Artificial intelligence-based uhpc material formulation optimization method

Abstract

The application discloses an UHPC material formula optimization method based on artificial intelligence, and belongs to the technical field of artificial intelligence assisted material formula optimization. The method comprises the following steps: acquiring basic physical parameters of raw materials such as cement, slag, magnesium oxide and admixtures, constructing an initial formula space and generating a uniform design formula set; performing hydration heat kinetics simulation on each candidate formula and outputting hydration reaction progress characteristics; screening an initial formula subset meeting hydration heat control requirements; inputting the initial formula subset into a trained graph convolutional neural network for formula-performance mapping and outputting predicted performance parameters of each formula; constructing a multi-objective optimization model and solving the model through a multi-objective optimization algorithm to output a Pareto optimal formula set; and scoring the process adaptability of each formula and selecting a target optimized formula. The application combines hydration heat simulation, graph neural network and multi-objective optimization, realizes efficient and accurate design of the UHPC formula, takes into account performance and processability, and effectively shortens the research and development cycle.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted material formulation optimization technology, specifically to a method for optimizing UHPC material formulations based on artificial intelligence. Background Technology

[0002] Formulation optimization for ultra-high performance concrete (UHPC) is a typical multi-component, multi-objective decision-making problem. The interactions between cement, slag, magnesium oxide, and various admixtures are complex, and even minor adjustments in proportions can have significant nonlinear effects on hydration heat release behavior, mechanical properties, and durability. Current technologies generally employ a trial-and-error approach based on experimental design for formulation selection. This involves generating a finite number of formulations through orthogonal experiments or simplex centroid design, followed by specimen preparation and standard curing, and then conducting mechanical and durability tests on each formulation. Finally, the optimal formulation is manually selected based on the test results. This method heavily relies on the physical experimental cycle for each iteration, resulting in low experimental throughput and difficulty in covering the combinatorial distribution characteristics in the high-dimensional formulation space.

[0003] Existing experiment-driven optimization methods have two main shortcomings. First, the formulation screening process lacks a pre-judgment mechanism for the early hydration reaction process of materials. The peak hydration exothermic rate and total heat release directly affect the temperature cracking risk of large-volume UHPC components, and if the delayed expansion characteristics of magnesium oxide are mismatched with the content of highly active tricalcium aluminate, it will introduce long-term volume stability risks. Traditional methods only test the final performance after curing, and cannot eliminate candidate combinations with hydration heat runaway or delayed expansion risks at the formulation generation stage, resulting in a large amount of subsequent experiments being spent on infeasible formulations. Second, the comprehensive performance prediction stage usually uses empirical regression models or conventional machine learning algorithms. These methods treat the formulation as a set of independent component vectors, ignoring the non-Euclidean spatial interaction structure formed by differences in chemical compatibility between components, resulting in insufficient mapping accuracy of durability indicators such as chloride ion diffusion coefficient under sparse experimental data conditions. Furthermore, when solving multi-objective optimization problems, existing methods often directly use the optimization results as output without considering the actual feasibility of the formulation in the Pareto front solution set in processes such as pumping and mixing. As a result, the theoretically optimal formulation obtained is not manufacturable within the engineering operation window. Summary of the Invention

[0004] This invention aims to provide an artificial intelligence-based method for optimizing UHPC material formulations, addressing the problems of low formulation screening efficiency and poor process adaptability of optimization results in existing technologies. By introducing hydration thermodynamics simulation into the initial formulation screening stage, utilizing graph convolutional neural networks to capture inter-component interaction features during the performance mapping stage, and incorporating process adaptability scoring in the final decision-making stage, intelligent optimization across the entire chain from formulation generation to engineering verification is achieved.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a UHPC material formulation optimization method based on artificial intelligence, the method comprising the following steps.

[0006] The basic physical properties of UHPC raw materials, including cement, slag, magnesium oxide, and admixtures, are obtained. As one technical solution of this invention, when obtaining these basic physical properties, X-ray fluorescence spectroscopy is performed on the cement raw material to obtain its oxide composition mass fraction, including the mass fractions of calcium oxide, silicon dioxide, aluminum oxide, and ferric oxide; laser particle size analysis is performed on the slag raw material to obtain its particle size distribution curve, and the specific surface area and average particle size of the slag are calculated accordingly; thermogravimetric analysis is performed on the magnesium oxide raw material to obtain its loss on ignition and activity index; and infrared spectroscopy is performed on the admixture raw material to obtain its functional group type and solid content. This allows for a comprehensive and accurate characterization of the physicochemical properties of each raw material, providing a data foundation for subsequent formulation design.

[0007] An initial formulation space is constructed based on the acquired basic physical property parameters, and a homogeneous design formulation set is generated based on this initial formulation space. Preferably, this process is implemented as follows: the hydration activity coefficient of cement is calculated based on the mass fraction of its oxide composition, and the filling efficiency coefficient of slag is calculated based on its specific surface area and average particle size; the dosage ranges of cement and slag are determined based on these hydration activity coefficients and filling efficiency coefficients; the dosage range of magnesium oxide is determined based on its loss on ignition and activity index, and the dosage range of admixtures is determined based on their functional group type and solid content; these dosage ranges are combined to construct a multidimensional initial formulation space; a homogeneous design algorithm is used to sample within this multidimensional initial formulation space, thereby generating a homogeneous design formulation set that can uniformly cover the entire space with fewer samples. This method can ensure the representativeness and comprehensiveness of the formulation combinations while reducing the number of experiments.

[0008] Hydration thermodynamics simulation is performed on each candidate formulation in the homogeneous design formulation set, and the hydration reaction process characteristics of each candidate formulation are output. In a preferred embodiment, the specific simulation process includes: calculating the initial content of each mineral phase (tricalcium silicate, dicalcium silicate, tricalcium aluminate, and tetracalcium aluminoferrite) in the cement based on the mass fraction of oxide composition in the cement of the candidate formulation; establishing a set of multiphase hydration reaction kinetic equations based on the initial content of these mineral phases and the activity index of magnesium oxide in the formulation; numerically solving the set of equations to obtain the hydration degree curves of each mineral phase over time; extracting the peak value of the hydration heat release rate and the total hydration heat release of each candidate formulation from these curves, and outputting these two as the hydration reaction process characteristics. This step can deeply reveal the heat release behavior during the hydration process of different formulations without conducting a large number of physical experiments.

[0009] Based on the output hydration reaction process characteristics, a preliminary subset of formulations meeting the hydration heat control requirements is selected. In one specific embodiment of the invention, the selection process includes: comparing the peak hydration heat release rate of each candidate formulation with a preset upper limit for the heat release rate, and eliminating formulations exceeding the limit; comparing the total hydration heat release of each candidate formulation with a preset upper limit for the total heat release, and similarly eliminating formulations exceeding the limit; furthermore, calculating the delayed expansion risk coefficient of each formulation based on the ratio of the activity index of magnesium oxide in each candidate formulation to the tricalcium aluminate content in the cement, and eliminating formulations whose risk coefficient exceeds a preset risk threshold; and forming a preliminary subset of formulations from the remaining candidate formulations after the above multiple eliminations. This process effectively avoids the cracking risk caused by concentrated release of hydration heat or delayed expansion, improving the reliability and safety of the preliminary formulations.

