physical parameters are obtained in real time, and each item is evaluated for DFT stability. std Check if all computational mechanical stability threshold conditions are met. If all conditions are met, the structure is determined to have computational mechanical stability, and the subsequent initial state snapshot construction process begins; if any condition is not met, the user is informed of the specific structural quality issue, and the user is prompted to correct the input structure and resubmit.

[0081] After determining that the structure possesses computational mechanical stability, based on the retrieved domain knowledge entries K... rag Based on large language model reasoning, three key computational parameters required for the current computational task are identified: the first is the pseudopotential file type, namely the VASP pseudopotential library (PAW). PBE or PAW LDA The selection of pseudopotential files corresponds to the elements of the target system. The choice of pseudopotential file should comprehensively consider the element types, valence electron configuration, and target computational accuracy. Secondly, the cutoff energy, i.e., the kinetic energy cutoff value (in eV) used for plane-wave basis set expansion, should be set with reference to the recommended values ​​in the selected pseudopotential file and appropriately increased according to the system complexity. Thirdly, the k-point sampling accuracy, i.e., the k-point grid density used for Brillouin zone integration, needs to achieve a balance between computational cost and accuracy based on the cell size and target computational accuracy. After determining these three types of parameters, they are compared with the standardized crystal structure object D. std The atomic index mapping relationship and supercell information are encapsulated together to complete the construction of the initial state snapshot of the simulation task. The initial state snapshot serves as a unified physical benchmark for all subsequent computational subtasks, ensuring that the structure and parameters used by each subtask in the entire workflow remain globally consistent.

[0082] The process of determining the atomic-level tool call sequence adopts the ReAct (Reasoning and Acting) iterative reasoning execution mechanism of the large language model. The ReAct mechanism refers to the large language model alternately outputting "Reasoning" (analysis of the current state and inference of the next action) and "Acting" (specific tool call instructions) in each reasoning step, and incorporating the execution result returned by the tool as "Observation" into the reasoning context of the next step, forming a complete iterative closed loop of "Reasoning → Acting → Observation → Rethinking", which makes the reasoning process interpretable and traceable. Based on an initial state snapshot and domain-specific knowledge entries, the large language model uses the ReAct mechanism to sequentially infer the calling order, calling parameters, and physical constraints between each step of all atomic-level tools (including structure analysis tools, geometry optimization tools, symmetry analysis tools, Ewald energy calculation tools, VASP input generation tools, task submission tools, and result extraction tools) required to complete the current doping task. These constraints include the calling order, calling parameters, and physical constraints between each step (e.g., geometry optimization must be completed before Ewald screening, and the selection of POTCAR must strictly correspond to the element order in POSCAR). This ultimately forms a complete and executable atomic-level tool calling sequence L. tools This serves as the operational blueprint for the execution of subsequent subtasks.

[0083] Example 3

[0084] like Figure 2 As shown, in this embodiment, structural optimization is performed on the normalized crystal structure in the initial state snapshot. An initial set of candidate doped configurations is generated through a symmetry analysis program. Multi-level screening is then performed according to the physicochemical rules in the domain expertise entries to obtain a screened set of candidate structures, including:

[0085] Step S21: Based on the atomic-level tool call sequence, the geometry optimization program is invoked to start the initial state structure optimization calculation of the standardized crystal structure, and the energy convergence curve and atomic force vector are captured in real time. After the preset convergence threshold is met, the parallel computing subtask set covering multiple different desodium components is executed in a high-concurrency manner through the concurrent scheduling engine to obtain the relaxed optimized structure.

[0086] Step S22: Perform symmetry analysis on the relaxed optimized structure using PyMatgen to generate all non-equivalent sodium removal configurations and independent doping configurations, forming the original candidate doping configuration set.

[0087] Step S23: For the original candidate doped configuration set, apply the Ewald electrostatic energy reverse order screening rule, the interatomic spacing threshold elimination rule, and the equivalent configuration deduplication rule to perform a three-level screening to obtain the screened candidate structure set.

[0088] Specifically, firstly, the geometry optimization program refers to a structural relaxation calculation program based on quantum mechanical density functional theory (DFT) or classical force fields (in this invention, it mainly refers to the ion relaxation function controlled by the IBRION parameter of the VASP software). Its core objective is to converge the total energy of the system to a local minimum by iteratively adjusting the atomic coordinates (and optional cell shape and volume) under a given atomic composition and unit cell constraints, while simultaneously converging the Hellmann-Feynman force on all atoms to a level close to zero. Based on the initial state snapshot, an INCAR file containing all control parameters required for structural optimization is generated using a large language model (specifying key parameters such as IBRION, ISIF, NSW, EDIFF, and EDIFFG). The INCAR file, along with the POSCAR, KPOINTS, and POTCAR files corresponding to the standardized crystal structure, is then submitted to the computing server via a file upload interface to initiate the initial state structural optimization calculation.

[0089] During computation, the energy convergence curve and atomic force vectors output by the geometry optimization program at the end of each ionic step are captured in real time by polling the computation node status. The energy convergence curve is defined as follows: with the ionic step number as the horizontal axis and the total system energy E of the current ionic step as the vertical axis. tot The energy evolution curve plotted on the vertical axis reflects the gradual decrease in system energy as atomic positions are iteratively adjusted during structural optimization. The atomic force vector refers to the three-dimensional force component (F) experienced by each atom in the system. x , Fᵧ, F z The magnitude of the modulus reflects the degree to which the atom deviates from its mechanical equilibrium position. When the maximum component of the force on all atoms is lower than the preset convergence threshold EDIFFG (usually set to -0.02 eV / Å, with the negative sign indicating the use of force convergence criteria), the structure optimization is considered to have reached convergence.

[0090] After the structural optimization meets the preset convergence threshold, the relaxed atomic coordinates and unit cell parameters are read from the CONTCAR file output by VASP, serving as the high-precision reference structure for subsequent sodium removal configuration generation. Then, based on domain-specific knowledge item K... rag The professional specifications for sodium removal simulation protocols in China, and the large language model inference generation covering various sodium removal components x (Na) {1-x} The parallel computing subtask set {T1,T2, ...,T2} in MO2 (where x can take values ​​such as x = 0, 1 / 6, 1 / 3, 1 / 2, 2 / 3, 5 / 6, 1, etc.) nThe concurrent scheduling engine refers to the scheduling component in the system workflow orchestration layer responsible for parallel task allocation and resource management. Its core function is to maintain a pool of available computing resources (including the number of CPU cores, memory quota, and queue status information of the computing server). When multiple subtasks are simultaneously executable, the scheduling engine dynamically allocates computing threads based on the current resource availability and task priority, ensuring that the structural optimization tasks of each desodium component achieve high-concurrency execution in a logically decoupled state, and obtaining the relaxed optimized structure set corresponding to each desodium component.

[0091] Secondly, symmetry analysis refers to using the SpacegroupAnalyzer module in the pymatgen library to perform crystal symmetry analysis on the relaxed optimized structure, identifying equivalent Wyckoff sites (i.e., sets of atomic positions that are mutually equivalent under the action of crystal symmetry operations). Regarding the generation of desodium configurations, based on the relaxed optimized structure, all possible combinations of desodium atoms are enumerated, and symmetry analysis is used to identify and merge crystallographically equivalent (i.e., mutually convertible through crystal symmetry operations) desodium configurations, retaining only non-equivalent independent desodium configurations. This significantly compresses the candidate set size while ensuring the completeness of the structure space. Non-equivalent desodium configurations refer to the set of desodium configurations that are not mutually convertible under the action of crystal symmetry operations; each non-equivalent desodium configuration represents a physically independent and irreducible structural variant. In terms of doping configuration generation, using the relaxed optimized structure as the parent structure, and based on the doping element, doping site, and doping concentration information in the structured intent metadata, all possible doping atom substitution schemes are enumerated. Similarly, symmetry analysis is used to generate all crystallographically independent doping configurations, i.e., the independent doping configurations. All the aforementioned non-equivalent sodium removal configurations and independent doping configurations together constitute the original candidate doping configuration set S. raw This provides a complete and non-redundant set of candidate structures in a crystallographic sense for subsequent multi-level screening.

[0092] Finally, regarding the original candidate doping configuration set S raw Perform the following three levels of filtering in sequence:

[0093] The first-level screening uses the Ewald electrostatic energy reverse order screening rule. The Ewald electrostatic energy refers to the total energy of electrostatic interactions between all ions in a crystal structure, calculated using the Ewald summation method (a classical electrostatic energy calculation method that decomposes long-range Coulomb interactions into the sum of short-range terms in real space and long-range terms in reciprocal space). Its physical significance lies in approximating the electrostatic rationality of the ion arrangement in the structure—the lower the Ewald electrostatic energy, the more favorable the spatial distribution of positive and negative ions in the structure is for electrostatic stability, and the corresponding configuration is relatively more thermodynamically stable. The EwaldSummation module of pymatgen is called to screen S...raw The Ewald electrostatic energy of each candidate configuration is calculated, and the configurations are sorted from low to high according to their Ewald electrostatic energy. The top-K1 candidate configurations are retained, and the structures with excessively high Ewald electrostatic energy (i.e., significant electrostatic instability) are eliminated. This completes the first stage of rapid coarse screening, and the overall execution time is in the second range.

