AI material aided design and collaborative research method based on large language model

Abstract

This invention belongs to the interdisciplinary field of artificial intelligence and materials science, and relates to an AI-aided material design and collaborative R&D method based on a large language model. This method extracts high-quality corpus through TCS scoring, integrates material crystallographic features, and performs CPT and SFT training on the base model, selecting models with SCS>1.5 and CCS>1.5 as reinforcement bases. Subsequently, it performs reinforcement learning in the direction of material reinforcement, introducing a lightweight physical simulation module to ensure the output conforms to physical laws. It deploys a four-agent collaborative embedded model and DFT simulation tool, receiving R&D goals, breaking down tasks, generating material schemes, and completing novelty verification and performance verification. If verification fails, parameters are adjusted and redesigned based on JSON negative feedback until a verifiable experimental scheme is output. This invention achieves intelligent and collaborative material design, effectively shortening the R&D cycle, reducing costs, and providing a new, efficient R&D path.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and materials science, and relates to an AI-aided design and collaborative research and development method for materials based on a large language model. Background Technology

[0002] Materials science research and development is undergoing a transformation from the traditional experimental trial-and-error method to a data-driven artificial intelligence paradigm. In recent years, large language models (LLMs) have demonstrated powerful logical reasoning and natural language processing capabilities in general domains, but in the highly specialized field of materials science, existing models still face significant limitations.

[0003] First, general-purpose models lack "symbolic competence." Materials science heavily relies on specific symbolic systems, such as SMILES encoding for describing molecular structures, CIF files for crystal structures, and IUPAC nomenclature. Research shows that existing general-purpose models often fail to maintain syntactic coherence when generating complex SMILES strings, resulting in chemically invalid or physically unrealizable molecular structures. In complex 3D lattice structure recognition and space counting tasks, the models also perform poorly, frequently failing to accurately identify the number of atoms or symmetries within a unit cell.

[0004] Secondly, materials research and development tasks face a "latent solvability" threshold. While reinforcement learning (RL) can amplify the reasoning ability of a model, this requires that the base model has already allocated non-zero probabilities to the correct answer within the "long tail" of its output distribution. Due to the highly specialized nature of materials science data and its extremely low proportion in general internet corpora, the base model often fails to generate effective reward signals, causing reinforcement learning to stagnate.

[0005] Furthermore, most existing materials AI systems are isolated, single-point tools, lacking a systematic collaborative R&D architecture. Materials design involves the entire process, from initial literature mining and performance index setting to precursor selection, process parameter optimization, and laboratory verification. In traditional R&D models, information silos are severe at each stage, lacking a command system capable of deeply understanding the underlying logic of materials and coordinating multiple specialized models (such as simulation models and prediction models).

[0006] Taking aerogel materials as an example, their development involves the nonlinear coupling of multiple parameters, such as precursor ratio (e.g., the ratio of vinyltriethoxysilane to water glass), catalyst selection, aging time, and drying process (atmospheric pressure drying or supercritical drying). Existing AI methods struggle to simultaneously optimize thermal conductivity, hydrophobicity, and porosity while ensuring the stability of the aerogel framework (avoiding collapse during atmospheric pressure drying). Summary of the Invention

[0007] The purpose of this invention is to provide an AI-aided material design and collaborative research and development method based on a large language model, comprising:

[0008] Step S100: High-quality corpus is extracted using TCS scoring and fused with the crystallographic features of the material. Based on the corpus, the base model is trained using CPT and SFT, and the model output is assigned accordingly. <think>and <answer>The structuring ability; finally, the model with symbolic ability score SCS>1.5 and chemical ability score CCS>1.5 is input into the next stage as the basis for subsequent reinforcement;

[0009] Step S200: Perform material-enhanced reinforcement learning on the base point models selected in S100, with the format reward... Character matching reward Building upon this foundation, a lightweight physics simulation module is introduced to generate physics simulation rewards. This ensures that the sequences generated by the model conform to the objective laws of the physical world;

[0010] Step S300: Based on the core model optimized in step S200, deploy orchestration agents and design agents, and call embedded models and DFT simulation tools to deploy retrieval agents and simulation agents;

[0011] In step S400, after receiving the material research and development goal input by the user, the orchestration agent calls the model optimized in step S200 to decompose the task, collaboratively calls the design agent to generate a material generation scheme with structured tags, uses the retrieval agent to perform novelty verification, and uses the simulation agent to automatically configure computational chemical parameters to perform physical performance verification.

[0012] In step S500, if the verification in step S400 fails, the simulated agent serializes the unconverged structural defects, force gradients, or physical error data and constructs a negative feedback sample in JSON format. The design agent receives this proof, dynamically adjusts the generation parameters and redesigns in the current dialogue context until the target performance is met, and finally outputs a verifiable experimental scheme.

[0013] Furthermore, the specific process of step S100 includes:

[0014] Step S101: Obtain multi-source raw data from large-scale datasets such as MatMech and MateriMiner; first, perform deduplication, non-text element parsing, and basic format cleaning on the raw data; then, introduce a dynamic filtering mechanism based on Text Chemistry Score (TCS) to filter data based on word frequency ratio. Weighted summation is performed to remove low-information text and noisy text paragraphs from non-chemistry fields, retaining samples with TCS>4 to obtain high-density chemical knowledge text.

[0015] Step S102: Use a Named Entity Recognition (NER) tool to scan the high-density chemical knowledge text filtered in step S101, locate the molecular names or crystal structure description entities, call RDKit (for SMILES) or Pymatgen (for CIF), and hard-encode the extracted entities into a standardized linear symbol sequence; then, forcibly insert preset exclusive delimiters ([START_SMILES] and [END_SMILES], or [START_CIF] and [END_CIF]) before and after the symbol sequence, and interweave them with the original natural language context.

[0016] Step S103: Continuous pre-training (CPT) is performed on the base model based on the corpus obtained in step S102, which is unsupervised training based on "next word source prediction". General domain data is mixed into the CPT training dataset at a predetermined ratio of 5% to inject prior knowledge of chemical common sense and material properties (such as band gaps and formation energy) while preventing catastrophic forgetting and preserving its basic natural language understanding ability.

[0017] Step S104: After CPT training is completed, the model is fine-tuned using an instruction dataset containing high-value question-answer pairs and scientific chain of thought (CoT) trajectories.

[0018] Step S105: The model’s mastery of chemical symbols and knowledge is judged by the symbolic ability score (SCS) and chemical ability score (CCS). Qualified models with SCS>1.5 and CCS>1.5, that is, those with high symbolic resolution and high mastery of chemical facts, are selected as the benchmark models for subsequent reinforcement learning.

[0019] Furthermore, the TCS in step S101 is calculated using the following formula: ;

[0020] Weight The nonlinear mapping between the indicator function and the frequency ratio is employed, specifically as follows:

[0021] ;

[0022] In this formula:

[0023] This represents the total number of lemmas in the text segment to be evaluated;

[0024] It is a word element Relative frequency in chemical text corpora;

[0025] It is a word element Relative frequency in ordinary text corpora;

[0026] It is a word element The weights;

[0027] Furthermore, in step S104, when performing the SFT, the following operations should be performed:

[0028] 1. Master three specific paradigms through the SFT-guided model: (i) symbolic understanding of molecular or material structures; (ii) structure-aware scientific question answering; and (iii) chain of thought (CoT) stepwise reasoning in the fields of chemistry and materials.

[0029] 2. To accommodate complex material design tasks and lengthy crystal structure files, the maximum context window length of the model is extended in the SFT stage (from 4096 words to 8192 words) to accommodate long sequences of physical inference steps;

[0030] 3. Train the model using a structured feedback dataset containing standard formats, forcing it to respond to scientific instructions in specific... <think>The tag outputs the underlying physical / chemical logical derivation process (such as valence balancing and space group symmetry analysis), followed immediately by... <answer>The label outputs the final material symbol sequence.

[0031] Furthermore, the SCS and CCS in step S105 are described as follows:

[0032] SCS is primarily used to measure a model's mastery of the syntax of chemical symbols (such as SMILES or CIF). The calculation consists of two steps:

[0033] Step 1: Calculate the mean log-likelihood of the sequence. The formula is as follows:

[0034] ;

[0035] In this formula:

[0036] For a given model;

[0037] The number of tokens;

[0038] For a given sequence of symbols .

