A method and system for autonomous calculation of material phase diagrams based on a multi-layered intelligent agent collaborative architecture

Abstract

This invention relates to the fields of materials design and thermodynamics calculation in materials science, specifically to a method and system for autonomous calculation of material phase diagrams based on a multi-layered intelligent agent collaborative architecture. The method, through the collaborative work of error-correcting agents, historical dialogue merging agents, requirement rewriting agents, requirement enhancement agents, structured recognition agents, and code generation agents, receives and identifies multi-round design and calculation dialogue requirements, corrects input errors, breaks down complex requirements into executable sub-tasks and enhances instruction explicitness, and generates a precise set of calculation scripts after structuring the requirements. This application, through the deep integration of multi-layered intelligent agents and large models, achieves fully intelligent thermodynamic phase diagram calculation driven by natural language, significantly improving code generation robustness and requirement processing accuracy, lowering the technical threshold, and providing efficient tool engine support for intelligent materials design.

Description

Technical Field

[0001] This invention relates to the field of materials design and thermodynamics calculation technology, specifically to an autonomous calculation method and system for material phase diagrams based on a multi-layer intelligent agent collaborative architecture. Background Technology

[0002] Material phase diagrams, as tools revealing the interactions between material composition, microstructure, and external conditions (temperature, pressure), are known as the "maps" of materials design, playing a crucial role in alloy composition design, process parameter optimization, material performance prediction, and material failure analysis. However, relying on experimental determination of phase diagrams requires significant human and material resources, is time-consuming, and has technical limitations under extreme conditions such as ultra-high temperature and ultra-high pressure. The thermodynamic calculation method and supporting software based on CALPHAD (Calculation of Phase Diagrams) technology, leveraging a database of Gibbs free energy expressions for low-component alloy systems, can accurately extrapolate the phase transformation behavior and thermophysical properties (such as specific heat capacity and diffusion coefficient) of complex multi-component alloys under extreme service environments such as high temperature and high pressure. This technology has become a core tool for material composition optimization design, phase transformation kinetic mechanism analysis, and phase diagram construction, demonstrating significant application value in high-end materials research and development fields such as aerospace and energy equipment.

[0003] Currently, the deep integration of artificial intelligence (AI) technology with the CALPHAD method, using AI to achieve high-throughput CALPHAD calculations and combining it with machine learning (ML) algorithms to mine data correlation patterns, has become a key direction for promoting the transformation of materials research and development models. This integration path can construct an AI-driven, mechanism-supported framework for forward design and reverse optimization of materials. It can not only rely on the thermodynamic mechanism of CALPHAD to ensure the scientific nature of the design, but also use the efficient data processing capabilities of AI to break through the efficiency bottleneck of traditional experimental trial and error methods, providing a new paradigm for the rapid research and development of high-performance materials.

[0004] However, existing commercial thermodynamic calculation software still faces significant application barriers and technical limitations, hindering its widespread adoption in engineering scenarios and AI-driven materials research and development. Firstly, the software requires users to have a high level of thermodynamic expertise, and for complex tasks such as high-throughput property calculations, users need to be proficient in specific scripting languages ​​(such as TC-Python for Thermo-Calc and API scripts for JMatPro), increasing the difficulty for researchers without a professional background. Secondly, the software's interaction mechanism lacks flexibility. Users must strictly adhere to the specific script formats of different software when initiating calculation requests, and if key parameters (such as temperature range, alloy element types and contents) need to be adjusted during batch calculations, the entire script often needs to be rewritten, making rapid parameter iteration and efficient data production impossible. Thirdly, the lack of a unified parameter consistency verification mechanism during calculations makes it difficult to guarantee the homogeneity of batch calculation data (such as thermodynamic database versions and calculation model assumptions). The output data often requires additional manual screening and standardization, failing to directly meet the input requirements of AI models for high-quality, highly consistent training data, thus affecting the accuracy and efficiency of AI-driven materials design. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method and system for autonomous calculation of material phase diagrams based on a multi-layer intelligent agent collaborative architecture. Through natural language interaction, it identifies user needs and decomposes user calculation tasks. Through the multi-intelligent agent collaborative architecture, it achieves high-efficiency parallel operation of calculation tasks, batch generation of high-quality calculation data, and automatic saving of task result data, thereby overcoming the problems existing in current technologies and commercial systems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] On the one hand, this application provides an autonomous calculation method for material phase diagrams based on a multi-layer intelligent agent cooperative architecture, including: Step S1: Receive natural language instructions input by the user, wherein the natural language instructions include phase diagram calculation requirements or high-throughput thermal property calculation requirements; wherein the natural language instructions are single-turn dialogue instructions or multi-turn dialogue instructions. Step S2: The error correction agent identifies the error type in the natural language instruction, and accurately modifies the correctable errors according to the error type. For the uncorrectable errors, the error cause summary is output. The historical dialogue merging agent integrates the context of the natural language instruction to generate a semantically consistent fused instruction. Step S3: By performing computation type splitting, enumeration condition splitting, and Cartesian product combination by the demand rewriting agent, the complex requirements of the fusion instruction are decomposed into independent sub-tasks. By using the demand enhancement agent to supplement and improve the computation type parameters and system information parameters of all the independent sub-tasks through hard coding enhancement and LLM enhancement, a standardized single-point computation task instruction is generated. Step S4: The standardized single-point calculation task instruction is parsed and parameters are extracted by the structured recognition agent to generate a structured task description containing element system, control variables and calculation target. Based on the structured task description, the corresponding thermodynamic database and script template are called by the code generation agent to generate a set of calculation scripts that conform to the CALPHAD calculation software. Step S5: Invoke the CALPHAD calculation software and execute the calculation script set to obtain the calculation results; wherein, the calculation results include: phase diagram image data or thermodynamic property numerical data; Step S6: Through the result processing intelligent agent, the display format of the calculation results is analyzed according to the user's needs, and the calculation results are summarized and visualized to show the user phase diagram, physical property curve and data table; Step S7: Obtain the user's new natural language modification instructions, return to step S2 to start the multi-round thought chain generation mechanism, iteratively execute the requirement processing and calculation process until the user's requirements are met.

