Big language model-based cementitious material composition-performance collaborative design method

Abstract

The invention provides a cementing material composition-performance collaborative design method and system based on a large language model, and aims to solve the problems that a traditional cementing material formula design process depends on experience, the test cost is high, the design efficiency is low and the performance interpretability is insufficient. According to the method, the ability of a large language model in the aspects of scientific literature knowledge synthesis and experimental data reasoning is utilized, multi-source material data and domain knowledge rules are fused, and automatic correlation modeling and optimization design among raw material composition, matching parameters and macroscopic performance indexes of the cementing material are achieved. Meanwhile, a special cue word and rule expression template in the field of cementing materials is constructed, a large language model is guided to extract key physical and chemical constraints, structure-performance relationships and empirical rules from literature and experimental data, and the key physical and chemical constraints, the structure-performance relationships and the empirical rules are converted into computable feature vectors; on the basis, prediction and reverse design of the strength, durability, environmental adaptability and other performance of the cementing material are completed in combination with an interpretable machine learning model. According to the method, a stable and reliable design result can still be obtained under the condition of few samples, and an efficient and explainable intelligent design path is provided for research and development of novel low-carbon cementing materials and high-performance composite cementing systems.

Description

Technical Field

[0001] This invention relates to the field of intelligent design and materials information technology for cementitious materials, specifically a composition-performance co-design method for cementitious materials based on a large language model. Background Technology

[0002] As a core component of concrete, mortar, and other engineering materials, cementitious materials directly determine the load-bearing capacity, service life, and stability and safety of engineering structures under complex environmental conditions through their composition, structure, and performance. In practical engineering applications, key indicators such as compressive strength, durability, and environmental adaptability of cementitious materials are often significantly coupled with each other. Therefore, the formulation design process is essentially a complex engineering problem involving multiple factors and multi-objective constraints.

[0003] (a) Limitations of traditional cementitious material design methods

[0004] Currently, the composition design of cementitious materials mainly relies on empirical formulas, recommended ranges in specifications, and extensive physical testing. Specifically, engineers typically evaluate and optimize material performance by adjusting compositional parameters such as cement content, water-cement ratio, and mineral admixture proportions, combined with long-term accumulated engineering experience. While this method is feasible to some extent in engineering practice, it still has the following significant shortcomings.

[0005] 1. High testing costs and long R&D cycles: The performance evaluation of cementitious materials typically involves multiple stages, including mixing, molding, curing, and testing. Some key performance indicators (such as durability and environmental adaptability) require lengthy curing and testing periods. This makes the material formulation optimization process often reliant on numerous repetitive experiments, resulting in high R&D costs and long cycles, making it difficult to meet the engineering needs of rapid design and iterative optimization.

[0006] 2. Formulation design heavily relies on expert experience, making it difficult to transfer and reuse: The optimal proportioning scheme for cementitious materials under different engineering conditions often depends on the experience and judgment of senior engineers. This kind of experience usually exists in the form of tacit knowledge, which is difficult to formalize through unified rules or models, and is also difficult to effectively transfer and reuse between different engineering scenarios or different R&D teams, thus limiting the generalizability of the design method.

[0007] 3. Multi-performance synergistic optimization is challenging and lacks systematic design methods: The mechanical properties, durability, and environmental performance of cementitious materials often exhibit interdependent relationships. For example, improving early strength may adversely affect long-term durability. Traditional testing methods often focus on a single performance index as the optimization target, making it difficult to conduct systematic synergistic design under multi-performance constraints, thus limiting the overall improvement of the material's comprehensive performance.

[0008] 4. Low data utilization rate and difficulty in directly utilizing literature knowledge: A large number of research results on the relationship between the composition and performance of cementitious materials are scattered in academic literature, technical reports, and engineering specifications. Their content is mostly in the form of qualitative descriptions or experience summaries, making it difficult to directly transform them into calculable and reusable design basis. This results in existing research results and engineering experience being difficult to integrate efficiently and apply to the design and optimization of new materials, leading to an overall low data utilization rate.

