An air-entrained concrete mix proportion intelligent optimization method and system for low-pressure and low-humidity environments

Abstract

The application discloses an air-entraining concrete mix proportion intelligent optimization method and system for a low-pressure and low-humidity environment, and belongs to the technical field of concrete material design and artificial intelligence optimization. The method comprises the following steps: constructing an original database, data enhancement, physical consistency screening, constructing a multi-task prediction model, training the multi-task prediction model, constructing a multi-objective optimization model, multi-objective optimization solving, and outputting a recommended mix proportion. The application can improve the efficiency, accuracy and multi-performance coordination capability of air-entraining concrete mix proportion design in a low-pressure and low-humidity environment.

Description

Technical Field

[0001] This invention belongs to the technical fields of concrete material design, engineering materials for low-pressure environments, artificial intelligence optimization, and building material durability prediction. Specifically, it relates to an intelligent optimization method and system for the mix proportion of air-entrained concrete in low-pressure and low-humidity environments. More particularly, it involves using a conditional denoising diffusion model for small-sample data augmentation, employing a composition-aware attention multi-task neural network to establish a mapping relationship between the composition of cementitious materials and multiple performance indicators of concrete, and combining a multi-objective optimization algorithm to achieve reverse design of the ternary cementitious material composition of cement, fly ash, and granulated blast furnace slag powder. This method is applicable to the synergistic optimization of the workability, bubble structure, mechanical properties, and durability of air-entrained concrete in special service environments such as high-altitude, cold regions, low-pressure, and low-humidity conditions. Background Technology

[0002] With the continuous expansion of the construction scale of railways, highways, bridges, tunnels, water conservancy projects, and infrastructure in cold regions, the service performance of concrete materials in special environments such as low pressure, low humidity, strong evaporation, large diurnal temperature range, and frequent freeze-thaw cycles has attracted widespread attention. Compared with normal pressure and humidity environments, high-altitude and low-pressure environments significantly alter the physicochemical behavior of concrete from mixing, pouring, hardening to long-term service. On the one hand, low pressure conditions may weaken the stability of micro-bubbles in fresh concrete, making them more prone to merging, floating, and coarsening. This leads to a decrease in the air content of fresh concrete, an increase in the bubble spacing coefficient after hardening, and an increase in the average bubble diameter, thereby reducing the pressure release capacity of concrete during freeze-thaw cycles. On the other hand, low humidity conditions accelerate early moisture loss in concrete, limiting the continuous hydration and hardening process of cementitious materials. This results in coarsening of the pore structure, loosening of the interface transition zone, increased drying shrinkage strain, and an increase in microcracks, which in turn exacerbates durability problems such as chloride ion intrusion, carbonation, and freeze-thaw damage.

[0003] Air-entrained concrete typically improves freeze-thaw resistance by introducing stable, uniform, and fine air bubbles, making it an important technique for durability design of concrete structures in cold and high-altitude regions. However, under low-pressure and low-humidity conditions, the mix design of traditional air-entrained concrete faces significant challenges. The composition of cementitious materials such as cement, fly ash, and granulated blast furnace slag powder not only affects the workability and strength development of concrete but also influences bubble stability, pore connectivity, shrinkage deformation, resistance to chloride ion penetration, carbonation resistance, and freeze-thaw durability. For example, an appropriate amount of fly ash can improve the flowability of the mixture, but under low-pressure conditions, it may affect the effectiveness of the air-entraining agent and reduce bubble stability; granulated blast furnace slag powder helps improve later compaction and resistance to chloride ion migration, but excessive amounts may increase the risk of shrinkage under low-humidity conditions. Therefore, the mix design of concrete under low-pressure and low-humidity conditions is no longer a problem of controlling only strength or workability, but a complex multi-objective optimization problem involving the synergistic balance of workability, bubble structure, mechanical properties, volume stability, and multiple durability indicators.

[0004] Existing concrete mix design methods mainly rely on empirical formulas, single-factor experiments, response surface methodology (RSM), or manual mix design. While these methods can reveal the relationship between cementitious material composition and concrete performance to some extent, they still have the following shortcomings: First, traditional empirical design methods struggle to simultaneously consider the coupled effects of low air pressure and low humidity on bubble stability, hardening process, and durability degradation. Second, single-factor experiments cannot effectively reflect the nonlinear interactions among the ternary cementitious materials of cement, fly ash, and granulated blast furnace slag powder. Third, RSM methods typically rely on finite-order polynomial fitting, which has limited expressive power for multi-index, strongly nonlinear, and multi-constraint problems. Fourth, low-pressure, low-humidity curing tests are costly, time-consuming, and have limited sample sizes, leading to overfitting in conventional machine learning or deep learning models and making it difficult to reliably generalize to untested mix proportions. Fifth, existing methods primarily focus on performance prediction and lack a complete intelligent design framework that combines small-sample data augmentation, simultaneous prediction of multiple performance parameters, and multi-objective back-optimization.

[0005] In recent years, deep learning methods have demonstrated strong nonlinear modeling capabilities in material performance prediction and mix design optimization. However, their application to air-entrained concrete in low-pressure, low-humidity environments still faces challenges such as data scarcity, numerous output indicators, complex coupling relationships between indicators, and difficulty in ensuring physical consistency. Especially under low-pressure curing conditions, there are clear intrinsic correlations among concrete performance indicators. For example, an increase in the bubble spacing coefficient leads to a decrease in relative dynamic elastic modulus and an increase in mass loss rate; the limited low-humidity hardening process causes coarsening of the pore structure, further affecting chloride ion migration and carbonation depth; changes in the proportions of fly ash, granulated blast furnace slag powder, and cement simultaneously alter workability, bubble structure, and later-stage durability. Therefore, using only ordinary neural networks for single-indicator prediction is insufficient to meet the engineering requirements for multi-performance synergistic control and optimal mix design.

[0006] Therefore, there is an urgent need to propose an intelligent optimization method for the mix proportion of air-entrained concrete in low-pressure and low-humidity environments. This method organically combines small-sample data augmentation, deep learning multi-task prediction, and multi-objective optimization algorithms to accurately establish the mapping relationship between the composition of ternary cementitious materials (cement, fly ash, and granulated blast furnace slag powder) and multiple performance indicators based on limited experimental data. Furthermore, it can rapidly search for the optimal cementitious material composition suitable for low-pressure and low-humidity environments under multiple constraints, including slump, air content, bubble spacing coefficient, compressive strength, drying shrinkage strain, unsteady-state chloride ion migration coefficient, carbonation depth, relative dynamic elastic modulus, and mass loss rate. This method has significant engineering application value for improving the durability of concrete structures in plateau and low-pressure areas, reducing experimental mix design costs, and enhancing the intelligence level of material design. Summary of the Invention

[0007] The purpose of this invention is to address the problems existing in the mix design of air-entrained concrete under low-pressure and low-humidity environments, such as high experimental costs, small sample size, complex coupling relationships of multiple performance indicators, limited nonlinear expression ability of traditional response surface methodology, and difficulty of ordinary machine learning models in taking into account physical consistency and multi-objective reverse optimization. This invention provides an intelligent optimization method and system for the mix proportion of air-entrained concrete under low-pressure and low-humidity environments.