[0039] Step 2: Calculate the distribution difference between the correct sequence and the corrupted sequence. The calculation formula is as follows:

[0040] ;

[0041] In this formula:

[0042] It is a set of completely correct normalized sequences;

[0043] It is a set of illegal sequences generated by randomly deleting characters (corruption rate ρ=0.2);

[0044] The average likelihood of the canonical sequence;

[0045] This represents the average likelihood of the corrupted sequence.

[0046] To combine the standard deviations, the formula is as follows: ;

[0047] In this formula:

[0048] It is a set of correct sequence samples. The sample size is the total number of canonical sequences.

[0049] It is a corrupted sequence sample set The sample size, i.e., the total number of illegal sequences.

[0050] It is a set of correct sequence samples. The standard deviation of the likelihood of each sequence.

[0051] It is a corrupted sequence sample set The standard deviation of the likelihood of each sequence.

[0052] CCS is used to measure the model's grasp of "chemical facts." Positive samples are authentic, scientifically accurate descriptions of chemicals / materials, while negative samples are texts containing scientific errors due to random sentence replacements. The formula is as follows:

[0053] ;

[0054] In this formula:

[0055] The average likelihood of the model for factually correct text;

[0056] This represents the average likelihood of the tampered text.

[0057] Furthermore, the specific process of step S200 includes:

[0058] Step S201: Input material cue words into the model to generate the corresponding candidate material symbol sequence CIF, and call the lightweight physics simulation module, namely the general machine learning potential function M3GNet model, to calculate the generated material configuration. Total system energy and atoms Force The calculation formula is: ;

[0059] In step S202, the module uses the conjugate gradient method to guide the atomic coordinate update based on the aforementioned force gradient, performing rapid structural relaxation. Its basic iterative update process can be represented as:

[0060] ;

[0061] in, The relaxation step size is determined until the maximum force on the largest atom in the system is less than a preset threshold of 0.05 eV / Å, at which point the relaxed energy is obtained. , This indicates the current iteration step.

[0062] Step S203: The module calculates the enthalpy of formation for each atom. (Also known as formation energy) To determine the thermodynamic stability of a material (i.e., whether it can be synthesized in reality), the calculation formula is:

[0063] ;

[0064] In this formula:

[0065] This represents the total energy after the unit cell is relaxed.

[0066] For elements The number of atoms;

[0067] For elements Chemical potential in the standard elemental state.

[0068] in, <0 indicates that the material has thermodynamic stability. The lower the value, the more stable the material.

[0069] Step S204: Map the above calculation results to the reward signal received by the reinforcement learning model GRPO algorithm. To avoid gradient explosion, an exponential decay function is used to normalize the energy feedback. The specific calculation formula is as follows:

[0070] ;

[0071] In this formula:

[0072] Preset hyperparameters to control the steepness of the penalty;

[0073] The expected material stability energy threshold;

[0074] The enthalpy of formation of the generated material is obtained from simulation calculations.

[0075] Step S205, combining the reward function mentioned in S200, the global reward obtained by the model in one training step (Step) can be quantified as follows:

[0076] ;

[0077] In this formula:

[0078] Formatted Rewards Weighting coefficients;

[0079] Character matching reward Weighting coefficients;

[0080] Rewards for physical simulation Weighting coefficients;

[0081] Rewards are given for adherence to format and structure;

[0082] Rewards for symbolic syntax and chemical legitimacy;

[0083] Rewards are given for physical simulation.

[0084] After the above optimization steps, the final core model is output for subsequent multi-agent systems to call.

[0085] Furthermore, in step S205 To evaluate whether the model strictly adheres to the reward function of the predefined structured output instructions, regular expressions are used to determine whether the model's output text contains a complete pair of... <think>Labels and a complete pair <answer>Tags, and <think>Must <answer>Previously, positive rewards were given for correct actions and negative punishments were given for incorrect actions;

[0086] To evaluate the model <answer>The system checks whether the generated symbol sequence in the label is a "valid code". RDKit (for SMILES) or Pymatgen (for CIF) will be used to determine the validity of the structure, awarding positive rewards for correct answers and negative penalties for incorrect answers.

[0087] Furthermore, the specific process of step S300 includes:

[0088] Step S301: Deploy orchestration agents using a Kubernetes (K8s) cluster, using LangGraph as its core framework to manage the workflow and state graph of all agents, using FastAPI to expose interfaces to call the core model, and using Redis to store the session context (memory) and current execution state of the entire R&D task to ensure that the task can be recovered after interruption.

[0089] Step S302: Deploy the design agent on the local A100 server. The inference backend is vLLM. The core model is called with high concurrency. Combine RDKit (for SMILES) or Pymatgen (for CIF) to perform fast hard-coded verification of the generated basic format. Receive the generation prompt words assigned by the orchestration agent through RESTAPI and return the generated structure sequence.

[0090] Step S303: Deploy the retrieval agent on the local A100 server, deploy the Milvus management vector database, and deploy a dedicated small chemical embedding model using HuggingFace's TEI (TextEmbeddingsInference).

[0091] Step S304: Deploy the simulation agent using a "cloud agent + HPC cluster" approach; deploy a scheduling agent on the K8s cluster, and build an asynchronous task queue using Celery and RabbitMQ to listen for simulation requests; the agent service uses the ASE (Atomic Simulation Environment) middleware to parse the large model output into a standardized input file for VASP; then, submit the task across the network to the high-performance computing cluster (HPC) for execution through the Slurm scheduling system, parse the output logs after completion, and asynchronously return the physical performance metrics to the orchestration agent through a Webhook callback mechanism.

[0092] Furthermore, the specific process of step S400 includes:

[0093] In steps S401 and S300, the orchestration agent receives the user-inputted material development goal (such as high-temperature resistant polymer materials) and transforms it into system prompts containing real chemical boundary conditions. In addition to basic settings such as temperature, pressure, and solvent environment, the system adaptively extracts microscopic constraints specific to the domain, such as extracting the "room temperature ionic conductivity threshold" for solid-state battery materials or the "singlet-triplet bandgap" as a rigid boundary condition for photoelectric luminescent materials. Subsequently, the orchestration agent persistently stores the session context and state machine of the current development task using an in-memory database (such as Redis), and then sends the prompts to the design agent via Remote Procedure Call (RPC).

[0094] Step S402: Design the agent to call the S200 core model to process the prompt word, forcing the model to output with... <think>and <answer>The response sequence is structured and tagged, and the generated result is published to the asynchronous message bus RabbitMQ. <think>The tags contain the chemical reasoning logic generated by the model, the expected electronic structure characteristics, and the steric hindrance analysis; <answer>The label contains only the generated candidate material structures (SMILES or CIF sequences).

[0095] Step S403, precise extraction of the retrieval agent. <answer>The symbol sequence within the label first calls the embedding model to vectorize the sequence, then initiates a high-dimensional similarity retrieval to the independently deployed vector database cluster Milvus, filters out known materials, and returns a list of novel materials that have passed the screening to the orchestration agent;

[0096] In step S404, the orchestration agent packages the validated material sequences into simulation tasks and pushes them to the task queue Celery. The simulation agent agent service listens to this queue and parses the data. <think>The chemical reasoning logic within the tags is based on a domain rule base to achieve adaptive mapping of computational tasks and automatic configuration of underlying computational parameters. Specific mapping rules include: for alloy catalytic materials, automatic slicing to build surface models and configuring density functional theory (DFT) for structural optimization to calculate the surface adsorption energies of their d-band centers and key intermediates; for solid electrolyte materials, automatic configuration of ab initio molecular dynamics (AIMD) and extraction of ion mean square shifts to verify macroscopic ionic conductivity; and for high-temperature resistant polymer tasks, configuration of all-atom molecular dynamics (MD) to calculate their free volume ratio and glass transition temperature. After configuration, the system generates corresponding standardized input files.