[0094] The second-level screening applies an interatomic spacing threshold elimination rule. For each configuration in the candidate configuration set that has passed the first-level screening, the Euclidean distance between all atomic pairs is calculated. Each configuration is checked to see if the distance between any two atoms is less than the threshold obtained by multiplying the sum of the covalent bond radii of the corresponding element pairs by a preset safety factor (default 0.75). If such a case exists, the configuration is determined to have an anomalous interatomic spacing, making it physically unreasonable (atoms being too close together will cause extremely high repulsive potential energy in the DFT calculation, resulting in computational divergence), and is therefore eliminated. This level of screening directly eliminates geometrically singular configurations that may arise during structure optimization or doping, further reducing the size of the candidate set.

[0095] The third-level screening applies an equivalent configuration deduplication rule. Although pymatgen has already merged obviously equivalent configurations during the symmetry analysis stage, after the first two levels of screening, there may still be equivalent configurations that have not been completely deduplicated due to differences in numerical precision. The candidate configuration set after the first two levels of screening is then subjected to an equivalence determination based on structural fingerprints: the structural fingerprint of each configuration (including a pairwise distribution function or a symmetry-based hash value) is calculated. Configuration pairs with identical fingerprints or similarity exceeding a preset threshold are determined to be equivalent configurations. Only the one with the lowest energy is retained, and the remaining duplicate configurations are eliminated, ensuring that the final selected candidate structure set S is obtained. filtered Each configuration is physically independent and geometrically non-redundant, providing high-quality candidate structure inputs for subsequent batch VASP accurate DFT calculations.

[0096] Example 4

[0097] In this embodiment, for the selected candidate structure set, a VASP input file is automatically generated based on the atomic-level tool call sequence and the domain expertise entries, and a multi-step computation workflow is constructed. The multi-step computation workflow is submitted to the computation server in batches, the computation convergence status is monitored in real time, and the computation parameters are adaptively adjusted based on large language model inference to obtain the DFT calculation results for each doped configuration, including:

[0098] Based on the selected candidate structure set, the geometric and electronic structure features of each candidate configuration are analyzed by the Agent. The matching physical models and expert experience are retrieved from the sodium electrochemical material simulation knowledge base by the RAG, and INCAR, POSCAR, KPOINTS and POTCAR files are automatically generated to form the VASP input file. A multi-step calculation workflow supporting serial, parallel and conditional branch execution is constructed according to the atomic-level tool call sequence, and the multi-step calculation workflow is submitted to the computing server in batches.

[0099] Based on the computing status of the computing server, the energy convergence curve, atomic force vector and charge density residual are captured in real time, and the convergence risk is determined by reasoning using a large language model. If it is determined that there is a local structural collapse anomaly, the subsequent DFT calculation of the corresponding candidate configuration is terminated. If it is determined that there is convergence difficulty, the calculation parameters in the INCAR file are automatically corrected and the self-consistent calculation request is reissued to obtain the DFT calculation results of each doped configuration.

[0100] Specifically, firstly, for the screened candidate structure set S filtered For each candidate configuration, the Agent sequentially analyzes its geometric and electronic structural features. The geometric features include lattice parameters (lattice constant and lattice angle), atomic coordinates, elemental composition, coordination environment (the types and distance distribution of neighboring atoms for each atom), and the degree of local structural distortion (quantified by comparing changes in bond length and bond angle before and after doping). The electronic structural features include magnetic configuration requirements (determining whether spin polarization calculation (ISPIN=2) needs to be enabled based on the types of transition metal elements in the system), strong correlation effect requirements (determining whether DFT+U correction needs to be introduced to handle the strong correlation effect of transition metal d orbitals), and predictions of the system's metallicity or insulatingness (affecting the selection of the Brillouin zone k-point sampling density and broadening method).

[0101] After feature analysis, the RAG is used to retrieve the physical model and expert experience entries that best match the elemental composition, structural type, and computational requirements of the current candidate configuration from the sodium electrochemical materials simulation knowledge base. Combined with large language model reasoning, the following four types of VASP input files are automatically determined and generated: The INCAR file is the core computational parameter control file for VASP, containing key parameters such as the exchange-correlation functional type (e.g., GGA-PBE), DFT+U parameters (Hubbard U value, for d orbitals of transition metals such as Mn, Fe, Ni, and Co), cutoff energy (ENCUT), self-consistent field convergence accuracy (EDIFF), ion relaxation convergence threshold (EDIFFG), aliasing algorithm (ALGO, a convergence acceleration scheme for controlling the self-consistent iteration of electron density), broadening method, and broadening width (ISMEAR and SIGMA, used for electronic state occupancy broadening processing for Brillouin zone integration); the POSCAR file is the crystal structure input file for VASP, containing the lattice matrix of the candidate configuration, a list of element types, and the atomic values ​​of each atom. Fractional coordinates are directly generated from the standardized crystal structure object; the KPOINTS file is the VASP k-point grid setting file, which automatically generates a k-point grid of appropriate density using the Monkhorst-Pack scheme based on the cell size and target calculation accuracy; the POTCAR file is the VASP pseudopotential file, which automatically assembles the pseudopotential files corresponding to each element from the VASP pseudopotential library based on the element type list in the POSCAR and the pseudopotential type determined in step S1.3, ensuring that the element order in the POTCAR is strictly consistent with that in the POSCAR. These four types of files together constitute a complete VASP input file set for each candidate configuration, providing all the necessary input data for subsequent DFT calculations.

[0102] After generating the VASP input files for all candidate configurations, the atomic-level tool call sequence L is followed. toolsA multi-step computational workflow is constructed. This workflow refers to a sequence of multi-stage DFT computation tasks arranged in the form of a directed acyclic graph, covering the entire process from structural optimization to formation capability computation. The workflow supports three execution modes: a serial execution mode for handling computation steps with strict dependencies (e.g., performing structural optimization on the same candidate configuration first, then performing static self-consistent computation using the optimized CONTCAR as the starting structure; file dependencies are automatically managed by the system); a parallel execution mode for handling independent candidate configuration computation tasks (DFT computations of multiple candidate configurations are independent and can be submitted to the computation server simultaneously for parallel execution, fully utilizing the multi-node computing capabilities of the high-performance computing cluster); and a conditional branch execution mode for dynamically determining subsequent computation paths based on intermediate computation results (e.g., if a candidate configuration is determined to have partial structural collapse during the structural optimization stage, all subsequent computation steps for that configuration are automatically skipped, and computational resources are allocated to other candidate configurations). The multi-step computational workflow is submitted to the computation server in batches through a job scheduling system (such as SLURM or PBS) to achieve fully automated parallel computation management of a large number of candidate configurations.

[0103] Second, during the execution of batch DFT calculation tasks, the running status of each computing node on the computing server is continuously monitored through a polling mechanism. The following three types of streaming physics observation data are captured in real time and fed back to the context window of the large language model: the energy convergence curve reflects the trend of the total energy evolution of the system in each ion step, which is used to determine whether the structural relaxation has converged normally; the atomic force vector reflects the magnitude and direction of the Herman-Feynman force on each atom in each ion step, which is used to determine whether the structure is close to mechanical equilibrium; the charge density residual refers to the norm of the difference between the charge densities of two adjacent iterations in the self-consistent field (SCF) iteration process, which reflects whether the electron density distribution has reached self-consistent convergence and is the core indicator for judging whether the DFT self-consistent calculation is proceeding normally.

[0104] Convergence risk assessment is one of the core intelligent reasoning steps of this invention. The large language model performs multi-dimensional evolutionary analysis on the above three types of real-time physical observation data, focusing on identifying the following two types of abnormal patterns: First, structural local collapse anomalies, specifically manifested as a sudden energy increase in the energy convergence curve at a certain ion step (the energy change relative to the previous step exceeds the preset mutation threshold), accompanied by an abnormal increase in the magnitude of the force vector of some atoms. The physical cause is usually excessive local lattice distortion introduced by doping atoms, resulting in some atom pairs being too close, generating extremely strong repulsive potential energy, causing the structure to disintegrate locally during relaxation. Second, difficulty in self-consistent convergence, specifically manifested as the SCF iteration step exceeding the preset upper limit (the maximum number of self-consistent iterations controlled by the NELM parameter) still failing to reach the charge density residual convergence criterion EDIFF, that is, the charge density residual oscillates rather than monotonically decreases during the iteration process. The physical causes usually include the existence of electronic state degeneracy in the system (partial occupied states near the Fermi surface in the metallic system cause numerical instability), improper value of the broadening width SIGMA, or inappropriate selection of the aliasing algorithm ALGO, etc.

[0105] For the two types of anomalies mentioned above, the following adaptive handling strategies are adopted respectively: For candidate configurations that are determined to be structural partial collapse anomalies, a termination command is immediately sent to the job scheduling system to stop all subsequent computing tasks of the candidate configuration on the computing server, and the resources of the corresponding computing nodes are released and reallocated to other normally operating candidate configurations to avoid consuming valuable computing resources on physically failed computing tasks. For candidate configurations deemed difficult to converge self-consistently, an adaptive parameter correction mechanism is triggered: the large language model infers the most likely effective parameter correction scheme based on the specific performance of the current SCF oscillations (such as the convergence trend of oscillation amplitude, oscillation period, and charge density residual). For example, the aliasing algorithm ALGO is switched from the default "Fast" to the more robust "All" (using the full-step Davidson iteration scheme), or the broadening width SIGMA is appropriately reduced from its original value (e.g., from 0.2 eV to 0.05 eV) to reduce the occupation contribution of spurious states near the Fermi surface. Accordingly, the corresponding parameters in the INCAR file are automatically corrected, and the self-consistent calculation request is reissued through the job scheduling system to restart the SCF calculation of the candidate configuration, achieving autonomous closed-loop optimization of the calculation path. After the above real-time monitoring and adaptive adjustment, the VASP output files of all successfully converged candidate configurations are finally collected to obtain the DFT calculation results of each doped configuration, providing a complete first-principles calculation data foundation for subsequent energy calculation inference and stability ranking.