[0008] Furthermore, in the method described above, step S2 includes: The error-correcting agent identifies erroneous instructions in the natural language commands and determines the type of all erroneous instructions. The error command is processed accordingly based on its type. The historical dialogue merging agent performs dialogue system identification on the natural language instructions after erroneous instruction processing to determine the interaction scenario to which the natural language instructions belong; By performing intent-context dual detection on the natural language instruction and determining whether the natural language instruction needs to be semantically fused with the historical context based on the interaction scenario to which the natural language instruction belongs; If so, the natural language instruction is subjected to multi-dimensional semantic execution fusion to generate a fusion instruction with standardized expression, and the fusion instruction is subjected to semantic stability verification and fusion accuracy verification.

[0009] Furthermore, in the methods described above, the types of erroneous instructions include at least: ingredient ratios that violate chemical laws, physically invalid operation instructions, improper use of units, deviations in numerical ranges, and semantic-logical contradictions; The step of processing the error command according to its type includes: Based on industry standards, the erroneous instructions regarding the violation of chemical laws in ingredient ratios, improper use of units, and deviations in numerical ranges have been modified. The error attribution algorithm is used to summarize the error causes of physically invalid operation instructions and semantically logically contradictory error instructions from the dimensions of semantic logic, domain rules, and operation feasibility, generate error descriptions, and send the natural language instructions and error descriptions to the user.

[0010] Furthermore, the method described above, which involves decomposing the complex requirements of the fusion instruction into independent subtasks by splitting the computation type, enumerating the conditions, and combining the Cartesian product through demand rewriting of the intelligent agent, includes: Determine whether the fusion instruction is a high-throughput computing requirement with multiple computing types or multiple component conditions. If so, the fusion instruction is split into computing types, enumerated conditions, and combined into Cartesian products by rewriting the agent according to the requirement, generating multiple independent subtasks. Perform iterative evaluations of semantic stability, expressive naturalness, and logical integrity for all the aforementioned independent subtasks.

[0011] Furthermore, the method described above, wherein the demand-enhancing agent supplements and improves the computation type parameters and system information parameters of all the independent subtasks through hard-coding enhancement and LLM enhancement, and generates standardized single-point computation task instructions, includes: The required enhancement agent performs hard-coding enhancement and LLM enhancement on all the independent subtasks, supplements and clarifies the computation type parameters, system information parameters and language style of the independent subtasks, and generates standardized single-point computation task instructions. All standardized single-point computation task instructions are tested using LLM evaluation and iterative judgment.

[0012] Furthermore, the method described above, wherein parsing the standardized single-point computation task instructions and extracting parameters through a structured recognition agent to generate a structured task description containing an element system, control variables, and computational objectives, includes: The structured recognition agent performs parameter semantic lookup, information extraction, model fine-tuning verification, and iterative judgment on the standardized single-point calculation task instructions to generate a structured task description in standardized JSON format; wherein, the structured task description includes element system field, temperature field, pressure field, component ratio field, calculation target field, and output requirement field.

[0013] Furthermore, the method described above, wherein the step of generating a set of computational scripts conforming to the CALPHAD computational software by calling the corresponding thermodynamic database and script template through a code generation agent based on the structured task description includes: The intelligent agent is generated through the code, and a material thermodynamics database that matches the description of the structured task is searched based on the element system field in the structured task. Based on the calculation target field and output requirement field in the structured task, a script template that matches the structured task is found, and a set of calculation scripts containing database reading, condition setting, calculation execution and result output is generated.

[0014] Furthermore, the multi-turn thinking chain generation mechanism described above includes: a multi-turn dialogue thinking chain and an error recognition thinking chain; Iterative optimization is achieved through the fusion of structured data and multi-round results via the aforementioned multi-round dialogue thought chain. The error identification thinking chain enables accurate diagnosis by extracting and analyzing fields of erroneous instructions, identifying the cause of errors, and inferring data contradictions.