[0008] This invention takes the composition of a ternary cementitious material consisting of cement, fly ash, and granulated blast furnace slag powder as input, and outputs multiple performance indicators, including slump, air content of fresh concrete, air content of hardened concrete, bubble spacing coefficient, average bubble diameter, 3-day, 7-day, and 28-day compressive strength, unsteady-state chloride ion migration coefficient, 3-day and 28-day drying shrinkage strain, carbonation depth, relative dynamic elastic modulus, and mass loss rate. It uses a conditional denoising diffusion model to enhance small sample data, establishes a nonlinear mapping relationship between the cementitious material composition and multiple performance indicators through a composition-aware attention multi-task neural network, and achieves reverse design of concrete mix proportions under low-pressure and low-humidity conditions through a multi-objective optimization algorithm. This method can obtain recommended mix proportions that meet the requirements of workability, mechanical properties, volume stability, chloride ion migration resistance, carbonation resistance, and freeze-thaw durability based on limited experimental data.

[0009] To achieve the above objectives, this invention provides a method for intelligent optimization of air-entrained concrete mix proportions for low-pressure and low-humidity environments, comprising the following steps: S1. Constructing the original database: Using the mass fraction of cement, fly ash, and granulated blast furnace slag powder in the air-entrained concrete cementitious material as input variables, and using multiple performance indicators of air-entrained concrete under low-pressure and low-humidity conditions as output variables, an original cementitious material composition-performance database is established. S2. Data Augmentation: The original database is augmented using a conditional denoising diffusion model. The composition of the ternary cementitious material is used as a conditional variable. Gaussian noise is added to the joint samples through a forward diffusion process, and then an augmented sample that initially satisfies the constraints of the ternary cementitious material composition is generated through a reverse denoising process. S3. Physical consistency screening: Physical consistency screening is performed on the enhanced samples. The physical consistency screening includes ternary composition constraints, performance boundary constraints and material evolution law constraints. Samples that do not meet the physical laws of concrete materials are eliminated to obtain effective enhanced samples. S4. Construct a multi-task prediction model: Construct a compositional perceptual attention multi-task neural network, which includes a compositional feature embedding layer, a shared feature extraction layer, a task attention layer, and a multi-task output layer; S5. Training a multi-task prediction model: Using the dataset after merging the original database and the effective enhanced samples as training data, construct a total loss function including prediction error loss and physical consistency loss, train the composition-aware attention multi-task neural network, and establish the mapping relationship between the composition of cementitious materials and the multiple performance indicators. S6. Construct a multi-objective optimization model: Use the trained composition-aware attention multi-task neural network as a surrogate model to construct a multi-objective function with workability, bubble structure, mechanical properties, volume stability, resistance to chloride ion migration, resistance to carbonization, and freeze-thaw durability. Set the composition constraints and performance boundary constraints of the ternary cementitious material according to the design requirements to establish a multi-objective optimization model. S7. Multi-objective optimization solution: Within the design domain of ternary cementitious material composition, the multi-objective optimization model is solved using a multi-objective optimization algorithm to obtain the Pareto optimal candidate mix proportion set based on the surrogate model prediction; S8. Output recommended mix ratios: Sort the Pareto optimal candidate mix ratio set according to the comprehensive satisfaction function, and output the recommended mix ratios and their corresponding prediction performance indicators.

[0010] The present invention also provides an intelligent optimization system for air-entrained concrete mix proportions in low-pressure and low-humidity environments, for implementing the aforementioned intelligent optimization method for air-entrained concrete mix proportions in low-pressure and low-humidity environments, comprising: The data input module is used to input the composition of cement, fly ash and granulated blast furnace slag powder ternary cementitious materials and concrete performance data under low pressure and low humidity conditions, forming the original cementitious material composition-performance database; The data augmentation module is used to augment the original database with the ternary cementitious material composition as the conditional variable and through a conditional denoising diffusion model to generate augmented samples that initially meet the ternary composition constraints. The physical screening module is used to screen the enhancement samples for physical consistency according to the ternary composition constraints, performance boundary constraints and material evolution law constraints, and output the effective enhancement samples; The multi-task prediction module is used to construct and train a perceptual attention multi-task neural network using the dataset after merging the original database and the effective enhanced samples as training data, and to simultaneously predict multiple performance indicators of concrete. The multi-objective optimization module is used to use the trained composition-aware attention multi-task neural network as a surrogate model to establish and solve a multi-objective optimization model that includes multi-objective functions, ternary cementitious material composition constraints and performance boundary constraints, and to obtain the Pareto optimal candidate mix ratio set based on the prediction of the surrogate model. The recommended mix ratio output module is used to sort the Pareto optimal candidate mix ratio set according to the comprehensive satisfaction function, and output the recommended mix ratio and its corresponding predictive performance index. Output the recommended optimal mix proportion and the corresponding predicted performance index from the Pareto optimal mix proportion set. ,in, For the first Satisfaction weights for each performance metric This is the set of optimal candidate combinations for Pareto; Single-index satisfaction function The definition is as follows: For performance metrics that are generally considered better the larger they are... ,in, For the proxy model to the first Predicted values ​​of each performance indicator For the first The target lower limit for each performance metric For the first The target upper limit for each performance indicator; For performance metrics that are as small as possible, the better ; For interval-type performance indicators ,in The center value of the target interval. The target interval half-width; And on Perform truncation: , The truncated value is limited to the interval [0,1]. Satisfaction with various performance metrics.

[0011] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention introduces the conditional denoising diffusion model into the mix design of air-entrained concrete under low-pressure and low-humidity conditions, which can generate physically reasonable enhanced samples under small sample test data conditions, thereby improving the training stability and generalization ability of deep learning models.

[0012] (2) The present invention embeds the composition constraints of cement, fly ash and granulated blast furnace slag powder into the entire process of data augmentation, performance prediction and proportion optimization, so as to ensure that the generated samples and optimization results meet the composition constraints of cement.

[0013] (3) The present invention constructs a multi-task neural network for perceptual attention, which can simultaneously predict multiple performance indicators such as slump, air content of fresh concrete, air content of hardened concrete, bubble spacing coefficient, average bubble diameter, compressive strength, unsteady chloride ion migration coefficient, drying shrinkage strain, carbonation depth, relative dynamic elastic modulus and mass loss rate, thus overcoming the shortcomings of single-index prediction methods that cannot reflect the performance coupling relationship.