[0009] (II) Advances and shortcomings of data-driven approaches

[0010] With the continuous accumulation of materials databases and the development of machine learning methods, data-driven methods for predicting the performance of cementitious materials have gradually attracted attention. Related studies have attempted to establish mapping relationships between the compositional parameters of cementitious materials and performance indicators using algorithms such as regression models and neural networks, which has improved the efficiency of performance prediction to some extent and reduced the reliance on a large number of physical experiments.

[0011] However, existing data-driven approaches still have significant limitations.

[0012] On the one hand, existing methods are mostly "black box models," and their prediction process lacks interpretability. Although the models can output performance prediction results, they are difficult to specify which material composition parameters or mechanistic factors play a key role in performance, making it difficult for engineers to make reliable formulation adjustments based on the model results.

[0013] On the other hand, existing methods are typically highly dependent on the number of samples. Given the high cost of acquiring experimental data for cementitious materials, it is often difficult to obtain large-scale, high-quality samples in actual research and development. Furthermore, black-box models are prone to overfitting or unstable predictions under small sample conditions, affecting their reliability in engineering applications.

[0014] Furthermore, existing data-driven methods struggle to effectively integrate mechanistic knowledge from the materials science field. A large number of known physicochemical constraints and empirical rules are not embedded in the model-building process, potentially leading to model predictions that contradict fundamental material mechanisms and thus limiting their application value in practical engineering design.

[0015] (III) Technological Development Needs

[0016] In summary, existing cementitious material design methods have significant shortcomings in terms of R&D costs, knowledge utilization efficiency, model interpretability, and multi-performance collaborative design. There is an urgent need for an intelligent design method that can work stably under small sample conditions and simultaneously integrate material knowledge and experimental data.

[0017] Therefore, how to systematically integrate the knowledge of cementitious materials scattered in literature and experimental data, and on this basis, construct a cementitious material design method that is interpretable, meets physicochemical constraints, and can be used for reverse design, has become a key technical problem that urgently needs to be solved in the field of cementitious materials engineering. Summary of the Invention

[0018] (a) Purpose of the invention

[0019] As a core component of concrete and related engineering materials, cementitious materials' raw material composition, chemical reaction processes, and microstructural evolution directly determine their mechanical properties, durability, and environmental adaptability. Current cementitious material formulation design mainly relies on empirical formulas or extensive physical experiments, resulting in problems such as long design cycles, high testing costs, difficulty in explaining performance mechanisms, and challenges in synergistic optimization of multiple properties.

[0020] With the development of materials databases and data-driven methods, machine learning-based methods for predicting the performance of cementitious materials have gradually emerged. However, most existing methods rely on a large number of experimental samples, and the internal mechanisms of the models are not transparent. They also have poor stability under small sample conditions, making it difficult to provide interpretable and constrained design basis for materials designers.

[0021] The purpose of this invention is to propose an intelligent design method for cementitious materials that integrates the knowledge synthesis capabilities of large language models with data reasoning capabilities. By introducing domain knowledge constraints and rule-based reasoning mechanisms, it achieves interpretable modeling and collaborative optimization between the composition parameters and macroscopic properties of cementitious materials. Specifically, it includes:

[0022] 1. Achieve interpretable modeling of the composition-property relationship of cementitious materials.

[0023] By transforming the material laws implicit in literature knowledge and experimental data into a set of explicit rules. This allows performance prediction results to be traced back to a clearly defined material composition range and physicochemical constraints.

[0024] 2. Achieve stable performance prediction under limited sample conditions

[0025] By leveraging the comprehensive knowledge of the materials domain through large language models, we construct knowledge-enhanced feature representations under limited experimental sample conditions to improve the generalization ability and robustness of performance prediction models.

[0026] 3. Achieving reverse material design oriented towards target performance

[0027] Guided by rule constraints and interpretable models, the composition parameter space of cementitious materials is optimized in reverse to obtain material design schemes that meet the target performance requirements.

[0028] By achieving the above objectives, this invention provides an efficient, interpretable, and transferable intelligent design method for the development of novel high-performance cementitious materials and low-carbon cementitious systems.