[0015] On the other hand, this application provides an autonomous material phase diagram calculation system based on a multi-layer intelligent agent cooperative architecture, including... The intelligent demand perception module is used to receive natural language commands input by the user and guide the user to refine the calculation requirements. The calculation requirements include material phase diagram calculation requirements or material thermophysical property calculation requirements. The natural language commands include single-turn dialogue commands and multi-turn dialogue commands. The material phase diagram calculation requirements include binary phase diagrams, ternary isothermal section diagrams, and temperature / composition-property phase diagrams. The material thermophysical property calculation requirements include basic property calculations, thermodynamic property calculations, and mechanical property calculations. The intelligent interaction enhancement module is used to identify the error types in the natural language instructions through the error correction agent, and to accurately modify correctable errors according to the error types, and to output a summary of the error causes for uncorrectable errors. The module integrates the context of the natural language instructions through the historical dialogue merging agent to generate semantically consistent fused instructions. The intelligent task splitting module is used to split the complex requirements of the fusion instruction into independent subtasks by performing computation type splitting, enumeration condition splitting, and Cartesian product combination through the requirement rewriting agent. The requirement enhancement agent supplements and improves the computation type parameters and system information parameters of all the independent subtasks through hard coding enhancement and LLM enhancement, and generates standardized single-point computation task instructions. The intelligent script generation module is used to parse the standardized single-point calculation task instructions and extract parameters through a structured recognition intelligent agent, generate a structured task description containing element system, control variables and calculation target, and generate a set of calculation scripts that conform to CALPHAD calculation software by calling the corresponding thermodynamic database and script template through the code generation intelligent agent according to the structured task description. The high-throughput thermodynamic calculation module is used to call the CALPHAD calculation software and execute the calculation script set to obtain calculation results; wherein, the calculation results include: phase diagram image data or thermodynamic property numerical data; The intelligent result visualization module is used to analyze the display format of the calculation results according to user needs through the result processing intelligent agent, perform data summarization and visualization processing on the calculation results, and display phase diagrams, physical property curves and data tables to users; The multi-round iteration and thought chain generation module is used to obtain new natural language modification instructions from the user, return to the intelligent interaction enhancement module to start the multi-round thought chain generation mechanism, iteratively execute the requirement processing and calculation process until the user's needs are met. The intelligent computing resource management module is used to deploy the adjusted natural language model and knowledge base, providing computing power support for reasoning tasks.

[0016] Furthermore, in the system described above, the intelligent demand perception module adopts a web interface interaction method, supporting the latest versions of Chrome, Edge, Firefox and Safari browsers, ensuring that the interface display and function operation are consistent across different browsers.

[0017] The beneficial effects of this invention are as follows: 1. By collaborating among various intelligent agents, the system achieves context merging and error correction of natural language instructions, decomposition and enhancement of complex requirements for fused instructions, and structured recognition of standardized single-point computing task instructions and precise code generation based on structured task descriptions. This enables precise processing of the entire process from input, execution to output, significantly reducing human error.

[0018] 2. The multi-turn dialogue thinking chain generation enables deep integration and iterative optimization of the context. The error identification thinking chain accurately locates and solves problems in data or code, improves system robustness, and supports an efficient "hypothesis-verification" scientific research closed loop.

[0019] 3. Natural language interaction combined with intelligent demand processing eliminates the need for users to have professional scripting or thermodynamics knowledge; batch task concurrent execution and multi-round iterative optimization significantly reduce repetitive operations and waiting time, improving scientific research efficiency.

[0020] 4. Supports single-turn / multi-turn dialogue, complex requirement breakdown, high-throughput computing and other application scenarios. The modular architecture design makes it easy to add intelligent agents or expand functions, and can adapt to the intelligent computing design requirements of new material systems. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart provided in one embodiment of the material phase diagram autonomous calculation method based on a multi-layer intelligent agent collaborative architecture of the present invention; Figure 2 This is a hardware and software architecture diagram provided in an embodiment of the material phase diagram autonomous calculation method based on a multi-layer intelligent agent collaborative architecture of the present invention. Figure 3 This is a schematic diagram of a structure provided by an embodiment of the material phase diagram autonomous calculation system based on a multi-layer intelligent agent collaborative architecture of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0024] Figure 1 This is a flowchart provided by an embodiment of the material phase diagram autonomous calculation method based on a multi-layer intelligent agent collaborative architecture of the present invention. Figure 2 This is a hardware and software architecture diagram provided in one embodiment of the material phase diagram autonomous calculation method based on a multi-layer intelligent agent collaborative architecture according to the present invention. Please refer to... Figure 1 and Figure 2 This embodiment may include the following steps: Step S1: Receive natural language instructions input by the user. The natural language instructions include phase diagram calculation requirements or high-throughput thermophysical property calculation requirements. The natural language instructions can be single-turn or multi-turn dialogue instructions. Step S2: The error correction agent identifies the error type in the natural language instruction, and accurately modifies the correctable errors according to the error type. For the uncorrectable errors, the error cause summary is output. The historical dialogue merging agent integrates the context of the natural language instruction to generate a semantically consistent fused instruction. Step S3: By rewriting the requirements, the intelligent agent performs computation type splitting, enumeration condition splitting, and Cartesian product combination, the complex requirements of the fusion instruction are decomposed into independent sub-tasks. The requirement enhancement agent supplements and improves the computation type parameters and system information parameters of all independent sub-tasks through hard coding enhancement and LLM enhancement, and generates standardized single-point computation task instructions. Step S4: The structured recognition agent parses the standardized single-point calculation task instructions and extracts parameters to generate a structured task description containing element system, control variables and calculation objectives. Based on the structured task description, the code generation agent calls the corresponding thermodynamic database and script template to generate a set of calculation scripts that conform to the CALPHAD calculation software. Step S5: Call the CALPHAD calculation software and execute the calculation script set to obtain the calculation results; wherein, the calculation results include: phase diagram image data or thermodynamic property numerical data; Step S6: Through the result processing intelligent agent, the display format of the calculation results is analyzed according to the user's needs. The calculation results are summarized and visualized to show the user phase diagram, physical property curve and data table. Step S7: Obtain the user's new natural language modification instructions, return to step S2 to start the multi-round thought chain generation mechanism, iteratively execute the requirement processing and calculation process until the user's requirements are met.

