Intelligent optimization method for mix proportion of mirror concrete based on machine learning

Abstract

The application discloses a mirror concrete mix proportion intelligent optimization method based on machine learning, and belongs to the technical field of concrete materials, and comprises the following steps: extracting the total light reflectance, surface pore distribution entropy and color difference root mean square of historical test pieces to construct an apparent quality index as an optimization target label; a decoupling type physical information neural network is constructed, a hydration heat release rate function and a physical residual function are constructed in combination with mix proportion parameters, a multi-source composite loss function is used for joint training to generate a proxy evaluation model; the reduced and reconstructed mix proportion parameters are embedded into a multi-objective optimization algorithm, the proxy evaluation model is called to predict the maximum temperature gradient to establish a temperature collaborative anti-cracking constraint, and a non-inferior solution set is generated to select and output a target mix proportion scheme. The application can fuse thermodynamic laws in the model, can change model prediction into physical constraint optimization conforming to hydration heat evolution, can avoid that excessive hydration heat causes structural microcracks, and can guarantee the structural integrity and apparent reflection effect of the mirror concrete.

Description

Technical Field

[0001] This application belongs to the field of interdisciplinary technology of artificial intelligence and materials big data, and specifically relates to a machine learning-based intelligent optimization method for the mix proportion of mirror concrete. Background Technology

[0002] Mirror-finish concrete, as a type of cast-in-place concrete with a smooth, flat surface and light-reflective properties, can directly present the original texture of a building structure and eliminate the need for secondary finishing processes. The forming quality of mirror-finish concrete is highly dependent on its mix proportions; the components and quantities of each raw material in the mix directly determine the hydration kinetics and final appearance of the concrete. With the development of artificial intelligence, materials informatics, and big data technologies, machine learning methods based on multi-source material data such as historical mix proportions, appearance images, temperature fields, and mechanical properties have been introduced into the field of building materials. These methods are used to construct nonlinear mapping models between raw material components and macroscopic properties, providing a data-driven technical path for the intelligent design of mirror-finish concrete mix proportions.

[0003] Current mix design methods in this field mainly include empirical trial mixing and pure data-driven machine learning prediction models. In terms of data-driven prediction, existing technologies construct datasets by collecting multi-source material data such as historical mix proportion parameters, apparent image features, temperature field data, and mechanical performance indicators. These datasets are then used to train regression evaluation models using machine learning algorithms such as artificial neural networks to replace traditional physical trial mixing. For example, invention patent CN121393693B discloses a method and system for performance prediction and mix proportion optimization of engineering cement-based composite materials based on physics-guided deep learning and multi-objective optimization. This method uses a physics-guided convolutional neural network model, PhyResNet, to predict the performance of tensile test data of engineering cement-based composite materials and utilizes the NSGA-II multi-objective optimization algorithm for mix proportion optimization. In the optimization design stage, existing technologies generally set a single material cost or a specific mechanical strength parameter as a fitness function, combining heuristic optimization algorithms to search for the theoretically optimal combination of raw material admixtures within given parameter boundaries.

[0004] However, the existing technologies still have some shortcomings: On the one hand, although existing multi-source material data is constantly accumulating, the data-driven models mainly focus on predicting the mechanical properties of materials, without incorporating the spatiotemporal temperature field evolution law generated by the exothermic hydration reaction inside concrete into the prediction framework. This lack of thermodynamic and physical constraints means that while the predicted mix proportions may probabilistically meet strength and cost requirements, they are prone to generating excessive heat of hydration in actual physical environments, causing microcracks in the structure and damaging the structural integrity and apparent reflective effect of the mirror concrete. On the other hand, the existing technologies still only characterize the mirror appearance features at the level of qualitative description or single indicators, failing to transform multi-dimensional appearance features such as total light reflectance, surface porosity distribution entropy, and root mean square color difference into numerical composite physical indicators. This results in a single optimization objective for the model, making it difficult for the final predicted scheme to achieve the expected mirror standard. Therefore, there is an urgent need for an intelligent design method for mirror concrete mix proportions that can integrate artificial intelligence, material data, and the physical laws of exothermic hydration, and achieve synergistic optimization of appearance quality and crack prevention. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a machine learning-based intelligent optimization method for the mix proportion of mirror concrete. By establishing a decoupled physical information neural network that integrates the hydration heat release law and spatiotemporal temperature prediction, a temperature gradient collaborative crack prevention constraint is introduced into the multi-objective optimization process, outputting a construction mix proportion scheme that conforms to the laws of thermodynamics and maximizes the apparent quality.

[0006] In a first aspect, embodiments of the present invention provide a machine learning-based intelligent optimization method for the mix proportion of mirror concrete, specifically including: The total light reflectance, surface pore distribution entropy, and root mean square color difference of historical mirror concrete specimens were collected to construct an apparent quality index. The apparent quality index and actual compressive strength were then injected into the multi-source material data training set as optimization target labels. A decoupled physical information neural network is constructed, comprising a material mapping subnetwork and a spatiotemporal temperature prediction subnetwork. The material mapping subnetwork is used to map the mix proportion parameters to hydration intrinsic parameters. A hydration heat release rate function and a physical residual function are constructed based on the mix proportion parameters and the hydration intrinsic parameters. The spatiotemporal temperature prediction subnetwork outputs a predicted spatiotemporal temperature field based on the hydration heat release rate function. The temperature statistical features of the predicted spatiotemporal temperature field are extracted and concatenated with the mix proportion parameters, and then input back into the material mapping subnetwork to output the standardized compressive strength prediction value and the apparent quality index prediction value. The multi-source material data training set and spatiotemporal point data are input into the decoupled physical information neural network. A multi-source composite loss function is constructed by combining the measured temperature field data, physical boundary conditions and the physical residual function. The decoupled physical information neural network is jointly trained based on the multi-source composite loss function to generate a proxy evaluation model that can reflect the spatiotemporal evolution law of hydration heat release. The mix proportion parameters are dimensionally reduced and reconstructed, and then embedded as optimization variables in a multi-objective optimization algorithm. The optimization iteratively aims to maximize the predicted value of the apparent quality index and minimize the material cost. The surrogate evaluation model is called to predict the maximum temperature gradient. Based on the maximum temperature gradient, temperature-coordinated crack prevention constraints are established, and a non-inferior mix proportion solution set is generated. A multi-attribute decision algorithm is used to select the target mix proportion individual that meets the critical temperature gradient threshold for crack prevention and has the highest apparent quality index from the non-inferior mix proportion solution set, and the mirror concrete construction mix proportion scheme is output.