[0097] In step S405, the simulation agent submits the computational task to a high-performance computing (HPC) cluster for execution through a cluster scheduling system (such as Slurm). During this period, the orchestration agent places the session state of the R&D task in an "asynchronous suspension" state, releases front-end computing resources, and waits for the physical simulation results.

[0098] Furthermore, the specific process of step S500 includes:

[0099] In step S501, after the HPC cluster completes the computation task, the simulated agent agent service captures and parses the underlying output logs (such as force gradient, system total energy, or bandgap data), and sends a completion signal through the Webhook callback mechanism or the result message queue. Upon receiving this event, the orchestration agent is awakened and reloads and restores the task's context state from the in-memory database.

[0100] In step S502, the system automatically compares the extracted physical calculation results with the initial R&D goals set in step S401. If all physical properties meet the target thresholds, the candidate material structure and its experimental scheme are output as the final result, and the process ends; if not, the online physical meltdown mechanism is triggered immediately.

[0101] In step S503, the simulated agent serializes the specific physical error data that leads to melting and packages it into a "failure proof" negative feedback load with strict JSON format. This proof deeply integrates microscopic failure characteristics. If it is thermodynamically unstable, it will extract the "imaginary frequency" and the corresponding geometric distortion vector that appears in the phonon dispersive spectrum calculation; if it is an electronic structure that does not meet the standards, it will clearly mark key parameters that determine the failure of material performance, such as "wide bandgap leading to limited light absorption" or "high adsorption free energy".

[0102] Step S504: The orchestration agent appends the "failure proof" as a new prompt to the current dialogue history and sends it back to the design agent via RPC. The design agent then... <think>In the inference space, the system is forced to first analyze the physical defects in the "failure proof," and then invoke relevant theorems in chemical materials science to dynamically adjust the atomic substitution strategy or spatial configuration. Specific correction strategies include: if the failure proof shows imaginary frequencies in the phonon spectrum or extreme force gradients causing structural non-convergence, the design agent will trigger a "steric hindrance elimination strategy," actively trimming large-volume side groups (e.g., replacing tert-butyl with methyl) or extending flexible chains to generate new configurations; if the failure proof shows electronic band mismatch or excessively high catalytic barriers, the design agent will trigger an "equivalent / isoelectronic element substitution strategy" or a "functional group induction strategy," dynamically adjusting the d-band center of the material or narrowing the band gap, thereby generating targeted corrections. <answer>Candidate sequences;

[0103] In step S505, the design agent pushes the revised new sequence back to the asynchronous message bus, re-triggering the verification process from steps S403 to S405. This asynchronous closed-loop process across computing power levels continues to iterate online until materials meeting the target performance are generated, or the preset maximum number of retries is reached.

[0104] Furthermore, the "failure proof" in the above steps specifically includes: 1. Target and result deviation module, including the R&D target setting value and the actual simulation calculation value, clearly indicating which physical indicator triggered the circuit breaker; 2. Microscopic physical and geometric feature module, including structural convergence features and spatial relaxation features, recording the physical microscopic reasons that led to the substandard performance; 3. Abnormal interruption diagnosis, if the density functional theory (DFT) calculation collapses in the electron step self-consistent field (SCF) loop, extracting the cause of the collapse;

[0105] Furthermore, the "proof of failure" in step S503 is constructed as follows:

[0106] Step S5031: When the physical simulation task on the HPC cluster ends (or terminates abnormally), the simulation agent agent service calls the embedded atomic simulation environment ASE to automatically scan and parse the unstructured output files generated by the underlying physical software VASP.

[0107] Step S5032: The agent service has a built-in "chemical anomaly expert rule tree". The system inputs the parsed log data into the rule tree for threshold comparison.

[0108] In step S5033, the proxy service encapsulates all extracted numerical deviations and abnormal features into JSON serialization, and on this basis, calls the template rendering engine to convert the JSON data into prompts that are easy for the large language model to understand.

[0109] In step S5034, the orchestration agent finally appends this serialized text as the latest "system observation" seamlessly to the end of the current dialogue context (Memory) and sends a retry request to the design agent. The design agent is forcibly triggered upon recognizing this "proof of failure". <think>Tags, execution of reflection and error-correcting reasoning.

[0110] The present invention has significant advantages over the prior art as follows:

[0111] 1. Enhance the "symbolic power" of large models to ensure the legitimacy of generated molecular and structural sequences.

[0112] Existing general-purpose large models often lack syntactic coherence when generating complex SMILES or CIF codes, resulting in outputs of chemically illegal or physically unrealizable structures. This invention introduces a TCS mechanism to extract high-density chemical corpora using word frequency ratio weighting, and then uses Named Entity Recognition (NER) and underlying cheminformatics tools (RDKit / Pymatgen) to force entities to be hard-coded into standardized symbol sequences. This is in line with the requirement for the model to force output... <think>Reasoning and <answer>The structured instruction fine-tuning (SFT) of the results, along with the rigorous pre-screening of symbolic ability score (SCS) and chemical ability score (CCS), effectively overcomes the cross-modal cognitive gap and ensures the legitimacy of the generated sequences in the basic chemical theory.

[0113] 2. Overcoming the "potential solvability" threshold of reinforcement learning to ensure that the generated materials possess true thermodynamic stability.

[0114] In the highly specialized field of materials science, conventional reinforcement learning often stagnates because the model fails to receive effective reward signals. This invention creatively introduces a lightweight physics simulation module based on the M3GNet machine learning potential function. This module performs fast structural relaxation on candidate networks generated from large models using the conjugate gradient method, instantly calculating the system's total energy, atomic force gradients, and enthalpy of formation. These physical quantities are mapped to physical simulation rewards using an exponential decay function. This leads to the exploration of material structures that follow objective physical laws and can be synthesized in reality by large language models.

[0115] 3. Adopt a multi-agent collaborative R&D architecture to improve R&D efficiency.

[0116] In traditional materials research and development, the various stages, from requirement definition and precursor design to laboratory simulation verification, are often fragmented. This invention effectively solves this problem by deploying a distributed multi-agent architecture with an orchestration agent as the central nervous system, linking design agents, retrieval agents, and simulation agents. This collaborative approach automatically transforms the user's macroscopic performance requirements into microscopic physical constraints, automatically performs molecular-level novelty checks using a vector database (Milvus), and adaptively maps chemical sequences to the molecular dynamics or first-principles (DFT) calculation parameters required by the underlying supercomputing cluster (HPC), thus realizing an efficient and automated AI-aided materials design process.

[0117] 4. Use negative feedback signals in JSON format to help the model converge to the correct result faster.

[0118] When faced with complex physical verification failures, large models often struggle to perform accurate secondary optimizations due to the ambiguity of error messages. The technical approach of this invention lies in the fact that when a supercomputing cluster simulation fails to meet standards, the simulation agent automatically scans the underlying logs and transforms these specific micro-failure feature sequences into a JSON-formatted "failure proof" negative feedback load. The structured JSON data eliminates semantic ambiguity, helping the design agent adopt targeted error correction strategies in the next iteration, enabling more precise dynamic error correction and parameter adjustment. Attached Figure Description

[0119] Figure 1 A flowchart for AI-assisted design and collaborative R&D based on a large language model.

[0120] Figure 2 A schematic diagram of the computational mechanism for symbolic capabilities.

[0121] Figure 3 System architecture diagram for multi-agent collaborative development.

[0122] Figure 4 3D surface plots for optimizing the properties of aerogel materials. Detailed Implementation

[0123] The present invention will be further described below with reference to the embodiments.

[0124] This embodiment uses aerogel-assisted design based on a large language model as an example to illustrate the specific implementation steps of the present invention in detail.

[0125] This invention relates to an AI-aided design and collaborative development method for materials based on a large language model. The specific steps of this method are as follows:

[0126] Step S100: The general base model is endowed with a structured cognitive ability to understand the complex sol-gel dynamics of aerogels, the drying mechanism of supercritical fluids, and the theory of polymer cross-linking and polymerization. Domain knowledge is filtered, symbolic injection is performed, and targeted fine-tuning is carried out on it.

