CN-122024965-A - Cementing material composition-performance collaborative design method based on large language model

CN122024965ACN 122024965 ACN122024965 ACN 122024965ACN-122024965-A

Abstract

The invention provides a large language model-based cementing material composition-performance collaborative design method and system, and aims to solve the problems of dependence on experience, high test cost, low design efficiency and insufficient performance interpretability in the traditional cementing material formula design process. The method utilizes the capability of the large language model in the aspects of scientific literature knowledge synthesis and experimental data reasoning, fuses multisource material data and domain knowledge rules, and realizes automatic association modeling and optimization design among cementing material raw material composition, proportioning parameters and macroscopic performance indexes. Meanwhile, the invention guides the large language model to extract key physical and chemical constraint, structure-performance relation and experience rule from literature and experimental data by constructing the special prompt word and rule expression template in the field of the cementing material, and converts the key physical and chemical constraint, structure-performance relation and experience rule into a characteristic vector which can be calculated, and on the basis, the prediction and reverse design of performances such as the strength, the durability, the environmental adaptability and the like of the cementing material are completed by combining an interpretable machine learning model. The method can still obtain stable and reliable design results under the condition of few samples, and provides an intelligent design path which is efficient and interpretable for the research and development of novel low-carbon gel materials and high-performance composite gel systems.

Inventors

ZHANG JUNFEI
LI JIANKUN
WANG LING

Assignees

河北工业大学
天津智控环宇科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (7)

1. A method for designing a cementing material composition-performance co-operation based on a large language model is characterized by comprising the steps of a, constructing multisource cementing material data, collecting raw material composition parameters, chemical component parameters and physical parameters of the cementing material, collecting corresponding mechanical properties, durability and environmental performance indexes, constructing a cementing material design data set, wherein the composition and materialization parameters at least comprise one or more of cement consumption, mineral admixture proportion, water-gel ratio, specific surface area or chemical oxide content, the performance indexes at least comprise one or more of compressive strength, durability indexes or environmental performance indexes, b, constructing domain knowledge cue words, constructing a domain knowledge cue word set according to a chemical reaction mechanism, hydration product formation rule and engineering application experience of the cementing material domain, wherein the domain knowledge cue word set is used for limiting the physical and chemical constraint conditions of the cementing material domain when the large language model generates material knowledge and limiting the relation expansion between the inference direction and surrounding the material composition, structure and macroscopic performance, c, generating the large language model knowledge and rule, generating the large language model knowledge cue word set and the cementing material design data, generating candidate rule sets by the relation between the large language model knowledge cue word set and the large language model design data, and the candidate rule is verified by the relation between the performance index and the physical and chemical property index, and the candidate rule is verified by the relation between the candidate rule and the performance index is verified, the method comprises the steps of (a) obtaining a set of effective rules by eliminating invalid rules which violate the physical and chemical constraint conditions of the cementing material or contradict the statistical relation of a cementing material design data set, step e) mapping the set of effective rules into a set of calculable rule functions by rule vectorization and feature mapping, wherein the set of rule functions at least comprises an indication function type rule function and a continuous response type rule function, and applying the rule functions to a cementing material sample to calculate a rule response value, thereby constructing a rule feature vector of the cementing material, wherein the rule response value is used for quantitatively representing the satisfaction degree or response strength of the cementing material sample to the corresponding rule, step f) establishing a cementing material performance prediction model by adopting an interpretable machine learning model based on the rule feature vector, wherein the interpretable machine learning model is a random forest model or a linear regression model, and obtaining the cementing material composition parameters meeting the target performance requirements by optimizing and solving on the basis of setting target performance constraint, thereby realizing the reverse optimization design of the cementing material composition parameters.
2. The method of claim 1 wherein the rules in the candidate rule set in step c include at least one of interval constraint rules for defining effective value intervals of the composition parameters or the physicochemical parameters and giving promotion or inhibition to the target performance index, performance inhibition rules for expressing inhibition of the composition parameters or the physicochemical parameters to the target performance index under specific value conditions, and interval segmentation rules for expressing different influence directions or different influence intensities of the composition parameters or the physicochemical parameters to the target performance index in different value intervals, so that influence of the composition parameters of the gel material on the performance index has interval constraint and directional expression.
3. The method of claim 1, wherein the consistency verification in step d comprises at least one of a rule validity verification based on a physical and chemical constraint condition in the field of the cementing material to eliminate a rule conflicting with a hydration reaction mechanism, a hydration product formation rule or engineering experience, and a rule inconsistent with or significantly deviating from a correlation direction of a composition parameter and a performance index in sample data based on a statistical consistency verification of the cementing material design data set.
4. The method according to claim 1, wherein the indicating function type rule function is used for outputting a binary response of whether the material sample meets the rule constraint, the continuous response type rule function is used for outputting continuous response strength of the material sample to the rule constraint, and a plurality of rule response values corresponding to the rule are spliced to form the rule feature vector.
5. The method of claim 1, wherein in step f, when the interpretable machine learning model is a random forest model, rule importance weights are further output, wherein the rule importance weights are used for characterizing the contribution degree of each rule feature to the prediction result and for performing interpretable analysis on the cement performance prediction result.
6. The method according to claim 1, wherein the inverse optimization design further takes the deviation degree of the material composition parameters from the effective rule set as a penalty term and incorporates the optimization objective function on the basis of meeting the target performance constraint, so as to search the material composition parameter space for an optimal solution and obtain a cement design scheme meeting the target performance constraint.
7. The method of claim 1, a large language model-based cement composition-performance co-design system, comprising (1) A data management module, (2) Prompt word construction module (3) Large language model reasoning module (4) Rule generation and verification module (5) Rule vectorization module (6) Performance prediction and reverse design module The system comprises a data management module, a prompt word construction module, a large language model reasoning module, a rule generation and verification module, a rule vectorization module and a performance prediction and reverse design module, wherein the data management module is used for storing and managing cementing material composition parameters, performance indexes and experimental data, executing missing value processing and normalization processing and training set and verification set division, the prompt word construction module is used for generating and managing a field knowledge prompt word set at least comprising physicochemical constraint conditions, variable definition and reasoning targets, the large language model reasoning module is used for generating a candidate language rule set based on the prompt word set and the training set, the rule generation and verification module is used for carrying out consistency verification on the candidate rule according to the physicochemical constraint conditions and the training set statistical consistency and outputting a final rule set, the rule vectorization module is used for converting the final rule set into a rule function set comprising an indication function type and a continuous response type and generating rule feature vector, and the performance prediction and reverse design module is used for training an interpretable machine learning model based on the rule feature vector and outputting rule contribution weights and reversely optimizing under the target performance constraint and composition parameter boundary constraint to output a cementing material design result.

