Solid-state refrigeration material intelligent screening method and device based on large model and agent

Abstract

The application provides a solid-state refrigeration material intelligent screening method and device based on a large model and an agent, which realizes efficient discovery and interpretable evaluation of a solid-state refrigeration material candidate through fusion of field-specific knowledge base construction, proxy model, physical constraint active learning and multi-fidelity calculation screening, shortens the period from research idea to high-trust candidate material screening report, and makes all recommended conclusions traceable to specific literature evidence, calculation parameters and model version, thereby providing a systematic solution for accelerated calculation discovery of a solid-state refrigeration material.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and device for intelligent screening of solid-state refrigeration materials based on large models and intelligent agents. Background Technology

[0002] Global refrigeration and air conditioning systems account for approximately one-fifth of total electricity consumption, and this figure is projected to double in the coming decades, driven by climate change. Traditional vapor compression refrigeration technologies heavily rely on refrigerants with high global warming potential (GWP), such as hydrofluorocarbons (HFCs) and hydrochlorofluorocarbons (HCFCs). The Kigali Amendment to the Montreal Protocol has incorporated HFCs into its regulations and is promoting a gradual global reduction, providing sustained policy support for low-carbon alternative refrigeration technologies.

[0003] Solid-state caloric effects (SCE) are a class of phenomena that utilize the reversible phase transition of functional materials under the influence of an external driving field, resulting in significant isothermal entropy changes (ΔS) and adiabatic temperature changes (ΔT). ad This provides the physical mechanism for the heat pump effect in refrigeration cycles. Based on the type of driving field, solid-state refrigeration effects can be subdivided into four categories: Electrocaloric Effect (ECE, driving field: electric field E, unit kV / cm), Magnetocaloric Effect (MCE, driving field: magnetic field H, unit T), Elasticocaloric Effect (eCE, driving field: uniaxial stress / strain σ / ε, unit MPa / %), and Barocaloric Effect (BCE, driving field: hydrostatic pressure p, unit GPa). Typical solid-state refrigeration material systems include: perovskite ferroelectric ceramics (BaTiO3-based, PMN-PT system, ECE), lead-based relaxor ferroelectrics (PZT-PT system, ECE), polyvinylidene fluoride (PVDF) polymer films (ECE), gadolinium-based alloys, and La(Fe,Si). 13 Solid-state refrigeration devices include shape memory alloys (MCE), NiTi shape memory alloys (eCE), and organic-inorganic hybrid perovskites (BCE). Compared to traditional vapor compression refrigeration, solid-state refrigeration devices operate entirely in a solid state with zero refrigerant leakage. They are naturally compatible with thin film deposition and micro / nano fabrication processes, giving them a unique competitive advantage in applications such as heat dissipation for microelectronic devices, cooling for flexible wearables, and precision temperature control in medical applications. They are widely recognized as one of the most promising new solid-state refrigeration technologies for industrialization.

[0004] However, current solid-state phase change refrigeration materials face multiple deep-seated physical and engineering bottlenecks:

[0005] (1) The constraint between strong effect and availability: Although a strong first-order phase transition can produce a huge entropy change near the phase transition temperature, it is often accompanied by a high driving field requirement (electric card: close to the material breakdown field strength; magnetic card: requires a strong magnetic field at the level of a superconducting magnet; spring card / pressure card: close to the material fatigue limit), a narrow effective operating temperature range (usually only ±5~10K), and material cyclic fatigue problems. How to maintain a high ΔT while widening the effective temperature range and reducing the driving field threshold and fatigue risk is the core constraint contradiction in the current solid-state refrigeration material design.

[0006] (2) It is difficult to achieve multiple performance objectives simultaneously: an ideal refrigeration material must simultaneously meet the following requirements: high ΔS / ΔT, low driving field threshold, low thermal hysteresis, wide operating temperature range, high cycle life, good thermal conductivity, and processability. There are complex mutual constraints among these indicators, forming a de facto "performance impossible triangle".

[0007] (3) The chemical space is huge: the possible combinations of elements, doping ratios, crystal structures and processing parameters form an astronomical number of candidate spaces. The traditional "trial and error method" often takes 10 to 20 years from conception to commercialization, which is extremely costly and inefficient.

[0008] (4) Domain knowledge is highly fragmented: key experimental data (ΔS, ΔT, thermal hysteresis, driving field, cycle life, test conditions, etc.) are scattered in a large number of unstructured documents, and the test conditions are different and difficult to compare directly, resulting in low efficiency of experience accumulation. Summary of the Invention

[0009] This application proposes a method and device for intelligent screening of solid-state cooling materials based on large models and intelligent agents, which can solve one of the problems existing in the background technology.

[0010] To achieve the above objectives, this application adopts the following technical solution:

[0011] Firstly, a method for intelligent screening of solid-state cooling materials based on a large model and intelligent agents is provided, including:

[0012] The task of screening solid-state phase change refrigeration materials was obtained.

[0013] Furthermore, based on the established cross-modal knowledge base of solid-state refrigeration materials, the constructed surrogate model is used to perform multi-level calculations and verifications of the screening task according to the directed acyclic graph workflow, thereby obtaining the processing results of the screening task. The directed acyclic graph workflow includes: candidate generation and step-by-step calculation, screening, evaluation, and recommendation. The output of candidate generation serves as the input for step-by-step calculation, screening, evaluation, and recommendation. The constraint mechanism of the surrogate model is implemented by embedding phase boundary sensing weights in the physical constraint optimization acquisition function. The phase boundary sensing weights are kernel functions of the normalized distance from the state point of the candidate material to the phase boundary of the phase diagram, and increase as the normalized distance decreases. The phase boundary of the phase diagram is determined by the change in stable phase labels of adjacent regions after the composition-state space is divided by the phase region classifier. The physical constraint optimization acquisition function is jointly determined by the prediction mean and prediction uncertainty of the surrogate model. The phase boundary sensing weights act multiplicatively on the standard Bayesian optimization acquisition function. Thus, before selecting the next batch of candidates for high-fidelity calculation in each round, the sampling priority of the candidates is redistributed.

[0014] In one possible design of the first aspect, the candidate generation process includes:

[0015] Based on the aforementioned screening task, keyword retrieval and dense vector retrieval are performed;

[0016] Material knowledge-constrained filtering is applied to the results of keyword retrieval and dense vector retrieval.

[0017] The ranking results of keyword retrieval and dense vector retrieval after filtering by material knowledge constraints;

[0018] Furthermore, a second ranking is performed on the merged Top-N results to obtain the candidate generation results.

[0019] In one possible design approach of the first aspect, the candidate generation adopts an iterative retrieval mechanism, which breaks down the complex problem into several verifiable sub-problems, and after each round of retrieval, automatically generates the next round of retrieval query based on evidence coverage, conflict status, and missing fields, until the iterative convergence condition is met.

[0020] In one possible design approach of the first aspect, the step-by-step calculation and filtering process includes:

[0021] A lightweight machine learning model is used for fast screening of the first-level L1.

[0022] Based on rapid screening, a second-level L2 precise screening is performed based on first principles.

[0023] Based on precise screening and manual approval, high-fidelity calculations of the third level (L3) are performed using molecular dynamics simulations.

[0024] In one possible design approach of the first aspect, a multi-fidelity machine learning method is used for bias correction between different levels of computational precision.

[0025] In one possible design approach of the first aspect, the agent model's multi-objective scoring function for:

[0026] in, The overall score for candidate materials ranges from 0 to 1. Score by performance; For stability; For preparability; Based on novelty; Cost-based; to Configurable weights.

[0027] In one possible design approach of the first aspect, the intelligent screening method for solid-state cooling materials based on large models and intelligent agents further includes:

[0028] The computational data generated from multi-level computational verification is written back to the cross-modal knowledge base of solid-state refrigeration materials, and the proxy model is updated based on the computational data to form a continuously evolving closed loop.

[0029] In one possible design approach of the first aspect, the physical constraint optimization acquisition function for:

[0030]

[0031]

[0032]

[0033] in, This is the physical constraint acquisition function; Optimize the acquisition function for standard Bayesian methods; For phase boundary sensing weight function; These are the composition points of the candidate materials; Candidate component points Distance to the phase diagram boundary (Boundary); For bandwidth parameters; Let be the normalized state vector of the candidate material, where As a component variable, For temperature, As the driving field, For the sake of pressure; in actual implementation, only the dimensions required by the target system are retained; For the phase boundary point set, For the first One phase boundary point; For the first The scale normalization parameter of the dimension can be taken as the allowable range of the dimension, the standard deviation of the training data, or a preset physical scale; To explore the lower bound, a non-zero sampling probability is retained in regions far from the currently estimated phase boundary. For the standard Bayesian optimized acquisition function after nonnegativity processing, the desired improvement is taken. Or the upper confidence bound after nonnegative transformation ; and distance The smaller, The larger.

[0034] In one possible design approach of the first aspect, the solid-state refrigeration material cross-modal knowledge base is constructed based on open databases, domain literature, and project-generated computational results; the solid-state refrigeration material intelligent screening method based on large models and intelligent agents further includes: preprocessing the data in the open databases, domain literature, and project-generated computational results to obtain the dataset of the solid-state refrigeration material cross-modal knowledge base, wherein the preprocessing includes: multimodal data preprocessing, entity relation extraction based on instruction fine-tuning of the large language model, and cross-validation and confidence scoring; the solid-state refrigeration material cross-modal knowledge base logically includes: a material data domain, a workflow record domain, and an evidence chain domain.

[0035] In a second aspect, an electronic device is provided, comprising: a processor and a memory coupled to the processor, the memory for storing a computer program; the processor for executing the computer program stored in the memory such that the electronic device performs the intelligent screening method for solid-state cooling materials based on large models and agents as described in any possible implementation of the first aspect.

[0036] Beneficial effects:

[0037] Based on the above technical solutions, by integrating domain-specific knowledge base construction, proxy models, physical constraint active learning, and multi-fidelity computational screening, we can achieve efficient discovery and interpretable evaluation of solid-state refrigeration material candidates. This shortens the cycle from research ideas to high-confidence candidate material screening reports and makes all recommended conclusions traceable to specific literature evidence, computational parameters, and model versions, providing a systematic solution for accelerated computational discovery of solid-state refrigeration materials. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the solid-state cooling phase change material smart body architecture provided in the embodiments of this application;

[0040] Figure 2 This is a schematic diagram of the API flow of solid phase change refrigeration materials provided in the embodiments of this application;

[0041] Figure 3 This is a schematic diagram of the dual-channel hybrid retrieval architecture provided in the embodiments of this application;

[0042] Figure 4 This is an iterative retrieval flowchart for complex scientific problems provided in the embodiments of this application;

[0043] Figure 5 This is a diagram of the overall architecture of the intelligent discovery system for card effect materials provided in this application embodiment;

[0044] Figure 6 This is a schematic diagram of the Langgraph state machine node definition provided in an embodiment of this application;

[0045] Figure 7 This is a flowchart of the active learning closed-loop process provided in the embodiments of this application;

[0046] Figure 8 This is a complete closed-loop flowchart of solid-state refrigeration phase change refrigeration materials provided in the embodiments of this application. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification and the above-mentioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0050] This embodiment aims to provide an intelligent discovery system and method for solid-state refrigeration materials based on a large language model and multi-agent collaboration. By integrating domain-specific knowledge base construction, hierarchical multi-agent collaboration, physical constraint active learning, and multi-fidelity computational screening, it achieves efficient discovery of solid-state refrigeration material candidates, interpretable evaluation, and generation of highly reliable computational screening reports. This shortens the cycle from "research idea to highly reliable candidate material screening report" from months to days, and makes all recommended conclusions traceable to specific literature evidence, computational parameters, and model versions.