[0127] Step S101: Obtain multi-source raw data from large-scale datasets such as MatMech and MateriMiner. First, perform deduplication, non-text element (such as garbled characters and formatting labels) parsing, and basic format cleaning on the raw data. Then, introduce a dynamic filtering mechanism based on Text Chemistry Score (TCS) to filter data by word frequency ratio. Weighted summation is performed to remove low-information text and noisy text paragraphs from non-chemistry fields, retaining samples with TCS>4 to obtain high-density chemical knowledge text.

[0128] In step S102, to enable the model to understand the complex cross-linked network topology of aerogels in three-dimensional space, the system uses a Named Entity Recognition (NER) tool to perform a deep scan of the high-density chemical corpus obtained in step S101, locating molecular names, functional group features, and crystal structure description entities. Subsequently, the system uses API calls to low-level cheminformatics tools such as RDKit (for organic polymers and precursors) or Pymatgen (for inorganic crystals and porous frameworks) to forcibly hard-encode the extracted natural language entities, converting them into standardized linear symbol sequences (SMILES strings or CIF file format).

[0129] In the design chemistry of aerogels, the choice of precursors and the spatial configuration of crosslinking agents directly and decisively affect the mechanical strength, thermal stability, and porosity characteristics of the final material. To enable the model to learn the topological nature of these material components, the system performs a detailed symbolic transformation on the core aerogel formulation system. Table 1 below shows the symbolic mapping of typical aerogel precursors and crosslinking agents in the system and their core physicochemical roles in material design: silicon-based precursors Tetraethoxysilane CCO(OCC)(OCC)OCC The basic tetrafunctional cross-linking center, after hydrolysis and condensation, forms a classic pure silica inorganic framework, exhibiting extremely high brittleness and hydrophilicity. silicon-based precursors Methyltrimethoxysilane C(OC)(OC)OC By introducing nonpolar methyl side groups, the crosslinking density is reduced through steric hindrance, giving silicon-based aerogels macroscopic superhydrophobic properties (contact angle >150°) and significant structural resilience and flexibility. polyimide precursor Pyromellitic dianhydride O=C1OC(=O)c2cc3c(cc12)C(=O)OC3=O It provides a highly rigid planar aromatic ring structure framework, giving polyimide (PI) aerogels extremely high initial thermal decomposition temperature and excellent creep resistance. polyimide precursor 4,4'-Diaminodiphenyl ether Nc1ccc(Oc2ccc(N)cc2)cc1 Containing flexible ether linkages, it introduces conformational rotational freedom into the polyimide backbone, making it a key monomer for improving the flexibility of macromolecular chain segments and inhibiting excessive shrinkage. Organic crosslinking agent 1,3,5-Triaminophenoxybenzene NC1=CC(N)=CC(N)=C1 Highly efficient trifunctional aromatic crosslinking agents can significantly improve the crosslinking density and Young's modulus of organic aerogels, but they can also lead to excessive rigidity of the three-dimensional network and increased brittleness. Hybrid crosslinking agent Octaaminophenyl polyhedral oligomeric silsesquioxane Nc1ccc(Oc2ccc(N)cc2)cc1 (Simplified illustration) It possesses a nanoscale inorganic siloxane cage-like core structure (POSS), providing eight polymerization reaction sites. Its large volume effect effectively expands the polymer chain segments, prevents pore collapse, and significantly improves thermal stability and dielectric properties.

[0130] After completing the above entity to SMILES / CIF conversion, pre-defined exclusive delimiters ([START_SMILES] and [END_SMILES], or [START_CIF] and [END_CIF]) are forcibly inserted before and after the symbol sequence, and interleaved with the original natural language context;

[0131] Step S103: Continuous pre-training (CPT) is performed on the base model based on the corpus obtained in step S102, i.e., unsupervised training based on "next word source prediction". In this process, the system not only deeply injects the physicochemical prior knowledge unique to aerogel materials (such as mesoporous distribution law, formation energy change trend, band gap characteristics) into the hidden layer weights of the model, but also, in order to prevent the model from "catastrophic forgetting" when deeply cultivating vertical domains, the system mixes general domain data into the CPT training dataset at a strictly preset ratio of 5% to retain its basic complex natural language logical reasoning and instruction understanding capabilities;

[0132] Step S104: After CPT training, the model enters the Supervised Fine-Tuning (SFT) stage. Addressing the extremely high complexity of the aerogel material design task, the system performs several deep, low-level configuration optimizations at this stage: First, considering the lengthy molecular dynamics analysis trajectory data of the aerogel framework and the extremely large porous crystal structure files (such as CIF files containing thousands of atoms), the system hard-extends the maximum context window length of the model during the SFT stage, significantly increasing it from the basic 4096 terms to 8192 terms, ensuring sufficient attentional memory capacity to accommodate long sequences of physical reasoning steps. Second, the system uses a carefully constructed instruction dataset containing high-value question-answer pairs and standard chain-of-thought (CoT) trajectories to force-guide the model. Through SFT, the model is forced to master a specific response paradigm: when receiving material design instructions, the model must first, in a specific... <think>The plain text output within the tag represents the underlying physical and chemical logic derivation process. For aerogels, this includes, but is not limited to, valence balancing calculations, space group symmetry analysis, steric hindrance analysis for specific crosslinking agents, and preliminary estimation of crosslinking density. After a rigorous logical derivation loop is completed, the model must be immediately followed by... <answer>Within the tag, output the final generated material symbol sequence code (SMILES or CIF) cleanly and without redundancy. It is strictly forbidden to add any natural language interpretation within this tag.

[0133] In step S105, after the SFT stage, the model's mastery of chemical symbols and knowledge is judged by the symbolic ability score (SCS) and chemical ability score (CCS). Qualified models with SCS>1.5 and CCS>1.5, that is, those with high symbolic resolution and high mastery of chemical facts, are selected as the benchmark models for subsequent reinforcement learning.

[0134] The positive samples of CCS are true descriptions of aerogels that conform to objective scientific common sense (e.g., "Using MTMS precursors can significantly improve the hydrophobic properties and mechanical flexibility of aerogels"), while the negative samples are texts that lead to scientific logical inversions through random word substitution (e.g., "Using MTMS will cause aerogels to become extremely hydrophilic and the pores to collapse instantly").

[0135] Step S200: For the base point model selected in S100, perform reinforcement learning with material enhancement, and apply the format reward. Character matching reward Building upon this foundation, a lightweight physics simulation module is introduced to generate physics simulation rewards. This ensures that the sequences generated by the model conform to the objective laws of the physical world;

[0136] In step S201, during the exploration phase of reinforcement learning, when the model outputs a sequence of candidate structure symbols (such as a CIF file representing a fragment of amorphous silica mesoporous network or SMILES representing polyimide crosslinked clusters) based on aerogel design cues, the system does not directly award a reward. Instead, it immediately calls the embedded lightweight physics simulation module for evaluation. To address the enormous computational overhead of traditional density functional theory (DFT) calculations in reinforcement learning, which often involves millions of iterations, this module uses the M3GNet model. Faced with the extremely complex porous network structure of aerogel, which intertwines long-range disorder and short-range order, M3GNet can quickly read the initial coordinate set for generating the configuration. And accurately calculate the total energy of the system. And acting on each specific atom Force gradient vector on :

[0137] Instantaneous force gradient It is a core quantitative indicator for assessing whether there are abnormal bond lengths, non-physical bond angle distortions, or severe atomic spatial overlaps within the three-dimensional framework of aerogels.

[0138] In step S202, the module uses the conjugate gradient method to guide the atomic coordinate update based on the aforementioned force gradient, performing rapid structural relaxation. Its basic iterative update process can be represented as:

[0139]

[0140] in, The relaxation step size is adaptively adjusted based on the local curvature until the maximum force on the largest atom in the system is less than a preset threshold of 0.05 eV / Å, at which point the relaxed energy is obtained. , This indicates the current iteration step.

[0141] In step S203, after structural relaxation is completed, the module immediately calculates the core parameter characterizing the thermodynamic stability of the material: the enthalpy of formation. (In computational materials science, this is usually equivalent to formation energy) This is to determine the thermodynamic stability of the material (i.e., whether it can be synthesized in reality).