[0010] The initially selected formula subset is input into a trained graph convolutional neural network for formula-performance mapping, outputting a set of predicted performance parameters for each initially selected formula. Preferably, this mapping process includes: constructing a formula composition graph structure by using the mass fraction of each raw material in the initially selected formula as node features and the chemical compatibility coefficient between raw materials as edge features; inputting this graph structure into the graph convolutional neural network, where the network sequentially performs node feature aggregation and edge feature update operations, ultimately outputting a formula graph embedding vector containing overall formula information; inputting this formula graph embedding vector into a multilayer fully connected network for performance regression prediction, thereby outputting a set of predicted performance parameters including predicted compressive strength, predicted flexural strength, predicted chloride ion diffusion coefficient, and predicted drying shrinkage rate. Specifically, the chemical compatibility coefficient is calculated based on the differences in pH value and ion dissolution rate among the raw materials. Through the graph convolutional neural network, the complex interactions between raw materials can be accurately captured, achieving high-precision synergistic prediction of multiple performance characteristics.

[0011] A multi-objective formulation optimization model is constructed based on the predicted performance parameter set, and the model is solved using a multi-objective optimization algorithm to output a Pareto optimal formulation set. In a preferred implementation, the optimization objectives are minimizing cement usage, maximizing compressive strength, maximizing flexural strength, and minimizing chloride ion diffusion coefficient, with constraints on the usage of each raw material and process conditions. The multi-objective formulation optimization model is constructed by non-dominated ranking of the multiple optimization objectives in the model and calculating the crowding distance for each candidate formulation. Based on the non-dominated ranking results and crowding distance, an evolutionary algorithm with an elite retention strategy is used to iteratively solve the model. Finally, the non-dominated solution set obtained after iterative convergence is output as the Pareto optimal formulation set. This method can find the optimal balance point among multiple conflicting performance objectives, providing designers with a set of flexible optimized formulation schemes.

[0012] The process suitability score of each formulation in the Pareto optimal formulation set is evaluated, and a target optimized formulation is selected from the set based on the score. Specifically, the process suitability score evaluation process is as follows: the water-cement ratio and slurry yield stress of each formulation are obtained; the fluidity index is calculated based on the water-cement ratio, and the pumpability index is calculated based on the slurry yield stress; the initial setting time and final setting time of each formulation are obtained, and the operable time window is calculated accordingly; the total content of slag and magnesium oxide in each formulation is obtained, and the early strength development rate index is calculated accordingly; the above fluidity index, pumpability index, operable time window, and early strength development rate index are weighted and summed to obtain the process suitability score of each formulation. When selecting the target optimized formulation, all formulations are ranked from highest to lowest according to their process compatibility score, and the formulation ranked first is selected as the initial target formulation. The mass fractions of each raw material in the initial target formulation are then checked against process boundary constraints to confirm whether they all fall within the corresponding allowable process range. If the check passes, it is output as the target optimized formulation; if the check fails, the next highest-ranked formulation in the ranking is selected for check, and so on, until a successful check is performed before outputting the target optimized formulation. This method ensures that the selected formulation not only has excellent performance but also good process feasibility in actual production and construction.

[0013] As a further technical solution of the present invention, the method also includes a verification test on the target optimized formulation. Specifically, UHPC specimens are prepared by weighing, mixing, and molding according to the mass fraction of each raw material in the target optimized formulation; the specimens are subjected to standard curing, and mechanical and durability tests are conducted when the standard curing age is reached to obtain their actual compressive strength, actual flexural strength, and actual chloride ion diffusion coefficient; these actual performance parameters are compared with the corresponding predicted performance parameters to calculate the performance deviation rate; if the deviation rate is less than a preset deviation threshold, the target optimized formulation is marked as a verified formulation and stored in the formulation database. This step forms a feedback loop through a small number of physical experiments, verifying the accuracy of the optimization results and realizing the continuous accumulation of data and experience.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By constructing a set of multiphase hydration reaction kinetic equations and numerically solving each candidate formulation in a homogeneous design formulation set, the peak hydration exothermic rate and total hydration exothermic amount can be obtained without relying on physical experiments. By comparing these hydration reaction process characteristics with preset upper limits for the exothermic rate and total exothermic amount, and combining this with the delayed expansion risk coefficient calculated from the ratio of magnesium oxide activity index to tricalcium aluminate content, candidate formulations with potential hydration heat accumulation or volume stability issues can be systematically eliminated in the early stages of formulation generation. This pre-screening mechanism based on thermodynamics ensures that the initial subset of formulations entering the subsequent performance mapping stage is fundamentally feasible in terms of thermo-mechanical compatibility, compressing the optimization search space from the source and avoiding the consumption of computational and experimental resources on formulation combinations with intrinsic defects, thus controlling the time cost and material waste of the overall optimization process.

[0015] A subset of initially selected formulations, screened by mechanism, is mapped to performance using a graph convolutional neural network. The mass fraction of each raw material is used as node features, and the chemical compatibility coefficient is used as edge features to construct a graph structure of the formulation components. The chemical compatibility coefficient is calculated based on the differences in pH and ion dissolution rates among the raw materials, quantifying the interaction strength between components. The graph convolutional neural network extracts graph embedding vectors representing the global and local interaction information of the formulation through node feature aggregation and edge feature update operations. These vectors are then output as predicted values ​​for compressive strength, flexural strength, and chloride ion diffusion coefficient via a multi-layer fully connected network. This approach explicitly models the non-Euclidean spatial coupling relationship between formulation components using a graph structure, replacing the assumption of independent and identically distributed components in traditional methods. This reduces the prediction bias for durability indicators significantly affected by component interactions under the same scale of experimental sample training, thereby improving the fidelity of objective function evaluation in multi-objective optimization.

[0016] After obtaining the Pareto optimal formulation set, the water-cement ratio, slurry yield stress, initial setting time, final setting time, and total slag and magnesium oxide content of each formulation were quantitatively evaluated, and a weighted process suitability score was obtained. This score integrates the fluidity index, pumpability index, workable time window, and early strength development rate index, characterizing the engineering construction adaptability of the formulation from three dimensions: rheological property maintenance, on-site operation time margin, and early strength development capability. Based on this score, the Pareto front solution set was sorted and process boundary constraint verification was performed. The final output target optimized formulation has both theoretical optimal performance and on-site feasibility, eliminating the gap between traditional multi-objective output results and direct application in engineering environments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of an AI-based UHPC material formulation optimization method. Figure 2 This is a flowchart of the UHPC formulation optimization design based on the synergy of hydration activity and filling efficiency; Figure 3 This is a flowchart of a multi-objective formulation optimization and process adaptability scoring method; Figure 4 This is a flowchart of the target optimization formulation validation experiment; Figure 5 This is a comparison chart of the UHPC compressive strength prediction results based on graph convolutional neural networks. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 This invention provides an artificial intelligence-based method for optimizing UHPC material formulations, comprising: obtaining basic physical property parameters of UHPC raw materials, including cement, slag, magnesium oxide, and admixtures; constructing an initial formulation space based on the basic physical property parameters, and generating a uniform design formulation set based on the initial formulation space; performing hydration thermodynamic simulation on each candidate formulation in the uniform design formulation set, and outputting the hydration reaction process characteristics of each candidate formulation; selecting a subset of preliminary formulations that meet the hydration heat control requirements based on the hydration reaction process characteristics; inputting the subset of preliminary formulations into a trained graph convolutional neural network for formulation-performance mapping, and outputting a set of predicted performance parameters for each preliminary formulation; constructing a multi-objective formulation optimization model based on the set of predicted performance parameters, and solving the multi-objective formulation optimization model using a multi-objective optimization algorithm to output a Pareto optimal formulation set; performing process adaptability scoring on each formulation in the Pareto optimal formulation set, and selecting a target optimized formulation from the Pareto optimal formulation set based on the process adaptability score.

[0021] Example 1 In practice, the basic physical properties of UHPC raw materials are obtained, including cement, slag, magnesium oxide and admixtures.