[0007] As a preferred embodiment of the intelligent optimization method for the mix proportion of mirror concrete based on machine learning described in this invention, the construction of the apparent quality index includes: Extract the individual pixel color difference of the original color image and calculate the root mean square of the color difference of the entire image; The original color image is subjected to grayscale and binarization processing, and the area ratio of pore pixels within the grid region is extracted to calculate the surface pore distribution entropy. A uniform random perturbation is injected into the extracted total light reflectance and the root mean square of the color difference to generate preliminary perturbation feature values. A physical boundary truncation operator is applied to the preliminary perturbation feature values ​​based on historical extreme values ​​to generate physical consistency enhancement feature values. The physical consistency enhancement feature value, the surface pore distribution entropy, the actual compressive strength, the mix proportion parameters, and the measured temperature field data are tensor-stitched to expand and generate a multi-source material data training set. The actual compressive strength and apparent characteristic parameters in the training set of the multi-source material data are subjected to range standardization. The feature weight coefficients are solved based on the individual information utility values. The standardized apparent characteristic parameters are then weighted and superimposed according to the feature weight coefficients to construct the apparent quality index.

[0008] As a preferred embodiment of the intelligent optimization method for the mix proportion of mirror concrete based on machine learning described in this invention, the method of mapping the input mix proportion parameters to hydration intrinsic parameters using the material mapping subnetwork includes: The mixing ratio parameters are input into the kinetic parameter mapping branch of the material mapping subnetwork to perform feature extraction. The positive definite property hydration exothermic shape parameter and hydration characteristic time parameter are output by the activation function as the hydration intrinsic parameters. Based on the amount of each raw material in the mix proportion parameters, the multiphase medium mixing law is introduced to establish the thermal property parameter equation of concrete. The mass specific heat capacity constant is extracted to calculate the equivalent specific heat capacity, the absolute apparent density is extracted to calculate the aggregate volume fraction, and the mixed thermal conductivity constant is calculated by combining the single-phase thermal conductivity. Based on the equivalent specific heat capacity and the mixed thermal conductivity constant, the equivalent density and thermal diffusivity are calculated.

[0009] As a preferred embodiment of the intelligent optimization method for mirror concrete mix proportion based on machine learning described in this invention, the method comprises constructing an explicit hydration heat release rate function and a physical residual function based on the mix proportion parameters and the intrinsic hydration parameters, including: Extract the historical sequence of the predicted spatiotemporal temperature field, and construct an integral mapping to calculate the equivalent age of spatial nodes; The cement content, fly ash content, and slag powder content in the mix proportion parameters are extracted, and an explicit hydration heat release rate function is constructed by combining the equivalent age with the intrinsic hydration parameters. The first-order partial derivatives and Laplace operator of the predicted spatiotemporal temperature field are obtained with respect to time and space, respectively. The physical residual function is constructed by combining the equivalent specific heat capacity, the equivalent density, the thermal diffusivity, and the explicit hydration heat release rate function.

[0010] As a preferred embodiment of the intelligent optimization method for the mix proportion of mirror concrete based on machine learning described in this invention, the output of the standardized compressive strength prediction value and the apparent quality index prediction value includes: The surface temperature field data at the moment of maximum temperature rise is extracted from the predicted spatiotemporal temperature field, and a gradient truncation operator is introduced to obtain the static surface temperature sequence. The static surface temperature sequence is input into a global attention pooling layer to extract the surface temperature standard deviation and the global maximum surface temperature difference, and then merged to generate a surface temperature statistical feature scalar vector. The mixing ratio parameter and the surface temperature statistical feature scalar vector are tensor-concatenated to generate a fused auxiliary feature tensor. The fusion auxiliary feature tensor is input into the performance prediction branch of the material mapping subnetwork for forward computation, and the normalized compressive strength prediction value and the apparent quality index prediction value are output using the activation function.

[0011] As a preferred embodiment of the intelligent optimization method for the mix proportion of mirror concrete based on machine learning described in this invention, the construction of the multi-source composite loss function includes: Extract the internal physical spatiotemporal coordinate point set, the boundary spatiotemporal coordinate point set, and the initial spatial coordinate point set; Calculate the mean square error between the apparent quality label value and the network prediction value corresponding to the multi-source material data training set, and construct the data-driven loss. Extract the physical residual values ​​corresponding to the internal physical spatiotemporal coordinate point set to construct the physical constraint loss; Calculate the mean square error between the predicted temperature and the measured temperature field data, and construct a temperature measurement compensation loss; The deviation between the network-predicted initial temperature and the initial temperature of the mold pouring is calculated to construct the initial condition loss. Based on the laws of heat conduction and cooling, the continuity residuals of the outward heat conduction flux and convective heat dissipation flux of the boundary are calculated on the set of spatiotemporal loci at the boundary to construct the boundary condition loss. An adaptive gradient adjustment coefficient is introduced to aggregate the data-driven loss, the physical constraint loss, the temperature measurement compensation loss, the initial condition loss, and the boundary condition loss to construct a multi-source composite loss function.