[0106] Example 5

[0107] In this embodiment, the total energy of each doped configuration is extracted from the DFT calculation results for each doped configuration, and formation energy calculation inference is performed to generate energy convergence curves and stability visualization charts, thus obtaining the stability ranking results of each doped configuration, including:

[0108] Based on the DFT calculation results of each doped configuration, the total energy, atomic force convergence trajectory and charge density residual of each doped configuration are extracted from the DFT calculation output file to generate the energy-ion step convergence curve.

[0109] Based on the energy-ion step convergence curve and the total energy, the formation energy calculation reasoning is performed using a large language model. The thermodynamic spontaneity is determined by comparing the energy difference between the doped system and the pure system, and the stability visualization chart is generated to obtain the stability ranking results of each doped configuration.

[0110] Specifically, firstly, after all batch DFT calculation tasks are completed or the preset convergence criteria are met, the data integration program is invoked to perform systematic data extraction on the VASP output files in the corresponding calculation node directories of each candidate configuration on the computing server. After the VASP calculation is completed, several key output files will be generated in the working directory. Among them, the OUTCAR file is a log file containing the most complete information of the entire calculation process, recording detailed data such as energy, force, and electron occupation for each ion step and each self-consistent field iteration step; the OSZICAR file is a dedicated convergence monitoring file that records the total energy of each ion step and the charge density residual of each SCF iteration step in a compact format; and the CONTCAR file stores the final optimized atomic coordinates and cell parameters after the structure relaxation is completed, serving as the starting structure input for subsequent static calculations.

[0111] The following three types of key physical quantities can be extracted from the above output file: the first is the total energy E of the system for each doped configuration. tot The first is the total electronic energy of the system output by VASP after structural relaxation convergence, expressed in eV, which is the core data input for subsequent formation energy calculation and inference; the second is the atomic force convergence trajectory, namely the three-dimensional component (F) of the Herman-Feynman force on each atom extracted step-by-step from the OUTCAR file. x , Fᵧ, F z The first is the charge density residual, which is the charge density residual value extracted from each SCF iteration step in the OSZICAR file. It reflects the dynamic evolution trajectory of each atom from its initial position to the mechanical equilibrium state during the structural relaxation process. By observing the change trend of the maximum atom force with the ion step, the convergence quality of the relaxation process and the mechanical stability of the structure can be intuitively judged. The second is the charge density residual, which is the charge density residual value extracted from each SCF iteration step in the OSZICAR file. It reflects the convergence speed and quality of the electron density distribution approaching the self-consistent solution during the self-consistent field iteration process. The monotonic decay curve of the charge density residual indicates that the SCF convergence is normal, while oscillation or stagnation indicates that numerical convergence is difficult.

[0112] After extracting the three types of physical quantities mentioned above, a visualization program is used to process and plot the energy and force data, automatically generating energy-ion step convergence curves. These curves are energy evolution curves plotted with the ion relaxation step number on the horizontal axis and the total system energy of the current ion step on the vertical axis, visually presenting the energy decrease trajectory of each doped candidate configuration during structural optimization. A normally converged energy-ion step curve should exhibit a monotonically decreasing trend and eventually stabilize. If the curve shows a sudden energy increase or fails to converge for a long period at a certain step, the computational quality of the corresponding candidate configuration is questionable, and it will be marked and excluded in subsequent analysis. The self-consistent field charge density residual convergence curves of each candidate configuration are simultaneously output, forming a complete computational quality assessment atlas together with the energy-ion step convergence curves, providing a reliable energy data foundation for subsequent energy calculation inference.

[0113] Secondly, after the energy data extraction is completed, the total energy data of all candidate configurations, the total energy data of the corresponding pure parent structure, and the reference state energy data of each dopant element (i.e., the total DFT energy of the dopant element in its most stable elemental or compound form, which is usually pre-stored in the sodium electrical material simulation knowledge base as a standard reference value) are input into the context window of the large language model to trigger the formation energy calculation reasoning process.

[0114] Formation energy calculation inference refers to the process by which a large language model automatically calculates the formation energy of each doped configuration using a structured reasoning approach based on the physical definition of thermodynamic formation energy. The physical meaning of formation energy is the energy released or absorbed when forming a target doped compound from the reference states of each constituent element under standard thermodynamic conditions. A negative formation energy indicates that the doping process is thermodynamically spontaneous, and the doped structure is relatively stable; the lower the negative value of the formation energy (the larger the absolute value), the higher the thermodynamic stability. A positive formation energy indicates that the doping is thermodynamically unfavorable, and the structure is relatively unstable. When performing formation energy calculation inference, the large language model first verifies the consistency between the atomic composition and total energy data of each candidate configuration, confirming that all necessary reference state energy data are ready. Then, following the standard thermodynamic path for formation energy calculation, it performs energy difference calculations on each doped candidate configuration one by one to obtain the formation energy E of each candidate configuration. form Finally, the formation energies of each candidate configuration are sorted from low to high to obtain a list of candidate configurations with high to low thermodynamic stability, which is the stability ranking result of each doped configuration.

[0115] After completing the formation energy calculation inference, a visualization program is invoked to automatically generate stability visualization charts. These charts include: a bar chart comparing the formation energies of different doping configurations (with the candidate configuration identifier on the horizontal axis and the formation energy value on the vertical axis, visually demonstrating the differences in thermodynamic stability between different doping schemes); a ranking list of doped structure stability (listing all candidate configurations and their key physical quantities, including formation energy, total energy, lattice distortion rate, etc., from low to high formation energy); and a summary chart of energy-ion step convergence curves (plotting the convergence curves of all candidate configurations in the same chart, facilitating horizontal comparison of the computational convergence quality of each candidate configuration). These multimodal visualization charts are then automatically generated by a large language model, combining the chart content to create a richly illustrated analysis report. The report includes a clear identification of the thermodynamically most stable doping configuration, a trend summary of the effects of different doping elements or concentrations, and preliminary suggestions for subsequent experimental verification directions, providing materials researchers with an intuitive and efficient basis for doped structure screening decisions.

[0116] Example 6

[0117] In this embodiment, a self-evolving knowledge graph update step is also included, comprising:

[0118] Based on the ontological concepts and domain relationships in the field of sodium electrode cathode doping, the initial knowledge graph is constructed by automatically extracting entity and relation triples from the sodium electrode material simulation knowledge base, with parent material, dopant element, doping site, phase structure, calculation parameters and physical quantity indicators as node types, and doping initiation relationship, stabilization effect relationship, recommended calculation parameter relationship and calculation success or failure relationship as edge types.

[0119] Based on the initial knowledge graph and the DFT calculation results of each doped configuration, whenever any DFT calculation subtask is completed, computational feature information is automatically extracted from the DFT calculation output file. The computational feature information is written into the initial knowledge graph in the form of standardized triples, and entity alignment and conflict resolution are performed to obtain the merged knowledge graph. If the calculation fails to converge, the failure reason label and corresponding calculation parameter snapshot are recorded to form a failure experience node and written into the merged knowledge graph to obtain the updated knowledge graph. If the calculation converges successfully, the merged knowledge graph is used as the updated knowledge graph.

[0120] Based on the updated knowledge graph, when a new doping task request is received, the system retrieves the set of completed computation cases and the failed experience nodes that match the current doping task. The distribution of computational parameters of successful cases in the set of completed computation cases is used as a strong prior for initializing the computational parameters of the current task. Computational parameters that overlap with the failure modes recorded in the failed experience nodes are skipped or corrected to obtain task planning results optimized based on historical experience. The task planning results are then written into the atomic-level tool call sequence and the domain knowledge entries are updated.

[0121] Specifically, firstly, the self-evolving knowledge graph is a dynamic domain knowledge representation structure constructed by this invention for sodium electrode cathode doping computation tasks. Its core design concept lies in automatically converting the execution result of each DFT computation task—whether successful or not—into structured knowledge accumulated in the graph. This allows the system to continuously expand its domain knowledge during the ongoing computation tasks, achieving autonomous enhancement of task planning capabilities. Compared to static knowledge bases, the self-evolving knowledge graph possesses three core capabilities: dynamic updating, entity association, and experience reuse. It can maintain long-term memory and effective retrieval of computational experience across tasks.

[0122] In the initial knowledge graph construction phase, based on key ontological concepts in the field of sodium-ion cathode doping, a node type system and an edge type system are defined for the knowledge graph. The node type system includes six types of nodes: parent material nodes store the identification information of the parent structure of sodium-ion cathodes (such as NaMnO2, Na2Mn3O6, Na2FeMnO4, etc.) and its crystal structure parameters (phase structure type, space group, lattice constant, etc.); dopant element nodes store the physicochemical properties of foreign dopant elements (ionic radius, electronegativity, common valence, etc.); doping site nodes store the lattice position information of doping (TM site, Na site and its Wyckoff symbol); phase structure nodes store common phase structure types of sodium-ion cathodes (P2 type, O3 type, P3 type, etc.) and their structural stability ranges; computational parameter nodes store the complete parameter combination used in VASP calculations (functional type, cutoff energy, K-point grid density, ALGO selection, SIGMA value, etc.); and physical quantity index nodes store the numerical values ​​of key physical quantities obtained from calculations (formation energy, total energy, lattice distortion rate, magnetic moment, etc.). The edge type system includes four types of directed relation edges: "Doping-Induced" relation edges connect dopant element nodes and parent material nodes, representing the type of structural change induced by the introduction of a specific dopant element into a specific parent material; "Stabilization Effect" relation edges connect doping configuration and physical quantity index nodes, storing the direction and magnitude of the influence of doping operations on the thermodynamic stability of the structure; "Recommended Calculation Parameters" relation edges connect parent material nodes (or doping configuration nodes) and calculation parameter nodes, storing the optimal combination of calculation parameters for the corresponding system; and "Calculation Success / Failure" relation edges connect calculation parameter nodes and execution result nodes, recording the execution result and failure reason of a specific parameter combination on a specific system. An entity extraction program is called from the preset sodium electrochemical material simulation knowledge base to automatically identify and extract entities and relation triples (subject-relation-object) that conform to the above node type and edge type definitions, filling them into the above ontology model to complete the construction of the initial knowledge graph G0, serving as the starting point for self-evolutionary updates.