[0029] (II) Technical Solution

[0030] To achieve interpretable modeling of the composition-performance relationship of cementitious materials, stable performance prediction under few-sample conditions, and reverse material design oriented towards target performance, this invention proposes an intelligent design technology scheme for cementitious materials that integrates the knowledge synthesis capabilities of large language models and data reasoning capabilities. Based on multi-source cementitious material data and domain knowledge, this scheme forms a complete collaborative design methodology system for cementitious materials through steps such as rule generation, feature construction, performance prediction, and reverse optimization.

[0031] 1. Knowledge Synthesis and Rule Generation of Cementitious Materials Based on Large Language Model

[0032] (1) First, a unified knowledge synthesis and reasoning analysis was conducted on the relevant literature data and experimental data of cementitious materials using a large language model. The literature data included textual information such as the composition of cementitious materials, hydration reaction mechanism, structural evolution law and engineering experience summary; the experimental data included raw material composition parameters, physicochemical property parameters and corresponding mechanical properties, durability performance and environmental performance indicators.

[0033] (2) Collect the raw material composition parameters, chemical composition parameters, and physical parameters of the cementitious materials, and collect the corresponding mechanical properties, durability properties, and environmental performance indicators to construct a cementitious material design dataset: .

[0034] in: : indicates the first The composition and physicochemical parameter vector of a cementitious material sample, wherein the parameters include at least one or more of the following: cement content, mineral admixture ratio, water-cement ratio, specific surface area, or chemical oxide content in the cementitious material;

[0035] : Represents a performance index vector corresponding to the cementitious material sample, wherein the performance index includes at least one or more of compressive strength, durability index, or environmental performance index.

[0036] (3) Based on the chemical reaction mechanism, hydration product formation rules, and engineering application experience in the field of cementitious materials, a set of domain knowledge prompts is constructed to guide the large language model in knowledge integration and data reasoning: .

[0037] Each prompt word This is used to constrain large language models to meet the physicochemical constraints in the field of cementitious materials when generating material knowledge, and to limit their reasoning direction to revolve around the relationship between material composition, structural evolution and macroscopic properties.

[0038] (4) Methods for generating rules: First, they are used to guide the large language model to extract hydration reaction mechanisms, material composition constraints, and empirical rules from relevant literature on cementitious materials, and to transform the qualitative relationships described in the literature into structured knowledge representations: .

[0039] Among them, the This indicates the positive correlation, negative correlation, or threshold relationship between component parameters and performance indicators.

[0040] Next, it is used to guide the large language model to combine experimental data of cementitious materials, identify potential composition-performance correlation patterns, and generate a set of candidate rules through statistical or inference methods: .

[0041] in, This represents the rule-based reasoning process based on experimental data.

[0042] Used to integrate literature knowledge sets With experimental reasoning knowledge set Generate the final rule set: ;

[0043] Finally, the generated rules are verified for consistency: This eliminates invalid rules that violate the physicochemical constraints of cementitious materials.

[0044] (5) The set of domain knowledge prompts With cementitious material design dataset A common input large language model is used to generate an explicit set of rules describing the composition-structure-performance relationship of cementitious materials through language reasoning: .

[0045] Each rule This refers to the empirical or mechanistic laws governing the influence of a specific component or physicochemical parameter of a cementitious material on its target performance within a specific value range. Its general form is as follows: .

[0046] Furthermore, the rules also include performance suppression rules: ;

[0047] Or interval segmentation rules: By using the above-mentioned rules, the influence of cementitious material composition parameters on performance indicators has clear range constraints and directional expression, thereby enhancing the interpretability and controllability of the material design process.

[0048] 2. Transformation of Language Rules into Computable Features and Knowledge-Driven Feature Engineering

[0049] (1) The generated set of language rules Mapped to a set of computable rule functions: .

[0050] Its basic forms include: indicator function type: Continuous response type: .

[0051] (2) And apply the rule function to the cementitious material sample. The corresponding rule response values ​​are calculated, thereby constructing the rule feature vector of the cementitious material: .

[0052] Wherein, the rule function Used for quantitative characterization of cementitious material samples Corresponding rules The degree of satisfaction or response intensity.