[0022] X-ray fluorescence spectroscopy analysis was performed on cement raw materials. After drying the cement raw material powder to constant weight at 105℃, the sample to be tested was prepared by pressing. The sample to be tested was placed in the sample chamber of the X-ray fluorescence spectrometer, and a rhodium target X-ray tube was used as the excitation source. Full elemental scanning was performed under vacuum conditions, and the scanning range covered the characteristic spectral lines of elements from fluorine to uranium. Quantitative analysis was performed by the basic parameter method to obtain the mass fraction of oxide composition of cement raw materials. The mass fraction of oxide composition included the mass fraction of calcium oxide, silicon dioxide, aluminum oxide, and ferric oxide.

[0023] Laser particle size analysis was performed on slag raw materials. The slag raw materials were treated in an ultrasonic disperser for three minutes using anhydrous ethanol as the dispersion medium, and then injected into the sample introduction system of the laser particle size analyzer. The laser particle size analyzer used a helium-neon laser with a wavelength of 633 nm as the light source and employed the Mie scattering theory model for particle size inversion. The particle size distribution curve of the slag raw materials was obtained from the instrument output. The particle size distribution curve included the volume frequency distribution and cumulative distribution of particles in each size range. The specific surface area and average particle size of the slag raw materials were calculated based on the particle size distribution curve. The specific surface area was calculated by discretizing the particle size distribution curve into... The mass fraction of slag particles within each particle size range was obtained. and average particle size ,in The interval number, , The total number of granularity intervals. Take the first The arithmetic mean of the upper and lower limits of particle size in each interval; the true density of the slag raw material was determined by gas displacement method. The specific surface area of ​​slag raw materials is calculated using the following formula. :

[0024] in, The specific surface area of ​​slag raw materials is expressed in square centimeters per gram. The true density of slag raw material is expressed in grams per cubic centimeter and is determined under constant temperature conditions by the helium displacement specific gravity bottle method. Indicates the first The mass fraction of slag particles within a particle size range is obtained by multiplying the volume frequency distribution of that range by the effective density of the corresponding particles and then normalizing it. Indicates the first The average particle size of slag particles within a specific particle size range, expressed in micrometers. The average particle size of the slag raw material is directly calculated using the volume-weighted average particle size. It is obtained by accumulating the volume percentage in each interval and multiplying it by the average particle size of the corresponding interval. The calculation process does not introduce any additional formulas.

[0025] Thermogravimetric analysis (TGA) was performed on magnesium oxide raw materials. The magnesium oxide sample was placed in a platinum crucible of a TGA analyzer, and under a high-purity nitrogen atmosphere at a flow rate of 50 mL / min, the temperature was uniformly increased from 30°C to 1000°C at a rate of 10°C / min. The TGA curve of the magnesium oxide sample mass versus temperature was recorded simultaneously. The loss on ignition (LOI) of the magnesium oxide raw material was defined as the percentage of mass loss within the temperature range of 600°C to 900°C on the TGA curve. The difference between the initial and final mass of this range, expressed as a percentage of the initial mass of the magnesium oxide raw material, was taken as the LIO. The activity index of the magnesium oxide raw material was determined by the thermogravimetric step of magnesium hydroxide dehydroxylation on the TGA curve. Specifically, the thermogravimetric step at 350°C to 450°C, which corresponds to the decomposition of magnesium hydroxide, was identified on the TGA curve. The mass loss corresponding to this step was calculated, converted to the content of active magnesium oxide based on the stoichiometry of magnesium hydroxide, and the ratio of the active magnesium oxide content to the total mass of the magnesium oxide raw material was taken as the activity index of the magnesium oxide raw material.

[0026] Infrared spectroscopy analysis was performed on the admixture raw materials. The admixture stock solution was directly coated onto the crystal surface using an attenuated total reflectance attachment. The sample was scanned at four wavenumber resolutions on a Fourier transform infrared spectrometer, with a scanning range of 4000 to 400 wavenumbers, for a total of 32 scans to obtain the infrared absorption spectrum of the admixture raw materials. The infrared absorption spectrum was compared with a standard spectral library to identify the functional group types contained in the admixture raw materials. These functional groups included at least one or more of carboxyl, sulfonic acid, and hydroxyl groups. The solid content of the admixture raw materials was determined using a drying method. A certain mass of the liquid sample of the admixture raw materials was accurately weighed and placed in a weighing bottle. The sample was dried in a 105℃ forced-air drying oven until the difference between two consecutive weighings did not exceed 0.2 mg. The percentage of the residual solid mass after drying to the initial mass of the liquid sample of the admixture raw materials was calculated as the solid content of the admixture raw materials.

[0027] Example 2 In specific implementation, please refer to Figure 2 The hydration activity coefficient of cement is calculated based on the mass fraction of its oxide composition. The obtained mass fractions of calcium oxide, silicon dioxide, aluminum oxide, and ferric oxide in the cement are substituted into the Bogue formula to obtain the potential mass fraction of tricalcium silicate in the cement. Potential mass fraction of dicalcium silicate Potential mass fraction of tricalcium aluminate Potential mass fraction of tetracalcium aluminoferrite The hydration activity coefficient of cement Calculated using the following formula:

[0028] in, The hydration activity coefficient of cement is dimensionless and is used to reflect the strength of the overall hydration reaction trend of cement. The potential mass fraction of tricalcium silicate is expressed as a percentage. Indicates the potential mass fraction of dicalcium silicate; Indicates the potential mass fraction of tricalcium aluminate; Indicates the potential mass fraction of tetracalcium aluminoferrite; The value of the activity contribution coefficient corresponding to tricalcium silicate was determined by regression analysis based on pure mineral hydration heat test data and was set to 0.92. The activity contribution coefficient corresponding to dicalcium silicate is set to 0.38; The activity contribution coefficient corresponding to tricalcium aluminate was set to 0.24; The contribution factor to the activity of tetracalcium aluminoferrite was set to 0.11; , , , The setting is based on the relative activity ratio constant obtained by normalizing the cumulative heat release after isothermal calorimetry tests on each pure mineral phase under standard hydration conditions.

[0029] The slag filling efficiency coefficient is calculated based on the specific surface area and average particle size of the slag. In specific implementation, the specific surface area of ​​the slag measured in Example 1 is recorded as the slag specific surface area. The average particle size of slag is denoted as the average particle size of slag. Simultaneously, the specific surface area of ​​the cement used was measured according to the laser particle size analysis method in Example 1 and recorded as the cement specific surface area. The average particle size of cement is denoted as the cement average particle size. Slag filling efficiency coefficient By adjusting the average particle size of cement With average particle size of slag The ratio of slag specific surface area Specific surface area of ​​cement The square root of the product of the ratios is obtained.

[0030] The dosage ranges of cement and slag are determined based on the hydration activity coefficient of cement and the filling efficiency coefficient of slag. In specific implementation, upper and lower limits of the target total amount of cementitious materials for UHPC materials are set. The hydration activity coefficient of cement is used as a quantitative indicator of the cement's active contribution, and the filling efficiency coefficient of slag is used as a quantitative indicator of the physical filling effect of slag. The dosage ranges of cement and slag are allocated according to the principle of synergistic design between activity and filling: when the hydration activity coefficient of cement is high, a lower cement dosage range is allowed; when the filling efficiency coefficient of slag is high, a higher slag dosage range is allowed. The lower limit of the cement dosage range is set as the larger of the lower limit of the total amount of cementitious materials and the minimum cement dosage corresponding to the traditional water-cement ratio empirical range. The upper limit of the cement dosage range is linearly mapped to the range of 60% to 85% of the upper limit of the total amount of cementitious materials based on the hydration activity coefficient of cement. The slag dosage range is determined proportionally in the remaining space of cementitious materials based on the comparison between the slag filling efficiency coefficient and the preset filling efficiency threshold.