[0142] Generally speaking, <0 indicates that the material has thermodynamic stability. The larger the absolute negative value, the less likely the aerogel's skeletal structure is to undergo phase transition or degradation when exposed to high-temperature baking or severe mechanical stress.

[0143] Step S204: Map the above calculation results to the reward signal that the reinforcement learning model GRPO algorithm can receive. To avoid gradient explosion, an exponential decay function is used to normalize the energy feedback. The specific calculation formula is as follows:

[0144]

[0145] In this formula:

[0146] Preset hyperparameters to control the steepness of the penalty;

[0147] The expected stable energy threshold is set for specific aerogel applications (for example, when designing special silica aerogels that can withstand high temperatures of 1200°C, the threshold will be lowered accordingly).

[0148] Step S205: Based on the reward function mentioned in S200, calculate the global reward obtained by the final model in one training step (Step), and incorporate it into the reinforcement learning process.

[0149] After the above optimization steps, a core model that meets the requirements for aerogel synthesis is output for subsequent use by multi-agent systems.

[0150] Step S300: Based on the optimized core model from step S200, deploy the orchestration agent and the design agent, and call the embedded model and DFT simulation tools to deploy the retrieval agent and the simulation agent. The specific deployment details are shown in Table 2 below.

[0151] Orchestrator Agent LangGraph, FastAPI, Redis in-memory database Kubernetes (K8s) cloud-native cluster management node Serving as the neural center of the system, it manages the global state graph. Redis is used to persistently store the session context of the entire long-cycle aerogel R&D task, ensuring that the task progress can be recovered without loss after an asynchronous simulation interruption lasting several days, and the task delivery API is uniformly exposed to the outside world. Design Agent vLLM inference backend, RDKit / Pymatgen chemical hard-coded verification engine Local A100 / H100 high-performance GPU server array High-throughput concurrent calls to the core large model, which has undergone physical alignment optimization in the S200 stage. It receives constraint boundaries issued by the orchestration agent, performs chemical topological inference, and outputs candidate aerogel molecule sequences subject to strict syntax constraints. Retrieval Agent HuggingFaceTEI (TextEmbeddingsInference), Milvus vector database Local A100 GPU server node A "duplicate detection mechanism" for new material discovery is proposed. By using an embedding model, the generated sequence is transformed into a high-dimensional feature vector, and a millisecond-level similarity search is initiated in a vast material database to accurately filter out known configurations that have been disclosed in patents or literature, thus ensuring the novelty of the research and development. Simulation Agent Scheduling Layer Celery asynchronous task queue, RabbitMQ message bus, ASE middleware K8s internal cluster scheduling agent service node Upon receiving the compliant sequence from the orchestration agent, the agent service's built-in ASE middleware automatically converts the linear symbols into three-dimensional coordinates and calculation parameter initialization files supported by physics software such as LAMMPS or VASP, and pushes them into the message queue. Simulation Agent Computation Layer Slurm cross-network job scheduling system, physical simulation software stack (VASP, LAMMPS) Large, independently deployed high-performance computing (HPC) clusters Perform extremely CPU / GPU-intensive first-principles calculations and large-scale molecular dynamics tests (such as obtaining phonon spectra and heat flux autocorrelation functions). Upon completion, the extracted physical performance residuals are asynchronously awakened by a Webhook callback mechanism, waking up the suspended orchestration agent.

[0152] Step S400, user inputs industrial requirements: "Design a novel polyimide (PI) aerogel composite material with high temperature resistance, low thermal conductivity, and excellent mechanical flexibility."

[0153] Step S401: After receiving this macroscopic natural language objective, the orchestration agent translates it into system-level prompts containing real chemical boundary conditions using the built-in knowledge graph. To address this requirement, the system explicitly extracts "glass transition temperature ( The following performance boundary conditions are set as insurmountable: "C > 380°C", "Solid-state thermal conductivity at room temperature < 0.018 W / m·K", and "Ultimate compressive strain recovery rate > 80%". The orchestration agent sends the constructed state machine and constraint boundaries to the design agent via the RPC protocol.

[0154] In step S402, the intelligent agent calls the optimized core model to process the instruction. Since the model underwent mandatory paradigm fine-tuning in stage S100, it will now... <think>The model derivation involves in-depth chemical logic chain reasoning within the tags. For the design of balancing flexibility and heat resistance in PI aerogels, the derivation trajectory includes the following dimensions:

[0155] 1. Selection of backbone crosslinking network: Evaluate the properties of traditional diamine monomers (such as p-phenylenediamine pPDA) and ether-containing diamines (such as 4,4'-diaminodiphenyl ether ODA). It is reasoned that, to meet flexibility requirements, flexible ether bonds must be introduced into ODA to increase the rotational conformational freedom of the macromolecular chain segments.

[0156] 2. Prediction of Spatial Hindrance and Pore Distribution: The model evaluates the contribution of different multifunctional crosslinking agents to the three-dimensional network. While traditional small-molecule aromatic triamine crosslinking agents (such as TAB) can increase crosslinking density, they easily lead to high rigidity and brittle fracture. The model derivation tends to introduce octahedral oligomeric silsesquioxanes (OAPS) with cage-like nanostructures as nodes. The eight reactive arms of OAPS not only form a highly interpenetrating network, but their large three-dimensional volume effect can also play a "steric hindrance relief" role during gelation, effectively opening up the polymer backbone and preventing severe pore collapse during supercritical drying or high-temperature service, thus ensuring the ultra-high mesoporosity required for thermal insulation.

[0157] Step S403, after deep inference is completed, the model in <answer>The tag outputs the polymerization chemical equation and SMILES sequence of the best candidate network precursor. The search agent then vectorizes these sequences and initiates a search in Milvus. After confirming that there is no completely identical patent formulation in the library, this highly novel candidate solution is submitted to the validation queue.

[0158] In step S404, the orchestration agent transfers the validated sequences to the Celery task queue. The scheduling agent of the simulated agent, upon detecting this event, parses... <think>Analysis of the chemical reasoning logic left by the model in the tags. <think>The chemical reasoning logic within the tags utilizes a domain rule base to achieve adaptive mapping of computational tasks and automatic configuration of underlying computational parameters. For aerogels, this stage primarily involves the extremely complex simulation construction across two dimensions: macroscopic thermodynamics and microscopic mechanics.

[0159] A. Verification of ultra-low thermal conductivity using molecular dynamics (based on Green-Kubo theory of LAMMPS)

[0160] The reason why aerogels can approach the thermal conductivity limit of air lies in the fact that their dense nanoporous network induces extremely strong phonon scattering and severely restricts the mean free path of gas molecules (Knudsen Effect). To quantitatively verify the true thermal conductivity of the aerogel generated by the model, the proxy service calls AtomicSimulationEnvironment (ASE) to automatically generate an extremely complex LAMMPS molecular dynamics script.

[0161] In the generated LAMMPS script, the system first sets the atoms to achieve full thermal equilibrium for tens of thousands of steps in the NVT ensemble (isothermal and isochoric, controlled by the Nosé-Hoover heat bath), ensuring that the system's kinetic energy distribution conforms to the Maxwell-Boltzmann distribution. Once equilibrium is reached, the system seamlessly switches to the NVE ensemble (microcanonical ensemble, cutting off heat exchange with the external environment) and calls the core instruction computeheat / flux to continuously collect microscopic information. This instruction extracts and normalizes the transient macroscopic heat flux vector component of the system from the transient kinetic energy, potential energy distribution, and stress tensor component of each atom. The final thermal conductivity tensor Instead of simply calculating the mean, the result is obtained by performing an extremely lengthy integration operation in the time domain on the extracted Heat Current Auto-Correlation Function (HCACF) using a post-processing script (based on Python / Pandas) automatically written by the system. The calculation formula is as follows:

[0162]

[0163] In this formula, Represents the Boltzmann constant. Set the ambient temperature (e.g., 150K for extremely cold environments, or 800K for heat resistance tests). This represents the total volume of the constructed three-dimensional simulated cell. Because heat conduction within the aerogel is dominated by low-frequency phonons and follows an extremely tortuous path, the fluctuation decay of HCACF is very slow. Therefore, the simulation intelligently and adaptively extends the LAMMPS correlation length and sampling interval integration step size to tens or even hundreds of picoseconds (ps) to ensure the final convergence of the integral curve and the accuracy of prediction, avoiding the illusion of excessively low thermal conductivity.