[0053] This feature vector not only preserves information from the experimental data, but also explicitly embeds domain knowledge constraints, thereby improving the stability and generalization ability of the model under small sample conditions, and enhancing the physical and engineering interpretability of the features.

[0054] 3. Performance prediction based on interpretable machine learning models

[0055] Based on rule feature vectors The performance prediction function is constructed using a random forest model or a linear regression model: .

[0056] The prediction form of the random forest model is as follows: And further output the rule importance weights: ,in, Representation rules In the The information gain contribution in each decision tree. Interpretable analysis of the predicted cementitious material properties is achieved through the weighting of the rule importance.

[0057] 4. Reverse Design of Cementitious Materials Based on Rule Importance

[0058] After completing the performance prediction model construction, this invention further utilizes the rule importance weights output by the model to carry out reverse design of cementitious materials.

[0059] By setting target performance constraints: And combined with the importance weight of the rules Construct the inverse optimization objective function: .

[0060] in, Indicates the relationship between material composition parameters and rules The degree of deviation. The optimal solution is searched in the material composition parameter space. To obtain a cementitious material design scheme that meets the target performance constraints.

[0061] By combining rule importance information, a targeted optimization search is performed on the composition parameter space of the cementitious material to obtain a material formulation that meets the target performance requirements. This reverse design process, guided by rule constraints, avoids blind searching, effectively narrows the design space, improves design efficiency, and ensures that the obtained material solution conforms to existing physicochemical laws and engineering experience.

[0062] 5. Overall synergistic effect of the technical solution

[0063] Through the above technical solution, this invention constructs a knowledge synthesis and intelligent design framework for cementitious materials with a large language model as its core, realizing a complete technical closed loop from literature knowledge and experimental data to interpretable rules, from rule features to performance prediction, and then to reverse material design. This solution effectively solves the problems of high experimental costs, difficulty in knowledge reuse, uninterpretable models, and difficulties in multi-performance synergistic optimization in traditional cementitious material design methods, and has good prospects for engineering applications.

[0064] III. Beneficial Effects

[0065] Compared with existing cementitious material design methods, this invention introduces the knowledge integration and data reasoning capabilities of large language models, combined with interpretable machine learning models, to achieve the collaborative design of cementitious material composition and performance. It has significant benefits in terms of experimental cost, model interpretability, small sample adaptability, and the ability to develop new cementitious materials.

[0066] 1. Reduce testing costs and shorten the R&D cycle. This invention integrates literature knowledge with existing experimental data to predict the performance of cementitious materials and optimize their formulations. It can effectively evaluate performance during the material design stage, reducing reliance on a large number of repetitive physical tests and long-term curing tests, thereby significantly reducing testing costs and shortening the R&D cycle.

[0067] 2. Enhancing the interpretability of the material design process: This invention generates explicit material composition-performance rules through a large language model and introduces them into the performance prediction model, enabling the prediction results to be traced back to specific material composition parameters and rule constraints. This avoids the problem of traditional black box models being difficult to interpret and improves the engineering credibility of the design results.

[0068] 3. Suitable for stable modeling under conditions of few samples

[0069] By introducing knowledge and mechanistic constraints from the field of cementitious materials, this invention effectively reduces the model's dependence on large-scale data samples. Even with limited experimental data, it can still obtain stable and reliable performance prediction results, which are significantly better than pure data-driven methods.

[0070] 4. Support the efficient design of novel cementitious materials

[0071] This invention does not rely on a specific material system and is applicable to the design needs of novel materials such as low-carbon cementitious materials and multi-component composite cementitious systems. Through rule-guided reverse design, it helps to discover innovative material formulations that meet target performance requirements. Attached Figure Description

[0072] Figure 1 This is a flowchart of the collaborative design method of the present invention.

[0073] Figure 2 This is a diagram of the collaborative design system architecture of the present invention. Detailed Implementation

[0074] Example 1: Design of cement-mineral admixture composite cementitious material

[0075] This implementation example takes a typical cement-mineral admixture composite cementitious system as the research object to verify the feasibility and effectiveness of the proposed method for collaborative design of cementitious material composition and performance based on large language model knowledge synthesis and data reasoning in practical engineering material design.