[0051] This embodiment provides an intelligent discovery system for solid-state refrigeration materials based on a large language model and multi-agent collaboration, comprising the following five core modules:

[0052] (i) A cross-modal knowledge base for solid-state refrigeration materials (Caloric-DB) and a retrieval-enhanced generation engine;

[0053] (ii) Hierarchical multi-agent cooperative network;

[0054] (III) Physically Constrained Bayesian Active Learning and Multi-Objective Optimization Module;

[0055] (iv) Multi-fidelity calculation and screening production line;

[0056] (v) Computational closed loop and multi-fidelity data-driven knowledge continuous update mechanism.

[0057] By constructing such Figure 1 The solid-state cooling phase change material intelligent agent shown in the "Knowledge Base - Retrieval Enhanced Generation (RAG) Retrieval - Intelligent Agent Collaboration - Physical Constraint Active Learning - Multi-fidelity Simulation Screening - Closed-Loop Update" is also a hierarchical model architecture. This embodiment can reduce the number of low-confidence candidates entering the high-cost calculation stage during the candidate generation stage, and prioritize the selection of high-potential material components close to the solid-state phase change boundary through the phase boundary sensing acquisition function, thereby increasing the number of high-potential candidate materials obtained per unit computing power under a fixed computing power budget.

[0058] like Figure 1As shown, the overall system workflow is as follows: Users submit exploration tasks to the system, which are received and executed by the planner and budget management agent, including task decomposition, budget allocation, and priority scheduling. Subsequently, four candidate source modules are concurrently scheduled: a local PDF RAG search engine, an active literature inspection and journal keyword subscription module, and a thermal management materials database. The thermal management materials database (TMMD database) is a self-built and continuously maintained structured database for thermal functional materials such as solid-state refrigeration. This method primarily uses its solid-state refrigeration material data, which is organized by solid-state refrigeration effect type, covering categories such as electrocardiogram effect (ECE), magnetic card effect (MCE), pressure card effect (BCE), and spring card effect (eCE). The structured data of solid-state refrigeration materials in the thermal management materials database constitutes one of the components of the material data domain of the solid-state refrigeration material cross-modal knowledge base. For each type of solid-state refrigeration material, TMMD uniformly uses the following structured field system for storage:

[0059] (i) Effect characteristic quantities: The core performance of the corresponding solid-state refrigeration effect, including isothermal entropy change ΔS, adiabatic temperature change ΔTad and the corresponding driving field amplitude—for electric card materials, it is the applied electric field strength; for magnetic card materials, it is the applied magnetic field strength; for pressure card materials, it is the applied hydrostatic pressure; and for spring card materials, it is the applied stress or strain.

[0060] (ii) Phase transition and structural information: phase transition temperature (e.g., Curie temperature), phase transition type, crystal system, space group (symbol and number), number of atomic sites and crystal structure file (CIF); further distinguish and store the structural identifiers and structural files of the ferroelectric and paraelectric phases of the electrochemical materials;

[0061] (iii) Thermodynamic stability: formation energy, energy above hull, and stability prediction markers are used for preliminary stability determination during the candidate recall stage;

[0062] (iv) Functional properties: cell volume, density, band gap, dielectric constant, as well as spontaneous polarization, remanent polarization and coercive field of electric card materials, magnetic order type and total magnetic moment and elastic modulus of magnetic card materials, etc.

[0063] (v) Engineering availability and literature traceability: syntheticity markers, thermal hysteresis width and fatigue cycle life of the spring card material, and data governance fields such as the title, source, DOI, number and data review status of the source documents.

[0064] Records of various solid-state refrigeration materials are linked by material identifier, chemical formula, and open computational database number (such as MaterialsProject number). A unified Representational State Transition (REST) ​​interface is provided to offer retrieval services based on effect type, chemical formula keywords, and material constraints, serving as one of the data sources for candidate material recall in this method. A constraint-based new material structure generation module (exploration bypass, accounting for approximately 10% to 20% of the total computational budget) based on MatterGen is also included. The outputs of these four modules converge at the candidate material pool construction node, forming a unified candidate pool encompassing literature candidates, database candidates, and generated candidates.

[0065] Candidate materials enter the L1 fast screening node, where geometric optimization and thermodynamic stability screening are performed. Candidates that fail the L1 screening flow into the failure attribution recording node for bottleneck analysis, and the failure information is returned to the planner and budget management agent via a priori feedback loop to update the subsequent candidate generation strategy. Candidates that pass the L1 screening enter the L2 fine screening node, where VASP DFT calculations are performed to obtain key indicators such as polarization, dielectric constant, and energy barrier. Candidates that fail the L2 screening also flow into the failure attribution recording node.

[0066] Candidates selected through the L2 screening process enter the manual approval gate, where researchers perform manual approval, quota control, and circuit breaker strategy decisions. Candidates rejected by the approval process are returned to the failure attribution record node, while approved candidates enter the L3 high-fidelity computation node, where DPGEN, DeepMD, and LAMMPS are invoked to calculate statistical thermodynamic properties under temperature and driving field conditions. After the L3 calculation is completed, the sorting and report generation module performs multi-objective scoring on the candidates, including ΔT, ΔS, stability, prepareability, novelty, and cost, and generates an automated report containing uncertainty estimates and counterfactual interpretations. The final results are displayed on the system's visualization platform, providing a candidate pool dashboard, task queue monitoring, result interpretation pages, reproducibility experiment suggestion cards, and end-to-end evidence traceability (Trace ID) functionality. The system then enters the end or next iteration phase. Researchers can also adjust their exploration strategies by viewing platform feedback and resubmit exploration tasks, forming a continuously optimized outer iterative closed loop.

[0067] Through physical mechanism-driven synthetic benchmark simulation verification, on a synthetic benchmark where the quality factor of the giant thermal effect is distributed in a ridge pattern near the phase boundary, compared with random screening or traditional high-throughput traversal screening, this embodiment increases the number of high-potential candidates per unit computing power by no less than 2 times, and reduces the proportion of low-value high-fidelity evaluations (referring to the proportion of high-fidelity evaluations where the actual quality factor is lower than the set low quantile threshold, rather than tasks that fail to run) by no less than 30%; achieving full-link traceability: all recommended conclusions are accompanied by literature evidence, calculation parameters and model versions, ensuring 100% auditability and reproducibility;

[0068] Breaking the performance trilemma: Through multi-objective optimization and physical constraint acquisition functions, systematically explore the Pareto frontier of "high ΔS / ΔT × low driving field × low thermal hysteresis × high stability";

[0069] Significantly shorten the R&D cycle: The cycle from "research idea" to "high-confidence candidate material calculation and screening report" is shortened by no less than 50%;

[0070] It possesses self-evolution capabilities: through a computational closed-loop active learning mechanism, the proxy model continuously improves with the accumulation of multi-fidelity computational data, forming a positive feedback flywheel that accelerates computational material discovery.

[0071] This embodiment proposes an intelligent R&D system and method for discovering solid-state refrigeration materials. The system uses the Caloric-DB cross-modal knowledge base for solid-state refrigeration materials as its data foundation, a Retrieval Enhanced Generation (RAG) engine as its knowledge reasoning core, a hierarchical multi-agent collaborative network as its task orchestration framework, and combines physically constrained Bayesian active learning, multi-objective optimization, and a multi-fidelity computational screening pipeline to achieve the generation, screening, evaluation, traceability, and continuous learning and updating of solid-state refrigeration material candidates.

[0072] Solid-state refrigeration materials cross-modal knowledge base (Caloric-DB) and RAG engine. In this embodiment, the system constructs a data acquisition system with at least three layers to form a multi-source heterogeneous data set covering open databases, domain literature, and project-generated calculation results.

[0073] (1) First layer: Open computing databases and structured databases:

[0074] The first layer of data sources includes open computational databases or structural databases such as Materials Project, JARVIS, NOMAD, and OQMD. The system acquires computational data such as crystal structure, density functional theory (DFT) energy, elastic constants, and phonon spectrum in batches through a standard application programming interface (API) to provide structural candidates and basic property labels.

[0075] like Figure 2As shown, the Caloric intelligent discovery system exposes a unified REST API gateway. Requests initiated by clients or external schedulers are routed to various functional endpoints via the API gateway. The system provides six core interfaces: (1) Candidate query interface, which retrieves and returns a Top-N candidate list containing comprehensive scores, confidence intervals, and chains of evidence from the Caloric-DB based on the type and quantity of the Caloric heat effect; (2) Workflow submission interface, which receives candidate material identifiers, evaluation levels (L1 / L2 / L3), and priority parameters, creates workflow instances, and uses the workflow engine to perform hierarchical scheduling of L1 fast screening, L2 fine screening, and L3 high-precision evaluation. Workflow status and cost information are synchronously written to the Caloric-DB; (3) Workflow status query interface, which returns the status in real time. (3) Determine the execution progress of the workflow, the computational cost consumed, and the result preview; (4) Evaluation report interface, return a complete evaluation report of the specified candidate material, including multi-objective scores, explanations, evidence chains, and experimental suggestions; (5) Evidence intake interface, receive the literature path and metadata, trigger the evidence parser (NLP / IE pipeline) to extract structured information from the literature, and update the extraction results to the evidence chain and structured data table of Caloric-DB; (6) Active learning suggestion interface, call the active learning engine to read the model uncertainty and value function, and return the next batch of high-value experimental or computational suggestion points. The solid arrows in the figure represent the request call relationship, and the dashed arrows represent the data read and write relationship. Caloric-DB, as the core knowledge base of the system, uniformly stores candidate materials, scores, evidence chains, workflow status, and model intermediate quantities.

[0076] In one embodiment, the data acquisition module can be represented as:

[0077] 1. DataIngestion(

[0078] 2.sources=["materials_project","jarvis_dft","nomad","oqmd"],

[0079] 3. query_filters={

[0080] 4."elements":["Ba","Ti","Pb","Bi","Fe","Ni","K","Na",...],

[0081] 5."crystal_system":["cubic","tetragonal","orthorhombic","monoclinic"],

[0082] 6. "has_bandstructure": True

[0083] 7.},

[0084] 8.output_schema=CaloricCandidateSchema 9.)