[0025] Preferably, step S2 includes: The error-correcting agent identifies erroneous instructions in natural language commands and determines the type of all erroneous instructions. The error command should be handled accordingly based on its type. By merging historical dialogues with intelligent agents, the dialogue system of natural language instructions after erroneous instructions is identified, and the interaction scenario to which the natural language instructions belong is determined. By performing intent-context dual detection on natural language instructions and determining whether natural language instructions need to be semantically fused with historical context based on the interaction scenario to which the natural language instructions belong; If so, the natural language instructions are subjected to multi-dimensional semantic execution fusion to generate standardized fused instructions, and the semantic stability and fusion accuracy of the fused instructions are verified.

[0026] Preferably, the types of erroneous instructions include at least: ingredient ratios that violate chemical laws, physically invalid operation instructions, improper use of units, deviations in numerical ranges, and semantic and logical contradictions; Based on the type of the error command, the error command is handled accordingly, including: Based on industry standards, erroneous instructions that violate chemical laws, include incorrect component ratios, improper use of units, and deviations in numerical ranges have been modified. By using an error attribution algorithm, the algorithm summarizes the error causes of physically invalid operation instructions and semantically contradictory error instructions from the dimensions of semantic logic, domain rules, and operation feasibility, generates error descriptions, and sends the natural language instructions and error descriptions to the user.

[0027] Understandably, the core responsibility of the "error-correcting agent" is to identify and handle instructions containing various errors (such as ingredient ratios that violate chemical laws or operational instructions that are physically invalid). The agent first performs an "error type judgment" on the input instruction and then intelligently triages it based on the judgment result: if the "error identification model" determines that "there are no errors," the instruction is directly allowed; if it detects "only correctable errors" (such as improper unit usage or numerical range deviation), it invokes a "domain knowledge-driven instruction modification engine" to perform "precise instruction modification" based on industry standards (such as material synthesis ratio specifications or parameter thresholds in physical experiments); if there are "fundamental errors that cannot be modified" (such as settings that violate basic scientific principles) or uncovered anomalies, the system will use an "error cause attribution algorithm" to "summarize the error cause" from dimensions such as "semantic logic," "domain rules," and "operational feasibility," forming an explainable error description.

[0028] The "Historical Dialogue Merging Agent" focuses on the integrated processing of multi-turn dialogue inputs. This agent determines the interaction scenario (such as temperature control commands or parameter configurations) of the current command through "dialogue system recognition," and analyzes whether the current command needs semantic fusion with the historical context using "intent-context dual detection." Once fusion is determined, a "multi-dimensional semantic execution fusion" operation is initiated. For example, the temperature setting "800℃" and the adjustment command "increase by 200K" are merged into a standardized expression "1073.15K." Subsequently, the fusion result is comprehensively tested through multi-dimensional evaluation criteria such as "semantic stability verification" (ensuring no semantic shift after fusion) and "fusion accuracy verification" (such as unit conversion accuracy and logical consistency), ultimately generating a valid command that integrates historical context, is semantically accurate, and conforms to interaction logic.

[0029] Preferably, by rewriting the computation type, enumerating the conditions, and combining the Cartesian product of the intelligent agent's execution, the complex requirements of the fusion instruction are decomposed into independent subtasks, including: Determine whether the fusion instruction is a high-throughput computing requirement with multiple computing types or multiple component conditions. If so, the fusion instruction is split into computing types, enumerated conditions, and combined into Cartesian products by the requirement rewriting agent to generate multiple independent subtasks. Iterative evaluations of semantic stability, expressive naturalness, and logical integrity are performed on all independent subtasks. Understandably, the "requirement rewriting agent" bears the crucial responsibility of deconstructing the original complex requirements. Taking an instruction like "Calculate the density and enthalpy of the FECrNi system at 800℃ with Cr mass fractions of 0.1 and 0.2 and Ni mass fraction of 0.1" as an example, which involves multiple calculation tasks and combinations of conditions, it decomposes the original requirement into a series of independent and specific subtasks, such as "Calculate the density of the FECrNi system at 800℃ with Cr mass fractions of 0.1 and 0.2 and Ni mass fraction of 0.1," through operations such as "calculation type splitting" (separating density and enthalpy calculations independently), "enumeration condition splitting" (enumerating different Cr and Ni mass fractions one by one), and "Cartesian product combination" (combining calculation types with various component conditions). This process will be optimized through "effectiveness evaluation" (assessing whether the splitting effect fully covers the needs, whether the semantic stability is accurate and unambiguous, and whether the expression is natural and easy to understand) and "iterative judgment" to ensure that the split task list is semantically stable, naturally expressed and logically complete.

[0030] Preferably, the demand-enhancing agent supplements and improves the computation type parameters and system information parameters of all independent subtasks through hard-coding enhancement and LLM enhancement, generating standardized single-point computation task instructions, including: By using a demand-enhancing agent, all independent subtasks are hard-coded and LLM-enhanced, supplementing and clarifying the computation type parameters, system information parameters, and language style of the independent subtasks, and generating standardized single-point computation task instructions. All standardized single-point computation task instructions are tested using LLM evaluation and iterative judgment.