[0076] (1) Data acquisition and construction of multi-source cementitious materials

[0077] A composite cementitious system composed of cement, mineral admixtures, and water was selected as the research object, and its composition parameters and mechanical property data were collected. The composition parameters include, but are not limited to, water-cement ratio, cement dosage, type and proportion of mineral admixtures, etc.; the 28-day compressive strength was selected as the target mechanical property.

[0078] Each group of cementitious material samples is represented as follows: ,in, This refers to the water-to-glue ratio. This refers to the amount of cement used. For the first The proportion of each mineral admixture. The corresponding target output is: That is, the 28-day compressive strength. Based on this, a materials design dataset is constructed:

[0079] (2) Domain knowledge prompt word construction and large language model rule generation

[0080] Based on existing research findings and engineering experience in the field of cementitious materials, domain-specific cue words for cement-mineral admixture systems are constructed to guide the large language model to focus on the proportion of mineral admixtures, the degree of hydration reaction, and their impact mechanism on strength development. By inputting these cue words and the dataset into the large language model, the model automatically generates display rules describing the relationship between material composition and strength performance. For example, material rules of the following form are generated: When the proportion of mineral admixtures At that time, the degree of hydration increases, and the strength contribution increases. This type of rule comprehensively reflects the filling effect of mineral admixtures, the pozzolanic reaction effect, and their influence on the structural compactness of hydration products.

[0081] (3) Rule vectorization and feature construction

[0082] For the generated rule set, the rules in linguistic form are transformed into computable rule functions: The rule function is used to describe the sample The degree of response under the corresponding rule constraints. By mapping multiple rules using functions, a rule-based material feature vector can be constructed: This feature represents the fusion of material composition information with domain knowledge constraints, providing highly interpretable input features for subsequent performance prediction.

[0083] (4) Intensity prediction based on interpretable models

[0084] The constructed rule feature vector As input, a random forest model was used to predict the 28-day compressive strength of cementitious materials: .

[0085] in, This represents a random forest prediction model. Through model training, the contribution weights of different rule features to the intensity prediction results can be obtained, thereby identifying the relative importance of mineral admixture ratio and hydration-related rules in the intensity formation process.

[0086] (5) Allocation optimization and reverse design based on rule importance

[0087] After completing the strength prediction model construction, the target compressive strength constraint conditions are set: By combining the rule importance information output by the random forest model, the key component parameters such as the proportion of mineral admixtures and the water-cement ratio are targeted for adjustment and optimization search, thereby obtaining a cementitious material mix design that meets the target strength requirements. This reverse design process is carried out under rule constraints, avoiding blind trial and error, improving design efficiency, and ensuring that the optimization results conform to the hydration mechanism of cementitious materials and engineering experience.

[0088] (6) Explanation of implementation results

[0089] The method described in this embodiment enables stable prediction and proportion optimization of the 28-day compressive strength of cement-mineral admixture composite cementitious system with a small number of experimental samples, verifying the feasibility, effectiveness, and engineering application value of this invention in practical cementitious material engineering design.

[0090] Example 2: Synergistic design of "strength-chloride salt resistance" based on alkali-activated fly ash / slag cementing system

[0091] This embodiment focuses on an alkali-activated fly ash-slag (AAM) cementing system. Under the requirements of marine chloride salt environments, it simultaneously optimizes the 28-day compressive strength and chloride ion migration / permeability, verifying the applicability of the invention in the "multi-objective durability collaborative design" scenario.

[0092] (1) Dataset construction: The composition and physicochemical parameters of the AAM system samples were collected, including the fly ash mass fraction. Slag mass fraction The following parameters were collected: alkali equivalent AE (as Na2O%), water glass modulus Ms, liquid-to-solid ratio L / S, curing temperature T, and curing regime; corresponding performance indicators were also collected: 28-day compressive strength. Rapid chloride ion migration coefficient Or electric flux Q, resistivity ρ, etc. Construct sample vectors. Output .