[0014] (4) In the model training process, the present invention introduces physical consistency constraints such as strength growth, shrinkage growth and freeze-thaw durability evolution, so that the prediction results not only have data fitting accuracy, but also conform to the basic material law of air-entrained concrete in low pressure and low humidity environment.

[0015] (5) This invention couples a deep learning surrogate model with a multi-objective optimization algorithm to realize the transformation from positive performance prediction of cementitious material composition to reverse intelligent design of mix proportions, and can quickly obtain the Pareto optimal candidate mix proportion set based on the prediction of the surrogate model.

[0016] (6) The present invention can simultaneously consider the effects of low pressure and low humidity environment on bubble stability, hardening process, pore structure, transport performance and freeze-thaw durability, and is applicable to intelligent mix design of air-entrained concrete in high-altitude, cold, low-pressure and low-humidity engineering environments.

[0017] (7) This invention can reduce the number of repeated mix design tests under low pressure and low humidity conditions, shorten the material design cycle, reduce test costs, and improve the scientificity, reliability and engineering applicability of concrete mix design. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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. Wherein: Figure 1 The flowchart shows the intelligent optimization method for air-entrained concrete mix proportions in low-pressure and low-humidity environments provided by this invention.

[0019] Figure 2 This is a module framework diagram of the intelligent optimization system for air-entrained concrete mix proportions for low-pressure and low-humidity environments provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this invention provides an intelligent optimization method for the mix proportion of air-entrained concrete in low-pressure and low-humidity environments. This method uses the composition of a ternary cementitious material consisting of cement, fly ash, and granulated blast furnace slag powder as design variables, and takes the workability, bubble structure, mechanical properties, chloride ion migration resistance, carbonation resistance, volume stability, and freeze-thaw durability of air-entrained concrete under low-pressure and low-humidity conditions as optimization objectives. The method includes the following steps: constructing an initial database, data augmentation, physical consistency screening, constructing a multi-task prediction model, training the multi-task prediction model, constructing a multi-objective optimization model, solving the multi-objective optimization problem, and outputting a recommended mix proportion. The specific steps are as follows: S1. Construct the original database; First, the cementitious material system of air-entrained concrete is determined to consist of cement, fly ash, and granulated blast furnace slag powder. Using the mass fraction of each component in the total cementitious material as input variables, a composition vector is established. : ; In the formula, This represents the cement mass fraction. This refers to the mass fraction of fly ash. This refers to the mass fraction of granulated blast furnace slag powder.

[0022] Since the three cementing materials together constitute a complete cementing material system, the input variables satisfy the following composition constraints: ; ; In a preferred embodiment, to ensure that the air-entrained concrete has sufficient early hardening capacity and durability reserve under low-pressure and low-humidity conditions, the design range of cement, fly ash, and granulated blast furnace slag powder can be further limited as follows: ,0 , ; Under low-pressure and low-humidity conditions, multiple performance indicators of concrete with different cementitious material compositions were tested, and an output vector was constructed. : ; In the formula, Slump, in mm; The air content of freshly mixed concrete is expressed as % (%). Air content of hardened concrete, expressed as % This is the bubble spacing coefficient, in μm; The average bubble diameter is expressed in μm. , and These are the compressive strengths at 3d, 7d, and 28d, respectively, in MPa. This is the unsteady-state chloride ion migration coefficient, in units of 10. -12 m² / s; and The drying shrinkage strains at 3 days and 28 days are respectively, in units of 10. -6 ; Carbonization depth, in mm; The relative dynamic elastic modulus (RDME) after freeze-thaw cycles is expressed in percentages (%). The mass loss rate after freeze-thaw cycles is expressed in percent.

[0023] This forms the original experimental database. : ; in, This represents the original number of experimental samples. For the first Composed of cementitious materials, For the first The results of multiple performance tests of the cementitious material composition under low-pressure and low-humidity environmental conditions.

[0024] To eliminate the differences in the dimensions of different performance indicators, the input variables... and output variables Perform normalization. The normalization expression for the input variables is: ; The normalized expression for the output variable is: ; in, For the first In the nth sample Normalized values ​​of the input variables; For the first In the nth sample Normalized values ​​of each output metric; and The first The minimum and maximum values ​​of the input variables; and The first The minimum and maximum values ​​of each output indicator.

[0025] S2, Data Augmentation; Because low-pressure, low-humidity curing tests are lengthy and costly, and the number of original test samples is usually limited, directly training a deep learning model can easily lead to overfitting. Therefore, this invention employs a conditional denoising diffusion model to augment the original cementitious material composition-performance samples. This step can generate physically reasonable augmented samples under small sample test data conditions, improving the training stability and generalization ability of the deep learning model.

[0026] First, the first The combination of input composition and output performance forms a joint sample: ; Right now: ; in, This represents the real joint samples without noise. For the first In the nth sample The normalized value of each output metric.

[0027] Conditional denoising diffusion models include forward diffusion processes and reverse denoising processes.

[0028] During the forward diffusion process, from real joint samples Starting from this point, Gaussian noise is gradually added to obtain the first... Step-noisy samples Its expression is: ; In the formula, For the first Step noise variance parameter It is the identity matrix. =1,2,…, , The total number of diffusion steps, This represents a Gaussian distribution.

[0029] make , ,in, Indicates from step 1 to step 2. Step by step The result of the cumulative product, If it is a product term, then it can be directly derived from the real samples. Sampling at any time Noisy samples: ; The corresponding reparameterization expression is: ; in, It is standard Gaussian noise.

[0030] In the reverse noise reduction process, it is composed of cementitious materials. As a conditional variable, train the denoising network Predict noise from noisy samples. The reverse process is represented as: ; In the formula, The mean value predicted by the denoising network. Let covariance matrix be the variance matrix. For network parameters, For network parameters θ A defined inverse denoising probability distribution.

[0031] The denoised mean can be further expressed as: ; The model training objective is to minimize the mean square error between the actual noise and the predicted noise. ; To enhance the model's ability to identify mixing ratio conditions, the composition conditions of cementitious materials are embedded into a denoising network: ; At the same time, the number of diffusion steps Perform time embedding: ; The input to the denoising network can then be expressed as: ; in, Embed the function to match the ratio conditions. For time step embedding functions, For denoising neural networks, For the first The time embedding vector of each diffusion step is the conditional embedding vector for the composition of ternary cementitious materials.

[0032] After the diffusion model is trained, new candidate proportions are sampled in the proportion space that satisfies the ternary composition constraints. : ; in, This represents the mass fraction of cement in the new candidate mix proportions. The mass fraction of fly ash in the new candidate mix proportion. The mass fraction of granulated blast furnace slag powder in the new candidate mix; And satisfy: , ; by As a condition, from standard Gaussian noise Begin gradual noise reduction: ; in, Standard Gaussian noise, For the first Step sampling standard deviation, A new candidate matching vector to satisfy the ternary composition constraint.