[0012] As a preferred embodiment of the intelligent optimization method for the mix proportion of mirror concrete based on machine learning described in this invention, the joint training of the parameters of the decoupled physical information neural network includes: Calculate the partial derivative matrix of each sub-loss function with respect to the shared fully connected hidden layer in the material mapping sub-network independently, and calculate the instantaneous gradient balance coefficient based on the gradient 2 norm ratio of the partial derivative matrix; The extreme value truncation operator is applied to generate truncation instantaneous coefficients, and the adaptive gradient adjustment coefficients are updated in combination with the exponential moving average filtering mechanism. The parameters of the training set of the multi-source material data are extracted to construct a physical empirical estimation model of the intrinsic hydration parameters. A preheating loss function is constructed using the physical empirical initial value penalty term, and preheating training is performed on the material mapping sub-network. Freeze the trainable parameters of the material mapping subnetwork, perform local forward propagation calculation using the target alignment benchmark corresponding to the temperature measurement compensation loss, extract the average value of the adaptive gradient adjustment coefficient before convergence, and construct a relative proportion prior vector. Unfreeze the trainable parameters of the material mapping subnetwork, initialize the adaptive gradient adjustment coefficients using the relative proportion prior vector, initiate the global parameter joint training of the decoupled physical information neural network, and output the surrogate evaluation model.

[0013] As a preferred embodiment of the intelligent optimization method for the mix proportion of mirror concrete based on machine learning described in this invention, the dimensionality reduction and reconstruction of the mix proportion parameters includes: The mix proportion parameters are reconstructed using the absolute volume equation. The amount of fine aggregate is selected as the dependent variable uniquely determined by the absolute volume equation, while the remaining mix proportion parameters are retained as independent optimization variables. Together, they form a set of dimensionality-reduced control variables that are embedded in a non-dominated sorting genetic algorithm. Apply hard boundary constraints on the upper and lower limits of quality, apply reasonable range constraints on the dependent variable quality based on historical data distribution, and apply water-cement ratio inequality constraints in combination with hydration conditions to establish physical optimization boundaries; The target compressive strength is derived based on the statistical standard deviation of the sample, and a mandatory lower limit constraint on the engineering strength is applied.

[0014] As a preferred embodiment of the intelligent optimization method for mirror concrete mix proportion based on machine learning described in this invention, the output of the mirror concrete construction mix proportion scheme includes: The spatial discrete node set and the temporal discrete node set of the combined specimen are used to generate a sparsed reduced-order gradient evaluation domain matrix. The three-dimensional temperature gradient norm 2 is calculated based on the partial derivative vector of the predicted spatiotemporal temperature field with respect to spatial coordinate variables, and the global maximum three-dimensional temperature gradient norm 2 is extracted within the sparsified reduced-order gradient evaluation domain matrix. The critical temperature gradient threshold for crack prevention is calculated based on the constitutive equation, and an absolute constraint on the upper limit of the critical temperature gradient for crack prevention is constructed. Based on the background apparent quality level and the temperature gradient sensitivity coefficient, an exponential mathematical relationship between the predicted apparent quality index and the normalized temperature gradient is constructed as a cooperative empirical inequality constraint equation. The overall constraint violation degree of the optimization individual is calculated to prioritize the retention of feasible solution individuals, and a static Pareto front set is generated based on the non-dominated ranking and crowding distance as the non-dominated solution set; The objective function values ​​of individuals in the non-inferior mix ratio solution set are normalized, and objective weight parameters are generated by combining the objective function dispersion to construct a weighted normalized decision matrix. Based on the weighted normalized decision matrix, the relative proximity coefficient between each feasible solution individual and the absolute positive and negative ideal solutions is calculated. The individual ranked first is determined as the target mix proportion individual, and the mirror concrete construction mix proportion scheme is output.

[0015] Secondly, embodiments of the present invention provide a machine learning-based intelligent optimization system for the mix proportion of mirror concrete, comprising: The label building module is used to construct the apparent quality index and inject the apparent quality index and actual compressive strength as optimization target labels into the multi-source material data training set; The network construction module is used to build a decoupled physical information neural network. It uses the mix proportion parameters to map the hydration intrinsic parameters to output the predicted spatiotemporal temperature field and constructs the physical residual function. It extracts temperature statistical features and fuses them with the mix proportion parameters to output the standardized predicted values ​​of compressive strength and apparent quality index. The joint training module is used to input multi-source material data training set and spatiotemporal locus data into a decoupled physical information neural network, and combine measured temperature field data, physical boundary conditions and physical residual functions to construct a multi-source composite loss function, and train to generate a proxy evaluation model; The multi-objective optimization module is used to embed the reconstructed mix proportion parameters into a multi-objective optimization algorithm for optimization iteration, call the proxy evaluation model to predict the maximum temperature gradient to establish temperature-coordinated crack prevention constraints, generate a non-inferior mix proportion solution set, and select and output the mirror concrete construction mix proportion scheme.

[0016] Compared with the prior art, this application has the following advantages: 1. This invention is based on multi-source material data, extracting the total light reflectance, surface pore distribution entropy, and root mean square color difference of historical specimens of mirror concrete. Through material big data processing techniques such as uniform random perturbation, physical boundary truncation, range standardization, and information entropy weight calculation, an appearance quality index is constructed, transforming the subjective visual mirror effect into a numerical physical indicator. This index serves as the target label for model optimization, thereby solving the problems of existing technologies lacking quantitative mathematical expression of mirror appearance features and having a single optimization target. It also provides high-quality, physically consistent multi-source training data for subsequent artificial intelligence agent models.

[0017] 2. This invention establishes a decoupled physical information neural network composed of a material mapping subnetwork and a spatiotemporal temperature prediction subnetwork. It constructs a thermochemically coupled explicit hydration heat release rate function based on mix proportion parameters. A physical residual function characterizing the deviation from the physical laws of transient heat conduction is introduced into the prediction architecture. The output includes a predicted apparent quality index and a predicted spatiotemporal temperature field. During the multi-objective optimization stage, the global maximum three-dimensional temperature gradient L2 norm within the sparsed reduced-order gradient evaluation domain matrix is ​​extracted. Absolute constraints on the critical upper limit for crack prevention and the cooperative empirical inequality constraint equations between apparent quality and normalized temperature gradient are established. By integrating material data and thermodynamic physical laws into the data-driven model, mix proportion prediction is transformed into a physical constraint optimization that conforms to the spatiotemporal evolution law of hydration heat release. This avoids the predicted mix proportion scheme from causing excessive hydration heat and microcracks in the actual physical environment, ensuring the structural integrity and apparent reflective effect of the mirror concrete.