[0093] (2) Cue word and rule generation: Based on the main line of "reaction degree - gel structure - pore connectivity - chlorine transport", a domain cue word set is constructed to guide the large language model to extract and generate rules from literature and data. For example: ① When AE is in [8%, 12%] and Ms is in [1.0, 1.6], a dense N-(C)-ASH gel is formed. Significantly reduced; ② When L / S is too high (>0.45), the porosity increases, leading to ③ Increasing the proportion of slag can improve early strength but may increase the risk of self-shrinkage (as a constraint rule).

[0094] (3) Rule vectorization: Mapping the above language rules into computable rule functions. For example, using indicator functions With continuous penalty function: ; And form regular feature vectors. .

[0095] (4) Multi-objective performance prediction: Train interpretable models separately for each objective. and Make a prediction and obtain , and output the importance of the rules. This is used to explain "which rules drive the increase in strength / chlorine resistance".

[0096] (5) Reverse design: setting target constraints and And set the feasible region of the project (e.g., AE∈[6%,14%], Ms∈[0.8,2.0], L / S∈[0.30,0.50]).

[0097] Define the optimization goal:

[0098] ,

[0099] Get the recommended formula And give the corresponding key rule satisfaction level.

[0100] (6) Effect description: Compared with the experimental or black box regression method alone, this embodiment can clearly explain the source of chlorine resistance improvement (such as Ms window and L / S upper limit rule) through the "rule-feature-importance" link, and stably give the formulation scheme that meets the dual constraints of strength and durability under small sample conditions.

[0101] Example 3: Strength-Carbon Emission Synergistic Design for Low-Carbon LC3 (Calcinated Clay-Limestone) Cementitious Systems

[0102] This embodiment focuses on the low clinker cement system (LC3) to minimize carbon emissions per unit of cementitious material while ensuring strength and workability, thus achieving a reverse design with dual objectives of "performance and environment".

[0103] (1) Data acquisition: Collect the composition parameters of the LC3 system, including the clinker ratio. Calcination ratio of clay Limestone ratio gypsum ratio Specific surface area Additive dosage Water-to-binder ratio (w / b), etc.; performance indicators include 28-day compressive strength. Spreadability / flowability F, setting time Wait. Simultaneously establish the carbon factor of each raw material. (kg CO2-eq / kg) database.

[0104] (2) Construction of carbon emission index: Calculate the carbon emission of a unit cementitious material for any sample. And a transportation or energy consumption correction term ΔE can be added. E is taken as one of the outputs to be optimized, forming... .

[0105] (3) Rule generation: Based on the mechanism of "alumina-silicon reaction - aluminocarbonate / aluminosilicate generation - pore structure optimization", the following are examples of generated rules: ① When In [25%, 35%] and When the concentration is [10%, 20%], a stable aluminate carbide phase is formed, and the strength increase is accelerated; ② w / b > 0.45 will significantly reduce the strength and increase the indirect cost of E (more cementing material is needed to meet the strength requirement); ③ when The response is insufficient when the temperature is too low, and it needs to be controlled. .

[0106] (4) Interpretive prediction and importance: Input the rule vector z into the random forest / linear model to obtain the following results respectively. , , (Or modified by the model), and output the rule importance to explain the key interval of "low carbon without reducing power".

[0107] (5) Reverse optimization: Set the goal: , and minimize Weighted sums or ε-constraints can be used: Output a set of Pareto candidate formulations and provide the key rule satisfaction and carbon emission decomposition (clinker contribution, calcined clay contribution, etc.).

[0108] (6) Effect Description: This embodiment demonstrates that the present invention can incorporate environmental indicators into the material design closed loop in an interpretable manner, automatically search for low-carbon formulations under given strength and workability constraints, and indicate the upper limit of clinker ratio, The dominant role of factors such as the synergistic window in intensity / carbon emissions.

[0109] Example 4: Synergistic Design of Multiple Durability Indicators for "Sulfate Resistance, Chloride Resistance, and Freeze Resistance" in the Face of Composite Degradation in Marine Engineering

[0110] This embodiment addresses the common degradation caused by chloride erosion, sulfate erosion, and freeze-thaw cycles under marine engineering service conditions. It constructs a joint prediction and reverse design process for multiple durability indicators to verify the scalability of the invention under complex environmental constraints.