[0086] The query_filters are used to limit the set of elements, the crystal system range and the target physical property conditions, and the output_schema is used to unify the output data structure to facilitate subsequent database entry, alignment and retrieval.

[0087] (2) Second layer: Domain literature database:

[0088] The second layer of data source is a literature database in the field of solid-state refrigeration materials. The system uses compliant crawlers to systematically collect full-text and supplementary materials of Chinese and English literature related to solid-state refrigeration effects (SCE) from the past thirty years, while adhering to robots.txt, copyright rules, and access policies. The solid-state refrigeration effects mentioned include, but are not limited to, electrocard effect (ECE), magnetic card effect (MCE), e-card effect (eCE), and pressure card effect (BCE).

[0089] In this embodiment, the system uses tools such as GROBID and PyMuPDF to parse the layout of PDF documents, extract paragraphs, tables, and mathematical formulas, and separate visual content such as graphs, curves, and phase diagrams for subsequent visual model processing.

[0090] (3) Third layer: Project-generated calculation data:

[0091] The third layer of data sources consists of self-generated computational data produced by the system during operation through a multi-fidelity computational pipeline, including DFT data, phonon computational data, and molecular dynamics (MD) data. This type of data is continuously written to Caloric-DB to fill the gaps in the chemical space covered by existing literature and public databases, and to continuously improve the generalization ability and prediction accuracy of subsequent surrogate models.

[0092] In response to the highly coupled "numerical value-condition-material" expression characteristics in materials literature, this embodiment designs a structured extraction process based on a large language model.

[0093] (1) Multimodal data preprocessing:

[0094] The system introduces a multimodal large language model (VLM) with visual understanding capabilities to analyze and vectorize key spectra such as differential scanning calorimetry (DSC) curves, X-ray diffraction (XRD) patterns, pressure-temperature phase diagrams, and hysteresis loops in the literature. It extracts quantitative information such as peak values, phase transition temperatures, and integral areas, and converts them into structured text descriptions before sending them to the extraction pipeline.

[0095] (2) Entity relation extraction based on instruction fine-tuning large language model (LLM):

[0096] The system employs a language model fine-tuned on a corpus of materials science to perform few-shot prompting on the literature content, automatically identifying and extracting core scientific entities and their relationships related to solid-state refrigeration materials. The extracted records can adopt a unified data structure as shown in Table 1.

[0097] Table 1 is the field definition table for structured extraction records of solid-state refrigeration materials.

[0098]

[0099] Among them, evidence_source is used to record evidence information such as document DOI, page number, and paragraph position; uncertainty is used to indicate the data credibility, error range, and source type.

[0100] (3) Cross-validation and confidence scoring:

[0101] For the same numerical value or attribute, the system employs multiple prompting strategies for independent extraction, and improves extraction accuracy through majority voting and logical consistency checks. The logical consistency checks may include, but are not limited to: checking the thermodynamic constraints between entropy change, adiabatic temperature change, and heat capacity; checking the format consistency between material composition and space group; and checking the consistency between driving field units and effect types.

[0102] In one embodiment, each extracted record is required to include evidence fragment location information, including DOI, page number, and paragraph coordinates, thereby ensuring data traceability and auditability. The accuracy of the extraction results can be verified to meet a preset target, such as above 90%, through manual sampling or standard test set verification.

[0103] To ensure that knowledge base data meets the FAIR principles of searchability, accessibility, interoperability, and reusability, Caloric-DB uses a standardized metadata schema for all records. Each data record includes at least the following fields:

[0104] data_version: Database version number;

[0105] extraction_model: Extraction model name and version identifier;

[0106] extraction_timestamp: Extracts the timestamp;

[0107] evidence_chain: The complete chain of evidence, including document identifiers, page numbers, paragraph coordinates, and excerpts from the original text;

[0108] uncertainty: The uncertainty field includes the source type, confidence level, and error range;

[0109] conflict_resolution: Resolution record of conflicting data from multiple sources.

[0110] Through the above design, the system can uniformly manage and resolve conflicts of records of the same material from different literature sources, different extraction versions, and different calculation batches.

[0111] At the logical level, Caloric-DB can be divided into three interrelated data domains:

[0112] (1) Materials Domain:

[0113] It is used to store the composition information, crystal structure files (such as CIF), predicted properties, and performance data reported in the literature of candidate materials, along with version numbers and chain of evidence fields.

[0114] (2) Workflow Domain:

[0115] It is used to store the status of computing tasks, running logs, resource consumption (such as core hours), failure reasons, and retry records. Each record is bound to a globally unique trace_id to support full-process auditing.

[0116] (3) Evidence Domain:

[0117] It is used to store document fragments, page numbers, paragraph coordinates, confidence scores, and citation chain information, thereby supporting the reverse tracing of any recommended conclusion back to the original document basis.

[0118] The three data fields mentioned above can be linked through fields such as candidate_id and trace_id, so that any calculation conclusion can be traced back to the original literature basis along the evidence chain field, and traced back to the corresponding calculation parameters and execution process along the workflow record field.

[0119] Enhance search quality using the hybrid search RAG engine:

[0120] (1) Semantic segmentation strategy:

[0121] To address the strong coupling between "composition, process, structure, and performance" in materials literature, this embodiment does not employ a fixed word count segmentation method. Instead, it uses intelligent block segmentation based on semantic logic units and chemical entity boundaries. Specifically, the doping ratio, annealing conditions, phase transition temperature, and performance indicators of the same material are encapsulated in the same knowledge block to avoid contextual fragmentation and improve the accuracy of retrieval and question answering.

[0122] (2) Dual-channel hybrid retrieval architecture:

[0123] like Figure 3 As shown, the retrieval enhancement generation engine in this embodiment adopts a hybrid retrieval architecture of "dual-channel recall, multi-constraint filtering, ranking fusion, and evidence backlink". The original documents, databases, and computational records are first processed by the document parsing and paragraph segmentation module to form structured paragraphs. Then, index construction is performed in two parallel channels: the left channel constructs a sparse inverted index using the full-text keyword indexing module, and the right channel encodes paragraphs into high-dimensional dense vectors using the embedding vector generation module. Simultaneously, the user's research objective or agent retrieval question is parsed by the query parsing and constraint generation module, and the generated retrieval requests are input into the keyword retrieval module and the dense vector retrieval module, respectively. The keyword retrieval module performs exact matching based on the sparse index, outputting candidate evidence set two; the dense vector retrieval module performs semantic matching based on vector similarity, outputting candidate evidence set one. Two candidate sets are simultaneously fed into the materials knowledge constraint filtering module, which eliminates fragments that do not conform to physicochemical rules based on prior constraints in the materials field. The filtered candidate sets are then merged into a unified ranking list by the inverse ranking fusion module, and further refined by the candidate evidence re-ranking module according to relevance and credibility. Finally, the evidence chain output module generates a structured evidence chain for use by the subsequent candidate material generation, screening, and reporting modules. The document segmentation and indexing module is used to perform unified preprocessing on solid-state refrigeration material literature, materials database records, and system-generated calculation records. The system first parses PDF documents into paragraph-level text blocks and binds each text block with document number, page number, paragraph number, material name, material chemical formula, space group, phase transition type, driving field type, performance indicators, and source information. For each text block... The system generates text embedding vectors. This information, along with paragraph text, page numbers, and evidence metadata, is written into the knowledge base. In the engineering implementation, this module corresponds to the document ingestion process and the agent.document_chunks data table, used to store fields such as chunk_text, embedding, page, and paragraph_id.

[0124] The dense vector retrieval module is used to solve the recall problem in the materials science field where there are "similar expressions and mechanism descriptions but different keywords." This applies to retrieval problems generated by user queries or upstream intelligent agents. The system obtains the query vector through a domain embedding model:

[0125]

[0126] in, This represents a text embedding model in the materials science domain. Let represent the query vector. For the i-th text block in the knowledge base, its semantic similarity score is defined as:

[0127]

[0128] in, For query With the text blocks Semantic similarity score between them; For query The embedding vector; For the first text blocks The embedding vector; · It is the dot product of two vectors; and These are the moduli of the corresponding vectors.

[0129] When using a vector database or PostgreSQL pgvector implementation, cosine distance can also be used. Sort the data and let:

[0130]

[0131] This module is mainly used to recall literature fragments that are semantically related to "ferroelectric-paraelectric phase transition", "electrical adiabatic temperature change", "thermal hysteresis reduction mechanism" and "BaTiO3-based doped system".

[0132] The keyword retrieval module addresses the need for precise matching of material chemical formulas, space groups, numerical indicators, and experimental conditions. This module establishes a lexical index on the paragraph text and calculates keyword relevance based on BM25 or a full-text search ranking function. BM25 score. It can be represented as:

[0133]

[0134] in, The keyword 't' represents the keyword in the query; IDF(t) represents the inverse document frequency of keyword 't', used to measure the discriminative power of keyword 't' in the entire knowledge base, and can be expressed as:

[0135]

[0136] in, This represents the total number of text blocks in the knowledge base. Indicates the presence of keywords Number of text blocks; keywords The fewer text blocks appear in the knowledge base, the larger its IDF(t) value, and the higher the corresponding retrieval weight; This indicates that the keyword 't' is in the text block. Frequency of occurrence in Keywords In text block Frequency of occurrence in Indicates the length of the text block. This indicates the average text block length. and Adjusting parameters for BM25. This module is primarily used to accurately recall evidence fragments containing specific material composition, space group, driving field strength, isothermal entropy change, adiabatic temperature change, thermal hysteresis, and experimental temperature range.

[0137] The materials knowledge constraint filtering module further narrows the general text search results to candidates that match the solid-state refrigeration materials task. Based on the materials knowledge graph or structured fields, the system performs entity consistency filtering, driving field type filtering, performance field integrity filtering, and phase transition type filtering on the recalled results. For example, when the research target is an electric card material, the system prioritizes retaining evidence fragments containing information on electric field, polarization change, ferroelectric phase transition, adiabatic temperature change, or isothermal entropy change; when the user specifies an element range or excludes elements, the system constrains based on the material's chemical formula. This filtering process can be represented as:

[0138]

[0139] in, is a material knowledge constraint filtering function, a binary indicator function with a value of 0 or 1; q represents a user query or a retrieval request generated by an upstream intelligent agent, from which the query parsing and constraint generation module parses material system constraints, phase transformation type constraints, driving field type constraints, and performance field integrity constraints; Indicates the first in the candidate evidence set A text block. Represents a text block If all of the above constraints are met, the item will be retained. Represents a text block If any of the constraints are not met, the item is removed.