[0031] Understandably, the core mission of the "demand-enhanced agent" is to further improve the explicitness of instructions to ensure they meet executable requirements. It employs a dual-track strategy of "hard-coded enhancement" and "LLM (Large Language Model) enhancement," focusing on supplementing and clarifying key information such as "computation type" (e.g., whether it's a density or enthalpy calculation), "system information" (e.g., properties of the FECrNi system, parameters at 800℃), and "language style" (e.g., the way instructions are expressed). Through "LLM evaluation" (assessing dimensions such as explicitness of computation type, completeness of system information, naturalness of language expression, and consistency of computational goals) and "iterative judgment," it ensures that the enhanced instructions maintain consistency in computational goals and achieve completeness in system information.

[0032] Preferably, the standardized single-point computation task instructions are parsed and parameters are extracted by a structured recognition agent to generate a structured task description containing an element system, control variables, and computational objectives, including: By using a structured recognition agent, standardized single-point calculation task instructions are processed through parameter semantic lookup, information extraction, model fine-tuning verification, and iterative judgment to generate a standardized JSON-formatted structured task description. The structured task description includes fields for element system, temperature, pressure, component ratio, calculation target, and output requirements.

[0033] Understandably, after receiving a natural language query input by the user (such as "single-point calculation of the Fe-Cr-Ni system"), the "structured recognition agent" analyzes, evaluates, and iteratively judges the query intent through the processes of "parameter semantics / search" and "information extraction" and with the help of auxiliary tools such as "fine-tuning model / format verification" until a standardized structured JSON object is generated. This object precisely encapsulates all the parameters required for the calculation (such as elements, components, temperature, etc.).

[0034] Preferably, based on the structured task description, a code-generating agent calls a thermodynamic database and script templates to generate a set of computational scripts conforming to the OpenCalphad syntax specification, including: The intelligent agent is generated by code, and a material thermodynamics database that matches the structured task description is searched based on the element system field in the structured task. Based on the calculation target fields and output requirement fields in the structured task, a script template that matches the structured task is found, and a set of calculation scripts that includes database reading, condition setting, calculation execution, and result output is generated.

[0035] Understandably, structured JSON is used as input to the "code generation / agent". The code generation agent calls "auxiliary tools" (TDB thermodynamic database and code template library), and after the steps of "evaluating the effect" and "generating code", it finally compiles and generates executable OCM (OpenCalphadMacro) code that precisely corresponds to the computational requirements.

[0036] The preferred multi-turn thinking chain generation mechanism includes: multi-turn dialogue thinking chain and error recognition thinking chain; Iterative optimization is achieved through the fusion of structured data and the fusion of results from multiple rounds of dialogue thinking chains. Accurate diagnosis is achieved by extracting and analyzing fields of erroneous instructions, identifying the cause of errors, and inferring data contradictions through an error identification thinking chain.

[0037] Understandably, the multi-turn dialogue thought chain generation begins with "analyzing the extraction process of key fields from supplementary data," first analyzing the fusion process of multi-source data, then constructing a code mapping process to establish logical connections, and finally synthesizing a multi-turn dialogue thought chain. It interacts with the ordinary COT generation module through "structured data fusion," and also involves a cyclical optimization mechanism for "multi-turn result fusion." Both ends are associated with "guided generation of standard code," emphasizing the iterative reasoning process of supplementary data fusion and code generation in multi-turn interaction scenarios.

[0038] The error identification thinking chain generation process focuses on error investigation scenarios. Starting with the "analysis of key field extraction process," it identifies the root cause of the error to pinpoint the problem's origin, then infers the core root cause of data contradictions, and finally synthesizes a thinking chain for error diagnosis. Based on field extraction and analysis, it progressively completes error identification and data contradiction analysis, providing a clear reasoning path for problems in data or code, helping to accurately locate and solve problems.

[0039] In specific practice: Step 1: Command Reception: The user inputs natural language commands through the system's web interface, supporting single-turn or multi-turn dialogue. For example, a user previously inputted: "Calculate the density and heat of fusion of the three components FE, C, R, and N, at a temperature of 600℃, with a Cr mass fraction of 0.2% and a Ni mass fraction of 0.1%"; the current input is: "Increase the ambient temperature by 200K." The system receives this multi-turn dialogue command and enters the intelligent interaction enhancement processing.

[0040] Step 2: Intelligent Interaction Enhancement Processing Historical dialogue merging agent processing: Dialogue system recognition: Determine that the current query belongs to the "phase diagram calculation parameter adjustment" interaction scenario; Intent-Context Dual Detection: Analyzing the current "temperature increases by 200K" requires fusion with the historical "600℃" context; Multi-dimensional semantic fusion: 600℃ (373.15K) is converted into thermodynamic temperature, and 200K is added to obtain 1073.15K. The composition parameters are merged to generate a preliminary fusion command: "Please calculate the density and heat of FE, C, and N for me. The temperature is 1073.15K, and the composition is 0.2% Cr and 0.1% Ni by mass." Semantic stability verification and fusion accuracy verification: Confirm that the temperature units are consistent, the component parameters are not offset, and the logic is consistent after fusion.

[0041] Error-correcting agent processing: Assuming a previous user input contained an incorrect version: "Calculate the density and heat of fusion of the three components FE, C, R, N, and N, at a temperature of 1073.15 K, with a Cr mass fraction of 1.1 and a Ni mass fraction of -0.1," the error-correcting agent would execute: Error type judgment: The domain rules detected a Cr mass fraction of 1.1 (outside the 0-1 range) and a Ni mass fraction of -0.1 (negative, violating the chemical ratio rules), which are "correctable errors". Precisely modify the query: Based on the material composition ratio specification, correct the Cr mass fraction to 0.1 (within a reasonable range) and the Ni mass fraction to 0.1 (non-negative and in compliance with the total constraint), and finally output the effective fusion command: "Please calculate the density and heat of FE, CR, and NI for me. The temperature is 1073.15K, and the composition is Cr mass fraction 0.1 and Ni mass fraction 0.1". If there is an uncorrectable error (such as "Calculate the phase diagram of the Fe-Cu system at -100K"), the error summary will be output as follows: "Temperature of -100K violates the common sense of physics (temperature cannot be negative), and it is recommended to adjust it to a reasonable value in the range of 800-1600K".