[0033] go through After the reverse denoising step, the generated sample is obtained: ; in, The joint samples generated by the reverse denoising process. The normalized performance index generated for the diffusion model. After inverse normalization, we obtain: ; in, This is the actual value of the k-th generation performance metric after inverse normalization; Thus, a diffusion-enhanced sample set is obtained: ; In the formula, To increase the sample size, Let be the generated performance vector corresponding to the j-th augmented sample.

[0034] S3, Physical Consistency Screening; To avoid the diffusion model generating samples that violate material properties, this invention enhances the dataset. Physical consistency screening is performed. This step ensures that the generated samples and optimization results meet the constraints of cementitious material composition and the evolution law of material properties, so that the prediction results conform to the basic material laws of air-entrained concrete under low-pressure and low-humidity environments.

[0035] First, compositional constraints are used for screening. For any augmented sample... ,Require: ; ; in, Numerical calculation tolerances for composition and constraints.

[0036] Secondly, performance boundary screening is performed. For the first... One performance indicator, requirement: ; in, and The first The reasonable lower and upper limits of each performance index under low pressure and low humidity conditions.

[0037] Next, the age-related evolution patterns are screened. The compressive strength should meet the following requirements: ; The drying shrinkage strain should meet the following requirements: ; in, For the first Predicted 28-day compressive strength of a reinforced sample. For the first Predicted 7-day compressive strength of a reinforced sample. For the first Predicted 3D compressive strength values ​​for each reinforced sample. For the first Predicted 28-day drying shrinkage strain values ​​for one enhanced sample. For the first Predicted 3-day drying shrinkage strain values ​​for an enhanced sample.

[0038] Furthermore, a basic correlation should exist between bubble structure and freeze-thaw durability: as the bubble spacing coefficient increases, the overall freeze-thaw resistance of concrete decreases, while the overall mass loss rate increases. For the enhanced sample set, this can be expressed through local correlation constraints as follows: ; ; in, For the correlation coefficient function, To generate the bubble spacing coefficient for the sample, To generate the RDME of the sample, The quality loss rate of the generated samples.

[0039] Gradient filtering can also be used: ; ; After filtering, the effective augmented dataset is obtained. : ; And merge with the original database: ; in, This is the union of the original database and the effectively augmented dataset, which is used as training data for subsequent deep learning prediction models.

[0040] S4. Construct a multi-task prediction model; To simultaneously predict multiple performance indicators of concrete under low-pressure and low-humidity conditions, this invention constructs a composition-aware attention-based multi-task neural network. This network can simultaneously predict multiple performance indicators, including slump, air content of fresh concrete, air content of hardened concrete, bubble spacing coefficient, average bubble diameter, compressive strength, unsteady-state chloride ion migration coefficient, drying shrinkage strain, carbonation depth, relative dynamic elastic modulus, and mass loss rate, overcoming the limitation of single-indicator prediction methods in reflecting the coupling relationship of performance. The network includes a composition feature embedding layer, a shared feature extraction layer, a task attention layer, and a multi-task output layer.

[0041] First, compositional feature embedding is performed on the original ternary composition input. Considering the nonlinear interaction between cement, fly ash, and granulated blast furnace slag powder, the input vector is expanded as follows: ; in, , and This indicates the coupling effect between any two cementitious materials. , and This represents the nonlinear effect of a single component.

[0042] Furthermore, higher-order compositional embeddings can be constructed. : ; The embedded vectors are input into the shared feature extraction network: ; ; in, This is the basic composition embedding vector consisting of the ternary cementitious material composition and its second-order interaction terms. The initial input features of the shared feature extraction network are the high-order compositional embedding vectors u. , For the first Each shared layer outputs features. and The first Layer weight matrix and bias vector, It is a non-linear activation function. This represents the number of shared layers.

[0043] To improve training stability, normalization and residual connections can be added to the shared feature layer: ; When the dimensions are inconsistent, a linear mapping can be used to make the dimensions of the residual terms consistent: ; in, This is a dimension mapping matrix.

[0044] Because different performance indicators are affected differently by the composition of the cementitious materials, this invention sets up an independent attention mechanism for each output task. For the first... For each performance metric, its attention vector is: ; in, To share the final output of the feature extraction network, and The first Attention weight matrix and bias vector for each task.

[0045] No. The attention-weighted features for each task are: ; in, This indicates element-wise multiplication.

[0046] Then, Enter the first Task output headers: ; ; ; in, For the first Predicted values ​​for each performance metric For the first Each task outputs the feature vector of the first hidden layer. For the first Each task outputs the feature vector of the second hidden layer. It is a non-linear activation function. This is the weight matrix of the first hidden layer. This is the bias vector of the first hidden layer. This is the weight matrix of the second hidden layer. This is the bias vector for the second hidden layer. The weight matrix of the output layer. This is the bias term for the output layer.

[0047] The final model output is: ; when At that time, there were: ; in For carbonization depth, To generate the bubble spacing coefficient for the sample, To generate the RDME of the sample, The quality loss rate of the generated samples.

[0048] S5. Train a multi-task prediction model; To ensure that the composition-aware attention-based multi-task neural network not only possesses high prediction accuracy but also conforms to the evolutionary laws of concrete material properties, this invention uses the dataset resulting from the merging of the original database and the effective enhanced samples as training data. A total loss function, including prediction error loss and physical consistency loss, is constructed to train the composition-aware attention-based multi-task neural network, establishing a mapping relationship between the cementitious material composition and the various performance indicators. This step introduces physical consistency constraints such as strength growth, shrinkage growth, and freeze-thaw durability evolution during model training, ensuring that the prediction results not only have data fitting accuracy but also conform to the basic material laws of air-entrained concrete under low-pressure and low-humidity environments.

[0049] First, define the multi-task prediction error loss: ; in, The number of training samples. The total number of performance indicators. For the first Loss weights for each performance metric To standardize the prediction error, For the first The scale parameters of each performance indicator.

[0050] To reflect the basic law of concrete compressive strength increasing with age under low pressure and low humidity conditions, a strength evolution loss is set: ; To reflect the development of drying shrinkage strain over time, a shrinkage evolution loss is set: ; To reflect the intrinsic relationship between bubble structure and freeze-thaw performance, a freeze-thaw consistency ranking loss is set: ; in , For set The number of elements in the middle.

[0051] In addition, to ensure that the prediction results meet a reasonable physical range, a boundary penalty loss is set: ; Taking into account the above losses, the total loss function of the model is: ; in, , , ,and These are the weighting coefficients for the losses of each physical constraint.

[0052] By minimizing To obtain the trained deep learning agent model: ; Right now: ; in, These are all the parameters after the model has been trained.