[0123] Second, whenever any DFT calculation subtask (structural optimization subtask, static self-consistent calculation subtask, or formation energy calculation subtask) is completed, the asynchronous update process of the knowledge graph is immediately triggered. The calculation feature information of this calculation is extracted from the VASP output file of the corresponding calculation node, including: parent material identifier, dopant element and site combination, doping concentration, complete combination of calculation parameters used (functional / cutoff energy / K-point / ALGO / SIGMA, etc.), convergence status flag (successful convergence or failure and failure type), and the final total energy of the system E. tot The calculated formation energy E form(If the formation energy calculation subtask has been completed), lattice distortion rate (calculated by comparing lattice parameters before and after relaxation), and iteration step statistics (ion relaxation steps and SCF iteration steps, reflecting computational efficiency). The above computational features are encapsulated into a standardized triplet set ΔG according to the predefined node and edge type patterns of the knowledge graph, constituting the knowledge update package generated by this computational task.

[0124] Entity alignment refers to checking whether the entities involved in ΔG (such as specific parent materials, specific dopant elements, etc.) already have corresponding nodes in the existing knowledge graph when writing the knowledge update package ΔG into the knowledge graph. If they already exist, the new information is merged into the existing node instead of creating a duplicate node, ensuring that the same entity is not registered repeatedly in the knowledge graph. Conflict resolution refers to the process where, when the attribute values ​​of new calculation results are inconsistent with those of existing nodes in the knowledge graph (such as slight differences in formation energy values ​​obtained from different calculation tasks of the same system), the conflicting values ​​are weighted based on the calculation accuracy indicators (such as cutoff energy, K-point density, and other quantitative accuracy indicators). The calculation result with higher accuracy is retained as the authoritative value, while the lower-accuracy result is retained as a reference record, preventing low-quality data from overwriting high-quality data, thus completing the construction of the merged knowledge graph.

[0125] If the current DFT calculation subtask fails to converge, an additional "failure experience node" is created in the knowledge graph. This node contains a failure reason label (labeled according to a predefined failure type classification system, such as "partial structural collapse," "SCF charge density oscillation," "POTCAR element order error," "memory overflow," etc.) and a snapshot of the calculation parameters at that time (completely recording all parameter values ​​in the INCAR file when the failure was triggered). The failure experience node is then connected to the corresponding parent material node, doping configuration node, and calculation parameter node via "computation success / failure" relationship edges. Finally, this information is written into the merged knowledge graph, resulting in an updated knowledge graph G containing the failure experience. {t+1} If the calculation converges successfully, the merged knowledge graph will be used directly as the updated knowledge graph, without the need to create failed experience nodes.

[0126] Third, upon receiving a new doping task request, the intent recognition agent, after extracting structured intent metadata, simultaneously triggers a historical experience retrieval process based on the updated knowledge graph. Through the graph query interface, using the parent material type, doping element list, and doping site of the current task as primary keys, it retrieves a set of completed computational cases C from the updated knowledge graph that have the highest structural similarity to the current task. hist (Sorted by a weighted average of parent material consistency, dopant element overlap, and dopant site matching), and the set of failure experience nodes F that are similar to the current dopant combination in elemental composition or concentration range.hist .

[0127] If C hist Not empty, extract C hist The statistical information on the distribution of computational parameters from all successful cases (most frequent cutoff energy value, optimal K-point density, most commonly used ALGO selection, optimal SIGMA value, etc.) is used as a strong prior for initializing the INCAR parameters of the current task. When generating the INCAR file for the current task through large language model inference, historically successful parameter combinations are used as default recommended values, and deviations are only allowed when there is sufficient physical justification. If a parameter combination in the current task planning (such as a specific ALGO selection and a specific SIGMA combination) is detected to be inconsistent with F... hist If the parameter snapshots of a certain failed experience node recorded in the database highly overlap (the overlap exceeds a preset threshold), the agent proactively intervenes to perform preventative corrections on that parameter combination—for example, automatically replacing ALGO=Fast with ALGO=All in failed cases that cause SCF oscillations, or lowering the high initial SIGMA value that leads to structural collapse—avoiding historical failure paths during the task planning phase and obtaining task planning results optimized based on historical experience. The task planning results are then written into an atomic-level tool invocation sequence L in the form of tool invocation instructions. tools Furthermore, new domain knowledge entries generated during this retrieval and reasoning process (such as "For P2-type systems with Mn-Mg co-doping, it is recommended to use ALGO=All and SIGMA=0.05 to ensure stable convergence of SCF") are updated to the domain knowledge entries in the sodium electrochemical materials simulation knowledge base, thereby realizing the continuous autonomous enhancement of system planning and reasoning capabilities with the accumulation of computational experience.

[0128] Example 7

[0129] In this embodiment, after obtaining the filtered candidate structure set and before submitting the multi-step computation workflow to the computation server in batches, a multi-fidelity filtering funnel execution step is further included, comprising:

[0130] Based on the aforementioned domain expertise items, a base model is loaded from a pre-trained general interatomic potential model library. The root mean square error between the machine learning interatomic potential prediction of the base model and the DFT reference value is verified for the target parent material system. If the root mean square error is lower than a preset error threshold, the base model is used as the system-specific machine learning interatomic potential model. Otherwise, the base model is transferred and fine-tuned using the pre-acquired DFT reference calculation results of the target parent material system as the fine-tuning dataset to obtain the system-specific machine learning interatomic potential model.

[0131] Based on the selected candidate structure set, the system-specific machine learning interatomic potential model is invoked to perform fast structure relaxation and energy sorting on each candidate structure, eliminating structures with abnormal interatomic spacing or energy exceeding a preset energy window threshold, thereby obtaining a machine learning interatomic potential fine-screening candidate structure set. Each candidate structure in the machine learning interatomic potential fine-screening candidate structure set carries a machine learning relaxed structure after machine learning interatomic potential relaxation.

[0132] Based on the machine learning interatomic potential screening candidate structure set, the machine learning relaxation structure corresponding to each candidate structure is used as the starting structure for DFT calculation to submit accurate DFT calculation. The accurate formation energy is calculated back from the total DFT energy to obtain the stability ranking of the doped structure. The stability ranking of the doped structure is then input into the multi-step calculation workflow to prioritize the scheduling of a preset number of candidate structures with the highest stability ranking.

[0133] Specifically, firstly, the multi-fidelity screening funnel is an intermediate screening mechanism introduced in this invention between the physical / chemical rule coarse screening (the three-level screening in Example 3) and the high-precision DFT calculation. Its design goal is to utilize the characteristics of Machine Learning Interatomic Potential (MLIP) which is far faster than DFT (2 to 3 orders of magnitude faster) and significantly more accurate than pure classical force fields. It performs fast energy sorting and quality assessment on the numerous candidate structures that still exist after the physical / chemical rule coarse screening, further reducing the number of candidate structures that need to be submitted for precise DFT calculation. This forms a three-level multi-fidelity calculation funnel of "physical rule coarse screening (seconds) → machine learning interatomic potential intermediate screening (minutes) → DFT calculation (hours)," achieving optimal allocation of computing resources while ensuring the completeness of the screening.

[0134] The pre-trained general interatomic potential model library refers to a collection of general machine learning interatomic potential models pre-trained on large-scale first-principles datasets, covering most elements of the periodic table. Models used in this invention include CHGNet (based on a graph neural network architecture, pre-trained on hundreds of thousands of DFT structures in the Materials Project database) and MACE-MP-0 (based on a message-passing atom cluster extension architecture, with higher expressive power and generalization). The base model refers to the original pre-trained model loaded from the aforementioned general interatomic potential model library that has not yet been specifically adapted for the current target system. While it has a wide elemental coverage, its prediction accuracy for specific systems (such as layered oxide sodium cathode systems containing Na-Mn-O) may be insufficient. Based on domain-specific knowledge item K... ragInformation about the elemental composition of the target parent material system is used to select a base model that covers commonly used elements in sodium cathodes such as Na, Mn, Fe, Ni, Co, Ti, Mg, Al, Zn, and Cu. The base model is then encapsulated into a standardized structural relaxation tool through the ASE (Atomic Simulation Environment) unified interface, allowing it to directly receive pymatgen standardized crystal structure objects as input.