[0042] Step 3: Refining Intelligent Requirements The fused valid command is: "Calculate the density and heat of fusion of the three components FE, C, R, and N, at a temperature of 1073.15 K, with Cr mass fractions of 0.1% and 0.2%, and Ni mass fraction of 0.1%." This leads to further refinement of the intelligent requirements. Rewrite the intelligent agent processing requirements: Calculation type splitting: The two calculation types "density" and "heat capacity" are separated into independent calculation types; Enumeration condition decomposition: Enumerate the Cr mass fractions of 0.1 and 0.2, and the Ni mass fraction of 0.1 one by one; Cartesian product combination: Combining computation type and component conditions, it is broken down into 4 independent subtasks: ① Calculate the density of the FECrNi system at 1073.15 K with a Cr mass fraction of 0.1 and a Ni mass fraction of 0.1. ② Calculate the heat capacity of the FECrNi system at 1073.15 K with a Cr mass fraction of 0.1 and a Ni mass fraction of 0.1. ③ Calculate the density of the FECrNi system at 1073.15 K with a Cr mass fraction of 0.2 and a Ni mass fraction of 0.1. ④ Calculate the heat capacity of the FECrNi system at 1073.15 K with a Cr mass fraction of 0.2 and a Ni mass fraction of 0.1. Effect evaluation and iteration judgment: The effect of the split is evaluated from the aspects of semantic stability, naturalness of expression and logical integrity. After confirming that there are no omissions, the sub-task list is output.

[0043] Demand-enhancing agent processing: Hard-coded enhancements: supplement system properties (FECrNi is a common alloy system), default pressure value (1e5Pa), calculation step size (default 50K); LLM Enhancement: Clearly define the calculation type (density unit kg / m³, heat capacity unit J / (mol)). K), standardizing language expression style; Enhanced effect evaluation: By verifying the consistency of calculation objectives and the accuracy of parameters, standardized single-point calculation task instructions are generated, such as: "Single-point calculation: FECrNi alloy system, temperature 1073.15K, pressure 1e5Pa, Cr mass fraction 0.1, Ni mass fraction 0.1, calculate density (unit kg / m³)".

[0044] Step 4: Implementing and processing intelligent requirements: Structured recognition agent processing: Parameter semantic lookup: Identifying core parameters in standardized instructions based on a fine-tuning model; Information extraction: Extract the elemental system ["Fe","Cr","Ni"], temperature 1073.15K, pressure 1e5Pa, composition {"Cr":0.1,"Ni":0.1}, calculate the target "density", and unit "kg / m³"; Fine-tuning model verification and iterative judgment: Verify the logical validity of parameters (total component ≤ 1, temperature non-negative), and generate a standardized JSON-formatted structured task description.

[0045] Code generation agent processing: Auxiliary tool call: Locate the TDB thermodynamic database (fecrni.TDB) and density calculation script set template that are compatible with the Fe-Cr-Ni system; Integrating "code-description pair" construction and error identification (COT): Ensuring script syntax compliance through field analysis and error cause prediction; Generate .ocm script: Fill the template with structured parameters to generate a complete set of calculation scripts.

[0046] Step 5: Perform thermodynamic calculations. The script executes four subtasks in parallel using the OpenCalphad framework via the Python subprocess module (maximum parallelism of 10). Execution logs are captured in real time. For tasks that fail (e.g., database incompatibility), the cause is analyzed using an error identification logic chain (e.g., "TDB database does not cover this component range"), and up to three automatic retries are initiated. If the task still fails, the reason is noted.

[0047] Step Six: Result Processing and Display The results processing and visualization module processes the results of successfully executed subtasks: Phase diagram / property data: Generate density numerical tables (CSV format) and trend comparison charts; Structured processing: Output a computation summary, including parameter settings and execution status of each subtask; Results Display: The web interface displays image previews and data tables to users, and supports exporting in PDF / PNG / CSV formats.

[0048] Step 7: Multi-round iterative optimization After viewing the results, the user enters the modification command: "Increase the Cr mass fraction by 0.3 and calculate the heat capacity." The system then returns to step S2 to initiate the multi-round COT generation mechanism. Multi-turn dialogue thought chain: integrate historical context, merge newly modified parameters, and generate new fusion instructions; Repeat steps two through six: perform error correction, requirement breakdown (add subtasks corresponding to Cr=0.3), structured recognition, code generation and calculation in sequence; Results Update: Displays the calculation results of newly added subtasks to users, supports comparison of results before and after modification, and enables iterative optimization.