[0140] Candidate set after constraint filtering for:

[0141]

[0142] in, This represents the set of candidate evidence output by the dense vector retrieval module and the keyword retrieval module. This represents the set of candidate evidence retained after being filtered by material knowledge constraints.

[0143] The reciprocal ranking fusion module is used to merge the ranking results of dense vector retrieval and keyword retrieval. Let... Represents a text block Ranking in vector search results The fusion score indicates its ranking in keyword search results. Defined as:

[0144]

[0145] in, To smooth out the parameters, the excessive influence of top-ranked items on the fusion results is reduced. If a text block appears only in one channel, only the lowest-ranked item in that channel is calculated. Through this fusion method, the system can prioritize highly credible evidence fragments that are simultaneously hit by semantic retrieval and keyword retrieval, reducing the risk of missed detections or false positives caused by a single retrieval method.

[0146] The candidate evidence re-ranking module performs secondary ranking on the fused Top-N results. This module can employ a cross-coding re-ranking model or a lightweight text overlap model to refine the query. and candidate text blocks Relevance scoring was performed to obtain the relevance score results. :

[0147]

[0148] in, This represents a reordering scoring function used to compare query q with candidate text blocks. The constructed text pairs are subjected to fine-grained relevance scoring, and the output is normalized to... The correlation score of the interval; the higher the score, the better. The stronger the correlation with q, the better. When using a cross-coding reordering model, (·) Combine query q with candidate text blocks After concatenation, the data is input into a pre-trained language model for joint encoding, and the model's output layer provides a semantic relevance score between the two. When implemented using a lightweight text overlap model... Based on query q and candidate text blocks A relevance score is calculated based on the degree of overlap between the word sets, for example, by taking the ratio of the intersection to the union of the word sets. This represents the score for the completeness of evidence, and the final retrieval score. It can be represented as:

[0149]

[0150] in, The score for the completeness of evidence takes into account whether it includes material composition, performance values, test conditions, page numbers, and source information. , , The weighting coefficients are used. Through this re-sorting process, the system prioritizes outputting document fragments that are "semantically relevant, have keyword hits, satisfy material constraints, and have complete evidence fields."

[0151] The evidence chain output module encapsulates the final search results into traceable evidence. Each piece of evidence includes at least a document number, document title, page number, paragraph number, evidence fragment, confidence level, and corresponding candidate material information. The confidence level can be obtained by normalizing the fusion score and re-ranking score.

[0152]

[0153] in, This indicates the confidence level of the i-th piece of evidence. This represents the weight between the merged ranking score and the re-ranking score. This chain of evidence is then provided to the candidate material generation, L1 rapid screening, L2 fine screening, and report generation modules, enabling each candidate material output by the system to be traced back to the original document's page number and paragraph.

[0154] The technical advantages of the aforementioned dual-channel hybrid retrieval architecture are as follows: dense vector retrieval enhances the recall capability of similar content; keyword retrieval ensures accurate matching of material composition and performance values; material knowledge constraint filtering reduces false recalls across domains; inverse ranking fusion improves the ranking priority of evidence in multi-channel one-fatal cases; the re-ranking module further improves the relevance of evidence; and the evidence chain output module ensures that the recommendation results are auditable and reproducible.

[0155] (3) Iterative retrieval mechanism:

[0156] like Figure 4As shown, for complex multi-hop scientific questions, such as "compared to BaTiO3-based materials, which type of relaxor ferroelectric has a higher adiabatic temperature change and lower thermal hysteresis under similar driving electric fields," the system employs an iterative retrieval mechanism based on the collaboration of a planner agent and a RAG analysis agent. The complex scientific question is first input into the planner agent, which analyzes the research objectives and constraints, and breaks down the original question into several independently verifiable sub-problems. Each sub-problem then enters a dual-channel hybrid retrieval module, where dense vector retrieval and keyword retrieval are used in parallel to retrieve candidate evidence sets. After the search results are evaluated by the evidence field integrity check module, they enter the "Field completeness" judgment node: If the field is incomplete, it enters the missing field identification module to identify the specific missing field type and generate a targeted query. After the query rewriting module reconstructs the search request, it returns to the dual-channel hybrid search module for the next iteration; If the field is complete, it enters the evidence conflict detection module to check the consistency of evidence from different sources and branches at the "Conflict exists" judgment node: If a conflict exists, the conflict verification search module is triggered to supplement the search for the conflict point. After query rewriting, it also returns to the search module for iteration; If there is no conflict, all evidence is sent to the evidence merging and comparison table generation module to summarize the evidence of each sub-problem and construct a structured comparison table, finally outputting a traceable conclusion with literature source annotation. Specifically, the system first uses the planner agent to perform structured analysis of the user problem to obtain the comparison object, constraints, and target indicators. For the above problem, the system can parse the following search constraints:

[0157] Comparison objects: BaTiO3-based materials, relaxor ferroelectrics;

[0158] Constraints: Similar driving electric fields;

[0159] Target indicator: Adiabatic temperature change Thermal hysteresis, operating temperature range;

[0160] Evidence required: material composition, driving electric field, Numerical values, thermal hysteresis values, and document page numbers.

[0161] The system then breaks down the original problem into several subqueries:

[0162] Q1: What is the electric field performance of BaTiO3-based electrochemical materials within a given electric field range? What is the thermal hysteresis?

[0163] Q2: Relaxor ferroelectrics such as PMN-PT, PZN-PT, and PLZT operate within a similar electric field range. What is the thermal hysteresis?

[0164] Q3: Which candidate materials simultaneously meet the higher requirements? And conditions with lower thermal hysteresis?

[0165] Q4: Is the above conclusion supported by a chain of documentary evidence?

[0166] Each subquery invokes a dual-channel hybrid retrieval module, performing dense vector retrieval, keyword retrieval, material knowledge constraint filtering, RRF fusion, and reordering to obtain an evidence set with page numbers, paragraph numbers, and confidence levels. The system performs an evidence field integrity check on each round of retrieval results to determine whether necessary fields such as material composition, driving field strength, adiabatic temperature change, thermal hysteresis, test temperature range, and source literature are covered.

[0167] The evidence completeness score S_cover(e_i) can be expressed as:

[0168]

[0169] in, This represents the i-th piece of evidence, which is a candidate text block retrieved through retrieval. It is encapsulated with metadata such as the document number, page number, paragraph number, and source information it is bound to; This represents the j-th required field; Present evidence Does the fragment text or its bound metadata contain fields? If it contains the value, take 1; otherwise, take 0. This indicates the field weights. For electric card material comparison tasks, necessary fields must include at least material composition, driving electric field, and... Thermal hysteresis, test temperature range, and literature source. If a field is missing in a search result, the system will automatically modify the next search based on the missing field. For example, if the search result "PMN-PT has high thermal hysteresis" is found, the system will automatically modify the next search result based on the missing field. However, if "thermal lag" data is missing, the system generates the next round of queries:

[0170] "PMN-PT electric card thermal hysteresis-driven electric field adiabatic temperature change";

[0171] If the search results show inconsistencies in the numerical descriptions of the same material's properties across different publications, the system will initiate a conflict verification search, prioritizing the examination of test conditions, sample morphology, differences between thin films and bulk materials, driving field amplitude, and measurement methods to determine the source of the discrepancy. The degree of conflict will be assessed. It can be represented as:

[0172]

[0173] in, and Indicates performance values ​​from different sources of evidence. and This indicates the corresponding uncertainty or error range. To avoid extremely small constants with a denominator of zero. When When the threshold is exceeded, the system continues to search for supplementary evidence and does not directly generate a conclusion.

[0174] The system updates the search status after each iteration. :

[0175]

[0176] Where t is the iteration step.

[0177] Next round of queries Generated by the following strategy:

[0178]

[0179] Here, Rewrite refers to the query rewriting model, which can be generated by a large language model based on missing fields, conflicting fields, and discovered candidate materials, or it can be generated by a rule template. For material comparison tasks, the query rewriting template should at least include the following formats: "material name + performance index + driving conditions", "material system + phase transition type + thermal hysteresis", and "candidate material + test temperature range + literature evidence".

[0180] The stopping conditions for iterative retrieval include:

[0181] 1. Key field coverage reaches the preset threshold;

[0182] 2. The confidence levels of all Top-K evidence are higher than the preset threshold;

[0183] 3. No new high-confidence evidence was found in two consecutive rounds of searching;

[0184] 4. Reach the maximum number of iterations .

[0185] Finally, the RAG analysis agent merges the evidence obtained from multiple rounds of retrieval, forms a comparison table, and submits it to the evaluation and critique agent. The output includes at least candidate materials, material type, driving electric field, and... The system incorporates parameters such as thermal hysteresis, test temperature range, evidence page number, confidence level, and recommended conclusion. This iterative retrieval mechanism avoids large language models from making direct inferences based solely on a single retrieval result, thereby reducing the risk of evidence omissions and reasoning illusions in complex material comparison problems.

[0186] (4) Retrieval downgrade and hallucination protection mechanisms:

[0187] When the scores of all candidate documents obtained from the hybrid retrieval are lower than the preset confidence threshold (confidence_threshold), the system enters a pure model inference degradation mode, ceasing the injection of low-confidence document fragments to avoid low-quality evidence interfering with the inference process. This logic can be expressed as:

[0188] 1. IF max(retrieval_score) <confidence_threshold:

[0189] 2.mode="pure_llm_fallback"

[0190] 3. context_injection=None

[0191] 4. uncertainty_flag=True

[0192] 5. response_prefix="[Note: The current answer failed to retrieve any literature evidence directly related to this question from the knowledge base. The following content is generated based on prior knowledge of the model and has limited credibility. Please make a judgment in conjunction with the opinions of domain experts.]"

[0193] 6. ELSE:

[0194] 7. mode="rag_grounding"

[0195] 8.context_injection=top_k_chunks_with_citation

[0196] Through the above design, the system will not generate false citations when the knowledge base coverage is insufficient. Instead, it will replace them with explicit uncertainty statements, thereby enabling researchers to accurately judge the degree of literature support for their current conclusions.