[0135] For the target parent material system, 20 to 50 representative reference configurations (covering different desodium components and different doping levels) were sampled from the historical calculation dataset of the sodium electrochemical material simulation knowledge base. The energy and stress of these reference configurations were independently predicted using the aforementioned base model and existing VASP high-precision DFT calculations. The root mean square error (RMSE) between the base model prediction value and the DFT reference value was calculated, and the energy root mean square error (RMSE) was obtained. E (Unit: meV / atom) and root mean square error of force (RMSE) F (Unit: meV / Å). If RMSE E If the error is below the preset error threshold (default 5 meV / atom), it is determined that the base model has directly usable prediction accuracy for the current target parent material system, and it can be directly used as a system-specific machine learning inter-atom potential model for subsequent candidate structure screening without additional training; if RMSE E If the error exceeds the preset error threshold, it is determined that the base model has insufficient systematic accuracy in the current system and migration fine-tuning needs to be performed.

[0136] Transfer learning fine-tuning refers to using a pre-acquired DFT reference calculation result (including structure, energy, and force ternary data pairs) of the target parent material system as the fine-tuning dataset. It employs a transfer learning strategy that freezes low-level parameters (preserving the weights of the underlying network responsible for element embedding and basic geometric feature extraction in the base model) and only updates high-level interaction parameters (adjusting only the weights of the upper-level network responsible for energy prediction) to fine-tune the base model using mini-batch gradient descent. The fine-tuning dataset typically consists of 100 to 500 system-specific DFT calculation structures, saving over 95% of data and computational costs compared to training from scratch. After fine-tuning, the RMSE of the fine-tuned model is verified on an independent test set. E The target is to keep the potential within 2 meV / atom. Once the target is met, it will be used as a dedicated machine learning interatomic potential model for subsequent screening.

[0137] Secondly, for the selected candidate structure set S filteredFor each candidate structure, the dedicated machine learning interatomic potential model for the system is invoked via the ASE interface. Starting from the initial coordinates of each candidate structure, fast structural relaxation is performed. The number of relaxation steps is set to 10 to 50 ionic steps (far fewer than the 100 to 500 steps typically required for DFT structure optimization). The relaxation calculation time for a single structure is on the order of seconds to minutes. In each ionic step, the machine learning interatomic potential model replaces DFT self-consistent field calculation with neural network forward inference, directly predicting the system energy and atomic forces from atomic coordinates and elemental information. The atomic coordinates are updated through conjugate gradients or the FIRE (Fast Inertial Relaxation Engine) algorithm, gradually approaching the local energy minimum, to obtain the structure relaxed by machine learning interatomic potential (the machine learning relaxed structure) and the corresponding total machine learning interatomic potential energy E. MLIP .

[0138] Energy sorting and quality screening were then performed: based on E MLIP For S filtered All candidate structures are sorted by energy from low to high, and the machine learning interatomic potential formation energy E of each candidate structure relative to the lowest energy reference structure is calculated. form_MLIP The following two types of quality checks are performed to determine the rejection criteria: First, an atomic spacing anomaly check, which examines whether, after machine learning atomic potential relaxation, any two atomic spacings still exist that are less than the sum of the covalent bond radii of the corresponding element pairs multiplied by a preset safety factor (default 0.75). If such a case exists, the structure is determined to still possess geometric singularity after machine learning atomic potential relaxation and is therefore rejected. Second, an energy window threshold check, which checks whether the energy window threshold of a candidate structure is less than the sum of the covalent bond radii of the corresponding element pairs multiplied by a preset safety factor (default 0.75). form_MLIP The energy required to form a candidate structure exceeds the minimum energy level plus a preset energy window threshold ΔE. window If the energy level is 0.5 eV / atom (default), the thermodynamic stability of the structure is determined to be significantly lower than that of the optimal candidate structure, and it is eliminated. Through the above two types of checks, candidate structures with high energy ranking (usually 10% to 20% of the total number of candidate structures after the three-level screening in Example 3) and no geometric singularities are retained and stored together with their corresponding machine learning relaxation structures, resulting in the machine learning interatomic potential fine screening candidate structure set S. MLIP This provides a high-quality initial structural input for subsequent accurate DFT calculations.

[0139] Finally, the candidate structure set S is finely screened using machine learning interatomic potential. MLIPThe candidate structures are used, with the machine learning relaxation structures (rather than the original initial configurations) corresponding to each candidate structure serving as the POSCAR input files for VASP precise DFT calculations. Following the batch VASP workflow generation and submission process in Example 4, corresponding INCAR, KPOINTS, and POTCAR files are automatically generated for each refined candidate structure and submitted to the computing server for high-precision DFT structure optimization and static self-consistent calculation. The core advantage of using machine learning relaxation structures as the starting structures for DFT is that machine learning interatomic potential relaxation has eliminated the main geometric distortions in the structure, reducing the number of ion relaxation steps required for DFT starting from the starting structure by 50% to 80%. This significantly shortens the computation time of a single DFT task and also reduces the risk of DFT calculation divergence due to poor initial structure quality.

[0140] After the DFT calculation is completed, the high-precision total DFT energy E is extracted from the VASP output files of each candidate structure. DFT By combining the total DFT energy of each reference state element pre-stored in the sodium electrochemical material simulation knowledge base, the precise formation energy E of each finely screened candidate structure is calculated. form_DFT , for S MLIP Candidate structures in E form_DFT The doped structures are sorted from low to high to obtain a ranking of their stability based on first-principles precise energies. This stability ranking is then input into the task scheduling queue of the multi-step computation workflow as a priority weight. This instructs the scheduling engine to prioritize scheduling candidate structures with the highest stability ranking (default 5 to 10) in subsequent fine-grained computation stages (such as electronic structure analysis, density of states calculation, differential charge density calculation, etc.). Candidate structures with lower rankings are deferred or skipped from subsequent fine-grained computations. This allows for the focused analysis of the thermodynamically most stable doping schemes with minimal computational budget, automatically focusing computational resources on the most valuable candidate structures.

[0141] Example 8

[0142] In this embodiment, after the step of refining the candidate structure set based on the machine learning inter-atomic potential and submitting the accurate DFT calculation using the machine learning relaxation structure corresponding to each candidate structure as the starting structure for DFT calculation, the method further includes a dynamic calibration step for the prediction error of the machine learning inter-atomic potential, comprising:

[0143] Based on each candidate structure that simultaneously completes machine learning interatomic potential pre-screening and precise formation energy back-calculation, the machine learning interatomic potential formation energy prediction error value of each candidate structure is calculated, and an error feature mapping model between the error value and the corresponding structural features is established; based on the error feature mapping model, the retention ratio of subsequent candidate structures is dynamically adjusted to obtain the calibrated screening threshold.

[0144] Based on the calibrated screening threshold, when the cumulative error of a certain type of doping combination exceeds a preset deviation threshold, candidate structures that have been eliminated by the system-specific machine learning interatomic potential model are automatically extracted and forcibly included in the DFT calculation queue for cross-validation to obtain cross-validation results. Based on the cross-validation results, the system-specific machine learning interatomic potential model is incrementally fine-tuned to obtain an updated system-specific machine learning interatomic potential model, and the updated system-specific machine learning interatomic potential model replaces the system-specific machine learning interatomic potential model for fast structure relaxation of subsequent candidate structures.

[0145] Specifically, firstly, the machine learning inter-atomic potential prediction error dynamic calibration mechanism is an online adaptive correction mechanism designed by this invention to address the core problem of systematic accuracy deviation in a general machine learning inter-atomic potential base model on a specific sodium-ion cathode system. Its basic design idea is as follows: during the execution of the multi-fidelity screening funnel described in Example 7, whenever a candidate structure simultaneously completes machine learning inter-atomic potential pre-screening (obtaining E...)... form_MLIP ) and exact DFT forming energy inverse calculation (to obtain E form_DFT The calculations in these two stages yield a pair of measured data points that can be used to evaluate the prediction error of inter-atomic potentials in machine learning. As such data points accumulate with the advancement of the computational task, the assessment of the reliability of inter-atomic potential predictions in machine learning can be continuously corrected, thereby dynamically adjusting the screening strategy and avoiding the erroneous elimination of truly stable doped structures in the intermediate and fine screening stages due to systematic biases in the inter-atomic potentials of machine learning.

[0146] For each candidate structure that has undergone simultaneous machine learning interatomic potential pre-screening and precise formation energy inverse calculation, its machine learning interatomic potential formation energy prediction error value δᵢ is calculated, which is the absolute deviation between the machine learning interatomic potential formation energy prediction value and the DFT precise formation energy. The correlation data between δᵢ and the structural features of the candidate structure is continuously recorded in the local error database. The structural features are multi-dimensional feature vectors used to characterize the geometric and chemical properties of the candidate doping configuration, specifically including the following dimensions: dopant element combination type (represented by one-hot encoding or element embedding vectors), doping concentration (the proportion of doped atoms to the total number of corresponding lattice sites), degree of local coordination distortion (quantified by the root mean square deviation of the bond lengths adjacent to the doped sites before and after relaxation), difference in ionic radius between the doped element and the parent transition metal, and electronegativity deviation of the doped element. These structural features constitute the input feature space of the error feature mapping model.

[0147] Error feature mapping (EMM) models are lightweight regression prediction models built upon the aforementioned error-feature correlation datasets. Their function is to predict the expected value of the machine learning inter-atomic potential for an unevaluated candidate configuration with given structural features, and then estimate the prediction confidence score of the machine learning inter-atomic potential on that candidate configuration. conf The error feature mapping model is constructed using lightweight machine learning models suitable for small sample scenarios, such as Gaussian Process Regression or Random Forest. Whenever a few data points are added to the local error database, the model is incrementally updated to maintain the model's ability to track the error distribution of the current computation task in real time.