[0164] B. Mesoscopic mechanical modulus and simulation of high-temperature structural collapse failure

[0165] To verify the mechanical and thermal stability boundary conditions, the proxy service simultaneously configures deformation testing schemes based on first-principles calculations or coarse-grained molecular dynamics. The system applies uniaxial compression to the generated 3D unit cell model, monitors the stress response of the framework network, and plots a complete nonlinear stress-strain curve. By fitting the slope of the linear elastic region, the system extracts Young's modulus. Compared to Poisson's ratio. The physical simulation rule base pre-sets calibration rules, such as verifying whether the modulus strictly follows the power-law relationship of aerogel characteristics ( This allows the system to determine the validity of the skeleton connection. Furthermore, the system incorporates a phased heating program into the computational engine to track the evolution of the radial distribution function (RDF) in real time as the material heats from 300K to 1200K, accurately capturing potential macropore wall softening, fracture, and catastrophic pore collapse leading to material densification.

[0166] After all the underlying configurations are assembled into standardized cluster scheduling commands for VASP or LAMMPS, the simulated agent pushes this massive computational load to the HPC supercomputing center through the Slurm system. At the same time, the orchestration agent marks the node of this project in the memory graph as "suspended pending verification," releasing control of the resources it occupies.

[0167] In step S500, if the verification in step S400 fails, the simulated agent serializes the unconverged structural defects, force gradients, or physical error data, constructing a negative feedback sample in JSON format. The design agent receives this proof, dynamically adjusts the generation parameters and redesigns within the current dialogue context until the target performance is met, ultimately outputting a verifiable experimental scheme.

[0168] In step S501, when the HPC cluster computing task is completed or terminates abnormally due to non-convergence of the physical model, the agent service in the simulated agent is activated. The ASE middleware within the agent service begins automated deep scanning and reverse parsing of the unstructured output stacks generated by the underlying physical software (such as the verbose OUTCAR and OSZICAR logs of VASP, or the log.lammps thermodynamic output sequence of LAMMPS).

[0169] At this point, the agent service invokes its built-in core module—the "Expert Rule Tree for Chemical Anomalies." The system substitutes the extracted thermodynamic and mechanical values ​​into the rule tree for threshold cross-boundary comparison. If any parameter (such as thermal conductivity or tensile limit) fails to meet the performance boundary set in S401, or if mathematical divergence occurs during the calculation, the system will immediately trigger the "Online Physical Fusing Mechanism."

[0170] To ensure that error messages are sent back to the large language model without causing any parsing ambiguity, the proxy service packages all extracted physical deviations, crash point geometric features, and numerical matrices into a "failure proof" payload (JSONPayload) strictly bound by JSONSchema strong typing constraints through a template rendering engine. This failure proof is structurally divided into three deep attribution modules, as shown in Table 3 below:

[0171] Target vs. Result Deviation Module "deviation_metrics":{"target_conductivity":0.018,"actual_conductivity":0.035,"triggered_fuse":true} Record the macroscopic physical characteristics that trigger the fuse, and clearly indicate whether the design failure is caused by thermal conductivity (e.g., actual value of 0.035 far exceeding the 0.018 threshold) or mechanical properties (e.g., premature brittle fracture). Microscopic physical and geometric feature module "microstructural_failure":{"failure_type":"pore_collapse_at_HT","critical_temp_K":850,"structural_distortion_vector":[...]} This is the core of the proof. It doesn't just report an "error," but also explains "why it's wrong." For example, it records what was captured by MD simulation: due to intensified high-temperature thermal vibration, the macropore walls around the 850K threshold undergo fusion and densification, leading to severe irreversible shrinkage of the nanoporous network (Pore Collapse), causing the material to lose its thermal insulation mechanism. Abnormal interruption diagnosis module "computational_abort":{"dft_scf_convergence":false,"sloshing_detected":true,"homo_lumo_gap_eV":0.02} This is used to handle numerical crashes in the computing platform itself. For example, if VASP fails to converge in an electronically self-consistent field (SCF) calculation, the feedback will indicate that it is due to severe charge oscillations (ChargeSloshing) caused by an excessively small HOMO-LUMO bandgap resulting from doping, reminding the large model to adjust the doping sites or rethink the electronic structure symmetry.

[0172] In steps S502-S504, after the orchestration agent receives the serialized and verified JSON text, it appends it as the latest "SystemObservation" in the current progress line to the end of the long-range dialogue history (Memory) of the development task in Redis, and then sends a retry request to the sleeping design agent via the RPC protocol.

[0173] After the intelligent agent perceives a severe "proof of failure" in the historical context, it will trigger another thought chain with depth. <think>The system uses tags to perform highly targeted source tracing, reflection, and parameter adjustments. When faced with various failure proofs, it dynamically attaches a matching underlying chemical error correction strategy.

[0174] Response Strategy A: Equivalent / Isoelectronic Substitution and Introduction of Flexible Segments

[0175] If the deviation module of the failure proof indicates that the tensile mechanics verification shows that the material undergoes brittle fracture at only 8% strain (far below the target flexibility requirement), and the micro-geometry module marks the stress concentration points as being distributed in high-density aromatic ring cross-linked clusters, the design intelligence will prompt a profound reflection on the problem of excessive stiffness in the skeleton. In the new... <think>During inference, the model triggers an equivalent substitution and conformational adjustment strategy: decisively reducing the molar proportion of extremely rigid traditional small-molecule crosslinking agents (such as TAB) in the formulation, and instead introducing a large number of OAPS crosslinking agents with spatial octahedral cage structures into the network nodes. The core of OAPS is the inorganic Si-O-Si flexible bond with a high degree of freedom of bond angle rotation. The model determines through logical deduction that this atomic-level substitution can not only release localized stress concentration points, but its eight divergent reactive arms can also effectively expand the three-dimensional skeleton at the mesoscale to buffer mechanical impacts, thereby fundamentally reversing brittleness and generating a new generation of improved material sequences with excellent strain recovery rates.

[0176] Response Strategy B: Spatial Obstacle Elimination and Side Base Trimming Strategy

[0177] If the failure proves that although the porosity of the feedback material meets the standard, MD simulation shows a huge positive deviation in its thermodynamic formation energy, and VASP phonon dispersive spectroscopy calculations capture severe "imaginary frequency" accompanied by geometric distortion of specific atomic groups, the design intelligence will immediately identify that this stems from excessive steric repulsion generated locally during skeletal polymerization. To eliminate this instability, the model will trigger a "steric hindrance elimination strategy": actively performing atomic-level "surgery" on the generated polymer side chains, such as precisely cutting and replacing large-volume side groups that cause severe compression (such as tert-butyl) with smaller, more stable methyl groups (-CH3), or inserting a longer carbon chain flexible "bridge" group between two large-volume crosslinking hubs, completely eliminating the bonding potential barrier anomaly caused by steric hindrance.

[0178] Response Strategy C: Dynamic Adaptive Fine-Tuning of Precursor Cocondensation Ratio

[0179] If the failure indicates that the high humidity of the testing environment caused capillary tension-induced pulverization and collapse (sharp reduction in pore volume) of the traditional pure tetraethyl orthosilicate (TEOS) framework, the design intelligence will invoke the domain theorem related to material interfacial tension, significantly reducing the absolute proportion of TEOS in the formulation and instead incorporating methyltrimethoxysilane (MTMS), which contains nonpolar groups, as a co-condensation precursor. The derivation shows that the high density of methyl (-CH3) side groups in MTMS coating the pore wall surface provides strong superhydrophobicity and a spring-like repulsive force from the pore wall, perfectly resisting the densification shrinkage caused by surface tension, thereby protecting the stable retention of the ultra-low thermal conductivity properties.