[0018] 3. This invention deeply integrates artificial intelligence with material data, constructs a training dataset using multi-source material data, performs joint modeling of data-driven and physical constraints through a decoupled physical information neural network, and employs a multi-objective optimization algorithm to achieve intelligent optimization of the mix proportion of mirror concrete. Compared with traditional experience-based mix design and purely data-driven models, this invention can fully explore the nonlinear relationship between historical mix proportions, apparent characteristics, temperature fields, and mechanical properties, reduce the number of physical mix design trials, shorten the mix design cycle, improve prediction accuracy and engineering applicability, and realize data-driven intelligent design of mirror concrete mix proportions.

[0019] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the intelligent optimization method for the mix proportion of mirror concrete based on machine learning is shown. Figure 2 A flowchart of step S2 in this invention is shown. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Example 1 Existing pure data-driven machine learning models suffer from a black box effect when predicting mix proportions. The predicted mix proportions may meet the mirror requirements in terms of mathematical probability, but in terms of actual physical laws, they are prone to causing excessive heat of hydration and microcracks. Furthermore, existing technologies lack a quantitative mathematical expression of the mirror effect, resulting in a single optimization objective and making it difficult for the final prediction scheme to achieve the expected mirror standard.

[0024] To solve the above technical problems, such as Figure 1As shown, this embodiment provides a machine learning-based intelligent optimization method for the mix proportion of mirror concrete. In the specific implementation process, the total light reflectance, surface pore distribution entropy, and root mean square color difference of historical mirror concrete specimens are collected to construct an apparent quality index. The apparent quality index and actual compressive strength are then injected into a multi-source material data training set as optimization target labels. A decoupled physical information neural network is constructed, comprising a material mapping subnetwork and a spatiotemporal temperature prediction subnetwork. The material mapping subnetwork maps the mix proportion parameters to hydration intrinsic parameters. Based on the mix proportion parameters and hydration intrinsic parameters, a hydration heat release rate function and a physical residual function are constructed. The spatiotemporal temperature prediction subnetwork outputs a predicted spatiotemporal temperature field based on the hydration heat release rate function. The temperature statistical features of the predicted spatiotemporal temperature field are extracted and concatenated with the mix proportion parameters before being input back into the material mapping subnetwork. The output is the standardized predicted compressive strength and apparent quality. The index prediction value is obtained by inputting the multi-source material data training set and spatiotemporal locus data into a decoupled physical information neural network. A multi-source composite loss function is constructed by combining measured temperature field data, physical boundary conditions, and physical residual functions. The decoupled physical information neural network is then jointly trained based on the multi-source composite loss function to generate a surrogate evaluation model. The mix proportion parameters are dimensionality-reduced and reconstructed, and then embedded as optimization variables into a multi-objective optimization algorithm. The optimization iteratively optimizes the apparent quality index prediction value and minimizes material cost as objective functions. The surrogate evaluation model is called to predict the maximum temperature gradient. Based on the maximum temperature gradient, temperature-coordinated crack prevention constraints are established and a non-inferior mix proportion solution set is generated. A multi-attribute decision algorithm is used to select the target mix proportion individual from the non-inferior mix proportion solution set and output the mirror concrete construction mix proportion scheme.

[0025] Furthermore, to better illustrate the technical solution of Embodiment 1 of the present invention, a detailed description of the intelligent optimization method for the mix proportion of mirror concrete based on machine learning is provided, specifically including the following steps: S1. Extract the light reflectance, pore distribution entropy, and root mean square color difference of the mirror-finished concrete to construct a positively represented appearance quality index, which serves as the optimization target label for model learning, thus transforming the mirror effect from subjective visual perception to numerical physical indicators. This includes the following sub-steps: S11. Collect multidimensional appearance feature quantification and multi-source feature data, which is specifically divided into the following sub-steps: S111. The total light reflectance of the surface of the historical spectrophotometer with an integrating sphere was measured using a d / 8 geometric condition spectrophotometer. (Dimensionless). Under constant illuminance conditions with a standard D65 light source, orthogonal images of the historical specimen surface were captured using an industrial camera to obtain original color images containing red, green, and blue channels. The original color images were then converted to the CIELAB color space, using chromaticity coordinates of a standard reference grayscale. Using a reference frame, calculate each individual pixel in the image. Color difference And aggregate the root mean square of color difference of the entire image. The calculation formula is: ; ; In the formula, , , For the first The lightness and chromaticity coordinates of each pixel in the CIELAB color space and These are the pixel width and pixel height of the image, respectively.

[0026] The original color image is converted to grayscale. The Otsu's maximum inter-class variance method is used to calculate the globally optimal binarization threshold, segmenting the grayscale image into black pixels representing pores and white pixels representing the substrate to identify surface pore features. The binarized image is then divided into... Each grid region is divided into three independent grid regions of the same size. The area ratio of the aperture pixels to the total number of pixels in the grid region is extracted. Calculate the surface pore distribution entropy The calculation formula is: ; In the formula, For the first The proportion of pore area in each grid region, dimensionless; The constant is dimensionless to prevent the logarithmic function from reaching extreme points and causing computational overflow.

[0027] S112. Extract the actual compressive strength of the corresponding historical specimens. (Unit: MPa) and mix proportion parameters (Dimensionless). Mixing ratio parameters Including cement content in a single cubic meter of concrete Water consumption fly ash dosage Slag powder dosage Fine aggregate dosage , Coarse aggregate dosage Dosage of water-reducing agent All raw material usage units are kg / m³.

[0028] S113. Pre-embedded temperature sensors are evenly distributed along the shortest heat dissipation path from the geometric center to the surface of the historical specimen at fixed spatial intervals, and the three-dimensional spatial coordinate vectors of each pre-embedded temperature sensor are recorded. (Unit: m). Using the moment the historical specimen was poured into the mold as the zero point, real-time temperature data were synchronously acquired at a fixed sampling frequency during the period of intense hydration. Real temperature time series data The dimension of is Kelvin.