[0197] This embodiment employs a hierarchical multi-agent architecture driven by a LangGraph Directed Acyclic Graph (DAG) state machine. Specifically, the DAG state machine sequentially arranges the following nodes: intent parsing node, candidate recall node, L1 fast screening node, L2 fine screening submission and retrieval node, multi-objective scoring node, L3 high-fidelity submission and retrieval node, final fusion scoring node, and report generation node. The state machine sets interruption points at four key locations: the first interruption point is before L2 fine screening, used for researchers to manually approve L1 candidates; the second interruption point is before L2 job retrieval, used to wait for the first-principles calculation job to complete on the high-performance computing cluster; the third interruption point is before L3 high-fidelity, used for researchers to select candidates for high-fidelity verification; and the fourth interruption point is before L3 job retrieval, used to wait for the molecular dynamics job to complete. By leveraging the aforementioned interruption points, the system enables the pauseable and resumable orchestration of manual approval processes, budget quota control calculations, and task recovery. Upon resumption, the system continues the subsequent process based on researcher decisions or task results. Through clear role division, message routing, and a reflective error-correction mechanism, the system automates the entire discovery process of solid-state refrigeration materials. For example... Figure 5 As shown, the system includes a planner and budgeter, a document reconnaissance and retrieval enhancement generation module, a generation agent, a simulation orchestrator, an evaluator and critic, as well as a human review checkpoint and a result publisher. Each module is scheduled and data is transferred via a directed acyclic graph (DAG). The planner and budgeter, located at the top level of the architecture, is responsible for task decomposition, budget allocation, priority management, and the setting of human review checkpoints. It distributes subtasks in parallel to the document reconnaissance and retrieval enhancement generation module, the generation agent, and the simulation orchestrator based on the DAG. The document reconnaissance and retrieval enhancement generation module is responsible for document mining and knowledge base querying, the generation agent is responsible for candidate material structure generation, and the simulation orchestrator is responsible for scheduling and executing the simulation workflow. The outputs of the three modules are uniformly fed into the evaluator and critic, which perform multi-objective scoring and reflective error correction. After evaluation, the final result is reviewed by the human review checkpoint and output by the result publisher. The planner agent is responsible for global task orchestration and resource budget control.

[0198] The functional responsibilities include: receiving research objectives in natural language, such as "discovering high-performance solid-state cooling materials with driving fields below a set threshold in the 280K to 340K temperature range," and using chained reasoning logic to decompose the macro-objective into an executable DAG workflow; maintaining the global task state machine to manage the execution progress, dependencies, and resource consumption of each sub-agent; allocating computing power budgets according to three levels: L1 fast screening, L2 fine screening, and L3 high-fidelity screening, to prevent candidates that fail the lower-level screening from directly entering the high-cost computation stage; triggering a human gate at high-cost nodes, such as L2 to L3 promotion or the generation of a large number of candidates for the generative model; and triggering a circuit breaker strategy when the failure rate of the candidate set tasks exceeds a preset threshold, suspending the relevant tasks and performing failure attribution analysis. When the system detects that the task failure rate exceeds the preset threshold and sets the circuit breaker status to "on," the L2 fine screening submission node and the L3 high-fidelity submission node will directly jump to the report generation node, skipping the subsequent high-cost computation stage, thereby avoiding the continuous investment of ineffective computing power in abnormal states. The failure attribution is achieved through failure mode recording: candidates that fail the screening trigger failure mode recording. Failure modes that are approved by researchers will be subject to a soft penalty in subsequent L1 rapid screening, with a corresponding score deducted for each failure mode hit (e.g., 0.15 points) and an alarm label added. This enables the prior feedback loop to update the candidate generation and screening strategy.

[0199] The specific interaction sequence is as follows: Figure 6As shown, the diagram contains six participant lifelines: researcher, agent API, discovery graph, candidate and evidence sources, computational workflow, and ranking and reporting. Researchers submit research objectives and constraints via the POST / workflow / submit interface. The Agent API creates corresponding thread_id and trace_id and forwards the request to Discovery Graph. Discovery Graph executes the intent_parser node sequentially to parse the research objectives, data sources, and computational budget. It then initiates candidate_retrieval calls to candidate and evidence sources, recalling candidate pools and literature evidence from sources such as RAG search, TMMD database, Materials Project, and MatterGen, and performing source deduplication. After the candidate pool is returned, Discovery Graph executes the l1_screener node to perform low-cost rapid screening, forming an L2 candidate list, and sends an L2 Gate approval request to the researcher. The researcher returns an approve / reject / selected_ids decision. After manual approval, the system enters the computational verification stage, submitting an L2 VASP job to the computational workflow and waiting for the results. After the L2 results are returned, the L3 stage can be triggered to submit DPA-3 / LAMMPS high-fidelity kinetic verification. After the computation is completed, Discovery... Graph invokes the `scorer` and `scorer_final` nodes to perform multi-objective scoring based on the Pareto front, and sends the results to the ranking and reporting module. The report generator outputs a structured report containing "why selected, why rejected, and next steps recommendations," which is then returned via the Agent API as a Top-N candidate list, workflow status, and a complete report, ultimately presented to the researcher. Solid arrows in the diagram represent active invocations, while dashed arrows represent result feedback. Literature Scout & RAG Analyst Agent:

[0200] (1) Document reconnaissance sub-agent:

[0201] The literature reconnaissance sub-agent is responsible for proactively scouring literature related to solid-state refrigeration materials. It can perform periodic searches based on combinations of keywords, authors, journals, etc., for example:

[0202] "electrocaloric BaTiO3";

[0203] "lead-free electrocaloric";

[0204] "magnetocaloric effect";

[0205] “giant magnetocaloric”;

[0206] “elastocaloric NiTi”;

[0207] "barocaloric effect";

[0208] "solid-state caloric cooling" etc.

[0209] For newly retrieved documents, the system performs deduplication, credibility scoring, and compliance labeling, and automatically triggers a deep extraction process for high-value documents to incrementally update Caloric-DB.

[0210] (2) RAG analysis sub-agent

[0211] The RAG analysis sub-agent is used to answer mechanistic questions posed by planners and generate referable prior knowledge. For example, when a planner asks, "What is the main contributing mechanism of the electrocaloric entropy change of a PMN-PT relaxor ferroelectric near the Curie temperature?", the RAG analysis sub-agent calls the RAG engine to retrieve relevant literature from the knowledge base and generate a structured answer with page-numbered citations.

[0212] In addition, the RAG analysis sub-agent can also:

[0213] 1. Provide component design priors for structure-generating intelligent agents;

[0214] 2. To provide a performance benchmark and current best-in-class (SOTA) reference for similar materials for evaluating intelligent agents.

[0215] The structure-generating agent is used to construct a pool of candidate materials, and a dual-track parallel strategy is employed.

[0216] (1) Track 1: Candidate generation of literature / databases:

[0217] This track uses database interfaces such as the Materials Project to batch acquire CIF files of known structures under constraints of composition, space group, and stability. Simultaneously, it extracts high-performance candidates reported in the literature from Caloric-DB and generates derived candidates based on composition substitution strategies, such as equivalent element substitution and solid solution serial scanning.

[0218] (2) Track 2: Generative candidate generation:

[0219] This track utilizes diffusion-based crystal generation models such as MatterGen to generate novel candidate structures under given conditions, including element pool, valence state constraints, banned elements, target space group, and stability requirements. Further processing is performed on the generated results.

[0220] Synthesibility filtering ensures that the synthesis feasibility score exceeds a preset threshold.

[0221] The novelty score is used to ensure that the similarity to known structures is below a preset threshold.

[0222] In one embodiment, this track accounts for 10% to 20% of the total computational budget in the early stages of the system to verify the hit rate improvement effect of generative candidates relative to literature candidates; after the effect is verified, its budget share can be gradually increased.

[0223] The simulation coordination agent serves as the central interface between the system and the external computing toolchain, and is responsible for the unified scheduling of queries, modeling, task submission, result parsing, and failure retries.

[0224] (1) Toolbox design:

[0225] In one embodiment, the simulation tool interface may take the following form:

[0226] 1.tools={

[0227] 2."query_materials_project":query_mp_api,

[0228] 3."generate_vasp_input":generate_vasp_incar,

[0229] 4."submit_hpc_job":submit_to_slurm,

[0230] 5."run_lammps_md":run_lammps_simulation,

[0231] 6."run_dpgen_workflow":run_dpgen,

[0232] 7."parse_vasp_output":parse_vasprun_xml,

[0233] 8."calculate_phonon":run_phonopy,

[0234] 9."run_phase_field":run_phase_field_simulation

[0235] 10.}

[0236] (2) Fault tolerance and retry mechanism:

[0237] All computational tasks are accompanied by a unique end-to-end tracing identifier (trace_id) to support end-to-end auditing. For failed tasks, such as those due to energy non-convergence, structural collapse, or timeouts, the system automatically performs failure attribution analysis and adjusts the appropriate parameters based on the failure type before retrying. Tasks exceeding the maximum number of retries are automatically escalated to a manual intervention queue and accompanied by a failure log summary.

[0238] The evaluation agent is responsible for multi-objective scoring, uncertainty quantification, and interpretable output of the computation results.

[0239] (1) Multi-objective scoring function:

[0240] The overall score for each candidate material is obtained by weighting its performance, stability, prepareability, novelty, and cost scores.

[0241]

[0242]

[0243] in, The overall score for candidate materials ranges from 0 to 1. For performance, For stability, For the purpose of preparation, For novelty, Cost-based; to The weights are configurable and satisfy the following conditions:

[0244]

[0245] In a preferred embodiment, the weights can be set as follows:

[0246]

[0247] Among them, the performance component is used to characterize the cooling capacity of a material and is defined as:

[0248]

[0249]

[0250] in, For thermal temperature change, It is an isothermal entropy change. For cooling capacity, For the effective operating temperature range width, For driving field strength; This represents a normalization function that maps indices of different dimensions to 0 to 1. to Assign internal weights to performance, and satisfy the following conditions: Since a lower driving field is more advantageous for engineering applications, the driving field term adopts... express.

[0251] The normalization function can use threshold normalization or interval normalization. For metrics where larger values ​​are better, the normalization function... for:

[0252]

[0253] For metrics where smaller is better, the normalization function... for:

[0254]

[0255] Where x represents the original value of the performance metric to be normalized, i.e. , , , or The measured or calculated value of any one of the indicators; For the target value, and These are boundary values ​​determined based on literature statistics, historical calculation data, or user input.

[0256] Stability score is used to characterize the reliability of candidate materials in thermodynamic, kinetic, and cyclic processes, and is defined as:

[0257]

[0258] in, Classified by thermodynamic stability, For phonon dynamics stability, The cycle life is defined as H, where H represents thermal hysteresis. This represents the normalization function defined above, where smaller is better. The score after normalizing the thermal hysteresis H, where the target value of the thermal hysteresis is... Boundary values and Similarly, the smaller the thermal hysteresis, the better, as determined by literature statistics, historical calculation data, or user input. The closer to 1; to The internal weighting of stability is determined by the energy level above the convex hull value. Sure:

[0259]

[0260] in, The maximum convex hull energy threshold allowed to enter the screening. The phonon stability score can be defined as:

[0261]

[0262] When L3 high-fidelity dynamic simulation results are available, the stability score can be further integrated with the L3 dynamic stability score. :

[0263]

[0264] in, For L2 stage stability, For L3 dynamic stability, For fusion weights. In a preferred embodiment, .

[0265] Preparability The feasibility of transforming a material from a computational candidate into a real sample is defined as:

[0266]

[0267] in, Based on the simplicity of the synthesis route, Based on the availability of raw materials, It is a low-toxicity component. Score based on structural and technological maturity; to This is the internal weight. For candidate materials containing rare and precious elements or highly toxic elements such as platinum, iridium, osmium, rhodium, ruthenium, and rhenium, the system reduces... or For structural systems with established preparation experience, such as perovskite, tetragonal, cubic, or hexagonal structures, systematic improvement is needed. .