[0148] The prediction confidence score for each candidate configuration is based on the output of the error feature mapping model. conf The candidate structure retention ratio K2 in the machine learning interatomic potential fine screening stage of Example 7 is dynamically adjusted: for regions determined by the error feature mapping model to be of high confidence (i.e., score) conf For a subset of candidate configurations with high confidence and low expected δ, the system appropriately tightens K2 (reducing the retention ratio) and fully trusts the ranking results of the inter-atomic potentials obtained from machine learning to improve screening accuracy and reduce the number of candidate structures fed into the DFT; for regions determined by the error feature mapping model to be low confidence areas (i.e., score) conf For a subset of candidate configurations with low expected δ, K2 is appropriately relaxed (increasing the retention ratio). The ranking results of the inter-atomic potentials obtained through machine learning are carefully considered to avoid erroneously eliminating truly stable candidate structures due to energy ranking distortion caused by systematic biases in the inter-atomic potentials of machine learning. After the above dynamic adjustment of partitions, a calibrated screening threshold is obtained, replacing the static screening strategy with a fixed retention ratio in Example 7, allowing the overall screening behavior to be continuously optimized with the accumulation of computational experience.

[0149] Secondly, at the error monitoring level, cumulative errors are calculated separately for each doping combination category (classified by dopant element pairs or dopant element-concentration ranges). The cumulative error refers to the mean δᵢ of all candidate structures that have completed two-stage evaluation within the same doping combination category, reflecting the systematic prediction bias level of the machine learning interatomic potential for this type of doping combination. When the cumulative error of a certain doping combination category continues to rise and exceeds a preset bias threshold (default 20 meV / atom), it is determined that the current system-specific machine learning interatomic potential model has a significant systematic accuracy deficiency for this type of doping combination, triggering a supplementary sampling and cross-validation process.

[0150] Cross-validation refers to the process of forcibly extracting a number of structures belonging to the current high cumulative error doping combination category from candidate structures that have been eliminated in the machine learning interatomic potential screening stage based on energy window thresholds or interatomic distance thresholds (the number of structures extracted is 20% to 30% of the total number of eliminated structures in that category, with priority given to E). form_MLIP Structures that just exceed the energy window threshold (because these structures are most likely to be incorrectly eliminated due to machine learning overestimating the formation energy of the interatomic potential) are forcibly included in the DFT precise calculation queue. A complete high-precision DFT structure optimization and static self-consistent calculation are then performed to obtain their true and accurate formation energy E. form_DFT By comparing the actual DFT formation energies of these candidate structures with the formation energies predicted by machine learning interatomic potentials, cross-validation results are obtained, which are used to directly quantify the directional bias (overestimation or underestimation) and magnitude of the machine learning interatomic potentials under this doping combination category.

[0151] If the cross-validation results confirm that the interatomic potential of machine learning systematically overestimates this type of doped combination (i.e., the E of most of the eliminated structures), form_DFT Significantly lower than E form_MLIP This indicates that a truly stable structure has been incorrectly eliminated. The new structure-energy-force ternary data pairs obtained from the cross-validation are written into the machine learning inter-atomic potential fine-tuning dataset, triggering incremental fine-tuning of the dedicated machine learning inter-atomic potential model for the system. Incremental fine-tuning refers to performing small-batch online updates on the model with new data without resetting existing model parameters. It only locally corrects the feature space regions corresponding to high-cumulative-error doping combinations, avoiding catastrophic forgetting of other regions with good accuracy, thus achieving a targeted improvement in the accuracy of machine learning inter-atomic potentials. After incremental fine-tuning, the accuracy of the updated model is re-evaluated on an independent validation set. If the RMSE is... E Once the target accuracy requirement is met (default within 2 meV / atom), the updated system-specific machine learning inter-atom potential model replaces the currently used system-specific machine learning inter-atom potential model. This updated model is then used for the rapid structural relaxation task of all subsequent candidate structures, forming a complete closed-loop adaptive mechanism of "real-time monitoring of screening errors → identification of high-error regions → supplementary cross-validation → incremental fine-tuning → targeted accuracy improvement → replacement and deployment." This mechanism effectively solves the problem of insufficient accuracy of general pre-trained machine learning inter-atom potential base models in the local combinatorial space of specific sodium-ion cathode doped systems. Without the need to recollect large-scale DFT training data, it achieves continuous accuracy optimization of machine learning inter-atom potentials on the target system with minimal additional computational cost, ensuring the screening reliability of the multi-fidelity screening funnel throughout the entire computational task execution process.

[0152] Example 9

[0153] In this embodiment, after obtaining the filtered candidate structure set and before submitting the multi-step computation workflow to the computation server in batches, an active sequential search step based on contextual Bayesian optimization of a large language model is further included, comprising:

[0154] Based on the screened candidate structure set, the parent material type, dopant element list, element occupancy ratio, and doping site are extracted for each candidate doping configuration and concatenated into a standardized natural language description string. A Latin hypercube sampling strategy is used to select a preset number of initial candidate doping configurations from the screened candidate structure set to complete the DFT calculation, obtain the formation energy label corresponding to each initial candidate doping configuration, and construct an initial observation set by combining the standardized natural language description string with the corresponding formation energy label.

[0155] Based on the initial observation set, the initial observation set is formatted into a list of few samples and incorporated into the prompt word template of the large language model. The variance is calculated through multiple inferences using temperature sampling to obtain the predicted distribution of all candidate doped configurations in the selected candidate structure set by the current surrogate model. Based on the predicted distribution, a weighted confidence upper bound strategy is used to select a preset batch of candidate doped configurations with the highest collected values ​​from the selected candidate structure set in each iteration and submit them to the DFT calculation to obtain the new formation energy observation values ​​corresponding to each preset batch of candidate doped configurations.

[0156] Based on the new formation energy observation, the standardized natural language description string corresponding to each of the preset batch of candidate doping configurations and the corresponding new formation energy observation are added to the initial observation set to trigger the next round of iteration until a stable doped structure with a formation energy lower than a preset target threshold is found, the formation energy change of the optimal candidate doping configuration is less than a preset convergence threshold when a preset number of rounds is reached, or the DFT computing resources reach a preset upper limit. The thermodynamically most stable doping configuration recommendation set is obtained, and the doping configuration recommendation set is input into the multi-step computing workflow to prioritize the scheduling of the candidate doping configuration with the lowest formation energy.

[0157] Specifically, firstly, the proactive sequential search based on contextual Bayesian optimization of a large language model is an intelligent sampling strategy designed to address the core challenge of limited DFT computational resources in high-dimensional doped combinatorial spaces. Traditional full-scale DFT computation schemes require submitting precise calculations for each of the screened candidate structures, resulting in extremely high computational resource and time costs when the number of candidate structures reaches hundreds or even thousands. The core idea of ​​the proactive sequential search is to transform the doped structure stability prediction problem into a sequential decision-making problem under a Bayesian optimization framework: in each round, only a small number of candidate configurations selected by the intelligent strategy are subjected to DFT computation, and the obtained formation energy observations are updated to the surrogate model in real time. The predictions of the surrogate model are used to guide the sampling decision in the next round. Through multiple iterations, the sampling is gradually concentrated on the thermodynamically most stable doped configuration region, thereby completing the efficient exploration of the candidate space with far fewer DFT computations than full-scale evaluation.

[0158] During the initialization phase, the selected candidate structure set S... filtered For each candidate doping configuration, the following four types of attribute information are extracted from its structured metadata and normalized crystal structure object: parent material type (e.g., "P2 type Na2Mn3O6"), dopant element list (e.g., "Mg, Al, Ti"), the occupancy ratio of each element at the corresponding lattice site (e.g., "Mg:5%, Al:5%, Ti:5%)", and doping site type (e.g., "TM site"). Several key physicochemical descriptors are further extracted, including the average difference in ionic radius between each dopant element and the replaced transition metal element, the average electronegativity of the dopant element combination, and the coordination environment preference description of the dopant element in common sodium cathode systems. All the above attribute information is then concatenated into a normalized natural language description string D according to a predefined template format. cand This serves as the material representation input for subsequent contextual learning by the large language model. The design principle of the standardized natural language description string is to adopt a natural language format that is easy for the large language model to understand and recognize patterns, while retaining sufficient physicochemical information, so that the large language model can extract the structure-property correlation rules related to formation energy from the description string.

[0159] The Latin hypercube sampling strategy is a sampling method that achieves uniform hierarchical sampling in a multidimensional parameter space. Its core principle is to uniformly divide the parameter space of the doped combination (a multidimensional space composed of element type encoding, element ratio, and doping site encoding as dimensions) into N parts according to the value range of each dimension. init The initial sample set is divided into three equally spaced intervals. Then, one sample is randomly selected from each interval in each dimension, ensuring that each interval is sampled only once across all dimensions. This ensures that the initial sample set has good uniform coverage and representativeness across the entire candidate space, avoiding bias towards a local region. Latin hypercube sampling is used to sample from S...filtered N is selected from init There are N initial candidate doping configurations (default 20), and the batch DFT calculation module is called to process these N... init Perform exact DFT structural optimization and formation energy calculation on each initial candidate configuration to obtain the formation energy label E corresponding to each initial candidate configuration. form_i The standardized natural language description string D for each initial candidate configuration is... cand_i With the corresponding energy tag E form_i They are constructed as ordered pairs, which together form the initial observation set O0, serving as the starting point for subsequent Bayesian optimization iterations.