[0031] The dosage range of magnesium oxide is determined based on its loss on ignition and activity index. In practice, the activity index of magnesium oxide is used as an indicator to measure the contribution of its expansion effect to densification; the higher the activity index, the larger the dosage of magnesium oxide can be within the specified range. The loss on ignition of magnesium oxide is used as the basis for excluding ineffective components; when the loss on ignition exceeds one percent, the upper limit of magnesium oxide dosage is reduced by ten percent. The lower limit of the magnesium oxide dosage range is set to zero, and the upper limit is eight percent of the total amount of cementitious materials.

[0032] The dosage range of admixtures is determined based on their functional group type and solid content. In practice, the recommended dosage range of the admixture is determined according to the type of functional groups it contains. When the functional group type is mainly carboxyl, the corresponding recommended dosage range is 0.2% to 0.8% of the mass of the cementitious material. The recommended dosage range is then corrected based on the solid content of the admixture. Dividing the recommended dosage range by the solid content of the admixture yields the actual dosage range of the admixture.

[0033] The dosage ranges of cement, slag, magnesium oxide, and admixtures are combined to construct a multidimensional initial formulation space. In practice, the dosages of cement, slag, magnesium oxide, and admixtures are each treated as four independent dimensions. Each dimension is bounded by the upper and lower limits of its dosage range, forming a four-dimensional hypercube region, which constitutes the multidimensional initial formulation space.

[0034] A uniform design algorithm is used to sample within a multidimensional initial formulation space to generate a set of uniformly designed formulations. In the specific implementation, the total number of sampling points is set to a preset value. , The value is an integer between 50 and 200, and was determined to be 60 after weighing the formulation space dimension and experimental resources. A uniform design table was selected. As a sampling framework, according to the level allocation of each factor in the uniform design table, the cement dosage dimension, slag dosage dimension, magnesium oxide dosage dimension and admixture dosage dimension are each divided into 60 levels. The four-dimensional coordinates corresponding to each experimental point are obtained by looking up the table, and each set of four-dimensional coordinates is converted into a candidate formula. All 60 candidate formulas constitute the uniform design formula set.

[0035] Hydration thermodynamics simulation was performed on each candidate formulation in the homogeneous design formulation set, and the hydration reaction process characteristics of each candidate formulation were output. Based on the mass fraction of cement oxide composition in the candidate formulations, the initial content of each mineral phase in the cement was calculated. The mineral phases included tricalcium silicate, dicalcium silicate, tricalcium aluminate, and tetracalcium aluminoferrite. In practice, the mass fraction of cement in the candidate formulations was multiplied by the mass fraction of cement oxide composition to obtain the mass of each oxide per unit volume of the cementitious system. Then, the Bogue formula was used to calculate the initial content of tricalcium silicate, dicalcium silicate, tricalcium aluminate, and tetracalcium aluminoferrite per unit volume of the cementitious system.

[0036] Based on the initial content of each mineral phase and the activity index of magnesium oxide in the candidate formulations, a set of multiphase hydration reaction kinetic equations was established. This set of equations includes differential equations for the hydration rates of tricalcium silicate, dicalcium silicate, tricalcium aluminate, tetracalcium aluminoferrite, and magnesium oxide. Each differential equation adopts a kinetic function form derived from a boundary nucleation and growth model, where the reaction rate constant is a function of temperature and material fineness. When establishing the equation set, the activity index of magnesium oxide was used as a correction factor for the magnesium oxide hydration reaction rate constant; a higher activity index indicates an earlier onset of the accelerated hydration period.

[0037] Numerical solutions were used to solve the kinetic equations for multiphase hydration reactions, yielding hydration degree curves for each mineral phase over time. In the specific implementation, the RK45 method with an adaptive step size was employed to integrate the kinetic equations. The initial integration time step was set to 0.01 hours, with a maximum step size limit of 0.5 hours, and a total integration time of 168 hours. During the solution process, the hydration degree of each mineral phase was used as the state variable, outputting the hydration degree values ​​for tricalcium silicate, dicalcium silicate, tricalcium aluminate, tetracalcium aluminoferrite, and magnesium oxide at each time step, thus generating hydration degree curves for each mineral phase.

[0038] The peak hydration exothermic rate and total hydration exothermic amount of each candidate formulation were extracted based on the hydration degree curves, and these values ​​were output as characteristics of the hydration reaction process. In practice, the hydration degree curves of each mineral phase were converted into hydration exothermic rate and total hydration exothermic amount curves using the standard hydration calorific values ​​of each mineral phase. The standard hydration calorific values ​​for tricalcium silicate were 517 kJ / kg, dicalcium silicate 262 kJ / kg, tricalcium aluminate 1144 kJ / kg, tetracalcium aluminoferrite 418 kJ / kg, and magnesium oxide 805 kJ / kg. The hydration exothermic rate curves were obtained by time-by-time difference analysis of the total hydration exothermic amount curve. The maximum exothermic rate value within 168 hours was searched on the hydration exothermic rate curve as the peak hydration exothermic rate; the total hydration exothermic amount was the cumulative heat release value over 168 hours. The peak hydration heat release rate and the total hydration heat release are used as the output characteristics of the hydration reaction process for each candidate formulation.

[0039] Based on the characteristics of the hydration reaction process, a preliminary subset of formulations meeting the requirements for hydration heat control was selected. The peak hydration heat release rate of each candidate formulation was compared with a preset upper limit for the heat release rate, and candidate formulations with peak heat release rates exceeding the upper limit were eliminated. The preset upper limit for the heat release rate was determined based on the minimum cross-sectional size of the UHPC component and the template type; for components with a thickness less than 300 mm, the upper limit was set to 8 kJ / kg / hour. The peak hydration heat release rate of each candidate formulation was then compared one by one with 8 kJ / kg / hour, and candidate formulations with peak hydration heat release rates exceeding 8 kJ / kg / hour were removed from the set.

[0040] The total heat release from hydration of each candidate formulation is compared with a preset upper limit for total heat release, and candidate formulations with a total heat release exceeding the upper limit are eliminated. The preset upper limit for total heat release is determined based on the concrete adiabatic temperature rise control target, which is set at 55 degrees Celsius. Based on the average heat capacity of the cementitious materials, the upper limit for total heat release is calculated to be 330 kJ / kg. The total heat release from hydration of the remaining candidate formulations after the previous elimination step is then compared one by one with 330 kJ / kg, and candidate formulations with a total heat release from hydration exceeding 330 kJ / kg are deleted.

[0041] Based on the ratio of the activity index of magnesium oxide in each candidate formulation to the content of tricalcium aluminate in the cement, the delayed expansion risk coefficient of each candidate formulation is calculated, and candidate formulations with a delayed expansion risk coefficient exceeding a preset risk threshold are eliminated. In specific implementation, for each remaining candidate formulation, the activity index of magnesium oxide in that candidate formulation is extracted and recorded as the magnesium oxide activity index. The initial content of tricalcium aluminate per unit volume of the gelation system of this candidate formulation is recorded as the tricalcium aluminate content. The delayed expansion risk coefficient is determined by the magnesium oxide activity index. Divide by tricalcium aluminate content Multiplying this by a correction factor of 0.5 yields the result. The correction factor of 0.5 is based on a semi-quantitative assessment of the inhibitory effect of tricalcium aluminate on magnesium oxide expansion. The preset risk threshold is set to 2.0. When the delayed expansion risk coefficient exceeds 2.0, the candidate formulation is eliminated.

[0042] The remaining candidate formulas after the above three eliminations will form the initial selection formula subset.

[0043] Example 3 In specific implementation, the subset of preliminary formulas obtained in Example 2 is input into the trained graph convolutional neural network for formula-performance mapping, and the set of predicted performance parameters for each preliminary formula is output.