[0268] Novelty Used to measure the degree to which candidate materials differ from known materials, defined as:

[0269]

[0270]

[0271] in, Indicates candidate materials With known material libraries The structural similarity of the most similar structures. Indicates component similarity, Indicates the novelty of the source; to The internal weights are determined by the source. If a candidate originates from a generative model or an actively learned exploration region, it can be assigned a higher source novelty; if the candidate is only material from a known database, the source novelty is relatively low.

[0272] Cost Used to characterize the elemental cost and preparation cost of materials, defined as:

[0273]

[0274] in, The cost of elements is estimated based on stoichiometry. To reduce the cost of the synthesis process, Elemental abundance; and These represent the normalization functions defined above, where smaller is better. The target values ​​for the normalized scores of element cost and synthesis process cost are... With boundary values Based on literature statistics, historical cost data, or user input, the lower the cost, the closer the normalization score is to 1. to These are the internal weights. The element cost can be expressed as:

[0275]

[0276] in, For the first The molar ratio of each element in the chemical formula of the material. This represents the unit cost or scarcity price factor for the element. For candidate materials containing high-cost elements such as gold, silver, platinum, iridium, osmium, rhodium, ruthenium, and rhenium, the system reduces the cost score.

[0277] After the candidate materials complete L3 high-fidelity verification, the system introduces a risk penalty item into the comprehensive score. :

[0278]

[0279] in, The risk penalty coefficient is determined by the potential function applicability risk, extrapolation risk, and kinetic anomaly risk, and its value ranges from 0 to 0.3. If the candidate material has a high extrapolation risk, a potential function applicability warning, or an anomaly in the kinetic simulation, the penalty coefficient will be increased accordingly. This will lower the final score.

[0280] When the following anomalies are detected, the evaluation agent actively triggers a feedback loop: The presence of imaginary frequencies in the phonon spectrum indicates that the candidate material may have insufficient kinetic stability; the phase transition thermal hysteresis exceeds a preset thermal hysteresis threshold, indicating that its practical coefficient of performance (COP) may not meet engineering application requirements; the predicted adiabatic temperature change or isothermal entropy change deviates from the statistical distribution of similar materials by more than two standard deviations, i.e., exceeds... ,in It represents the standard deviation of the performance index of similar materials under the same or similar driving field, operating temperature range and test conditions; the irreversible phase transition of the structure in molecular dynamics (MD) simulation indicates that its cycle stability or fatigue life may be insufficient.

[0281]

[0282] in, This represents the predicted performance value of candidate materials, such as the predicted thermal temperature change. Or predict isothermal entropy change ; This represents the average performance index of similar reference materials under similar test conditions; This represents the standard deviation of this performance index for similar reference materials; This represents a threshold of two standard deviations. When the above condition is met, the system considers the predicted value to deviate significantly from existing literature or historical calculation data, requiring a feedback loop for verification.

[0283] After the feedback loop is triggered, the system writes the anomaly type, candidate material number, anomaly index, trigger threshold and related evidence into the operation record, and performs operations such as re-searching literature evidence, improving calculation fidelity, adjusting candidate score weights, reducing candidate priority or resubmitting simulation verification according to the anomaly type.

[0284] Each candidate's recommendation report should ideally include at least the following fields: why_picked: the main reason for entering the current computational level, including key literature evidence and prediction performance range; why_not_passed: if filtered, provide the main shortcomings and corresponding indicators; what_next: the suggested computational or experimental conditions for the next step; confidence_interval: the 95% confidence interval of the key performance indicators; evidence_chain: a list of literature supporting the recommendation and the location of specific evidence.

[0285] This embodiment employs surrogate models with different granularities at different screening stages.

[0286] (1) Component stratification rapid screening model:

[0287] For early candidates with unknown or incompletely determined structures, an attention-based compositional model, such as CrabNet, is used to make preliminary predictions of target properties such as adiabatic temperature change and isothermal entropy change based solely on the chemical formula.

[0288] (2) Precise prediction model for structural layers:

[0289] For candidates with known structural information, a crystal graph neural network, such as ALIGNN, is used to explicitly introduce structural information such as bond angles by performing message passing on the bond graph and its line graph, thereby improving the prediction accuracy of phase transition properties and field response properties.

[0290] (3) Uncertainty quantification:

[0291] The system employs deep ensemble for uncertainty quantification. Specifically, it trains multiple independent prediction models (e.g., The system uses the ensemble mean as a point estimate and the ensemble variance as an uncertainty measure, and outputs 95% confidence intervals for key performance indicators. The system can also verify the reliability of the prediction uncertainty using calibration curves.

[0292] Since the significant Karl kinetic effect of solid-state refrigeration materials with first-order ferroelectric / antiferroelectric, magnetic structure, or martensitic phase transformation as the main thermal effect is usually concentrated at the phase coexistence line, phase transition line, and their critical neighborhood, this embodiment introduces phase boundary sensing physical constraints in the active learning loop. (1) Phase region classifier:

[0293] The computational data and literature data in the knowledge base are used as training samples. The input of each sample is the normalized state vector of the material. This includes dimensions such as composition, temperature, and driving field; the label indicates the stable phase category at that state point, such as paraelectric / ferroelectric / antiferroelectric phase, or austenitic / martensitic phase. The labels are derived from first-principles total energy comparisons, minimum free energy criteria, or phase icon annotations from literature. The classifier can employ supervised classification models such as gradient boosting trees, support vector machines, graph neural networks, or feedforward neural networks, outputting stable phase labels for the state space grid points. Phase boundary set The location is determined by the change in the stable phase label of adjacent grid points. When adjacent grid points , satisfy At that time, interpolation between the two yields the phase boundary point. And record When the free energy data for each phase is available, the condition of equal free energy between the two phases can also be used. , The phase boundary is determined. The hyperparameters of the classifier are determined through cross-validation on the training system and fixed on the test system that is not involved in the training, avoiding back-tuning based on test results. (2) Phase boundary sensing acquisition function:

[0294] A physical constraint acquisition function is constructed by superimposing phase boundary sensing weights on the standard Bayesian optimized acquisition function. :

[0295]

[0296]

[0297]

[0298] In the formula, Optimize the acquisition function for standard Bayesian methods. For phase boundary sensing weight function; Let x be a candidate material composition point; d(x,Boundary) represents the distance from candidate composition point x to the phase diagram boundary Boundary. This is the bandwidth parameter. Let be the normalized state vector of the candidate material, where As a component variable, For temperature, As the driving field, For pressure; in actual implementation, only the dimensions required by the target system can be retained; For the phase boundary point set, For the first A phase boundary point. For the first The scale normalization parameter of a dimension can be taken as the allowable range of that dimension, the standard deviation of the training data, or a preset physical scale. To explore the lower bound, a method is used to ensure that regions far from the current estimated phase boundary still retain non-zero sampling probabilities; in the current simplified implementation, the following approach can be taken: As a special case. For the standard Bayesian optimized acquisition function after nonnegation processing, the desired improvement can be obtained. Or the upper confidence bound after nonnegative transformation . and distance The smaller, The larger the distance, the higher the sampling weight assigned. When using the desired improvement:

[0299] ,

[0300] in , To predict the mean and standard deviation for the surrogate model, As the current best observation, To explore parameters, , The cumulative distribution and probability density function of the standard normal distribution; Non-negativity is constructed. All symbols are defined as before. Boundaries and anomalies: (When it falls on the boundary) ; hour ;when When the value becomes zero after nonnegation (e.g., in regions where the surrogate model determines no improvement), then... As a secondary sorting key, it avoids degradation caused by products always being zero; when the acquisition function may take a negative value (such as the untransformed UCB), it is first processed... Alternatively, a monotonically positive transformation can be used to map the result to a non-negative interval before multiplication, thus ensuring multiplicative gating semantics.

[0301] Significant Karl kinetic effects in solid-state cooling materials typically occur near phase transition boundaries where different crystalline phases coexist. Standard Bayesian optimization of the acquisition function, by uniformly exploring the composition space, disperses limited computational resources to stable phase regions far from phase boundaries. This embodiment modulates the standard acquisition function with phase boundary sensing weights: these weights are Gaussian decay functions of the distance from candidate composition points to phase diagram boundaries, with greater weights closer to the phase boundary, thus systematically guiding the sampling density to high-potential composition regions near phase transition boundaries. These weights are not general penalties or rewards, but rather explicit encodings of the following physical priors: For target solid-state cooling material systems where field-induced first-order ferroelectric / antiferroelectric phase transitions, magnetic structural phase transitions, or martensitic phase transitions are the primary sources of thermal effects, when an external field drives candidate state points to cross phase coexistence lines, phase transition lines, or their critical neighborhoods, latent heat release, order parameter mutations, or rapid penetration of polar nanoregions can occur, resulting in large isothermal entropy changes, adiabatic temperature changes, or Karl kinetic effect coefficients; therefore, phase boundary proximity can serve as a physical prior for the high-potential response region of such materials. It should be noted that the above priors are not presented as unconditional laws applicable to all material systems and all Karl effect mechanisms: for continuous phase transitions, diffuse phase transitions, or supercritical continuous polarization changes far from the phase boundary, the response may not be concentrated at the phase boundary. In this case, the ability to sample regions where the priors of the phase boundary are not significant is preserved by exploring the lower limit ε and uncertainty gating.

[0302] The introduction of the aforementioned phase boundary-sensing physical constraints stems from the following discovery process. First, domain physics and publicly available experimental / computational evidence indicate that the giant thermal effect of solid-state cooling materials is not uniformly distributed in the composition-state space, but rather concentrated at phase transition lines, phase coexistence lines, and their critical neighborhoods: In the BaTiO3 paraelectric-ferroelectric system, the latent heat below the critical point of the phase transition line enhances the electrocaloric response and shifts the response peak to a lower driving field; the directly measured adiabatic temperature change at the phase transition can reach approximately 1.6 K (driving field on the order of 30 kV / cm); in the fully ordered scandium-tantalate (PST) system, the maximum value of the electrocaloric coefficient appears near the critical electric field where the polar nanoregion begins to penetrate, and the calculated value of its maximum adiabatic temperature change is consistent with the experimental value, reaching approximately 4.7–4.8 K; finite adiabatic electrocaloric temperature changes accompanied by anisotropy and inverse electrocaloric effects are observed at all three ferroelectric phase transitions of BaTiO3. Therefore, the "proximity between candidate state points and phase boundaries" can be used as a physical prior for the high latent response region of this type of material. Secondly, the standard Bayesian optimization acquisition function explores uniformly in the composition-state space, without utilizing the aforementioned phase boundary structure information. This disperses limited high-fidelity computing power to stable phase regions far from the phase transition boundary. In this embodiment, even with standard Bayesian optimization, the proportion of low-value high-fidelity evaluations with true quality factors below the set low quantile threshold still reaches approximately 22% (approximately 41% for random / high-throughput traversal). To narrow the sampling to the high-potential phase boundary zone without increasing computing power, this embodiment explicitly introduces phase boundary proximity into the acquisition function using a multiplicative weighting method. Thirdly, to verify that this weighting is indeed different from conventional result reward / penalty items and that the effect truly originates from the phase boundary information itself, this embodiment sets up a placebo control with scrambled phase boundaries and a robustness test for phase boundary localization error.