[0029] S12. Data augmentation and multi-source material data training set expansion based on physical boundary truncation, specifically divided into the following sub-steps: S121. Extract the total light reflectance from historical dataset samples. and root mean square color difference Introduce independent uniformly distributed random variable factors and Let the space of values ​​for the random variable factors be defined as follows: To address the measurement error inherent in image appearance features, the total light reflectance... and root mean square color difference Inject uniform random perturbations with an amplitude of 5% of the original value. Generate corresponding preliminary perturbation eigenvalues. and The calculation formula is as follows: ; ; In the formula, The initial perturbation total optical reflectance after noise injection is dimensionless. The root mean square of the initial perturbation color difference after noise injection is dimensionless.

[0030] S122. Traverse the original historical sample dataset and extract the historical maximum value of total light reflectance. Compared with historical minimum value Historical maximum value of root mean square of color difference Compared with historical minimum value A physical boundary truncation operator is constructed, and historical maxima and minima are used to forcibly truncate the perturbed feature data to prevent data from going out of bounds. Physically consistent enhanced feature values ​​are then generated. and The expression is as follows: ; ; In the formula, To find the maximum value function, This is a function that minimizes the value. This bidirectional truncation operation ensures that the enhanced feature data is strictly bounded to the historical real physical space domain, preventing extreme out-of-bounds data from disrupting the mapping topology of subsequent activation functions.

[0031] S123. Extract the entropy of undisturbed surface pore distribution. Actual compressive strength Mix proportion parameters Compared with real temperature time series data Enhance physical consistency eigenvalues and Tensor splicing is performed along with the above parameters, maintaining the corresponding mix ratio parameters. Compared with real temperature time series data With the labels unchanged, generate physically consistent enhanced samples. Perform the above enhancement operations on all historical samples, then merge the generated physically consistent enhanced sample matrix with the original historical sample dataset along the sample dimension, thereby expanding the multi-source material data training set. The sample size. The expanded multi-source material data training set is used as complete input and uniformly passed to step S13 to perform range standardization and apparent quality index. Quantification of .

[0032] S13. Perform dimensionless normalization mapping based on dual extreme domains and construct appearance quality labels, which is divided into the following sub-steps: S131. Traversing the expanded multi-source material data training set Record the maximum actual compressive strength among all samples. and minimum value To prevent subsequent multi-objective optimization algorithms from encountering boundary saturation when exploring the space for better mix proportions, an expansion upper limit is set. With expanding the lower limit Training set of multi-source material data Actual compressive strength Range standardization is performed by extending the extreme value range, and the range is mapped to... The standardized compressive strength is obtained within the dimensionless range. The calculation formula is as follows: ; This mapping method ensures that when the intensity predicted exceeds that of historical samples during the optimization phase, the standardized prediction value still does not exceed the limit.

[0033] S132. Simultaneously extract the expanded multi-source material data training set. The historical limit of the apparent characteristic parameters, including the historical maximum of total reflectance. Compared with historical minimum value Historical maximum value of surface pore distribution entropy Compared with historical minimum value Historical maximum value of root mean square of color difference Compared with historical minimum value For total light reflectance Surface pore distribution entropy Root mean square color difference Range standardization is performed using historical extreme values, and the range is strictly mapped to... Within the dimensionless range, the normalized light reflectance was obtained respectively. Standardized stomatal distribution entropy Standardized root mean square color difference The calculation formula is as follows: ; ; .

[0034] S133, Extracting standardized light reflectance Standardized stomatal distribution entropy Compared with standardized root mean square color difference Based on the logic of weighted superposition of positively correlated elements and reverse mapping weighting of negatively correlated elements, an appearance quality index is constructed. The calculation formula is: ; In the formula, The weighting coefficient for light reflectance characteristics. For the surface pore distribution entropy characteristic weighting coefficient, The root mean square feature weight coefficients of color difference satisfy... And all are greater than The calculated apparent quality index and its corresponding standardized compressive strength Injecting multi-source material data training set In this context, it serves as the set of absolute labels for forward inference in the model.

[0035] As one possible implementation method, the light reflectance characteristic weighting coefficient Surface pore distribution entropy characteristic weighting coefficient Color difference root mean square feature weight coefficient The method for determining it is as follows: S1331. Extract the expanded multi-source material data training set. All covered in A number of historical samples, constructed with the following dimensions. Apparent feature index matrix Define the elements of the feature index matrix. When feature index hour Refers to the first The standardized light reflectance of each sample, when the feature index is... hour Refers to the first The inverse stomatal distribution entropy of a sample, when the feature index hour Refers to the first The root mean square of the inverse color difference of the nth sample. Calculate the nth... Feature index in the first The proportion of independent features in each sample The calculation formula is as follows: .

[0036] S1332. Based on the feature weight distribution of all samples, Shannon's information theory is introduced to calculate the first... Information entropy value of feature index The calculation formula is as follows: .

[0037] S1333, at the above information entropy value During the calculation process, if the data matrix contains The physical boundary limit state is obtained by directly applying L'Hôpital's limit theorem. This avoids logarithmic operation errors. Based on the information entropy value... Calculate the first The objective information utility value possessed by each characteristic indicator The calculation formula is as follows: ;Utilizing the utility value of individual information The mathematical proportion of the total information utility value of all feature indicators is used to quantitatively solve for the corresponding feature weight coefficients. The calculation formula is as follows: ; index Take values ​​respectively , , Output the corresponding light reflectance feature weight coefficients Surface pore distribution entropy characteristic weighting coefficient Color difference root mean square feature weight coefficient .