[0180] Step S505, when designing the intelligent agent in <think>After completing the above meticulous and complex diagnosis and plan reconstruction, it will be in a completely new... <answer> The tag outputs an iteratively revised and optimized sequence of chemical symbols (SMILES / CIF). This new sequence is then pushed to the system bus, reawakening the dormant retrieval and simulation agents, and re-executing the rigorous verification process from S403 to S405. This asynchronous closed loop continues to iterate online until a material meeting the target performance is generated, or the preset maximum number of retries is reached.< / answer> < / think> < / think> < / think> < / think> < / think> < / answer> < / think> < / answer> < / think> < / answer> < / think> < / think> < / answer> < / think> < / think> < / answer> < / answer> < / think> < / answer> < / think> < / answer> < / answer> < / think> < / answer> < / think> < / answer> < / think> < / answer> < / think>

Claims

1. A method for AI-aided material design and collaborative research and development based on a large language model, characterized in that, include: Step S100: High-quality corpus is extracted using TCS scoring and fused with the crystallographic features of the material. Based on the corpus, the base model is trained using CPT and SFT, and the model output is assigned accordingly. <think>and <answer> The structuring ability; finally, the model with symbolic ability score SCS>1.5 and chemical ability score CCS>1.5 is input into the next stage as the basis for subsequent reinforcement;< / answer> < / think> Step S200: Perform material-enhanced reinforcement learning on the base point models selected in S100, with the format reward... Character matching reward Building upon this foundation, a lightweight physics simulation module is introduced to generate physics simulation rewards. The specific calculation formula is as follows: ； In this formula: Preset hyperparameters to control the steepness of the penalty; The expected material stability energy threshold; The generation enthalpy of the material generated by the model is obtained through simulation calculation; the sequence generated by the model conforms to the objective laws of the physical world. Step S300: Based on the core model optimized in step S200, deploy orchestration agents and design agents, and call embedded models and DFT simulation tools to deploy retrieval agents and simulation agents; In step S400, after receiving the material research and development goal input by the user, the orchestration agent calls the model optimized in step S200 to decompose the task, collaboratively calls the design agent to generate a material generation scheme with structured tags, uses the retrieval agent to perform novelty verification, and uses the simulation agent to automatically configure computational chemical parameters to perform physical performance verification. In step S500, if the verification in step S400 fails, the simulated agent serializes the unconverged structural defects, force gradients, or physical error data and constructs a negative feedback sample in JSON format. The design agent receives this proof, dynamically adjusts the generation parameters and redesigns in the current dialogue context until the target performance is met, and finally outputs a verifiable experimental scheme.

2. The AI-aided material design and collaborative R&D method based on a large language model according to claim 1, characterized in that, The specific process of step S100 includes: Step S101: Obtain multi-source raw data from large-scale datasets such as MatMech and MateriMiner; first, perform deduplication, non-text element parsing, and basic format cleaning on the raw data; then, introduce a dynamic filtering mechanism based on text chemical scores, using word frequency ratios... Weighted summation is performed to remove low-information text and noisy text paragraphs from non-chemistry fields, retaining samples with TCS>4 to obtain high-density chemical knowledge text. Step S102: Use a named entity recognition tool to scan the high-density chemical knowledge text filtered in step S101, locate the molecular names or crystal structure description entities, call RDKit or Pymatgen, and hard-code the extracted entities into a standardized linear symbol sequence; then, forcibly insert preset exclusive delimiters [START_SMILES] and [END_SMILES], or [START_CIF] and [END_CIF] before and after the symbol sequence, and interweave them with the original natural language context; Step S103: Perform continuous pre-training on the base model based on the corpus obtained in step S102, i.e., unsupervised training based on "next word source prediction"; Step S104: After CPT training is completed, the model is fine-tuned using an instruction dataset containing high-value question-answer pairs and scientific thought chain trajectories. Step S105: The model’s mastery of chemical symbols and knowledge is judged by the symbolic ability score (SCS) and chemical ability score (CCS). Qualified models with SCS>1.5 and CCS>1.5, that is, those with high symbolic resolution and high mastery of chemical facts, are selected as the benchmark models for subsequent reinforcement learning.

3. The AI-aided material design and collaborative R&D method based on a large language model according to claim 2, characterized in that, The TCS in step S101 is calculated using the following formula: ; Weight The nonlinear mapping between the indicator function and the frequency ratio is employed, specifically as follows: ； In this formula: This represents the total number of all lexical units in the text segment to be evaluated; It is a word element Relative frequency in chemical text corpora; It is a word element Relative frequency in ordinary text corpora; It is a word element The weights; In step S104, the following operations should be performed during the SFT process: 1) Master three specific paradigms through SFT-guided models: (i) symbolic understanding of molecular or material structures; (ii) scientific question answering with structure awareness; and (iii) step-by-step reasoning in the field of chemistry and materials. 2) To accommodate complex material design tasks and lengthy crystal structure files, the maximum context window length of the model is extended during the SFT stage to accommodate long sequences of physical inference steps; 3) Train the model using a structured feedback dataset containing a standard format, forcing it to respond to scientific instructions in specific... <think>The tag outputs the underlying physical / chemical logic derivation process, followed immediately by... <answer> The label outputs the final material symbol sequence in its pure form;< / answer> < / think> Specifically, SCS and CCS in step S105 are as follows: SCS is used to measure the model's mastery of the syntax of chemical symbols, as follows: Step 1: Calculate the mean log-likelihood of the sequence. The formula is as follows: ； In this formula: For a given model; The number of tokens; For a given sequence of symbols ; Step 2: Calculate the distribution difference between the correct sequence and the corrupted sequence. The calculation formula is as follows: ； In this formula: It is a set of completely correct normalized sequences; It is a set of illegal sequences generated by randomly deleting characters; The average likelihood of the canonical sequence; The average likelihood of the corrupted sequence; To combine the standard deviations, the formula is as follows: ; In this formula: It is a set of correct sequence samples. The sample size, i.e., the total number of canonical sequences; It is a corrupted sequence sample set The sample size, i.e., the total number of illegal sequences; It is a set of correct sequence samples. Standard deviation of the likelihood of each sequence in the dataset; It is a corrupted sequence sample set Standard deviation of the likelihood of each sequence in the dataset; CCS is used to measure the model's grasp of "chemical facts." Positive samples are authentic chemical / material descriptions that conform to common scientific sense, while negative samples are texts that contain scientific errors due to random sentence replacements. The formula is as follows: ； In this formula: The average likelihood of the model for factually correct text; This represents the average likelihood of the tampered text.

4. The AI-aided material design and collaborative R&D method based on a large language model according to claim 1, characterized in that, The specific process of step S200 includes: Step S201: Input material cue words into the model to generate the corresponding candidate material symbol sequence CIF, and call the lightweight physics simulation module, namely the general machine learning potential function M3GNet model, to calculate the generated material configuration. Total system energy and atoms Force The calculation formula is: ; In step S202, the module uses the conjugate gradient method to guide the atomic coordinate update based on the above-mentioned force gradient, and performs rapid structural relaxation; its basic iterative update process can be expressed as: ； in, The relaxation step size is determined until the maximum force on the largest atom in the system is less than a preset threshold of 0.05 eV / Å, at which point the relaxed energy is obtained. , Indicates the current iteration step; Step S203: The module calculates the enthalpy of formation for each atom. To determine the thermodynamic stability of a material, i.e. whether it can be synthesized in reality, the calculation formula is: ； In this formula: This represents the total energy after the unit cell is relaxed. For elements The number of atoms; For elements Chemical potential in its standard elemental state; in, <0 indicates that the material has thermodynamic stability; The lower the value, the more stable the material. Step S204, the enthalpy of formation calculated in step S203 is... Mapped to the reward signal received by the reinforcement learning model GRPO algorithm To avoid gradient explosion, an exponential decay function is used to normalize the energy feedback. Step S205, combining the reward function mentioned in S200, the final global reward obtained by the model in one training step can be quantified as: ； In this formula: Formatted Rewards Weighting coefficients; Character matching reward Weighting coefficients; Rewards for physical simulation Weighting coefficients; Rewards are given for adherence to format and structure; Rewards for symbolic syntax and chemical legitimacy; Rewards are given for physical simulation.