[0044] The mass fraction of each raw material in the initial formula is used as the node feature, and the chemical compatibility coefficient between the raw materials is used as the edge feature to construct the formula composition graph structure. In the specific implementation, the initial formula includes four raw materials: cement, slag, magnesium oxide, and admixtures. The formula composition graph structure contains four nodes, corresponding to cement, slag, magnesium oxide, and admixtures, respectively. The node feature of the cement node is the mass fraction of cement in the initial formula; the node feature of the slag node is the mass fraction of slag in the initial formula; the node feature of the magnesium oxide node is the mass fraction of magnesium oxide in the initial formula; and the node feature of the admixture node is the mass fraction of admixtures in the initial formula. The formula composition graph structure is a fully connected graph, where there is an edge between any two different nodes, and the edge feature is the chemical compatibility coefficient between the two raw materials connecting the edge.

[0045] The chemical compatibility coefficient is calculated based on the differences in pH value and ion dissolution rate among the raw materials. In practice, the pH value and ion dissolution rate of the cement, slag, magnesium oxide, and admixture slurries are measured separately. The method for measuring the pH value of the slurry is as follows: each raw material is mixed with water at a mass ratio of 1:10, stirred, and allowed to stand for 24 hours. The pH value of the supernatant is measured using a pH meter to obtain the pH values ​​of the cement slurry, slag slurry, magnesium oxide slurry, and admixture slurry. The method for measuring the ion dissolution rate is as follows: each raw material is mixed with water at a mass ratio of 1:100, and the conductivity is continuously recorded over time using a conductivity meter under stirring conditions. The slope of the conductivity rise curve in the first 10 minutes is taken as the ion dissolution rate to obtain the ion dissolution rates of cement, slag, magnesium oxide, and admixtures.

[0046] For connection nodes in the formula composition diagram structure and nodes The chemical compatibility coefficient is calculated using the following formula:

[0047] in, Indicates the connection node With nodes The chemical compatibility coefficient corresponding to the edge is dimensionless and ranges from greater than 0 to less than or equal to 1. Represents a node The corresponding pH value of the raw material slurry; Represents a node The corresponding pH value of the raw material slurry; Represents a node The corresponding ion dissolution rate of the raw material, in microsiemens per centimeter per minute; Represents a node The ion dissolution rate of the corresponding raw material; This is a pH-scale parameter with a value of 1.5. The ion dissolution rate is a scale parameter, with a value of 10 microsiemens per centimeter per minute; The value of 1.5 is based on the fact that the typical standard deviation of pH difference among different components in a cement-based material system is about 1.5 pH units. Using this standard deviation as a scale parameter can effectively distinguish compatibility differences. The value of 10 micro-Siemens per centimeter per minute is based on statistical analysis of the differences in ion dissolution rates in a large number of slag-cement systems. 10 micro-Siemens per centimeter per minute can cover the range of activity differences among common cementitious materials. The chemical compatibility coefficients corresponding to each of the six edges are calculated using the above formula, serving as the edge features of the formulation composition diagram structure.

[0048] The formulation composition map structure is input into a graph convolutional neural network (Graph Convolutional Neural Network). The Graph Convolutional Neural Network sequentially performs node feature aggregation and edge feature update operations, outputting a formulation map embedding vector. The core architecture of the Graph Convolutional Neural Network consists of two graph convolutional layers and one graph readout layer. The first graph convolutional layer receives the node feature matrix and edge feature matrix of the formulation composition map structure as input. The node feature matrix has a dimension of 4×1, and the edge feature matrix has a dimension of 4×4. In the first graph convolutional layer, for any target node, the node feature aggregation operation multiplies the node features of all first-order neighbor nodes of the target node with the chemical compatibility coefficients on the corresponding edges, sums them, and then concatenates the aggregation result with the target node's own features. A linear transformation is performed through a learnable weight matrix and bias vector, and then processed by the ReLU activation function to obtain the node embedding vector output by the first graph convolutional layer. The dimension of the node embedding vector is set to 16. The second graph convolutional layer uses the node embedding vector output by the first graph convolutional layer as the new node feature, and processes it using the same message passing and aggregation method. In the aggregation operation, the chemical compatibility coefficient on each edge is reused. The dimension of the node embedding vector output by the second graph convolutional layer is set to 32. Edge feature update operations are performed simultaneously: during aggregation, the second graph convolutional layer performs a Hadamard product between the chemical compatibility coefficient on the edge and the L2 norm of the sending node feature vector, completing the edge feature update. The updated edge features are used in the next round of aggregation and are not output within the graph convolutional neural network. After two layers of graph convolution processing, each node obtains a node embedding vector of dimension 32. The graph readout layer uses average pooling to calculate the element-wise average of the node embedding vectors of the four nodes, resulting in a 32-dimensional graph embedding vector.

[0049] The training process of the graph convolutional neural network is as follows: At least 200 sets of historical experimental data for UHPC formulations are collected. Each set of historical experimental data includes the mass fraction of each raw material, as well as the corresponding measured values ​​of compressive strength, flexural strength, chloride ion diffusion coefficient, and drying shrinkage rate. The historical experimental data are randomly divided into training and validation sets at a ratio of 4:1. For each set of formulation data in the training set, a formulation composition diagram structure is constructed according to the method described above, and the chemical compatibility coefficient is calculated and used as input. The four corresponding measured performance values ​​are combined into a 4-dimensional label vector. All learnable parameters of the graph convolutional neural network are initialized using Xavier, the loss function is the mean squared error loss function, the optimizer is the Adam optimizer, the initial learning rate is set to 0.001, the batch size is set to 32, and the training epochs are set to 500. In each training round, forward propagation calculates the set of predicted performance parameters, and the mean squared error is calculated between the set of predicted performance parameters and the label vector. Backpropagation then updates the network parameters. After each complete training round, the validation loss is calculated on the validation set. Training is terminated early when the validation loss stops decreasing after 50 consecutive rounds. After training, the weight parameters of the graph convolutional neural network are saved.

[0050] The formulation diagram is embedded into a vector and input into a multilayer fully connected network for performance regression prediction, outputting a set of predicted performance parameters for each initially selected formulation. The multilayer fully connected network contains three hidden layers: the first hidden layer contains 64 neurons, the second contains 32 neurons, and the third contains 16 neurons. Each hidden layer uses the ReLU activation function. The output layer contains 4 neurons and does not use an activation function. The multilayer fully connected network and a graph convolutional neural network are concatenated and trained together, with parameter settings as described above. After the formulation diagram is embedded into the multilayer fully connected network, the output values ​​of the four neurons in the output layer correspond to the predicted values ​​of compressive strength, flexural strength, chloride ion diffusion coefficient, and drying shrinkage rate, respectively. These four predicted values ​​together constitute the set of predicted performance parameters for the initially selected formulation. The units for the predicted compressive strength and flexural strength are megapascals (MPA), the unit for the chloride ion diffusion coefficient is square millimeters per year, and the unit for the drying shrinkage rate is microstrain.

[0051] See Figure 5 In the figure, the horizontal axis represents the measured compressive strength (in MPa) of the UHPC specimen, and the vertical axis represents the predicted compressive strength (in MPa) of the corresponding formulation. In the scatter plot, hollow circles represent training set sample data, solid triangles represent validation set sample data, and dashed lines represent… Reference line. Overall, the predicted values ​​of both the training and validation sets cluster well. Near the reference line, the joint model of graph convolutional neural network and multilayer fully connected network demonstrates high accuracy and generalization ability in predicting the compressive strength of UHPC. The predicted compressive strength shows a clear linear positive correlation with the increase of the measured compressive strength. Within the measured strength range of 120 MPa to 200 MPa, the model's prediction error is small, with no obvious systematic bias. The distribution of predicted points on the validation set is similar to that on the training set, verifying the model's stability on unseen data.