[0038] In step S1 of this embodiment, the total light reflectance, surface pore distribution entropy, and root mean square color difference of the mirror concrete are extracted. After performing range standardization on the above appearance feature parameters, a weighted superposition is performed using feature weight coefficients calculated based on the utility value of individual information items to construct a positively represented appearance quality index. This transforms subjective visual features into objective numerical physical labels that can be learned by the model. Specifically, for example, assuming that the historical maximum value of the total light reflectance is extracted from the multi-source material data training set. Compared with historical minimum value Historical maximum of surface pore distribution entropy Compared with historical minimum value Historical maximum value of root mean square of color difference Compared with historical minimum value Total reflectance enhancement characteristic value Surface pore distribution entropy Root mean square enhancement eigenvalue of color difference First, perform dimensionless range standardization on the sample, and then substitute the values ​​into the formula to calculate the standardized light reflectance. Standardized stomatal distribution entropy Standardized root mean square color difference Furthermore, assuming that the feature weight coefficients determined based on the multi-source material data training set are as follows: Light reflectance feature weight coefficients... Surface pore distribution entropy characteristic weighting coefficient Color difference root mean square feature weight coefficient Substitute the above parameters into the apparent quality index formula to calculate: Through the above calculations and derivations, the multidimensional apparent features of this sample are uniformly transformed into apparent quality index label values ​​with extreme values ​​ranging from 0 to 1. It satisfies the dimensional consistency and physical logic rationality under the extreme value constraint conditions.

[0039] S2. Establish a decoupled physical information neural network consisting of a material mapping subnetwork and a spatiotemporal temperature prediction subnetwork, and output the predicted values ​​of standardized compressive strength and apparent quality index. For example... Figure 2 As shown, the specific steps include the following: S21. Construct a dual-branch decoupled architecture and perform physical fusion of thermal property parameters, which is divided into the following sub-steps: S211. Construct a dual-branch decoupled physical information neural network topology. This network consists of a feedforward material mapping subnetwork and a fully connected spatiotemporal temperature prediction subnetwork, which are decoupled and cascaded through a shared physical parameter tensor.

[0040] As one possible implementation, step S211 specifically includes: The input layer of the material mapping subnetwork is configured with 7 neurons to receive data in dimension 1. Mixing ratio parameters This includes the unit dosage of cement, water, fly ash, slag powder, fine aggregate, coarse aggregate, and water-reducing agent. A shared fully connected hidden layer with 128 neurons is cascaded after the input layer. This hidden layer uses the ReLU activation function for nonlinear feature extraction, and its output tensor is divided into two completely independent feedforward computation branches.

[0041] The input layer of the first branch, i.e., the performance prediction branch, is explicitly configured with 137 neurons. It receives the complete tensor obtained by concatenating the 128-dimensional basic comparability feature tensor output from the shared fully connected hidden layer with the 9-dimensional fusion auxiliary feature tensor generated in step S233 along the feature channels. The first branch cascades three consecutive hidden layers with the number of neurons decreasing sequentially to 64, 32, and 16, respectively. Each layer uses the ReLU activation function. The output layer of the first branch is configured with two neurons, corresponding to the standardized stress resistance prediction values. Compared with the predicted value of the apparent quality index .

[0042] The second branch, the kinetic parameter mapping branch, contains three consecutive hidden layers with neuron node configurations of 64, 32, and 16, using the ReLU activation function. The output layer of the second branch has two neurons, corresponding to the hydration exothermic shape parameters. With hydration characteristic time parameters The output tensor of the second branch is fed into the subsequent spatiotemporal temperature prediction subnetwork as a forward propagation static physical constant.

[0043] The spatiotemporal temperature prediction subnetwork is designed as a deep fully connected network with 5 hidden layers, each containing 64 neurons. The Tanh activation function is used to ensure the network's second-order differentiability with respect to spatiotemporal variables, satisfying the computational requirements of the Laplace operator. The input layer of the spatiotemporal temperature prediction subnetwork has 4 neurons to receive the three-dimensional spatial coordinate components. With time variable The spatiotemporal temperature prediction subnetwork utilizes the hydration exothermic shape parameters output from the second branch of the material mapping subnetwork. With hydration characteristic time parameters As a function for calculating the built-in heat of hydration release rate The pre-coefficients are used for forward propagation inference. The output layer of the spatiotemporal temperature prediction subnetwork is configured with a single neuron, using a linear activation function, and outputs the predicted spatiotemporal temperature field for the corresponding spatiotemporal node. .

[0044] S212, Mixing ratio parameters Import the second branch of the material mapping subnetwork for forward computation.

[0045] As one possible implementation, step S212 specifically includes: S212-1. Obtain the 128-dimensional basic mix proportion feature tensor from the shared fully connected hidden layer output in step S211, and import it losslessly into the second branch of the material mapping sub-network for forward computation. The feature tensor passes through three consecutive fully connected hidden layers with 64, 32, and 16 nodes respectively. Linear affine transformation and ReLU nonlinear mapping are performed sequentially within each hidden layer to complete the deep feature dimensionality reduction and feature representation from macroscopic mix proportion parameters to microscopic hydration exothermic properties.

[0046] S212-2. The 16-dimensional tensor after feature extraction is fed into the output layer of the second branch, which has two neurons. The output, an unbounded pre-activation variable in the real number domain, is calculated by multiplying the weight matrix of the fully connected layer and adding the bias. and .

[0047] S212-3, The second branch output layer forces the Softplus activation function to apply the unbounded pre-activation variable. and Nonlinear processing is performed to ensure the shape parameters of the final output hydration heat release. With hydration characteristic time parameters It must be a strictly positive definite value. The specific formula for calculating forced positive definite activation is as follows: , Outputs hydration exothermic shape parameters with strictly guaranteed physical positive definiteness. With hydration characteristic time parameters ,in The characteristic time scale characterizing the hydration heat release process of cement-based materials, in seconds (s).

[0048] S213, Based on input mix proportion parameters The amount of each individual raw material is used, and a dynamic update equation for the thermal properties of concrete is established by introducing the multiphase media mixing law, and the equivalent specific heat capacity is calculated. equivalent density With thermal diffusivity The nonlinear mapping relationship is calculated using the following formulas: , , ; In the formula, Corresponding cement dosage Water consumption fly ash dosage Slag powder dosage Fine aggregate dosage , Coarse aggregate dosage Dosage of water-reducing agent ; The specific heat capacity constant corresponding to each single-phase raw material is expressed in J / (kg·K); This is the mixed thermal conductivity constant estimated based on the aggregate volume fraction, in W / (m·K); Based on volume; equivalent specific heat capacity The unit is J / (kg·K); equivalent density The unit is kg / m³; thermal diffusivity The unit is m² / s.