[0303] Phase boundary sensing weight in this embodiment This differs fundamentally from conventional result-based penalty / reward items: conventional penalty / reward items affect the final score of candidate materials, used to adjust the evaluation of candidates after calculation; while the phase boundary sensing weights in this embodiment act multiplicatively on the Bayesian-optimized acquisition function. That is, before each round of active learning to select the next batch of high-fidelity computation candidates, the sampling priority of each component point in the chemical space is physically prior-redistributed. Because The mean of the predictions is derived from the proxy model. With uncertainty Decision, therefore The sampling decision is subject to multiplicative gating of predictive performance, cognitive uncertainty, and phase boundary physical priors: only candidate component points that possess both high expected value and high information gain, and are adjacent to the solid-state phase transition boundary, are given high sampling priority. The massive thermal effect of solid-state cooling materials is physically concentrated in a narrow component region near the first-order phase transition boundary. High-fidelity computation (L2's DFT / VASP, L3's DeepMD+LAMMPS) is costly; therefore, this multiplicative modulation is equivalent to actively shrinking the sampling quality to the high-potential phase boundary band under a fixed computing power budget, thereby increasing the number of high-potential candidate discoveries per unit computing power and reducing the proportion of low-value high-fidelity evaluations without increasing computing power. To clarify the essential difference between the phase boundary perception weight and the conventional result reward / penalty items, a comparison is shown in the table below; this difference should not be summarized simply as "multiplication or addition," but should be reflected simultaneously in four aspects: the timing of action, the source of information, the distribution of candidate sampling, and falsifiability.

[0304] Table 2 compares the technical differences between phase boundary sensing multiplicative weights and conventional reward / penalty items.

[0305]

[0306] To verify that the technical effect of phase boundary sensing weights indeed originates from the phase boundary information itself rather than the multiplicative operation itself, this embodiment conducted a fair ablation experiment on a physical mechanism-driven synthetic benchmark simulation. The simulation parameter configuration is shown in Table 3. Under the same high-fidelity computing budget, the same initial design, and the same random seed, the phase boundaries are all approximately given by a phase region classifier trained with finite labeled samples (non-truth value phase boundaries). The ablation results of the eight candidate strategies are shown in Table 4. The results show that compared with standard Bayesian optimization without phase boundary priors, the multiplicative phase boundary sensing acquisition function increases the number of high-potential candidates discovered per unit computing power by about 1.5 times, and reduces the proportion of low-value high-fidelity evaluations from about 21.9% to about 5.8%; it improves by about 2.9 times compared with random / high-throughput traversal; the key pairing differences and the approximately 95% confidence interval are shown in Table 5. To eliminate the loop caused by "encoding phase boundary information into the truth value" and to define the source of the effect, two further tests were conducted: First, a placebo control—after randomly shuffling the phase boundaries, the multiplicative strategy showed no significant difference from the standard Bayesian optimization (the average difference between paired strategies in Table 5 was +0.13, and the confidence interval crossed zero), indicating that the improvement came from the phase boundary information rather than the multiplicative operation itself; Second, a robustness test for phase boundary localization error—adding Gaussian displacement noise to the phase boundaries showed that the improvement monotonically decreased with increasing error and tended to disappear at a displacement of approximately 0.2 (normalized scale), as detailed in Table 6, indicating that the effect was positively correlated with the accuracy of phase boundary estimation and could be independently falsified. Furthermore, the sensitivity test of the number of labeled samples for the classifier is shown in Table 7: even with as few as 40 labeled samples, the advantage of the multiplicative phase boundary-aware strategy over the standard Bayesian optimization remained significant (the confidence intervals did not cross zero), indicating that the phase boundaries do not need to be precisely known, and a stable gain can be obtained by approximating the phase boundaries with a classifier trained on a suitable amount of labeled data, supporting the engineering feasibility of this method. It should be noted that the above is a verification of the synthesis mechanism; among them, the pure geometric phase boundary strategy can have a higher hit count than the multiplicative strategy under the synthesis setting of "high potential, i.e., close to the phase boundary". The incremental advantage of the multiplicative form over the additive / pure distance strategy needs to be further confirmed by offline replay of publicly available real data.

[0307] Table 3 shows the configuration table for synthetic baseline simulation parameters.

[0308]

[0309] Table 4 shows the comparison of ablation results using eight strategies (mean of 30 random seeds).

[0310]

[0311] Table 5 shows the comparison between key paired differences and approximately 95% confidence intervals.

[0312]

[0313] Table 6 shows the robustness test results of the impact of phase boundary localization noise on the multiplicative strategy advantage.

[0314]

[0315] Table 7 shows the feasibility test results of the impact of the number of labeled samples for the classifier on the advantage of the multiplicative strategy.

[0316]

[0317] The system employs multi-objective Bayesian optimization (MOBO) to maintain and update the Pareto front. Optimization objectives include, but are not limited to: maximizing adiabatic temperature change; maximizing isothermal entropy change; minimizing the driving field threshold; minimizing thermal hysteresis; and maximizing predicted cycle lifetime.

[0318] The system allows embedding engineering constraints, such as driving field upper limits, disabled elements, and toxicity constraints, into the acquisition function, thereby enabling constraint-aware batch candidate selection suitable for parallel task scheduling.

[0319] In each iteration, the system incorporates newly added computational data corresponding to regions with high uncertainty and high expected value into the training set and updates the surrogate model parameters, thereby gradually improving the model's prediction accuracy and convergence efficiency in high-potential regions.

[0320] like Figure 7 As shown, the active learning closed-loop process uses Caloric-DB literature data and Materials Project computational data as initial data. First, a surrogate model with graph neural network and uncertainty quantification (GNN+UQ) as its core is trained. Then, based on the combined acquisition function of expectation improvement (EI), confidence upper bound (UCB) and phase boundary perception weight, the scores of each candidate point are calculated. The top-K candidate batch with the highest scores is selected and put into the L1 / L2 / L3 computation and screening pipeline. After obtaining the computation or experimental results, the new data is written into Caloric-DB and the surrogate model is retrained. After completing one round of iteration, it returns to the training step along the dotted line to form a continuously optimized closed loop.

[0321] like Figure 8 As shown, the complete closed-loop process of Caloric-DB consists of six steps. Step 1 is candidate pooling, which involves uniformly incorporating candidate materials from literature sources, Materials Project sources, and generative methods into the Caloric-DB candidate pool. Step 2 uses GNN or compositional models (with uncertainty quantification) to perform surrogate prediction on the candidate materials, outputting the adiabatic entropy change ΔS and adiabatic temperature change ΔT. adThe predicted values ​​and confidence intervals are calculated; step 3 executes the phase boundary sensing acquisition function, using the product of EI / UCB and phase boundary weights as the scoring criterion, and selects Top-K high-value candidates from the candidate pool; step 4 executes multi-fidelity calculation, sequentially performing three levels of calculation: L1 machine learning potential, L2 DFT / VASP, and L3 DeepMD+LAMMPS, with each level's result appended with trace_id and written back to Caloric-DB; step 5 completes the surrogate model update, retraining the GNN or component model with the newly added L2 / L3 calculation data, and triggering a new round of acquisition function calculation and returning to step 3 along the dotted line after the validation set mean square error improves; step 6 updates the Pareto front, and the multi-objective scorer re-scores all candidates. If a significant update occurs, a human-machine collaborative analysis briefing is triggered, notifying researchers to review and intervene. The L1 layer uses lightweight machine learning potentials (such as M3GNet-like general potentials) or approximate DFTs (such as GGA-level) for fast calculation, including structural relaxation, formation energy calculation, and phase stability analysis (convex hull distance E). hull The algorithm first performs heuristic pre-scoring with no input or output on the recalled candidates; then, for candidates with structural identifiers, crystal structure files are retrieved from the structure database according to the identifiers to complete the structural information; finally, a machine learning force field (M3GNet-like general potential) is called to predict the energy per atom of the candidates and normalize it into a stability surrogate score, which is then compared with a preset threshold to complete the L1 fast screening.

[0322] Preferred screening criteria include: (E hull <0.1)eV / atom; No obvious collapse occurred during the structural optimization process.

[0323] The computational cost for this layer is typically around 0.01 to 0.1 core-hours per candidate.

[0324] The L2 layer employs the first-principles software VASP to perform high-precision structural relaxation and calculates corresponding key phase transition indicators based on different solid-state cooling effect types: For ECE: spontaneous polarization difference Born effective charge tensor, soft mode frequency, energy difference between ferroelectric and paraelectric phases, dielectric constant tensor; for MCE: magnetic moment difference Magnetic ordered phase energy difference and exchange parameters; for eCE: elastic constant tensor, energy difference between martensite and austenite phases, and phase transformation strain tensor; for BCE: equation of state parameters, pressure-induced phase transformation energy difference, and volume change rate. .

[0325] In addition, the system performs phonon dispersion calculations using Phonopy or DFPT to verify dynamic stability, and calculates field response tensors such as dielectric constant, permeability, and elastic modulus to evaluate surrogate indicators of driving field tolerance.

[0326] Preferred screening criteria include: no imaginary frequencies in the phonon spectrum; phase transition driving phase energy difference reaching a preset threshold; order parameter change reaching the corresponding branch threshold, for example: ECE: MCE: eCE: Phase transformation strain ≥1%; BCE: ΔV / V ≥1% .

[0327] in, This represents the change in polarization intensity of a material before and after the phase transition, used to characterize the polarization abrupt change amplitude of the ferroelectric-paraelectric phase transition in electrochemical materials, and is measured in units of... ; This represents the change in magnetization per unit chemical formula before and after the phase transition, used to characterize the magnitude of the abrupt change in magnetization during the magnetic phase transition of magnetic card materials. The unit is 1. (Bohr magnetons per chemical formula unit); Phase transformation strain represents the strain of the crystal lattice along the loading direction before and after the martensitic phase transformation; Indicates the change in unit cell volume before and after the phase transition. The ratio of the volume change to the unit cell volume V before the phase transition, i.e., the volume change rate, is a dimensionless quantity. The above quantities represent the phase transition order parameter changes corresponding to the effects of electric card, magnetic card, spring card, and pressure card, respectively. The larger the value, the more intense the phase transition and the greater the entropy change that can be obtained.