[0053] S6. Construct a multi-objective optimization model; After training, the trained perceptual attention multi-task neural network is used as a surrogate model. A multi-objective function is established, encompassing workability, bubble structure, mechanical properties, volumetric stability, resistance to chloride ion migration, resistance to carbonization, and freeze-thaw durability. Based on design requirements, ternary cementitious material composition constraints and performance boundary constraints are set, and a multi-objective optimization model is established to search for candidate cementitious material compositions within the ternary cementitious material composition design domain. This step couples a deep learning surrogate model with the multi-objective optimization algorithm, enabling a shift from forward performance prediction of cementitious material composition to reverse intelligent design of mix proportions. This allows for the rapid acquisition of a set of Pareto optimal candidate mix proportions predicted by the surrogate model.

[0054] The optimization variables are: ; The basic form of the optimization problem is: ; The constraints are: ; ; in, It is a vector of multi-objective functions.

[0055] Based on the prediction obtained from the deep learning agent model: ; The first type of objective is a workable objective. To ensure the slump is close to the center of the target range, the following definition is made: ; in, The target slump center value: ; In a preferred embodiment, , ,but: ; The second type of objective is the air-bubble structure objective. This objective considers the air content of fresh concrete, the air content of hardened concrete, the air-bubble spacing coefficient, and the average air-bubble diameter simultaneously. ; in, It is an interval-type penalty function. It is an upper bound type penalty function.

[0056] The interval penalty function is defined as: ; The upper bound penalty function is defined as: ; The lower bound penalty function is defined as: ; The third category of objectives is mechanical performance objectives. To improve 28-day compressive strength, the following is defined: ; in, This serves as a reference value for strength normalization. Since multi-objective optimization algorithms typically employ a minimization approach, the strength is negatively evaluated to maximize the 28-day compressive strength.

[0057] Both early strength and late strength can be considered simultaneously: ; in, , and These represent the intensity weights for different age groups.

[0058] The fourth category of targets includes resistance to chloride ion migration, resistance to carbonization, and volume stability. Definition: ; in, , and These are the weighting coefficients for chloride ion migration coefficient, carbonization depth, and drying shrinkage strain, respectively. , , These are the normalized reference values ​​for the corresponding performance indicators.

[0059] The fifth category of objectives is freeze-thaw durability. Definition: ; in, and These are the weighting coefficients for relative dynamic elastic modulus and mass loss rate, respectively. This is a normalized reference value for the quality loss rate.

[0060] At the same time, set multiple performance constraints: ; ; ; ; ; ; ; ; ; ; ; in, The non-steady-state chloride ion migration coefficient, This represents the upper limit of the allowable unsteady-state chloride ion migration coefficient. For carbonization depth, This represents the upper limit of the permissible carbonization depth. This represents the lower limit of the relative dynamic elastic modulus. This represents the upper limit allowed for the rate of quality loss. In a preferred embodiment, the multi-performance design constraints for air-entrained concrete under low-pressure and low-humidity conditions are taken as follows: ; ; ; ; ; ; ; ; .

[0061] S7. Multi-objective optimization solution; In a preferred embodiment, a multi-objective optimization algorithm is used to solve the multi-objective optimization model within the ternary cementitious material composition design domain to obtain a set of Pareto optimal candidate mix proportions predicted by the surrogate model.

[0062] First, generate the initial population: ; Where P is the population size. The first in the initial population Each individual component is a ternary cementitious material that satisfies compositional constraints. ; And satisfy: ; For each individual, a deep learning proxy model is invoked to predict performance: ; Then calculate the objective function: ; And calculate the degree of constraint violation: ; in, The number of inequality constraints. Indicates the first The constraints meet the requirements.

[0063] For upper limit constraints ,definition: ; For lower bound constraints ,definition: ; For interval constraints ,definition: ; ; When two solutions and When all constraints are satisfied, sort according to Pareto dominance. If the following conditions are met: ; And there is at least one target. Make: ; Then it is called Dominate .

[0064] For individuals within the same non-dominated stratum, calculate the crowding distance: ; in, For the first The crowded distance between individuals and The first The maximum and minimum values ​​of the objective function in the current undominated layer.

[0065] Then, offspring populations are generated through selection, crossover, and mutation. For two parent individuals... and Offspring can be generated using simulated binary crossover: ; ; in, This is the cross-expansion factor.

[0066] Perform polynomial mutation on the offspring after crossover: ; in, For variant perturbations, and The first The upper and lower limits of each variable.

[0067] To ensure that individuals after mutation and crossover still satisfy the ternary composition constraint, a normalization correction is performed on the individuals: ; ; ; After iteration After generation, the Pareto optimal candidate combination ratio set is obtained: ; in, This represents the number of Pareto optimal candidate solutions. It is the first candidate solution in the Pareto optimal candidate mix ratio set.

[0068] S8, Recommended output ratio.

[0069] Since the Pareto optimal candidate mix ratio set usually contains multiple candidate solutions, this invention further constructs a comprehensive satisfaction function to sort the Pareto solutions for easier engineering applications.

[0070] For performance indicators where a larger value is always better, such as 28-day compressive strength and relative dynamic modulus of elasticity, the satisfaction level of a single indicator is defined as follows: ; For performance indicators where smaller values ​​are better, such as the unsteady-state chloride ion migration coefficient, carbonization depth, drying shrinkage strain, and mass loss rate, the satisfaction level of a single indicator is defined as: ; For interval-type performance indicators, such as slump and air content of fresh concrete, the satisfaction level of a single indicator is defined as: ; in, The center value of the target interval. The target interval half-width: ; ; To avoid satisfaction levels being less than 0 or greater than 1, truncation is performed: ; The overall satisfaction function is defined as follows: ; in, For the first Satisfaction weights for each performance metric.

[0071] The final recommended ratio is: ; Right now: ; The system also outputs the predicted performance indicators corresponding to this recommended ratio: ; Right now: .

[0072] This enables the intelligent design of air-entrained concrete under low-pressure and low-humidity conditions, from "predicting performance based on given cementitious material composition" to "reverse optimizing mix proportions based on given performance targets".

[0073] like Figure 2 As shown, the present invention also provides an intelligent optimization system for the mix proportion of air-entrained concrete in low-pressure and low-humidity environments, comprising: Data input module 1 is used to input the composition of cement, fly ash and granulated blast furnace slag powder ternary cementitious materials and concrete performance data under low pressure and low humidity conditions, forming the original cementitious material composition-performance database; Data augmentation module 2 is used to augment the original database with the ternary cementitious material composition as the conditional variable and through a conditional denoising diffusion model to generate augmented samples that initially meet the constraints of the ternary cementitious material composition. The physical screening module 3 is used to screen the enhancement samples for physical consistency according to the ternary composition constraints, performance boundary constraints and material evolution law constraints, and output the effective enhancement samples; Multi-task prediction module 4 is used to construct and train a perceptual attention multi-task neural network using the dataset after merging the original database and the effective enhanced samples as training data, and to simultaneously predict multiple performance indicators of concrete. The multi-objective optimization module 5 is used to use the trained composition-aware attention multi-task neural network as a surrogate model to establish and solve a multi-objective optimization model containing multi-objective functions, ternary cementitious material composition constraints and performance boundary constraints, and to obtain the Pareto optimal candidate mix ratio set based on the prediction of the surrogate model. Recommended mix ratio output module 6 is used to sort the Pareto optimal candidate mix ratio set according to the comprehensive satisfaction function, and output the recommended mix ratio and its corresponding predictive performance index. Output recommended mix proportions and corresponding predicted performance indicators from the Pareto optimal mix proportion set. ,in, For the first Satisfaction weights for each performance metric This is the set of optimal candidate combinations for Pareto; The single-index satisfaction function The definition is as follows: For performance metrics that are generally considered better the larger they are... ; For performance metrics that are as small as possible, the better ; For interval-type performance indicators ,in , ; And for the above three situations Perform truncation: .