[0049] The present invention also provides a phase diagram computing device based on a multi-layer intelligent agent architecture for implementing the above method embodiments. Figure 3 This is a schematic diagram of the structure of an embodiment of the material phase diagram autonomous calculation system based on a multi-layer intelligent agent collaborative architecture according to the present invention. Figure 3 As shown, it includes: The intelligent demand perception module 1 is used to receive natural language commands input by the user and guide the user to improve the calculation requirements. The calculation requirements include material phase diagram calculation requirements or material thermal property calculation requirements. The natural language commands include single-turn dialogue commands and multi-turn dialogue commands. The material phase diagram calculation requirements include binary phase diagrams, ternary isothermal section diagrams and temperature / composition-property phase diagrams. The material thermal property calculation requirements include basic property calculations, thermodynamic property calculations and mechanical property calculations. The intelligent interaction enhancement module 2 is used to identify the error types in natural language instructions through the error correction agent, and to accurately modify correctable errors according to the error type, and output the error cause summary for uncorrectable errors. It integrates the natural language instruction context through the historical dialogue merging agent to generate semantically consistent fused instructions. The intelligent task splitting module 3 is used to split the computation type, enumerate the condition, and combine the Cartesian product by rewriting the requirements of the intelligent agent. It breaks down the complex requirements of the fusion instruction into independent sub-tasks. The requirement enhancement agent supplements and improves the computation type parameters and system information parameters of all independent sub-tasks through hard coding enhancement and LLM enhancement, and generates standardized single-point computation task instructions. The intelligent script generation module 4 is used to parse standardized single-point calculation task instructions and extract parameters through a structured recognition intelligent agent, generate a structured task description containing element system, control variables and calculation objectives, and generate a set of calculation scripts that conform to the CALPHAD calculation software by calling the corresponding thermodynamic database and script template through the code generation intelligent agent based on the structured task description. The high-throughput thermodynamic calculation module 5 is used to call the CALPHAD calculation software and execute the calculation script set to obtain the calculation results; the calculation results include: phase diagram image data or thermodynamic property numerical data; The intelligent result visualization module 6 is used to analyze the display format of the calculation results according to the user's needs through the result processing intelligent agent, perform data summary and visualization processing on the calculation results, and display phase diagrams, physical property curves and data tables to the user. The multi-round iteration and thought chain generation module 7 is used to obtain new natural language modification instructions from the user, return to the intelligent interaction enhancement module 2 to start the multi-round thought chain generation mechanism, iteratively execute the requirement processing and calculation process until the user's needs are met. The intelligent computing resource management module 8 is used to deploy the adjusted natural language model and knowledge base, providing computing power support for reasoning tasks.

[0050] Preferably, the intelligent demand perception module 1 adopts a web interface interaction method, which supports the latest versions of Chrome, Edge, Firefox and Safari browsers, ensuring that the interface display and function operation are consistent across different browsers.

[0051] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0052] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0053] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0054] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0055] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0056] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0057] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0058] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0059] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0060] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for autonomous calculation of material phase diagrams based on a multi-layered intelligent agent collaborative architecture, characterized in that, include: Step S1: Receive natural language instructions input by the user, wherein the natural language instructions include phase diagram calculation requirements or high-throughput thermal property calculation requirements; wherein the natural language instructions are single-turn dialogue instructions or multi-turn dialogue instructions. Step S2: The error correction agent identifies the error type in the natural language instruction, and accurately modifies the correctable errors according to the error type. For the uncorrectable errors, the error cause summary is output. The historical dialogue merging agent integrates the context of the natural language instruction to generate a semantically consistent fused instruction. Step S3: By performing computation type splitting, enumeration condition splitting, and Cartesian product combination by the demand rewriting agent, the complex requirements of the fusion instruction are decomposed into independent sub-tasks. By using the demand enhancement agent to supplement and improve the computation type parameters and system information parameters of all the independent sub-tasks through hard coding enhancement and LLM enhancement, a standardized single-point computation task instruction is generated. Step S4: The standardized single-point calculation task instruction is parsed and parameters are extracted by the structured recognition agent to generate a structured task description containing element system, control variables and calculation target. Based on the structured task description, the corresponding thermodynamic database and script template are called by the code generation agent to generate a set of calculation scripts that conform to the CALPHAD calculation software. Step S5: Invoke the CALPHAD calculation software and execute the calculation script set to obtain the calculation results; wherein, the calculation results include: phase diagram image data or thermodynamic property numerical data; Step S6: Through the result processing intelligent agent, the display format of the calculation results is analyzed according to the user's needs, and the calculation results are summarized and visualized to show the user phase diagram, physical property curve and data table; Step S7: Obtain the user's new natural language modification instructions, return to step S2 to start the multi-round thought chain generation mechanism, iteratively execute the requirement processing and calculation process until the user's requirements are met.

2. The method according to claim 1, characterized in that, Step S2 includes: The error-correcting agent identifies erroneous instructions in the natural language commands and determines the type of all erroneous instructions. The error command is processed accordingly based on its type. The historical dialogue merging agent performs dialogue system identification on the natural language instructions after erroneous instruction processing to determine the interaction scenario to which the natural language instructions belong; By performing intent-context dual detection on the natural language instruction and determining whether the natural language instruction needs to be semantically fused with the historical context based on the interaction scenario to which the natural language instruction belongs; If so, the natural language instruction is subjected to multi-dimensional semantic execution fusion to generate a fusion instruction with standardized expression, and the fusion instruction is subjected to semantic stability verification and fusion accuracy verification.