[0328] The computational cost for this layer is typically around 10 to 100 core-hours per candidate.

[0329] The L3 layer uses DPGEN / DeepMD-kit to train the material-specific machine learning potential and LAMMPS to perform large-scale, long-term molecular dynamics simulations. It can also be combined with phase field simulation software (such as MOOSE and MuPRO) to carry out phase transition domain evolution and reversibility verification.

[0330] This layer of calculations includes: performing large-scale, long-term MD simulations at near-DFT accuracy; and calculating reversible entropy changes under temperature-driving-field dual-parameter scanning conditions. and adiabatic temperature change The study aimed to evaluate the functional relationship between phase transition thermal hysteresis and loading / unloading rates; predict cyclic fatigue behavior; and verify the evolution of order parameter domains, phase transition reversibility, and critical behavior of the driving field based on phase field simulation.

[0331] The computational cost for this layer is typically between 1,000 and 10,000 core-hours per candidate.

[0332] The system employs a multi-fidelity machine learning approach to correct bias between different levels of computational precision, and the relationship can be expressed as follows:

[0333]

[0334] in, For high-fidelity calculation of predicted values ​​at the hierarchical level (such as L2 or L3); For low-fidelity computation of predicted values ​​(such as L1); The bias correction function is fitted by a machine learning model trained on high-fidelity data; This represents the characteristics of the candidate materials.

[0335] By training the bias correction model with a small amount of high-fidelity data, the prediction results of the low-cost level can be made closer to the high-precision calculation results, thereby reducing the overall computational cost.

[0336] To enable the write-back, benchmarking, and database storage of calculation results at different levels, this embodiment establishes standardized output specifications for the three calculation levels of L1, L2, and L3.

[0337] (1) L1 output specification:

[0338] Including: geometry-optimized crystal structure (CIF format), formation energy, and convex hull distance E. hull Estimated values ​​for bulk modulus and density, etc. When writing to Caloric-DB, it is preferable to include the machine learning potential version, input parameter set version, and computation time.

[0339] (2) L2 output specification:

[0340] It includes: high-precision relaxation structure, total energy, phonon dispersion data and imaginary frequency flags, as well as key phase transition quantities output according to different effect branches, such as polarization difference, magnetic moment difference, elastic tensor, state equation parameters and volume change rate, and also includes surrogate indicators such as field response tensor.

[0341] (3) L3 output specification:

[0342] This includes: machine learning potential files, MD trajectory files, and thermodynamic quantities obtained from trajectory statistics, such as order parameter hysteresis loops under temperature-driven field scanning, quasi-static reversible entropy change, adiabatic temperature change estimates, and kinetic thermal hysteresis widths. Phase field simulation outputs may also include sequence of order parameter domain evolution images, phase transition reversibility assessments, and results of critical behavior under driving fields.

[0343] When writing all calculation results to Caloric-DB, it is preferable to include complete calculation workflow version information, input file hash value, HPC job identifier, and calculation cost information to ensure traceability throughout the entire process.

[0344] In each round of active learning, the system updates the agent model based on the latest high-value computational data and dynamically adjusts the sampling strategy according to changes in the Pareto front and budget consumption. Preferred iteration termination conditions include: no significant updates to the Pareto front for N consecutive rounds; and reaching the preset computing power budget limit.

[0345] In the following scenarios, the system will forcibly pause the automated process and output an analysis brief to researchers to support human judgment: the emergence of a Pareto frontier breakthrough material that is significantly better than the current best candidate; the emergence of a novel candidate with high confidence but significantly deviating from the chemical space of the training set; anomalies such as strong imaginary frequencies or abnormally low phonon state densities appearing in L2 calculations; the machine learning potential training in the L3 stage failing to converge or the test error exceeding the threshold; and the difference between multifidelity predictions exceeding a set statistical threshold, such as exceeding 3σ.

[0346] The analysis brief preferably includes: the system's recommendation reasons and corresponding RAG evidence; the current uncertainty estimate; the suggested next calculation direction; and the corresponding expected information gain.

[0347] Through the aforementioned human-machine collaboration mechanism, researchers can intervene in system decision-making at critical junctures to improve R&D efficiency and reduce the ineffective investment of high-cost computing resources.

[0348] In summary, this embodiment discloses an intelligent discovery system and method for solid-state refrigeration materials based on a large language model and multi-agent collaboration. The system comprises five core modules: a cross-modal solid-state refrigeration material knowledge base (Caloric-DB) and a hybrid retrieval RAG engine (including evidence-free degradation paths and illusion protection mechanisms) oriented towards the solid-state refrigeration effect (SC); a hierarchical multi-agent collaborative network based on a LangGraph state machine (planner, literature scout, structure generator, simulation coordination, and evaluation / criticism agents); a Bayesian active learning multi-objective optimization module incorporating prior knowledge of ferroelectric-paraelectric phase boundaries; a multi-fidelity computational screening pipeline from L1 (ML potential fast screening) to L2 (DFT / VASP solid-state refrigeration fine screening) and then to L3 (DeepMD+LAMMPS+phase field simulation high-fidelity); and a computational closed-loop mechanism that writes the multi-fidelity computation results back to the knowledge base in real time to drive the continuous iteration of the agent model. This embodiment uses a phase boundary sensing physical constraint acquisition function to concentrate computational power on the high-potential region near the phase boundary of the ferroelectric-paraelectric phase diagram, and combines multi-objective Pareto optimization to achieve high ΔS / ΔT. ad This systematic exploration of Pareto frontiers through computational screening, characterized by "low driving electric field, low thermal hysteresis, and high cycling stability," ensures 100% traceability and reproducibility by providing complete literature / computational evidence chains and uncertainty estimates for all recommended conclusions. This embodiment is entirely computationally based and can run independently without physical experiments. Compared to traditional high-throughput computational screening, this embodiment can increase the number of high-potential solid-state cooling material candidates discovered with fixed computing power by at least 2 times and shorten the cycle from "research idea to high-confidence candidate material computational screening report" by at least 50%, providing a systematic solution for accelerated computational discovery of solid-state cooling materials.

[0349] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0350] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0351] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0352] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0353] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0354] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0355] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0356] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for intelligent screening of solid-state refrigeration materials based on large models and intelligent agents, characterized in that, include: The task of screening solid-state phase change refrigeration materials was obtained. Furthermore, based on the established cross-modal knowledge base of solid-state refrigeration materials, the constructed intelligent agent performs multi-level computational verification of the screening task according to the directed acyclic graph workflow to obtain the processing results of the screening task. The directed acyclic graph workflow includes: candidate generation and step-by-step computational screening and evaluation recommendation. The output of candidate generation serves as the input for step-by-step computational screening and evaluation recommendation. The constraint mechanism of the surrogate model is implemented by embedding phase boundary sensing weights in the physical constraint optimization acquisition function. The phase boundary sensing weights are kernel functions of the normalized distance from the state point of the candidate material to the phase boundary of the phase diagram, and increase as the normalized distance decreases. The phase boundary of the phase diagram is determined by the change in stable phase labels of adjacent regions after the phase region classifier divides the composition-state space. The physical constraint optimization acquisition function is jointly determined by the prediction mean and prediction uncertainty of the surrogate model. The phase boundary sensing weights act multiplicatively on the standard Bayesian optimization acquisition function. Thus, before selecting the next batch of candidates for high-fidelity computation in each round, the sampling priority of the candidates is redistributed.

2. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents as described in claim 1, characterized in that, The candidate generation process includes: Based on the aforementioned screening task, keyword retrieval and dense vector retrieval are performed; Material knowledge-constrained filtering is applied to the results of keyword retrieval and dense vector retrieval. The ranking results of keyword retrieval and dense vector retrieval after filtering by material knowledge constraints; Furthermore, a second ranking is performed on the merged Top-N results to obtain the candidate generation results.

3. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents as described in claim 2, characterized in that, The candidate generation adopts an iterative retrieval mechanism, which breaks down complex problems into several verifiable sub-problems. After each round of retrieval, the next round of retrieval query is automatically generated based on evidence coverage, conflict status, and missing fields, until the iterative convergence condition is met.

4. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents as described in claim 1, characterized in that, The step-by-step calculation and filtering process includes: A lightweight machine learning model is used for fast screening of the first-level L1. Based on rapid screening, a second-level L2 precise screening is performed based on first principles. Based on precise screening and manual approval, high-fidelity calculations of the third level (L3) are performed using molecular dynamics simulations.

5. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents as described in claim 4, characterized in that, A multi-fidelity machine learning approach is used to correct biases between different levels of computational precision.

6. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents as described in claim 1, characterized in that, The multi-objective scoring function of the proxy model for: in, The overall score for candidate materials ranges from 0 to 1. Score by performance; For stability; For preparability; Based on novelty; Cost-based; to Configurable weights.

7. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents as described in claim 1, characterized in that, The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents also includes: The computational data generated from multi-level computational verification is written back to the cross-modal knowledge base of solid-state refrigeration materials, and the proxy model is updated based on the computational data to form a continuously evolving closed loop.

8. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents as described in claim 1, characterized in that, The physical constraint optimization acquisition function for: in, This is the physical constraint acquisition function; Optimize the acquisition function for standard Bayesian methods; For phase boundary sensing weight function; These are the composition points of the candidate materials; Candidate component points The distance to the boundary (Boundary) of the phase diagram; For bandwidth parameters; Let be the normalized state vector of the candidate material, where As a component variable, For temperature, As the driving field, For the sake of pressure; in actual implementation, only the dimensions required by the target system are retained; For the phase boundary point set, For the first One phase boundary point; For the first The scale normalization parameter of the dimension can be taken as the allowable range of the dimension, the standard deviation of the training data, or a preset physical scale; To explore the lower bound, a method is used to ensure that regions far from the currently estimated phase boundary still retain non-zero sampling probabilities; For the standard Bayesian optimized acquisition function after nonnegativity processing, the desired improvement is taken. Or the upper confidence bound after nonnegative transformation ; and distance The smaller, The larger.

9. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents as described in claim 1, characterized in that, The solid-state refrigeration material cross-modal knowledge base is constructed based on open databases, domain literature, and project-generated calculation results. The intelligent screening method for solid-state refrigeration materials based on large models and intelligent agents further includes: preprocessing data from open databases, domain literature, and project-generated calculation results to obtain the dataset of the cross-modal knowledge base for solid-state refrigeration materials. The preprocessing includes: multimodal data preprocessing, entity relation extraction based on instruction fine-tuning of the large language model, and cross-validation and confidence scoring. The cross-modal knowledge base for solid-state refrigeration materials logically includes: a material data domain, a workflow record domain, and an evidence chain domain.

10. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory, so that the electronic device performs the intelligent screening method for solid-state cooling materials based on large models and intelligent agents as described in any one of claims 1-9.

Images