[0074] It should be further explained that the multi-objective optimization module 5 establishes the following multi-objective optimization model: ; Among them, the work-related objectives are: ; The target of the bubble structure is: ; The mechanical performance targets are: ; The targets for resistance to chloride ion migration, resistance to carbonization, and volume stability are: ; The freeze-thaw durability target is: ; in, The target slump center value, It is an interval-type penalty function. It is an upper bound type penalty function. , , , , These are the weighting coefficients for the chloride ion migration coefficient, carbonization depth, drying shrinkage strain, relative dynamic elastic modulus, and mass loss rate in the corresponding objective functions. This is a normalized reference value for the average bubble diameter. , , , , These are the normalized reference values ​​for the corresponding performance indicators; The multi-objective optimization model 5 satisfies the following performance boundary constraints: ; ; ; ; ; ; ; ; ; ; .

[0075] The present invention will be further illustrated below with specific embodiments. These embodiments are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0076] The implementation conditions of this invention are as follows: low-pressure and low-humidity environmental conditions: pressure 60 kPa, temperature 20℃, and relative humidity 60%; the cementitious material system is a ternary system of cement, fly ash, and granulated blast furnace slag powder. The compressive strength test ages are 3 days, 7 days, and 28 days; the drying shrinkage test ages are 3 days and 28 days; the unsteady-state chloride ion migration coefficient and carbonization depth test ages are 28 days; and the freeze-thaw performance is evaluated after 200 cycles.

[0077] Example 1 Air-entrained concrete was prepared using cement, fly ash, and granulated blast furnace slag powder as cementitious materials, with a cement mass fraction of 0.50%, a fly ash mass fraction of 0.50%, and a granulated blast furnace slag powder mass fraction of 0%. The concrete was mixed, poured, and cured in a low-pressure, low-humidity environment of 60 kPa, 20℃, and 60% relative humidity, and was cured to the corresponding performance test age.

[0078] Testing revealed that the cementitious material composition corresponds to a concrete slump of 198 mm, an air content of 4.1% in fresh concrete, an air content of 3.0% in hardened concrete, an air bubble spacing coefficient of 230 μm, and an average bubble diameter of 148 μm. Its 3-day, 7-day, and 28-day compressive strengths are 31.6 MPa, 36.4 MPa, and 40.0 MPa, respectively; the 28-day unsteady-state chloride ion migration coefficient is 5.3 × 10⁻⁶. －12 m² / s; drying shrinkage strains at 3d and 28d were 124 × 10⁻⁶ m² / s; －6 and 342×10 －6The carbonization depth after 28 days was 7.02 mm; after 200 freeze-thaw cycles, the relative dynamic elastic modulus was 66.70%, and the mass loss rate was 3.50%.

[0079] The above results show that, under conditions of high fly ash content, concrete exhibits good fluidity in low-pressure and low-humidity environments, but the bubble spacing coefficient and average bubble diameter are relatively large. After freeze-thaw cycles, the relative dynamic elastic modulus decreases and the mass loss rate increases. This cementitious material composition can serve as an example for the algorithm of this invention to identify unfavorable durability boundary samples under low-pressure and low-humidity environments.

[0080] Example 2 Another set of air-entrained concrete was prepared using cement, fly ash, and granulated blast furnace slag powder as cementing materials, with a cement mass fraction of 0.75%, a fly ash mass fraction of 0.25%, and a granulated blast furnace slag powder mass fraction of 0%. This concrete was also mixed, poured, and cured in a low-pressure, low-humidity environment of 60 kPa, 20℃, and 60% relative humidity, and was cured to the corresponding performance test age.

[0081] Testing revealed that the cementitious material composition corresponds to a concrete slump of 160 mm, an air content of 5.4% in fresh concrete, an air content of 3.9% in hardened concrete, an air bubble spacing coefficient of 220 μm, and an average bubble diameter of 143 μm. Its 3-day, 7-day, and 28-day compressive strengths are 40.9 MPa, 45.6 MPa, and 48.4 MPa, respectively; the 28-day unsteady-state chloride ion migration coefficient is 4.5 × 10⁻⁶. －12 m² / s; the drying shrinkage strains at 3d and 28d were 143 × 10⁻⁶ m² / s; －6 and 362×10 －6 The carbonization depth after 28 days was 5.63 mm; after 200 freeze-thaw cycles, the relative dynamic elastic modulus was 72.13%, and the mass loss rate was 2.76%.

[0082] Compared with Example 1, this cementitious material composition increased the cement ratio and decreased the fly ash ratio, exhibiting higher air content in fresh concrete, higher air content in hardened concrete, and higher compressive strength under low pressure and low humidity conditions. At the same time, the freeze-thaw mass loss rate decreased, indicating that appropriately increasing the cement ratio is beneficial to improving the comprehensive performance of air-entrained concrete under low pressure and low humidity conditions.

[0083] The test data of Examples 1 and 2 above can be used as representative samples in the original database of step S1 in the method of the present invention; the original database also includes other cementitious material composition-performance data obtained under the same raw materials, mix proportion parameters and low pressure and low humidity test conditions.

[0084] Based on the original database, the following optimization process is performed according to the method of the present invention.

[0085] First, a conditional denoising diffusion model is constructed according to step S2 for data augmentation. In this model, T represents the total number of diffusion steps. Let be the noise variance parameter at step t. A linear scheduling scheme is used. , ; Standard Gaussian noise. Composed of ternary cementitious materials. Using these as conditional variables, and following a forward diffusion and reverse denoising process, enhanced samples that initially satisfy the compositional constraints of the ternary cementitious material are generated. This step effectively expands the small sample dataset, improving the training stability and generalization ability of subsequent deep learning models.

[0086] Secondly, the enhanced samples are screened for physical consistency according to step S3. The tolerance for numerical calculation of composition and constraint in ternary composition constraints is also considered. Based on the accuracy setting of numerical calculations, the requirements are as follows: Furthermore, all components are non-negative. Simultaneously, a constraint on the age-dependent evolution of compressive strength is applied. Constraints on the Evolution of Drying Shrinkage Strain Age And the bubble spacing coefficient and freeze-thaw performance ranking constraint: for any two satisfying samples and Must meet and Samples that do not meet any of the above constraints are removed to obtain effective augmented samples, which are then merged with the original database to form the training dataset.