[0052] Example 4 In specific implementation, please refer to Figure 3 Based on the set of predicted performance parameters of each preliminary formulation obtained in Example 3, a multi-objective formulation optimization model is constructed, and the multi-objective optimization algorithm is used to solve the multi-objective formulation optimization model to output the Pareto optimal formulation set.

[0053] A multi-objective formulation optimization model is constructed with the optimization objectives of minimizing cement usage, maximizing compressive strength, maximizing flexural strength, and minimizing chloride ion diffusion coefficient, and with constraints on the usage of each raw material and process conditions. The objective function for minimizing cement usage is defined as follows: ,in Represents a vector of decision variables. , This represents the cement mass fraction. The mass fraction of slag. This represents the mass fraction of magnesium oxide. The admixtures represent their mass fractions, each expressed as a percentage of the total mass of the cementitious material. The objective function for maximizing compressive strength is defined as follows: , This represents the vector of decision variables. The predicted compressive strength output by the graph convolutional neural network in Example 3 is negativeened to transform maximization into minimization. The objective function for maximizing flexural strength is defined as follows: , This is the predicted flexural strength. The objective function for minimizing the chloride ion diffusion coefficient is defined as follows: , This is the predicted value for the chloride ion diffusion coefficient.

[0054] The constraints on the quantity of each raw material in the constraints include: cement mass fraction. The lower limit is the lower limit of the cement dosage range determined in Example 2, and the upper limit is the upper limit of the cement dosage range determined in Example 2; slag mass fraction The lower limit is the lower limit of the slag dosage range determined in Example 2, and the upper limit is the upper limit of the slag dosage range determined in Example 2; magnesium oxide mass fraction The lower limit is the lower limit of the magnesium oxide dosage range determined in Example 2, and the upper limit is the upper limit of the magnesium oxide dosage range determined in Example 2; admixture mass fraction The lower limit is the lower limit of the admixture dosage range determined in Example 2, and the upper limit is the upper limit of the admixture dosage range determined in Example 2. Process constraints include: total amount of cementitious materials constraint, i.e. The sum equals 100; water-cement ratio constraint, the lower limit of water-cement ratio is set to 0.14 and the upper limit of water-cement ratio is set to 0.22. The water-cement ratio is defined as the ratio of the mass of mixing water to the total mass of cementitious materials. In the optimization process, it is added as an independent variable to the decision variable vector to participate in constraint verification; slurry yield stress constraint, the slurry yield stress is set not to exceed 80 Pascals.

[0055] In a multi-objective formulation optimization model, multiple optimization objectives are ranked non-dominated, and the crowding distance for each candidate formulation is calculated. In the specific implementation, an evolutionary algorithm population is used as the candidate formulation set, with a population size of 100 individuals, each corresponding to a set of decision variable vectors. Four objective function values ​​are calculated for each individual in the population, forming an objective function value matrix. The non-dominated ranking process is as follows: traverse the population, find all individuals not dominated by other individuals, and define them as the first non-dominated layer; remove individuals from the first non-dominated layer from the population, and repeat the search for non-dominated individuals among the remaining individuals, defining them as the second non-dominated layer; repeat this process until all individuals are assigned to their corresponding non-dominated layers. The rule for determining the dominance relationship between two individuals is: when an individual... All objective function values ​​are less than or equal to the individual The corresponding objective function value, and at least one objective function value is strictly less than the individual At that time, determine the individual Dominant Individual Otherwise, the individual is judged. With individuals They do not control each other.

[0056] Crowding distance is calculated within the same non-dominated layer. For each objective function, individuals within that non-dominated layer are sorted in ascending order of objective function value. The crowding distances of the first and last individuals are set to infinity. The crowding distance components of the remaining individuals on that objective function are equal to the absolute value of the difference between the corresponding objective function values ​​of two adjacent individuals divided by the range of that objective function within the current non-dominated layer. The crowding distance components of an individual on all objective functions are summed to obtain the individual's crowding distance.

[0057] Based on the results of non-dominated sorting and crowding distance, an evolutionary algorithm with an elitist retention strategy is used to iteratively solve the multi-objective recipe optimization model. The evolutionary algorithm with elitist retention is a non-dominated sorting genetic algorithm with an elitist strategy. Its core architecture includes selection, crossover, mutation, and elitist retention operations. In each iteration, a mating pool is generated from the current parent population through a binary tournament selection. The binary tournament selection uses non-dominated level and crowding distance as comparison criteria, prioritizing those with lower non-dominated levels, and prioritizing those with larger crowding distances when levels are the same. Simulated binary crossover is performed on individuals in the mating pool, with a crossover probability of 0.9 and a crossover distribution exponent of 20, generating offspring individuals. Multinomial mutation is then performed on the offspring individuals, with the mutation probability set to the reciprocal of the decision variable dimension and the mutation distribution exponent set to 20. The parent and offspring populations are merged to form a mixed population of 200. The mixed population is then re-sorted for non-dominated hierarchy and crowding distance is recalculated. Based on the non-dominated hierarchy from low to high and the crowding distance within the same hierarchy from large to small, the top 100 individuals are selected to enter the next generation. This process is repeated 300 times. After the iterations, all individuals in the first non-dominated hierarchy of the final population are taken as the non-dominated solution set, which is then output as the Pareto optimal recipe set.

[0058] Process adaptability scores were performed on each formulation in the Pareto optimal formulation set. The water-cement ratio and slurry yield stress of each formulation were obtained. The flowability index of each formulation was calculated based on the water-cement ratio, and the pumpability index was calculated based on the slurry yield stress. The water-cement ratio was obtained by reading the water-cement ratio value from the corresponding decision variable vector for each formulation in the Pareto optimal formulation set. The slurry yield stress was obtained by using the slurry yield stress prediction branch, which was synchronously established during the training of the convolutional neural network shown in Example 3. This prediction branch added an extra neuron to the output layer of the multilayer fully connected network. During training, the same dataset and training configuration as those used for compressive strength and other properties were employed, and the predicted slurry yield stress value was output.

[0059] Liquidity Index The calculation formula is:

[0060] in, This represents a liquidity index, which is dimensionless and its value range is constrained by normalization to 0 to 1. ; The water-cement ratio corresponding to the formulation is given by the formulation decision variables; This represents the normalized lower limit of the water-to-binder ratio, with a value of 0.14, which is the lower limit of the water-to-binder ratio constraint. This represents the normalized upper limit of the water-to-binder ratio, with a value of 0.22, which is the upper limit of the water-to-binder ratio constraint. The fluidity index weighting factor is set to 1.0, based on the fact that the fluidity index participates with equal weight to other indices in the process adaptability score. A higher fluidity index value indicates better slurry fluidity.

[0061] Pumpability Index The pumpability index is obtained as follows: The predicted yield stress of the slurry is used as input. When the predicted yield stress is less than 20 Pascals, the pumpability index is set to a full score of 1.0; when the predicted yield stress is greater than 80 Pascals, the pumpability index is set to 0; when the predicted yield stress is between 20 Pascals and 80 Pascals, the pumpability index is calculated using linear interpolation, and the calculation formula is as follows: ,in This is the predicted yield stress of the slurry, in Pascals.

[0062] Obtain the initial setting time and final setting time for each formulation, and calculate the workable time window for each formulation based on the initial setting time and final setting time. The initial setting time and final setting time are determined using the penetration resistance test method, according to the method specified in GB / T50080, during the pre-mixing stage before the verification test, or using the setting time prediction value additionally output by the convolutional neural network in Example 3. The setting time prediction branch is trained in a similar manner to the slurry yield stress prediction branch, outputting the initial setting time prediction value. and predicted final setting time All units are in hours. Operable time window. The calculation formula is This calculation formula represents subtracting 30% of the difference between the initial setting time and the final setting time from the initial setting time to conservatively estimate the effective time period during which pouring operations can actually be carried out.