[0160] Second, in each round of Bayesian optimization iteration, the inference context of the large language model surrogate model is first constructed. The large language model surrogate model refers to the innovative design of this invention, which uses a large language model instead of the Gaussian process in traditional Bayesian optimization as the surrogate model. Traditional Gaussian process surrogate models require numerical encoding of input features and maintenance of the covariance matrix, resulting in limitations such as strong dependence on feature engineering and high computational overhead in high-dimensional spaces. In contrast, the large language model surrogate model directly uses natural language description strings as input, leveraging the materials science knowledge accumulated by the large language model during the pre-training phase to predict formation energy. It requires no model training or parameter updates, achieving the surrogate model function entirely through in-context learning, offering advantages such as zero-shot startup and flexible expansion.

[0161] Set the current observation set O t The standardized natural language description strings-formation energy pairs of all evaluated candidate configurations are sorted by formation energy from low to high and formatted into a few-shot example list (each example contains the description string of the candidate configuration and the corresponding formation energy value). This example list is then concatenated into a pre-defined large language model prompt word template. The prompt word instructs the large language model to evaluate the currently unevaluated candidate configuration D. unseen Estimate its formation energy prediction value E pred To obtain an estimate of the prediction uncertainty, a temperature sampling method is used for the same D. unseen Perform multiple independent inferences (default 10 to 20 times). By increasing the sampling temperature parameter of the large language model (Temperature, which controls the randomness of the output distribution; higher temperature results in more diverse outputs), each inference produces slightly different predicted values. The mean of the multiple inference results is calculated as E. pred Standard deviation as the prediction uncertainty σ pred Together, these two constitute the probabilistic prediction output of the current surrogate model for this candidate configuration. For S filteredRepeat the above inference process for all unevaluated candidate configurations to obtain the current surrogate model's predicted distribution P for all candidate configurations. t (Including E for each candidate configuration) pred With σ pred ).

[0162] The Weighted Upper Confidence Bound (W-UCB) strategy is an improved acquisition function designed specifically for doped structure search tasks, based on the standard upper confidence bound (UCB) acquisition function. The standard UCB acquisition function comprehensively considers the optimization of the predicted mean (exploitation, prioritizing E). pred The lowest configuration) and the exploration of prediction uncertainty (exploration, preference for σ) pred The two objectives (higher configuration) are balanced between utilization and exploration using a weighting coefficient β. An additional dimension of combinatorial space diversity is introduced on top of the standard UCB. This is achieved by calculating the semantic distance between the description strings of unevaluated candidate configurations and the set of description strings of evaluated configurations (using the cosine distance metric of the embedding vectors output by the large language model encoder). Candidate configurations with larger semantic distances (i.e., larger differences from the evaluated regions) are given additional diversity scores to prevent the algorithm from getting stuck in local optima and repeatedly sampling. The W-UCB acquisition function is used to evaluate P... t The collected value A is calculated for each unevaluated candidate configuration. i Select the preset batch of candidate doped configurations with the highest collected values ​​(default 5 to 10) and submit them to the batch DFT calculation module to perform accurate formation energy calculation, so as to obtain the new formation energy observation values ​​corresponding to each selected candidate configuration in this round.

[0163] Third, after this round of batch DFT calculations is completed, the newly obtained candidate configurations will be standardized into natural language description strings and corresponding new formation energy observations, forming a tag pair (D... cand_j E form_j Add them one by one to the current observation set O. t Updated to O {t+1} This re-triggers the aforementioned proxy model inference and W-UCB acquisition function calculation process, initiating the next iteration. Before each iteration begins, a termination condition check is performed synchronously to determine if one of the following three termination conditions is met: the first is the target discovery termination condition, i.e., in O... {t+1} There exists a formation energy lower than the preset target threshold E. target The candidate configuration (pre-defined by the user based on the thermodynamic stability requirements of the target material system, such as -0.5 eV / atom) indicates that a doped structure satisfying the thermodynamic stability requirements has been successfully discovered; the second is the convergence termination condition, namely, continuous K convergeIn the first iteration (default 3 to 5 iterations), the change in the formation energy of the optimal candidate doped configuration (the one with the lowest formation energy) in the observation set is less than the preset convergence threshold (default 10 meV / atom), indicating that the surrogate model has fully converged to the thermodynamically most stable region, and the marginal benefit of continuing the iteration is extremely low; the third is the resource limit termination condition, that is, the total number of candidate configurations that have completed accurate DFT calculations reaches the preset computational budget limit N. budget (Pre-set by the user based on available computing resources) to prevent the algorithm from iterating indefinitely beyond the constraints of computing resources.

[0164] After any termination condition is met, several candidate configurations with the lowest formation energy are selected from all candidate configurations that have completed DFT evaluation (the top 5 to 10 by default), forming a recommended set of thermodynamically most stable doped configurations, along with complete computational information for each candidate configuration (formation energy, total energy, structural parameters, computational convergence quality index, etc.). This recommended set of doped configurations is input into the task scheduling queue of the multi-step computation workflow with priority weights, instructing subsequent fine analysis stages (density of states calculation, differential charge density analysis, phonon spectrum calculation, etc.) to prioritize the candidate doped configurations with the lowest formation energy. This proactive sequential search mechanism achieves adaptive navigation in the high-dimensional doping combinatorial space through contextual Bayesian optimization of a large language model. It can significantly reduce the number of DFT evaluations required for high-entropy multi-component doped systems (such as ternary or even quaternary co-doped systems) from hundreds or even thousands to tens of times, significantly saving computational resources while ensuring a high confidence level in discovering the thermodynamically most stable doped configurations. This provides an efficient and intelligent computational acceleration solution for high-throughput doping computation screening of sodium cathodes.

[0165] Example 10

[0166] like Figure 3 As shown, the present invention also provides a smart agent-driven design and screening system for sodium battery cathode doped structures, comprising:

[0167] The instruction parsing and knowledge retrieval module 301 is used to receive natural language instructions for sodium electrolytic crystal doping requirements input by the user, search in the preset sodium electrolytic material simulation knowledge base through a similarity retrieval algorithm, and use a large language model to perform logical reasoning based on the retrieved domain professional knowledge entries to generate an atomic-level tool call sequence and an initial state snapshot.

[0168] The candidate structure generation and screening module 302 is used to perform structure optimization on the standardized crystal structure in the initial state snapshot, generate an original candidate doped configuration set through a symmetry analysis program, and perform multi-level screening according to the physicochemical rules in the domain expertise entries to obtain a screened candidate structure set.

[0169] The computational workflow construction and execution module 303 is used to automatically generate VASP input files and construct multi-step computational workflows for the screened candidate structure set based on the atomic-level tool call sequence and the domain knowledge entries. The multi-step computational workflows are submitted to the computational server in batches, the computational convergence status is monitored in real time, and the computational parameters are adaptively adjusted based on large language model inference to obtain the DFT calculation results of each doped configuration.

[0170] The stability analysis and ranking module 304 is used to extract the total energy of each doped configuration from the DFT calculation results of each doped configuration and perform formation energy calculation inference to generate energy convergence curves and stability visualization charts, thereby obtaining the stability ranking results of each doped configuration.

[0171] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for designing and screening intelligent agents driving a sodium battery cathode doped structure, characterized in that, include: The system receives a natural language command from a user specifying the doping requirements for sodium-ion crystals. It then searches a pre-defined sodium-ion material simulation knowledge base using a similarity retrieval algorithm. A large language model is used to perform logical reasoning based on the retrieved domain-specific knowledge entries, generating an atomic-level tool call sequence and an initial state snapshot. This process includes: receiving the natural language command; retrieving the domain-specific knowledge entries from the sodium-ion material simulation knowledge base using a cosine similarity algorithm; extracting doping elements, doping sites, doping concentrations, and parent materials from the natural language command using keyword matching and slot filling to obtain structured intent metadata; and mapping the structured intent metadata to a task orchestration agent. The structured task blueprint is broken down into subtasks: structure analysis, doping configuration generation, electrostatic energy screening, VASP input generation, and result analysis. A directed acyclic graph (DAG) execution plan is generated. Based on the DAG execution plan, the structure analysis program is called to transform the unstructured input data into a standardized crystal structure. It is then determined whether the standardized crystal structure reaches the computational mechanical stability threshold. If it does, the pseudopotential file, cutoff energy, and k-point sampling accuracy are determined based on the domain expertise entries. The initial state snapshot is then constructed, and an interpretable execution trajectory is formed using an iterative reasoning execution mechanism based on a large language model. The atomic-level tool call sequence and the physical constraint logic between each subtask are then determined. Structural optimization is performed on the standardized crystal structure in the initial state snapshot. An original candidate doped configuration set is generated by a symmetry analysis program. For the original candidate doped configuration set, a three-level screening is performed by sequentially applying the Ewald electrostatic energy reverse screening rule, the atomic spacing threshold elimination rule, and the equivalent configuration deduplication rule to obtain the screened candidate structure set. For the selected candidate structure set, VASP input files are automatically generated based on the atomic-level tool call sequence and the domain knowledge entries, and a multi-step calculation workflow is constructed. The multi-step calculation workflow is submitted to the computing server in batches, the calculation convergence status is monitored in real time, and the calculation parameters are adaptively adjusted based on the large language model inference to obtain the DFT calculation results of each doped configuration. For the DFT calculation results of each doped configuration, the total energy of each doped configuration is extracted and formation energy calculation inference is performed to generate energy convergence curves and stability visualization charts, thus obtaining the stability ranking results of each doped configuration. This includes: based on the DFT calculation results of each doped configuration, extracting the total energy, atomic force convergence trajectory, and charge density residual of each doped configuration from the DFT calculation output file to generate energy-ion step convergence curves; based on the energy-ion step convergence curves and the total energy, performing formation energy calculation inference using a large language model, determining thermodynamic spontaneity by comparing the energy difference between the doped system and the pure system, generating the stability visualization charts, and obtaining the stability ranking results of each doped configuration.