[0087] Construct a compositional perceptual attention-based multi-task neural network following step S4. The compositional feature embedding layer expands the ternary composition to... The shared feature extraction layer is represented as... , For the number of shared layers, For non-linear activation functions, the following can be used: function (defined as) Or other suitable activation functions. The task attention vector corresponding to each performance metric is: Attention-weighted features are The output layer consists of Give the first Predicted values ​​for each performance metric.

[0088] Construct a multi-task loss function with physical constraints and train the model following step S5. The prediction error loss is... ,in These are the loss weights for each performance metric. Physical consistency loss includes intensity growth loss. contraction and growth loss Freeze-thaw consistency ranking loss and boundary penalty loss The total loss function is ,in, , , and The weighting coefficients for each physical constraint loss can be determined by those skilled in the art through trial and error or grid search based on the magnitude of each loss term. A gradient descent-type optimizer is used for training, resulting in a deep learning surrogate model. .

[0089] Establish a multi-objective optimization model according to steps S6 and S7, and solve it using a multi-objective optimization algorithm. The optimization variables are: The feasible region is , 0≤ , , ≤1. Multi-objective function vector These correspond to the following targets: workability, bubble structure, mechanical properties, resistance to chloride ion migration, resistance to carbonization, volume stability, and freeze-thaw durability.

[0090] The performance boundary constraints are preferably set as follows: 160≤ ≤200mm, 5≤ ≤6%, ≤200μm, ≥50MPa ≤4.0×10 －12 m² / s, ≤400×10 －6 , ≤5.0mm, ≥75%, ≤ 3.0%. Air content of hardened concrete. and average bubble diameter Similarly, when boundary constraints are applied, those skilled in the art can determine their specific values ​​based on actual engineering requirements.

[0091] The parameters of the multi-objective optimization algorithm, such as population size, number of iterations, crossover probability, and mutation probability, can be determined by those skilled in the art based on the complexity of the optimization problem and computational resources, referring to conventional practices in the field of multi-objective optimization. After iteration, the Pareto optimal candidate mix ratio set is obtained. .

[0092] Finally, the overall satisfaction of each candidate solution in the Pareto optimal candidate mix design set is calculated according to step S8. ,in For the first Satisfaction weights for each performance metric For the first Individual performance indicators are used to determine the satisfaction function for each performance metric. The output is the recommended match ratio that yields the highest overall satisfaction. The system synchronously outputs the predicted performance index corresponding to the recommended combination ratio. .

[0093] The above embodiments provide representative original samples and implementation procedures for the method of the present invention. By combining conditional denoising diffusion data enhancement, physical consistency screening, composition-aware attention multi-task prediction, and multi-objective optimization, the workability, bubble structure, mechanical properties, and durability of different cementitious material compositions under low pressure and low humidity environments can be comprehensively analyzed, and the Pareto optimal candidate mix proportion set and recommended mix proportion based on the surrogate model prediction can be output.

Claims

1. A method for intelligent optimization of air-entrained concrete mix proportions for low-pressure, low-humidity environments, characterized in that, Includes the following steps: S1. Constructing the original database: Using the mass fraction of cement, fly ash, and granulated blast furnace slag powder in the air-entrained concrete cementitious material as input variables, and using multiple performance indicators of air-entrained concrete under low-pressure and low-humidity conditions as output variables, an original cementitious material composition-performance database is established. S2. Data Augmentation: The original database is augmented using a conditional denoising diffusion model. The composition of the ternary cementitious material is used as a conditional variable. Gaussian noise is added to the joint samples through a forward diffusion process, and then an augmented sample that initially satisfies the constraints of the ternary cementitious material composition is generated through a reverse denoising process. S3. Physical consistency screening: Physical consistency screening is performed on the enhanced samples. The physical consistency screening includes ternary composition constraints, performance boundary constraints and material evolution law constraints. Samples that do not meet the physical laws of concrete materials are eliminated to obtain effective enhanced samples. The constraints on the material evolution law include: Compressive strength age evolution constraints subscript To enhance the index numbering of samples, , , The first Predicted compressive strength values ​​at 3d, 7d, and 28d for each reinforced sample; Drying shrinkage strain age evolution constraints ,in , The first Predicted values ​​of drying shrinkage strain at 3d and 28d for one enhanced sample; And the ordering constraint between bubble spacing coefficient and freeze-thaw performance: for any two satisfying Enhanced samples and It should meet and subscript , To enhance the index numbering of samples, To generate the bubble spacing coefficient for the sample, To generate the RDME of the sample, The quality loss rate of the generated samples; S4. Construct a multi-task prediction model: Construct a compositional perceptual attention multi-task neural network, which includes a compositional feature embedding layer, a shared feature extraction layer, a task attention layer, and a multi-task output layer; S5. Training a multi-task prediction model: Using the dataset after merging the original database and the effective enhanced samples as training data, construct a total loss function including prediction error loss and physical consistency loss, train the composition-aware attention multi-task neural network, and establish the mapping relationship between the composition of cementitious materials and the multiple performance indicators. The prediction error loss is: ； in The number of training samples. The total number of performance indicators. For the first Loss weights for each performance metric For the first The scaling parameters of each performance metric and The first In the nth sample Measured and predicted values ​​of each performance indicator; The physical consistency loss includes: Intensity growth loss ; contraction growth loss ; Freeze-thaw consistency ranking loss ,in , For set The number of elements in the middle; Boundary penalty loss ,in and The first The lower and upper limits of each performance indicator; The total loss function is: ； in, , , , These are the weighting coefficients for intensity growth loss, shrinkage growth loss, freeze-thaw consistency ranking loss, and boundary penalty loss, respectively. S6. Construct a multi-objective optimization model: Using the trained composition-aware attention-based multi-task neural network as a surrogate model, construct multi-objective functions based on workability, bubble structure, mechanical properties, volume stability, resistance to chloride ion migration, resistance to carbonization, and freeze-thaw durability. Establish a multi-objective optimization model by setting compositional constraints and performance boundary constraints for the ternary cementitious material according to design requirements. ; Among them, the work-related objectives are: ; The target of the bubble structure is: ; The mechanical performance targets are: ; The targets for resistance to chloride ion migration, resistance to carbonization, and volume stability are: ; The freeze-thaw durability target is: ; in, The target slump center value, and , The collapse degree predicted by the surrogate model. This represents the upper limit of the target slump range. Â_f represents the lower limit of the slump target range; Â_f represents the air content of fresh concrete predicted by the surrogate model. This represents the lower limit of the target range for air content in freshly mixed concrete. This represents the upper limit of the target range for air content in freshly mixed concrete. The air content of hardened concrete predicted by the surrogate model. This represents the lower limit of the target range for air content in hardened concrete. This represents the upper limit of the target range for air content in hardened concrete. The bubble spacing coefficient predicted by the surrogate model. This represents the upper limit of the allowable bubble spacing coefficient. The average bubble diameter predicted by the surrogate model. This is a normalized reference value for the average bubble diameter; The 28-day compressive strength predicted by the surrogate model. This is a reference value for intensity normalization; The unsteady chloride ion migration coefficient predicted by the surrogate model. This is a normalized reference value for the unsteady-state chloride ion migration coefficient; The carbonization depth predicted by the surrogate model. This is a normalized reference value for carbonization depth; The surrogate model predicts the 28-day drying shrinkage strain. This is a normalized reference value for drying shrinkage strain; The relative dynamic elastic modulus after freeze-thaw cycles as predicted by the surrogate model; The mass loss rate after freeze-thaw cycles predicted by the surrogate model. This is a normalized reference value for the quality loss rate; , , , , These are the weighting coefficients for the chloride ion migration coefficient, carbonization depth, drying shrinkage strain, relative dynamic elastic modulus, and mass loss rate in the corresponding objective functions; The interval-type penalty function and the upper-level penalty function are as follows: ; ; The multi-objective optimization model satisfies the following performance boundary constraints: ; ; ; ; ; ; ; ; ; ; ； in, This represents the upper limit of the average bubble diameter. This represents the lower limit of the allowable compressive strength at 28 days. The upper limit of allowable shrinkage strain after 28 days of drying. The non-steady-state chloride ion migration coefficient, This represents the upper limit of the allowable unsteady-state chloride ion migration coefficient. For carbonization depth, This represents the upper limit of the permissible carbonization depth. This represents the lower limit of the relative dynamic elastic modulus. This is the upper limit of the allowable quality loss rate; S7. Multi-objective optimization solution: Within the design domain of ternary cementitious material composition, the multi-objective optimization model is solved using a multi-objective optimization algorithm to obtain the Pareto optimal candidate mix proportion set based on the surrogate model prediction. S8. Output recommended mix ratios: Sort the Pareto optimal candidate mix ratio set according to the comprehensive satisfaction function, and output the recommended mix ratios and their corresponding prediction performance indicators.