5. The AI-aided material design and collaborative R&D method based on a large language model according to claim 4, characterized in that, In step S205 To evaluate whether the model strictly adheres to the reward function of the predefined structured output instructions, regular expressions are used to determine whether the model's output text contains a complete pair of... <think>Labels and a complete pair <answer>Tags, and <think>Must <answer> Previously, positive rewards were given for correct actions and negative punishments were given for incorrect actions;< / answer> < / think> < / answer> < / think> To evaluate the model <answer> The system checks whether the generated symbol sequence in the label is a "valid code" segment. RDKit or Pymatgen will be used to determine whether the structure is valid, with positive rewards for correct answers and negative penalties for incorrect answers.< / answer> 6. The AI-aided material design and collaborative R&D method based on a large language model according to claim 1, characterized in that, The specific process of step S300 includes: Step S301: Deploy orchestration agents using a Kubernetes cluster, with LangGraph as its core framework to manage the workflow and state graph of all agents. Use FastAPI to expose interfaces to call the core model, and use Redis to store the session context and current execution state of the entire R&D task to ensure that the task can be recovered after interruption. Step S302: Deploy the design agent on the local A100 server. The inference backend is vLLM. The core model is called with high concurrency. The basic format of the generated data is quickly hard-coded and verified using RDKit or Pymatgen. The generated prompt words assigned by the orchestration agent are received through RESTAPI, and the generated structure sequence is returned. Step S303: Deploy the retrieval agent on the local A100 server, deploy the Milvus management vector database, and deploy a dedicated small chemical embedding model using HuggingFace's TEI. Step S304: Deploy the simulated agent using a "cloud agent and HPC cluster"; deploy a scheduling agent in the K8s cluster, and build an asynchronous task queue using Celery and RabbitMQ to listen for simulation requests; the agent service uses the ASE middleware to parse the large model output into a standardized input file for VASP; then, submit the task across the network to the high-performance computing cluster for execution through the Slurm scheduling system, parse the output logs after completion, and asynchronously return the physical performance indicators to the orchestration agent through the Webhook callback mechanism.

7. The AI-aided material design and collaborative R&D method based on a large language model according to claim 1, characterized in that, The specific process of step S400 includes: In steps S401 and S300, the orchestration agent receives the material research and development goal input by the user and transforms it into system prompts containing real chemical boundary conditions. In addition to the basic settings of temperature, pressure, and solvent environment, the system will also adaptively extract microscopic constraints. Subsequently, the orchestration agent uses an in-memory database to persistently store the session context and state machine of the current research and development task, and then sends the prompts to the design agent through remote procedure calls. Step S402: Design the agent to call the S200 core model to process the prompt word, forcing the model to output with... <think>and <answer>The response sequence is structured and tagged, and the generated result is published to the asynchronous message bus RabbitMQ; where, <think>The tags contain the chemical reasoning logic generated by the model, the expected electronic structure characteristics, and the steric hindrance analysis; <answer> The label contains only the generated candidate material structure;< / answer> < / think> < / answer> < / think> Step S403, precise extraction of the retrieval agent. <answer> The symbol sequence within the label first calls the embedding model to vectorize the sequence, then initiates a high-dimensional similarity retrieval to the independently deployed vector database cluster Milvus, filters out known materials, and returns a list of novel materials that have passed the screening to the orchestration agent;< / answer> In step S404, the orchestration agent packages the validated material sequences into simulation tasks and pushes them to the task queue Celery; the simulation agent agent service listens to this queue and parses the data. <think> The chemical reasoning logic within the tags is based on a domain rule base to achieve adaptive mapping of computational tasks and automatic configuration of underlying computational parameters. Specific mapping rules include: for alloy catalytic materials, automatic slicing to establish surface models and configuring density functional theory for structural optimization to calculate the surface adsorption energies of their d-band centers and key intermediates; for solid electrolyte materials, automatic configuration of ab initio molecular dynamics calculations and extraction of ion mean square shifts to verify macroscopic ionic conductivity; and for high-temperature resistant polymer tasks, configuration of all-atom molecular dynamics calculations to determine their free volume ratio and glass transition temperature. After configuration, the system generates corresponding standardized input files.< / think> In step S405, the simulated agent submits the computing task to the high-performance computing cluster for execution through the cluster scheduling system; during this period, the orchestration agent sets the session state of the research and development task to the "asynchronous suspension" state, releases the front-end computing resources, and waits for the physical simulation results.

8. The AI-aided material design and collaborative R&D method based on a large language model according to claim 7, characterized in that, The specific process of step S500 includes: In step S501, after the HPC cluster completes the computing task, the simulation agent agent service captures and parses the underlying output logs, and sends a completion signal through the Webhook callback mechanism or the result message queue; the orchestration agent is awakened after receiving the signal that the physical simulation calculation is complete, and reloads and restores the context state of the task from the memory database. In step S502, the system automatically compares the extracted physical calculation results with the initial R&D goals set in step S401. If each physical property meets the target threshold, the candidate material structure and its experimental scheme are output as the final result, and the process ends. If not, the online physical melting mechanism is triggered immediately. In step S503, the simulated agent serializes the specific physical error data that leads to melting and packages it into a "failure proof" negative feedback load with strict JSON format. This proof deeply integrates microscopic failure characteristics. If it is thermodynamically unstable, the "imaginary frequency" and the corresponding geometric distortion vector appearing in the phonon dispersion spectrum calculation will be extracted. If it is electronically unsatisfactory, key parameters that determine the material performance failure, such as "wide bandgap leading to limited light absorption" or "high adsorption free energy", will be clearly marked. Step S504: The orchestration agent appends the "failure proof" as a new prompt to the current dialogue history and sends it back to the design agent via RPC; the design agent then... <think>In the inference space, the system is forced to first analyze the physical defects in the "failure proof," and then invoke relevant theorems in chemical materials science to dynamically adjust the atomic substitution strategy or spatial configuration. Specific correction strategies include: if the failure proof shows imaginary frequencies in the phonon spectrum or extreme force gradients causing structural non-convergence, the design agent will trigger a "steric hindrance elimination strategy," actively trimming large-volume side groups or extending flexible chains to generate a new configuration; if the failure proof shows electronic band mismatch or excessively high catalytic barriers, the design agent will trigger an "equivalent / isoelectronic element substitution strategy" or a "functional group-induced strategy," dynamically adjusting the d-band center of the material or narrowing the band gap, thereby generating targeted corrections. <answer> Candidate sequences;< / answer> < / think> In step S505, the design agent pushes the revised new sequence to the asynchronous message bus again, triggering the verification process of steps S403 to S405 again, until a material that meets the target performance is generated, or the preset maximum number of retries is reached.

9. The AI-aided material design and collaborative R&D method based on a large language model according to claim 8, characterized in that, "Proof of Failure" specifically includes: a. Target-Result Deviation Module, including the R&D target setting value and the actual simulation calculation value, clearly indicating which physical indicator triggered the circuit breaker; b. Microscopic Physical and Geometric Feature Module, including structural convergence features and spatial relaxation features, recording the physical microscopic reasons that led to substandard performance; c. Abnormal Interruption Diagnosis, if the density functional theory calculation collapses in the electron step self-consistent field loop, extracting the cause of the collapse.

10. The AI-aided material design and collaborative R&D method based on a large language model according to claim 8 or 9, characterized in that, The "failure proof" in step S503 is constructed as follows: Step S5031: When the physical simulation task on the HPC cluster ends or terminates abnormally, the simulation agent agent service calls the embedded atomic simulation environment ASE to automatically scan and parse the unstructured output files generated by the underlying physical software VASP. Step S5032: The agent service has a built-in "chemical anomaly expert rule tree". The system inputs the parsed log data into the rule tree for threshold comparison. In step S5033, the proxy service encapsulates all extracted numerical deviations and abnormal features into JSON serialization, and on this basis, calls the template rendering engine to convert the JSON data into prompt words that are easy for the large language model to understand. In step S5034, the orchestration agent finally appends this serialized text as the latest "system observation" seamlessly to the end of the current dialogue context and sends a retry request to the design agent; the design agent is forcibly triggered upon recognizing this "proof of failure". <think> Tags, execution of reflection and error-correcting reasoning.< / think>

Images