[0049] As one possible embodiment, mixed thermal conductivity Based on the Maxwell-Eucken two-phase composite material model, the following were determined: Extracting mix proportion parameters Fine aggregate usage Coarse aggregate usage Combined with the absolute apparent density of fine aggregate Compared with the absolute apparent density of coarse aggregate Calculate aggregate volume fraction The calculation formula is as follows: ; In the formula, and The unit is ; and The unit is ;Calculated aggregate volume fraction Characterized by the absolute volume of coarse and fine aggregates in 1 cubic meter of concrete, it is dimensionless. The concrete system is considered as a composite heat transfer medium consisting of a continuous cement paste phase and a discrete aggregate phase. The Maxwell-Eucken composite heat transfer theory is introduced, combined with the thermal conductivity of a reference cement paste. thermal conductivity of reference aggregate Calculate the mixing thermal conductivity corresponding to the current mix ratio. The calculation formula is as follows: ; In the formula, This is the constant of thermal conductivity of continuous phase cement paste determined based on the water-cement ratio. The thermal conductivity constant of discrete phase aggregate mixtures, determined through physical experiments, is the unit of measurement for both. When the aggregate volume fraction Approaching At that time, the mixed thermal conductivity Strictly converges to the thermal conductivity of cement paste When aggregate volume fraction Approaching At that time, the mixed thermal conductivity Strictly converges to the thermal conductivity of the aggregate .

[0050] S22. Constructing the explicit hydration heat release rate function of thermochemical coupling and the physical residual function characterizing the degree of deviation from physical laws, specifically divided into the following sub-steps: S221. Extract the hydration exothermic shape parameters output from the second branch of the material mapping subnetwork in step S21. With hydration characteristic time parameters The shape parameters of the hydration exothermic reaction With hydration characteristic time parameters Together with three-dimensional spatial coordinates Time variables Tensor concatenation is performed along the feature dimension, and the concatenated tensors are used as joint feature inputs to be imported into the spatiotemporal temperature prediction subnetwork. The network nodes of the spatiotemporal temperature prediction subnetwork perform forward propagation calculations, outputting the predicted spatiotemporal temperature field corresponding to the input spatial coordinates and time variables. Among them, time variable The unit is seconds, predicting the spatiotemporal temperature field. The unit is Kelvin.

[0051] S222, Extract specific spatial coordinates From the start of hydration to the time variable Historical sequence of predicted spatiotemporal temperature field during the period An integral mapping calculation based on the historical temperature field is constructed to obtain the equivalent age of the corresponding spatial nodes. This is used to describe the actual hydration process under real thermo-coupling conditions. The calculation formula is as follows: ; In the formula, Apparent activation energy, expressed in J / mol; It is the ideal gas constant; For reference to absolute temperature, the international standard thermodynamic reference room temperature constant is set to a value of [value missing]. K.

[0052] As one possible implementation method, apparent activation energy The isothermal microcalorimetry method was used to obtain the data at two different but constant absolute temperatures. and The hydration heat release rate curves of the cementitious slurry with the corresponding mix proportions were determined. The characteristic time constants of hydration under two sets of temperature conditions were extracted. and Apparent activation energy calculated based on the logarithmic variant of the Arrhenius equation. The calculation formula is as follows: In the formula, Let be the ideal gas constant, and take the value of . J / (mol·K).

[0053] S223. Extract cement dosage from the feature matrix of the original mix proportion parameters. fly ash dosage Slag powder dosage And combined with the equivalent age calculated in step S222 And eigenparameters with positive definite properties and Construct an explicit hydration heat release rate function with thermochemical coupling. The calculation formula is as follows: ; In the formula, , , , respectively, represent the limiting heat of hydration release constants for cement, fly ash, and slag powder, in J / kg; explicit heat of hydration release rate function. The unit is W / m³.

[0054] As one possible implementation method, the limiting heat of hydration release constant , , The isothermal conduction calorimeter was used in conjunction with the Knudsen hyperbolic extrapolation model to determine: at a reference absolute temperature Under constant temperature conditions, the continuous cumulative heat release sequences of pure cement paste, pure fly ash paste, and pure slag powder paste were independently determined during the test period. The cumulative heat release sequence The reciprocal of the corresponding test time The reciprocal of the integer part is used to establish a univariate linear regression fitting model, and the calculation formula is as follows: In the formula, This is the half-life time constant when the cumulative heat release reaches half of the limit value. The ordinate of the fitted straight line for each group of single-phase materials is extracted, and the reciprocal of the ordinate is calculated to obtain the theoretical infinite-age limit cumulative heat release. The results were obtained by extrapolating the single-component systems of pure cement, pure fly ash, and pure slag powder, respectively. Absolute assignment to the limiting heat of hydration release constant , , .

[0055] S224. Using automatic differentiation techniques within a deep learning framework, the predicted spatiotemporal temperature field output by the spatiotemporal temperature prediction subnetwork is evaluated. Regarding time variables respectively Solving for first-order partial derivatives in spatial coordinates Solve the second-order spatial Laplace operator operation. Based on the physical laws of three-dimensional transient heat conduction, retrieve the parameters varying with the mix proportion in step S21. Dynamically updated equivalent specific heat capacity equivalent density With thermal diffusivity Construct a physical residual function to characterize the degree of deviation from physical laws. The calculation formula is as follows: ; In the formula, Characterize the predicted partial derivative terms, Characterizing the heat conduction term, The heat release term representing the internal heat source is uniformly normalized to Kelvin per second.

[0056] S23. Perform attention-based dimensionality reduction of temperature field features and causal mapping of apparent quality, which is divided into the following sub-steps: S231. The three-dimensional predicted spatiotemporal temperature field output by the spatiotemporal temperature prediction subnetwork. In the process, the extracted specimen was at a time point during the period of intense hydration (such as when the highest temperature rise was reached). The set of surface space coordinate nodes at time (time) The surface temperature field data is obtained by introducing a gradient truncation operator to block the chain-like backpropagation path of the surface temperature field data in the automatic differentiation mechanism, thus obtaining a static surface temperature sequence without gradient properties.