3. The method according to claim 2, characterized in that, The types of erroneous instructions include at least: ingredient ratios that violate chemical laws, physically invalid operation instructions, improper use of units, deviations in numerical ranges, and semantic and logical contradictions. The step of processing the error command according to its type includes: Based on industry standards, the erroneous instructions regarding the violation of chemical laws in ingredient ratios, improper use of units, and deviations in numerical ranges have been modified. The error attribution algorithm is used to summarize the error causes of physically invalid operation instructions and semantically logically contradictory error instructions from the dimensions of semantic logic, domain rules, and operation feasibility, generate error descriptions, and send the natural language instructions and error descriptions to the user.

4. The method according to claim 3, characterized in that, The process of decomposing the complex requirements of the fusion instruction into independent subtasks by splitting computation types, enumerating conditions, and combining Cartesian products through demand rewriting of the intelligent agent includes: Determine whether the fusion instruction is a high-throughput computing requirement with multiple computing types or multiple component conditions. If so, the fusion instruction is split into computing types, enumerated conditions, and combined into Cartesian products by rewriting the agent according to the requirement, generating multiple independent subtasks. Perform iterative evaluations of semantic stability, expressive naturalness, and logical integrity for all the aforementioned independent subtasks.

5. The method according to claim 4, characterized in that, The demand-enhanced agent, through hard-coded enhancement and LLM enhancement, supplements and improves the computation type parameters and system information parameters of all the independent subtasks, generating standardized single-point computation task instructions, including: The required enhancement agent performs hard-coding enhancement and LLM enhancement on all the independent subtasks, supplements and clarifies the computation type parameters, system information parameters and language style of the independent subtasks, and generates standardized single-point computation task instructions. All standardized single-point computation task instructions are tested using LLM evaluation and iterative judgment.

6. The method according to claim 5, characterized in that, The process involves parsing the standardized single-point computation task instructions and extracting parameters using a structured recognition agent to generate a structured task description containing an element system, control variables, and computational objectives, including: The structured recognition agent performs parameter semantic lookup, information extraction, model fine-tuning verification, and iterative judgment on the standardized single-point calculation task instructions to generate a structured task description in standardized JSON format; wherein, the structured task description includes element system field, temperature field, pressure field, component ratio field, calculation target field, and output requirement field.

7. The method according to claim 6, characterized in that, The step of generating a set of computational scripts conforming to the CALPHAD computational software by calling the corresponding thermodynamic database and script template through a code generation agent based on the structured task description includes: The intelligent agent is generated through the code, and a material thermodynamics database that matches the description of the structured task is searched based on the element system field in the structured task. Based on the calculation target field and output requirement field in the structured task, a script template that matches the structured task is found, and a set of calculation scripts containing database reading, condition setting, calculation execution and result output is generated.

8. The method according to claim 7, characterized in that, The multi-turn thinking chain generation mechanism includes: a multi-turn dialogue thinking chain and an error recognition thinking chain; Iterative optimization is achieved through the fusion of structured data and multi-round results via the aforementioned multi-round dialogue thought chain. The error identification thinking chain enables accurate diagnosis by extracting and analyzing fields of erroneous instructions, identifying the cause of errors, and inferring data contradictions.

9. A material phase diagram autonomous calculation system based on a multi-layer intelligent agent collaborative architecture, characterized in that, include The intelligent demand perception module is used to receive natural language commands input by the user and guide the user to refine the calculation requirements. The calculation requirements include material phase diagram calculation requirements or material thermophysical property calculation requirements. The natural language commands include single-turn dialogue commands and multi-turn dialogue commands. The material phase diagram calculation requirements include binary phase diagrams, ternary isothermal section diagrams, and temperature / composition-property phase diagrams. The material thermophysical property calculation requirements include basic property calculations, thermodynamic property calculations, and mechanical property calculations. The intelligent interaction enhancement module is used to identify the error types in the natural language instructions through the error correction agent, accurately modify correctable errors according to the error types, output the error cause summary for uncorrectable errors, and integrate the natural language instruction context through the historical dialogue merging agent to generate semantically consistent fused instructions. The intelligent task splitting module is used to split the complex requirements of the fusion instruction into independent subtasks by performing computation type splitting, enumeration condition splitting, and Cartesian product combination through the requirement rewriting agent. The requirement enhancement agent supplements and improves the computation type parameters and system information parameters of all the independent subtasks through hard coding enhancement and LLM enhancement, and generates standardized single-point computation task instructions. The intelligent script generation module is used to parse the standardized single-point calculation task instructions and extract parameters through a structured recognition intelligent agent, generate a structured task description containing element system, control variables and calculation target, and generate a set of calculation scripts that conform to CALPHAD calculation software by calling the corresponding thermodynamic database and script template through the code generation intelligent agent according to the structured task description. The high-throughput thermodynamic calculation module is used to call the CALPHAD calculation software and execute the calculation script set to obtain calculation results; wherein, the calculation results include: phase diagram image data or thermodynamic property numerical data; The intelligent result visualization module is used to analyze the display format of the calculation results according to user needs through the result processing intelligent agent, perform data summarization and visualization processing on the calculation results, and display phase diagrams, physical property curves and data tables to users; The multi-round iteration and thought chain generation module is used to obtain new natural language modification instructions from the user, return to the intelligent interaction enhancement module to start the multi-round thought chain generation mechanism, iteratively execute the requirement processing and calculation process until the user's needs are met. The intelligent computing resource management module is used to deploy the adjusted natural language model and knowledge base, providing computing power support for reasoning tasks.

10. The system according to claim 9, characterized in that, The intelligent demand perception module adopts a web interface interaction method, supports the latest versions of Chrome, Edge, Firefox and Safari browsers, and ensures that the interface display and function operation are consistent across different browsers.

Images