Description

Cementing material composition-performance collaborative design method based on large language model Technical Field The invention relates to the technical field of intelligent design and material informatization of cementing materials, in particular to a cementing material composition-performance collaborative design method based on a large language model. Background The cementing material is used as the core composition of concrete, mortar and other engineering materials, and the composition structure and performance level directly determine the bearing capacity, service life and stability and safety of the engineering structure under complex environmental conditions. In practical engineering application, the key indexes of the cementing material such as compressive strength, durability, environmental adaptation performance and the like usually have obvious mutual coupling relation, so that the formula design process is essentially a complex engineering problem related to multi-factor and multi-target constraint. (One) limitations of conventional cement design methods At present, the composition design of the cementing material mainly depends on an empirical formula, a specification recommended range and a large number of physical tests for verification. Specifically, engineering technicians typically evaluate and optimize material performance by adjusting the composition parameters such as cement usage, water-cement ratio, mineral admixture ratio, etc., in combination with long-term accumulated engineering experience. Such methods have certain feasibility in engineering practice, but still suffer from the following significant drawbacks. 1. The test cost is high, the research and development period is long, and the performance evaluation of the cementing material generally needs to be subjected to a plurality of links such as stirring, forming, maintenance and testing, wherein part of key performance indexes (such as durability, environment adaptation performance and the like) need a long maintenance and testing period. The optimization process of the material formula is often dependent on a large number of repeatability tests, so that the research and development cost is high, the period is long, and the engineering requirements of rapid design and iterative optimization are difficult to meet. 2. The formula design is highly dependent on expert experience, and is difficult to migrate and multiplex, and the optimal proportioning scheme of the cementing material under different engineering conditions often depends on experience judgment of deep engineering personnel. Such experience usually exists in the form of implicit knowledge, is difficult to formalize through unified rules or models, is difficult to realize effective migration and multiplexing among different engineering scenes or different research and development teams, and limits the popularization of the design method. 3. The difficulty of multi-performance collaborative optimization is high, and a systematic design method is lacked, so that the mechanical property, the durability and the environmental performance of the cementing material are generally in a mutually restricted relation. For example, increasing early strength may adversely affect long-term durability. The traditional test method mostly takes a single performance index as an optimization target, and systematic collaborative design is difficult to carry out under the constraint condition of multiple performances, so that the overall improvement of the comprehensive performance of the material is limited. 4. The data utilization rate is low, the literature knowledge is difficult to directly utilize, a large amount of research results about the relation between the composition and the performance of the cementing material are dispersed in academic literature, technical reports and engineering specifications, the content of the research results exists in a qualitative description or experience summary mode, and the research results are difficult to directly convert into a computable and reusable design basis. This results in existing research results and engineering experience that are difficult to integrate efficiently and for the design and optimization of new materials, with overall low data utilization. (II) progress and deficiencies of data-driven methods With the continuous accumulation of material databases and the development of machine learning methods, data-driven cement performance prediction methods are receiving attention. The related research attempts to establish the mapping relation between the cementing material composition parameters and the performance indexes by using algorithms such as regression models, neural networks and the like, so that the performance prediction efficiency is improved to a certain extent, and the dependence on a large number of physical tests is reduced. However, there are still significant limitations to existing data driven approaches. On