[0111] (1) Data set and indicators: Collect cementitious material composition parameters (w / b, cement dosage, type and proportion of mineral admixtures, air-entraining agent / water-reducing agent dosage, specific surface area, etc.) and durability performance indicators: sulfate erosion expansion rate. Relative dynamic elastic modulus of freeze-thaw cycles (RDM(N)) and chloride ion diffusion coefficient Or the electric flux Q.

[0112] Define the overall durability rating , where λ1~λ3 are the engineering weights.

[0113] (2) Composite Mechanism Hints: Constructing hints to constrain LLM simultaneously considers: ① Sulfate attack is related to the C3A / AFm phase and pore solution alkalinity; ② Freeze-thaw cycles are related to gas content, pore structure, and saturation; ③ Chlorine transport is related to pore connectivity and binding capacity. This guides the model to generate consistent rules across indices, such as "Increasing mineral admixtures can reduce..." However, if the air intake is insufficient, the RDM will decrease.

[0114] (3) Rule consistency verification: Verify the consistency of the candidate rule set with physical and chemical constraints, such as: gas content To avoid conflicts between freeze-thaw rules and workability / strength rules; constrain w / b ≤ 0.45 to ensure the upper limit of pore structure. Eliminate self-contradictory or mechanistic rules.

[0115] (4) Joint prediction: Using the regular feature vector z as input, a multi-output interpretable model is trained to obtain... , , The overall score DÎ is calculated. The output rule importance is used to explain the dominant factors for each durability metric.

[0116] (5) Reverse design and candidate selection: The main objective is to maximize DÎ, and baseline constraints on strength and workability are set (e.g. , Several candidate formulations were obtained through constrained optimization search. Sensitivity analysis was then performed on key parameters based on their importance, and the "2-3 composition parameters most sensitive to durability improvement" and suggested adjustment directions were given.

[0117] (6) Effect description: This embodiment proves that the present invention can achieve interpretable multi-durability synergistic optimization under the coupling of multiple degradation mechanisms. The output formula not only meets the comprehensive durability target, but also provides an engineering explanation at the rule level (e.g., air intake window, w / b upper limit, and admixture synergistic range).

[0118] This invention can be implemented using hardware comprising several different elements and a suitably programmed computer. Various component embodiments of the invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. The physical implementation of the hardware structure includes, but is not limited to, physical devices, including, but not limited to, transistors, computers, microprocessors, or digital signal processors. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of this invention can be implemented using various programming languages; the description of specific languages ​​herein is for the purpose of disclosing the best mode of implementation of the invention.

[0119] The above specific embodiments have provided a detailed description of the purpose, technical means, and beneficial effects of the present invention. It should be understood that the purpose of the detailed description is to enable those skilled in the art to better understand the present invention, and it is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A composition-performance co-design method for cementitious materials based on a large language model, characterized in that, The method includes: Step a, constructing multi-source cementitious material data, collecting raw material composition parameters, chemical composition parameters, and physical parameters of cementitious materials, and collecting corresponding mechanical properties, durability properties, and environmental performance indicators to construct a cementitious material design dataset. The composition and physicochemical parameters include at least one or more of cement content, mineral admixture ratio, water-cement ratio, specific surface area, or chemical oxide content. The performance indicators include at least one or more of compressive strength, durability indicators, or environmental performance indicators. Step b, constructing domain knowledge prompts, based on the chemical reaction mechanism, hydration product formation law, and engineering application experience in the cementitious material field, constructing a set of domain knowledge prompts. This set of prompts is used to constrain the large language model to meet the physicochemical constraints of the cementitious material field when generating material knowledge, and to limit the reasoning direction to revolve around the relationship between material composition, structural evolution, and macroscopic properties. Step c, large language model knowledge synthesis and rule generation, inputting the domain knowledge prompts set and the cementitious material design dataset into the large language model to generate a set of candidate rules describing the relationship between cementitious material composition parameters or physicochemical parameters and target performance indicators. Each rule in the candidate rule set... The set of rules includes at least the compositional or physicochemical parameters, the corresponding valid value ranges, and the direction of influence on the target performance indicators; Step d, rule consistency verification and filtering: perform consistency verification on the candidate rule set to eliminate invalid rules that violate the physicochemical constraints of the cementitious material or contradict the statistical relationship of the cementitious material design dataset, and obtain a valid rule set; Step e, rule vectorization and feature mapping: map the valid rule set into a computable rule function set, wherein the rule function set includes at least indicator function type rule functions and continuous response type rule functions, and apply the rule functions to the cementitious material samples to calculate the rule response values, thereby constructing the rule feature vector of the cementitious material, wherein the rule response values ​​are used to quantitatively characterize the degree of satisfaction or response intensity of the cementitious material samples with the corresponding rules; Step f, performance prediction and reverse design: based on the rule feature vector, establish a cementitious material performance prediction model using an interpretable machine learning model, wherein the interpretable machine learning model is a random forest model or a linear regression model, and obtain the cementitious material compositional parameters that meet the target performance requirements by optimizing the solution based on the set target performance constraints, thereby realizing the reverse optimization design of the cementitious material compositional parameters.