[0063] Obtain the total content of slag and magnesium oxide in each formulation, and calculate the early strength development rate index of each formulation based on the total content. Total content of slag and magnesium oxide The mass fraction of slag in the formula With magnesium oxide mass fraction The sum. Early intensity development rate index. Defined as , The upper limit of the total content of slag and magnesium oxide is empirically set at 25. This is based on the fact that when the total content of slag and magnesium oxide exceeds 25% of the total cementitious material content, the early strength development rate usually decreases significantly; therefore, 25 is used as a normalized reference value. hour Take 0.

[0064] The process suitability score for each formulation is obtained by weighting and summing the flowability index, pumpability index, operable time window, and early strength development rate index. In practical implementation, the operable time window... Normalization is performed to normalize the operable time window index. , The 4-hour timeframe is chosen based on the general engineering requirement that the operable time window should not be less than 4 hours. Early strength development rate index. Directly adopted. Process adaptability score. The calculation formula is:

[0065] in, The rating for process adaptability is dimensionless. Indicates the liquidity index; Indicates the pumpability index; This represents the normalized operable time window index; Indicating the early intensity development rate index; , , , The weighting coefficients are 0.3, 0.25, 0.25, and 0.2, respectively. The weighting is based on the fact that fluidity has the greatest impact on pumping and filling during construction, so the highest weight of 0.3 is taken. Pumpability and operating time window are equally important to ensure continuous construction, so each is taken as 0.25. Early strength development rate mainly affects the connection of subsequent processes, so it is taken as 0.2.

[0066] The target optimized formulation is selected from the Pareto optimal formulation set based on the process suitability score. The formulations in the Pareto optimal formulation set are sorted by process suitability score from highest to lowest, and the formulation with the highest score is selected as the initial target formulation. The mass fractions of each raw material in the initial target formulation are checked against process boundary constraints to confirm that the mass fractions of each raw material fall within the corresponding allowable process range. The allowable process range is determined based on the technical specifications of the specific construction equipment; for example, the allowable upper limit of slurry yield stress for pumping equipment is 75 Pascals, the allowable water-cement ratio deviation range for mixing systems is ±0.01, and the allowable weighing accuracy of admixtures for on-site metering systems is ±0.1%. The parameters corresponding to the initial target formulation are compared item by item with the allowable process range. If all parameters are within the allowable process range, the check passes, and the initial target formulation is output as the target optimized formulation. If the check fails, the next highest-scoring formulation in the ranking is selected for process boundary constraint check until the check passes, at which point the target optimized formulation is output.

[0067] Example 5 In specific implementation, please refer to Figure 4Verification experiments were conducted on the target optimized formula. UHPC specimens were prepared by weighing, mixing, and molding the raw materials according to their mass fractions in the target optimized formula. Weighing was performed using an electronic balance with an accuracy of 0.01 grams. The mass fractions of cement, slag, magnesium oxide, and admixtures were converted to their respective mass fractions based on the total amount mixed in a single batch. The mass of mixing water was also weighed according to the predetermined water-cement ratio. The mixing operation uses a 10-liter planetary forced mixer. First, the weighed cement, slag, and magnesium oxide are added to the mixing pot and dry-mixed at low speed for 120 seconds. The low-speed mixing speed is set to 140 rpm (rotation speed) and 62 rpm (revolution speed). Half of the weighed admixture and half of the mixing water are pre-mixed evenly and then slowly added to the mixing pot after the dry mixing is complete. Low-speed mixing continues for 90 seconds. Then, the remaining mixing water is added, and high-speed mixing is initiated for 120 seconds. The high-speed mixing speed is set to 285 rpm (rotation speed) and 125 rpm (revolution speed). After stopping, the adhering material to the mixing blades and pot walls is scraped off, and high-speed mixing is resumed for another 60 seconds. The molding operation uses a 40 mm × 40 mm × 160 mm triple steel mold. The mixed slurry is poured into the mold in two layers. After each layer is poured in, it is vibrated for 30 seconds on a vibrating table at a frequency of 50 Hz and an amplitude of 0.5 mm. The surface is then smoothed and covered with a polyethylene film.

[0068] The molded UHPC specimens were subjected to standard curing. After molding, the specimens were placed in a standard curing room at a temperature of 20±2 degrees Celsius and a relative humidity of not less than 95% for 24 hours while still in the mold. After demolding, the specimens were further immersed in a saturated calcium hydroxide aqueous solution at 20±1 degrees Celsius for curing until the standard curing age of 28 days. Upon reaching the standard curing age, the specimens were removed, the surface moisture was wiped dry, and mechanical property and durability tests were immediately conducted.

[0069] Mechanical property tests include flexural strength and compressive strength tests. The flexural strength test employs the three-point bending method on a universal testing machine with a support span of 100 mm and a loading rate of 0.05 MPa per second. The maximum load at which the specimen breaks is recorded, and the actual flexural strength is calculated using a formula. The arithmetic mean of the actual flexural strengths of the three specimens is taken as the actual flexural strength of the UHPC specimen. The compressive strength test is conducted on six half-section specimens after the flexural test, with a compression surface of 40 mm × 40 mm and a loading rate of 1.2 MPa per second. The failure load is recorded, and the actual compressive strength of each section is calculated. The arithmetic mean of the six data points is taken as the actual compressive strength of the UHPC specimen.

[0070] Durability testing was performed using the chloride ion diffusion coefficient method, following the NTBuild 492 standard. Three days before the standard curing age, specimens were removed from the curing solution, machined into cylindrical specimens with a diameter of 100 mm and a thickness of 50 mm, vacuum-saturated, and then installed in a migration tank. The anode tank was filled with 0.3 mol / L sodium hydroxide solution, and the cathode tank with 10% sodium chloride solution. A 30V external voltage was applied for 24 hours. After the test, the specimens were split and sprayed with 0.1 mol / L silver nitrate solution. The chloride ion penetration depth was measured, and the actual chloride ion diffusion coefficient of the UHPC specimen was calculated.

[0071] The actual compressive strength, actual flexural strength, and actual chloride ion diffusion coefficient are compared with the corresponding predicted performance parameters to calculate the performance deviation rate. The predicted performance parameters are provided by the set of predicted performance parameters corresponding to the target optimized formulation output in Example 4, including predicted values ​​for compressive strength, flexural strength, and chloride ion diffusion coefficient. The formula for calculating the performance deviation rate is:

[0072] in, This indicates the performance deviation rate, expressed as a percentage. This indicates the actual compressive strength of the UHPC specimen, expressed in megapascals (MPa), obtained through mechanical property testing. This represents the predicted compressive strength value corresponding to the target optimized formulation, in megapascals (MPa). This indicates the actual flexural strength of the UHPC specimen, expressed in megapascals (MPa). This represents the predicted flexural strength corresponding to the target optimized formulation, in megapascals (MPa). The actual chloride ion diffusion coefficient of the UHPC specimen is expressed in square millimeters per year. This represents the predicted chloride ion diffusion coefficient corresponding to the target optimized formulation, expressed in square millimeters per year.