2. The method according to claim 1, characterized in that, The process of performing structure optimization on the normalized crystal structure in the initial state snapshot, and generating an original candidate doping configuration set through a symmetry analysis program, includes: Based on the atomic-level tool call sequence, the geometry optimization program is invoked to start the initial state structure optimization calculation of the standardized crystal structure, and the energy convergence curve and atomic force vector are captured in real time. After the preset convergence threshold is met, the parallel computing subtask set covering multiple different desodium components is executed with high concurrency through the concurrent scheduling engine to obtain the relaxed optimized structure. The relaxed optimized structure is subjected to symmetry analysis using PyMatgen to generate all non-equivalent sodium removal configurations and independent doping configurations, forming the original candidate doping configuration set.

3. The method according to claim 2, characterized in that, For the selected candidate structure set, a VASP input file is automatically generated based on the atomic-level tool call sequence and the domain expertise entries, and a multi-step computation workflow is constructed. This multi-step computation workflow is then submitted in batches to the computation server. The computation convergence status is monitored in real time, and computation parameters are adaptively adjusted based on large language model inference to obtain the DFT calculation results for each doped configuration, including: Based on the selected candidate structure set, the geometric and electronic structure features of each candidate configuration are analyzed by the Agent. The matching physical models and expert experience are retrieved from the sodium electrochemical material simulation knowledge base by the RAG, and INCAR, POSCAR, KPOINTS and POTCAR files are automatically generated to form the VASP input file. A multi-step calculation workflow supporting serial, parallel and conditional branch execution is constructed according to the atomic-level tool call sequence, and the multi-step calculation workflow is submitted to the computing server in batches. Based on the computing status of the computing server, the energy convergence curve, atomic force vector and charge density residual are captured in real time, and the convergence risk is determined by reasoning using a large language model. If it is determined that there is a local structural collapse anomaly, the subsequent DFT calculation of the corresponding candidate configuration is terminated. If it is determined that there is convergence difficulty, the calculation parameters in the INCAR file are automatically corrected and the self-consistent calculation request is reissued to obtain the DFT calculation results of each doped configuration.

4. The method according to claim 1, characterized in that, It also includes self-evolving knowledge graph update steps, including: Based on the ontological concepts and domain relationships in the field of sodium-ion cathode doping, the initial knowledge graph is constructed by automatically extracting entity and relation triplets from the sodium-ion material simulation knowledge base, with parent material, doping element, doping site, phase structure, calculation parameters and physical quantity index as node types, and doping initiation relationship, stabilization effect relationship, recommended calculation parameter relationship and calculation success or failure relationship as edge types. Based on the initial knowledge graph and the DFT calculation results of each doped configuration, whenever any DFT calculation subtask is completed, computational feature information is automatically extracted from the DFT calculation output file. The computational feature information is written into the initial knowledge graph in the form of standardized triples, and entity alignment and conflict resolution are performed to obtain the merged knowledge graph. If the calculation fails to converge, the failure reason label and corresponding calculation parameter snapshot are recorded to form a failure experience node and written into the merged knowledge graph to obtain the updated knowledge graph. If the calculation converges successfully, the merged knowledge graph is used as the updated knowledge graph. Based on the updated knowledge graph, when a new doping task request is received, the system retrieves the set of completed computation cases and the failed experience nodes that match the current doping task. The distribution of computational parameters of successful cases in the set of completed computation cases is used as a strong prior for initializing the computational parameters of the current task. Computational parameters that overlap with the failure modes recorded in the failed experience nodes are skipped or corrected to obtain task planning results optimized based on historical experience. The task planning results are then written into the atomic-level tool call sequence and the domain knowledge entries are updated.

5. The method according to claim 1, characterized in that, After obtaining the filtered candidate structure set, and before submitting the multi-step computation workflow to the computation server in batches, the process further includes a multi-fidelity filtering funnel execution step, comprising: Based on the aforementioned domain expertise items, a base model is loaded from a pre-trained general inter-atomic potential model library. The root mean square error between the machine learning inter-atomic potential prediction of the base model and the DFT reference value is verified for the target parent material system. If the root mean square error is lower than a preset error threshold, the base model is used as the system-specific machine learning inter-atomic potential model. Otherwise, the base model is transferred and fine-tuned using the pre-acquired DFT reference calculation results of the target parent material system as the fine-tuning dataset to obtain the system-specific machine learning inter-atomic potential model. Based on the selected candidate structure set, the system-specific machine learning interatomic potential model is invoked to perform fast structure relaxation and energy sorting on each candidate structure, eliminating structures with abnormal interatomic spacing or energy exceeding a preset energy window threshold, thereby obtaining a machine learning interatomic potential fine-screened candidate structure set. Each candidate structure in the machine learning interatomic potential fine-screened candidate structure set carries a machine learning relaxed structure after machine learning interatomic potential relaxation. Based on the machine learning interatomic potential screening candidate structure set, the machine learning relaxation structure corresponding to each candidate structure is used as the starting structure for DFT calculation to submit accurate DFT calculation. The accurate formation energy is calculated back from the total DFT energy to obtain the stability ranking of the doped structure. The stability ranking of the doped structure is then input into the multi-step calculation workflow to prioritize the scheduling of a preset number of candidate structures with the highest stability ranking.

6. The method according to claim 5, characterized in that, After the candidate structure set is screened based on the machine learning inter-atomic potential, and the machine learning relaxation structure corresponding to each candidate structure is used as the starting structure for DFT calculation, the process further includes a dynamic calibration step for machine learning inter-atomic potential prediction error, including: Based on each candidate structure that simultaneously completes machine learning inter-atomic potential pre-screening and precise formation energy back-calculation, calculate the machine learning inter-atomic potential formation energy prediction error value for each candidate structure, establish an error feature mapping model between the error value and the corresponding structural features, and dynamically adjust the retention ratio of subsequent candidate structures based on the error feature mapping model to obtain the calibrated screening threshold. Based on the calibrated screening threshold, when the cumulative error of a certain type of doping combination exceeds a preset deviation threshold, candidate structures that have been eliminated by the system-specific machine learning interatomic potential model are automatically extracted and forcibly included in the DFT calculation queue for cross-validation to obtain cross-validation results. Based on the cross-validation results, the system-specific machine learning interatomic potential model is incrementally fine-tuned to obtain an updated system-specific machine learning interatomic potential model, and the updated system-specific machine learning interatomic potential model replaces the system-specific machine learning interatomic potential model for fast structure relaxation of subsequent candidate structures.

7. A system for designing and screening intelligent agents driving sodium battery cathode doped structures, used to implement the method for designing and screening intelligent agents driving sodium battery cathode doped structures according to any one of claims 1-6, characterized in that, include: The instruction parsing and knowledge retrieval module receives natural language instructions from the user regarding sodium electrolytic crystal doping requirements. It then searches a pre-defined sodium electrolytic material simulation knowledge base using a similarity retrieval algorithm. Based on the retrieved domain-specific knowledge entries, it performs logical reasoning using a large language model to generate an atomic-level tool call sequence and an initial state snapshot. This includes: receiving the natural language instruction; retrieving the domain-specific knowledge entries from the sodium electrolytic material simulation knowledge base using a cosine similarity algorithm; and extracting doping elements, doping sites, doping concentrations, and parent materials from the natural language instruction using keyword matching and slot filling to obtain structured intent metadata; and then processing the structured intent metadata through task encoding. The intelligent agent is mapped to a structured task blueprint, which is decomposed into subtasks of structure analysis, doping configuration generation, electrostatic energy screening, VASP input generation, and result analysis, generating a directed acyclic graph execution plan. Based on the execution plan, the structure analysis program is called to transform unstructured input data into a standardized crystal structure, and it is determined whether the standardized crystal structure reaches the computational mechanical stability threshold. If it does, the pseudopotential file, cutoff energy, and k-point sampling accuracy are determined based on the domain expertise entries, the initial state snapshot is constructed, and an interpretable execution trajectory is formed using an iterative inference execution mechanism with a large language model, determining the atomic-level tool call sequence and the physical constraint logic between each subtask. The candidate structure generation and screening module is used to perform structure optimization on the standardized crystal structure in the initial state snapshot, generate an original candidate doping configuration set through a symmetry analysis program, and perform a three-level screening on the original candidate doping configuration set by sequentially applying the Ewald electrostatic energy reverse screening rule, the atomic spacing threshold elimination rule, and the equivalent configuration deduplication rule to obtain the screened candidate structure set. The computational workflow construction and execution module is used to automatically generate VASP input files and construct multi-step computational workflows for the screened candidate structure set based on the atomic-level tool call sequence and the domain knowledge entries. The multi-step computational workflows are submitted to the computational server in batches, the computational convergence status is monitored in real time, and the computational parameters are adaptively adjusted based on large language model inference to obtain the DFT calculation results of each doped configuration. The stability analysis and ranking module is used to extract the total energy of each doped configuration from the DFT calculation results and perform formation energy calculation inference to generate energy convergence curves and stability visualization charts, thereby obtaining the stability ranking results of each doped configuration. This includes: extracting the total energy, atomic force convergence trajectories, and charge density residuals of each doped configuration from the DFT calculation output file based on the DFT calculation results, generating energy-ion step convergence curves; performing formation energy calculation inference using a large language model based on the energy-ion step convergence curves and the total energy, determining thermodynamic spontaneity by comparing the energy difference between the doped system and the pure system, generating the stability visualization charts, and obtaining the stability ranking results of each doped configuration.

Images