2. The method according to claim 1, characterized in that, The rules in the candidate rule set mentioned in step c include at least one of the following: interval constraint rules, which are used to limit the effective value range of the component parameters or material parameters and give the promoting or inhibiting effect on the target performance index; performance inhibition rules, which are used to express the inhibiting effect of the component parameters or material parameters on the target performance index under specific value conditions. Interval segmentation rules are used to express that the compositional parameters or physicochemical parameters have different directions or intensities of influence on the target performance indicators in different value ranges, so that the influence of the cementitious material compositional parameters on the performance indicators has interval constraints and directional expression.

3. The method according to claim 1, characterized in that, The consistency verification in step d includes at least one of the following: rule validity verification based on physicochemical constraints in the field of cementitious materials, to eliminate rules that conflict with hydration reaction mechanisms, hydration product formation rules, or engineering experience; statistical consistency verification based on the cementitious material design dataset, to eliminate rules that are inconsistent with or significantly deviate from the correlation direction between composition parameters and performance indicators in the sample data.

4. The method according to claim 1, characterized in that, The indicator function type rule function in step e is used to output the binary response of whether the material sample satisfies the rule constraints. The continuous response type rule function in step e is used to output the continuous response intensity of the material sample to the rule constraints, and the rule response values ​​corresponding to multiple rules are concatenated to form the rule feature vector.

5. The method according to claim 1, characterized in that, In step f, when the interpretable machine learning model is a random forest model, the rule importance weights are further output. The rule importance weights are used to characterize the contribution of each rule feature to the prediction result and to perform interpretable analysis on the prediction results of cementitious material performance.

6. The method according to claim 1, characterized in that, The reverse optimization design, based on satisfying the target performance constraints, further incorporates the deviation of material composition parameters from the effective rule set as a penalty term into the optimization objective function, so as to search for the optimal solution in the material composition parameter space and obtain a cementitious material design scheme that satisfies the target performance constraints.

7. The method according to claim 1, a composition-performance co-design system for cementitious materials based on a large language model, characterized in that, include (1) 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 data management module stores and manages the composition parameters, performance indicators, and experimental data of cementitious materials, and performs missing value processing, normalization, and training and validation set partitioning. The prompt word construction module generates and manages a set of domain knowledge prompt words that includes at least physicochemical constraints, variable definitions, and inference objectives. The large language model inference module generates a set of candidate language rules based on the prompt word set and the training set. The rule generation and verification module performs consistency verification on the candidate rules based on the statistical consistency between the physicochemical constraints and the training set, and outputs the final rule set. The rule vectorization module converts the final rule set into a set of rule functions including indicator functions and continuous response functions, and generates rule feature vectors. The performance prediction and reverse design module trains an interpretable machine learning model based on the rule feature vectors, outputs the rule contribution weights, and performs reverse optimization under the target performance constraints and composition parameter boundary constraints to output the cementitious material design results.

Images