2. The method according to claim 1, characterized in that, The input variables are represented as ,in This represents the cement mass fraction. This refers to the mass fraction of fly ash. The mass fraction of granulated blast furnace slag powder; the composition of the ternary cementitious material satisfies ,and .

3. The method according to claim 1, characterized in that, The aforementioned performance indicators include slump, air content of fresh concrete, air content of hardened concrete, bubble spacing coefficient, average bubble diameter, 3-day compressive strength, 7-day compressive strength, 28-day compressive strength, unsteady chloride ion migration coefficient, 3-day drying shrinkage strain, 28-day drying shrinkage strain, carbonation depth, relative dynamic elastic modulus after freeze-thaw cycles, and mass loss rate after freeze-thaw cycles.

4. The method according to claim 1, characterized in that, The conditional denoising diffusion model combines input composition and output performance into a joint sample. ,in and These are the normalized input and output variables, respectively; the forward diffusion process is as follows: ,in Indicates from step 1 to step 2. Step by step The result of the cumulative product, For the product term, Standard Gaussian noise, It is the identity matrix. For real joint samples, For the first Step-noisy samples, , The total number of diffusion steps; the reverse denoising process is based on the ternary cementitious material composition. As a condition variable, the inverse probability distribution is expressed as: ,in, The mean value predicted by the denoising network. Let covariance matrix be the variance matrix. For network parameters, For network parameters θ A defined inverse denoising probability distribution. This represents a Gaussian distribution.

5. The method according to claim 1, characterized in that, The ternary composition constraint is: and ,in Numerical calculation tolerances for composition and constraints, For the first The mass fraction of cement in each enhanced sample For the first Mass fraction of fly ash in each enhanced sample For the first The mass fraction of granulated blast furnace slag powder in each enhanced sample.

6. The method according to claim 1, characterized in that, The compositional feature embedding layer of the compositional perceptual attention multi-task neural network expands the ternary composition to... ,in, , and This indicates the coupling effect between any two cementitious materials. , and The nonlinear effect of a single component is represented; the shared feature extraction layer is represented as... , ,in, For the first Each shared layer outputs features. and The first Layer weight matrix and bias vector, It is a non-linear activation function. For the number of shared layers; the first The task attention vector corresponding to each performance metric is: ,in To share the final output of the feature extraction network, and The first The attention weight matrix and bias vector for each task; the attention weighted features are: ,in This indicates element-wise multiplication; the first... The predicted values ​​of each performance indicator in For the first The weight matrix of the output layer for each task For the first Each task outputs the feature vector of the second hidden layer. For the first Bias terms for the output layer of each task.

7. A system for intelligently optimizing the mix proportion of air-entrained concrete for low-pressure and low-humidity environments, used to implement the intelligent optimization method for air-entrained concrete mix proportions in any one of claims 1-6, characterized in that, include: The data input module is used to input the composition of cement, fly ash and granulated blast furnace slag powder ternary cementitious materials and concrete performance data under low pressure and low humidity conditions, forming the original cementitious material composition-performance database; The data augmentation module is used to augment the original database with the ternary cementitious material composition as the conditional variable and through a conditional denoising diffusion model to generate augmented samples that initially meet the ternary composition constraints. The physical screening module is used to screen the enhancement samples for physical consistency according to the ternary composition constraints, performance boundary constraints and material evolution law constraints, and output the effective enhancement samples; The multi-task prediction module is used to construct and train a perceptual attention multi-task neural network using the dataset after merging the original database and the effective enhanced samples as training data, and to simultaneously predict multiple performance indicators of concrete. The multi-objective optimization module is used to establish and solve a multi-objective optimization model that includes multi-objective functions, ternary cementitious material composition constraints and performance boundary constraints, by using the trained composition-aware attention multi-task neural network as a surrogate model, and obtain the Pareto optimal candidate mix ratio set based on the surrogate model prediction. Recommended mix ratio output module, used to calculate the overall satisfaction function. The Pareto optimal candidate mix proportions are sorted, and the recommended mix proportions are determined from the Pareto optimal candidate mix proportions. Output the recommended mix ratio and its corresponding predictive performance index, wherein, For the first Satisfaction weights for each performance metric This is the set of optimal candidate combinations for Pareto; Single-index satisfaction function The definition is as follows: For performance metrics that are generally considered better the larger they are... ,in, For the proxy model to the first Predicted values ​​of each performance indicator For the first The target lower limit for each performance metric For the first The target upper limit for each performance indicator; For performance metrics that are as small as possible, the better ; For interval-type performance indicators ,in The center value of the target interval. The target interval half-width; And on Perform truncation: , The truncated value is limited to the interval [0,1]. Satisfaction with various performance metrics.

Images