[0073] The calculated performance deviation rate is compared with a preset deviation threshold. The preset deviation threshold is set at 5%, based on the allowable deviation for material performance testing in the "Technical Specification for Ultra-High Performance Concrete Structures" and the engineering application requirement that the model prediction accuracy in the formulation design stage should be controlled within ±5%. If the performance deviation rate is less than 5%, the target optimized formulation is marked as a validated formulation and stored in the formulation database. The information stored in the formulation database includes the cement mass fraction, slag mass fraction, magnesium oxide mass fraction, admixture mass fraction, predicted compressive strength, predicted flexural strength, predicted chloride ion diffusion coefficient, predicted drying shrinkage rate, actual compressive strength, actual flexural strength, actual chloride ion diffusion coefficient, performance deviation rate, and process compatibility score corresponding to the validated formulation. If the performance deviation rate is greater than or equal to 5%, the target optimized formulation is marked as an unvalidated formulation and returned to the Pareto optimal formulation set in Example 4. The next formulation is reselected based on the second-order ranking of the process compatibility score to perform the validation test.

[0074] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for optimizing UHPC material formulation based on artificial intelligence, characterized in that, include: Obtain the basic physical properties of UHPC raw materials, including cement, slag, magnesium oxide, and additives; The basic physical properties include: X-ray fluorescence spectroscopy analysis of cement raw materials to obtain the mass fraction of oxide composition, which includes the mass fraction of calcium oxide, silicon dioxide, aluminum oxide, and ferric oxide; laser particle size analysis of slag raw materials to obtain the particle size distribution curve of slag, and calculation of the specific surface area and average particle size of slag based on the particle size distribution curve; thermogravimetric analysis of magnesium oxide raw materials to obtain the loss on ignition and activity index of magnesium oxide; and infrared spectroscopy analysis of admixture raw materials to obtain the functional group type and solid content of admixtures. An initial formulation space is constructed based on the aforementioned basic physical property parameters, and a uniform design formulation set is generated based on the initial formulation space. For each candidate formulation in the homogeneous design formulation set, a hydration thermodynamic simulation is performed, and the hydration reaction process characteristics of each candidate formulation are output, including: Based on the mass fraction of oxide composition of cement in the candidate formula, the initial content of each mineral phase of cement is calculated, and the mineral phases include tricalcium silicate, dicalcium silicate, tricalcium aluminate and tetracalcium aluminoferrite. Based on the initial content of each mineral phase and the activity index of magnesium oxide in the candidate formulation, a set of kinetic equations for multiphase hydration reaction was established. Numerical solution was performed on the set of multiphase hydration reaction kinetic equations to obtain the hydration degree curves of each mineral phase over time. The peak value of the hydration heat release rate and the total hydration heat release of each candidate formulation are extracted based on the hydration degree curve, and the peak value of the hydration heat release rate and the total hydration heat release are output as characteristics of the hydration reaction process. Based on the characteristics of the hydration reaction process, a preliminary subset of formulations meeting the requirements for hydration heat control was selected, including: Based on the comparison between the peak hydration exothermic rate of each candidate formulation and the preset upper limit of the exothermic rate, candidate formulations with a peak exothermic rate greater than the upper limit of the exothermic rate are eliminated. Based on the comparison between the total heat release from hydration of each candidate formulation and the preset upper limit of total heat release, candidate formulations with a total heat release exceeding the upper limit of total heat release are eliminated. Based on the ratio of the activity index of magnesium oxide in each candidate formulation to the content of tricalcium aluminate in cement, the delayed expansion risk coefficient of each candidate formulation is calculated, and candidate formulations with delayed expansion risk coefficients exceeding a preset risk threshold are eliminated. The remaining candidate recipes after elimination are used to form a preliminary selection subset of recipes; The subset of preliminary recipes is input into a trained graph convolutional neural network for recipe-performance mapping, outputting a set of predicted performance parameters for each preliminary recipe, including: The mass fraction of each raw material in the initial formula is used as the node feature, and the chemical compatibility coefficient between the raw materials is used as the edge feature to construct the formula composition graph structure. The formula composition graph structure is input into a graph convolutional neural network, which sequentially performs node feature aggregation and edge feature update operations to output a formula graph embedding vector. The formula diagram is embedded into a vector input to a multilayer fully connected network for performance regression prediction, and the predicted performance parameter set of each initially selected formula is output. The set of predicted performance parameters includes predicted values ​​for compressive strength, flexural strength, chloride ion diffusion coefficient, and drying shrinkage rate. The chemical compatibility coefficient is calculated based on the differences in pH value and ion dissolution rate among the raw materials; A multi-objective formulation optimization model is constructed based on the predicted performance parameter set, and the multi-objective formulation optimization model is solved by a multi-objective optimization algorithm to output the Pareto optimal formulation set. The multi-objective formulation optimization model takes minimizing cement dosage, maximizing compressive strength, maximizing flexural strength and minimizing chloride ion diffusion coefficient as optimization objectives, and uses the dosage constraints of each raw material and process constraints as constraints. The process suitability score of each formulation in the Pareto optimal formulation set is performed, and the target optimized formulation is selected from the Pareto optimal formulation set based on the process suitability score.

2. The UHPC material formulation optimization method based on artificial intelligence according to claim 1, characterized in that, An initial formulation space is constructed based on the aforementioned basic physical property parameters, and a uniform design formulation set is generated based on the initial formulation space, including: The hydration activity coefficient of cement is calculated based on the mass fraction of oxides in cement, and the filling efficiency coefficient of slag is calculated based on the specific surface area and average particle size of slag. The dosage range of cement and the dosage range of slag are determined based on the hydration activity coefficient and the filling efficiency coefficient. The dosage range of magnesium oxide is determined based on its loss on ignition and activity index, and the dosage range of the additive is determined based on the functional group type and solid content of the additive. By combining the dosage ranges of cement, slag, magnesium oxide, and admixtures, a multi-dimensional initial formulation space is constructed. A uniform design algorithm is used to sample within the multidimensional initial formulation space to generate a set of uniform design formulations.

3. The UHPC material formulation optimization method based on artificial intelligence according to claim 1, characterized in that, A multi-objective formulation optimization model is constructed based on the predicted performance parameter set, and the multi-objective optimization model is solved using a multi-objective optimization algorithm to output a Pareto optimal formulation set, including: The multiple optimization objectives in the multi-objective formulation optimization model are non-dominated and ordered, and the crowding distance of each candidate formulation is calculated. Based on the results of the non-dominated sorting and the crowding distance, an evolutionary algorithm with an elite retention strategy is used to iteratively solve the multi-objective recipe optimization model. The non-dominated solution set obtained after iterative convergence is output as the Pareto optimal recipe set.

4. The UHPC material formulation optimization method based on artificial intelligence according to claim 1, characterized in that, The process suitability of each formulation in the Pareto optimal formulation set is evaluated, including: Obtain the water-cement ratio and slurry yield stress of each formulation, calculate the fluidity index of each formulation based on the water-cement ratio, and calculate the pumpability index of each formulation based on the slurry yield stress. Obtain the initial setting time and final setting time of each formulation, and calculate the operable time window of each formulation based on the initial setting time and final setting time; Obtain the total amount of slag and magnesium oxide in each formulation, and calculate the early strength development rate index of each formulation based on the total amount. The process suitability score of each formulation is obtained by weighting and summing the fluidity index, the pumpability index, the operable time window, and the early strength development rate index.

5. The UHPC material formulation optimization method based on artificial intelligence according to claim 1, characterized in that, Based on the process suitability score, a target optimized formulation is selected from the Pareto optimal formulation set, including: The process suitability scores of each formulation in the Pareto optimal formulation set are sorted from high to low, and the formulation with the highest score is selected as the initial target formulation. The mass fraction of each raw material in the preliminary target formula is checked for process boundary constraints to confirm whether the mass fraction of each raw material in the preliminary target formula falls within the corresponding process allowable range. If the verification passes, the initially selected target formula will be output as the target optimized formula. If the verification fails, the next formula in the sorted sequence will be selected for process boundary constraint verification until the verification passes and the target optimized formula will be output.

Images