[0057] As an optional implementation, the moment when the highest temperature rise is reached in step S231. The method for determining it is as follows: extract the coordinate vector of the three-dimensional geometric center of the specimen. And the fixed discrete sampling time step set used when synchronously acquiring real temperature time series data in step S113. The geometric center coordinate vector With discrete sampling time step set The data are fed into the spatiotemporal temperature prediction subnetwork to perform local forward inference computation, outputting a one-dimensional center temperature prediction sequence for the geometric center in the corresponding time set. For the predicted core temperature sequence By applying an independent variable optimization operator, the discrete time step corresponding to the maximum value of the temperature scalar is dynamically extracted and assigned to the time of the highest temperature rise. The calculation formula is: The time of the highest temperature rise will be determined dynamically. With the set of surface space coordinate nodes Feature splicing is performed and the data is input again into the spatiotemporal temperature prediction subnetwork to obtain the three-dimensional surface predicted temperature field data at that specific moment. Then, the gradient truncation operator is applied according to the established logic.

[0058] S232. Input the truncated static surface temperature sequence into the global attention pooling layer. Define the set of surface spatial coordinate nodes. The total number of discrete nodes is , No. The static predicted temperature of each surface node is In the global attention pooling layer, an adaptive attention weight coefficient is introduced to focus on regions of temperature anomalies. The calculation formula is as follows: ; In the formula, For the trainable weight tensor of the attention pooling layer, These are the trainable bias term parameters. Based on attention weight coefficients. Calculate the attention-weighted mean surface temperature with surface temperature standard deviation Simultaneously extract the global maximum surface temperature difference. The calculation formula is as follows: ; ; ; The calculated surface temperature standard deviation With the global maximum surface temperature difference Merge the features across the feature channels to output a fixed two-dimensional surface temperature statistical feature scalar vector. The high-dimensional spatiotemporal temperature field is reduced to physical and statistical characteristics that directly affect the cracking and color unevenness of concrete surfaces.

[0059] S233, with dimension as Mixing ratio parameters With the output surface temperature statistical characteristic scalar vector Perform tensor splicing operations to generate tensors with A fusion auxiliary feature tensor of dimension is generated. This fusion auxiliary feature tensor is fed into the first branch of the material mapping subnetwork for forward computation. The features are sequentially passed through three cascaded fully connected hidden layers configured with 64, 32, and 16 neurons respectively to perform deep feature cross-mapping. In the output layer of the first branch, two output neurons are configured and the Sigmoid activation function is called to strictly constrain the computation results to the specified dimensions. Within the dimensionless range, output the standardized compressive strength prediction values ​​for the current mix proportion and physical temperature environment. Compared with the predicted value of the apparent quality index .

[0060] In step S2 of this embodiment, a decoupled physical information neural network is constructed, comprising a material mapping subnetwork and a spatiotemporal temperature prediction subnetwork. The material mapping subnetwork extracts mix proportion parameter features to map and generate hydration intrinsic parameters with positive physical definite properties. Based on the mixing law of multiphase media, a thermophysical property parameter equation is established. Then, a physical residual function characterizing the deviation from physical laws is constructed by combining the explicit hydration heat release rate function. Finally, temperature statistical features are fused to complete the physical constraint prediction of the apparent quality index and standardized compressive strength. Specifically, for example: assuming that the input set of mix proportion parameters corresponds to the amount of individual raw materials, such as cement dosage... Water consumption fly ash dosage Slag powder dosage Fine aggregate dosage , Coarse aggregate dosage Dosage of water-reducing agent The equivalent density is calculated using the mixing law of multiphase media, and then substituted into the formula to calculate the equivalent density. Meanwhile, it is assumed that the unbounded pre-activation variables obtained by the output layer of the dynamic parameter mapping branch of the material mapping subnetwork are... To satisfy the forced positive definite property of the physical parameters, the activation function is called to perform a nonlinear mapping, and the shape parameters of the hydration exothermic reaction are calculated by substituting them into the formula. This operation ensures that all network pre-parameters output by the model strictly conform to the positive definite laws of thermodynamics and the law of conservation of mass and volume, avoiding negative parameters or physical paradoxes of non-conservation of mass that may be caused by direct mapping from the network black box, so that the final output predicted spatiotemporal temperature field and physical residual function have objective real-world physical meaning.

[0061] S3. Construct a multi-source composite loss function and train the network using a staged adaptive adjustment strategy to output a surrogate evaluation model that conforms to the laws of thermodynamics. This includes the following sub-steps: S31. Perform multi-dimensional spatiotemporal physical collocation sampling and composite loss function topology construction, which is divided into the following sub-steps: S311. Define the three-dimensional solid space geometry of the mirror concrete specimen. and its outer surface boundary domain Set the time domain for the period of intense hydration The Latin hypercube sampling algorithm is used to extract an unbiased set of spatiotemporal collocations within a defined spatiotemporal domain: in the spatial geometric domain... With the time domain Random sampling within the Cartesian product domain A set of internal physical spatiotemporal coordinate points ; in the outer surface boundary domain With the time domain Extraction within the product domain A set of boundary spatiotemporal coordinate points In the spatial geometric domain initial moment Extraction An initial set of spatial collocation points Simultaneously, acquire the real temperature time-series data synchronously collected in step S113. The corresponding set of discrete spatial coordinates of the embedded temperature sensor With discrete sampling time step set The number formed is The set of spatiotemporal coordinate points for actual temperature measurements.

[0062] S312. Invoke the material mapping subnetwork and the spatiotemporal temperature prediction subnetwork to train the expanded multi-source material data set. middle Standardized compressive strength corresponding to each sample With apparent quality index Calculate the predicted value output by the network. and Mean squared error, constructing data-driven loss : .

[0063] S313, in A set of internal physical spatiotemporal coordinate points The automatic differentiation and partial differentiation calculations defined in step S224 are executed to obtain the thermo-coupling physical residual value corresponding to each collocation point. Construct physical constraint loss : .

[0064] S314. Extract the predicted temperature of the spatiotemporal temperature prediction subnetwork at the spatiotemporal coordinate point of the actual temperature measurement. Compared with real temperature time series data Calculate the mean square error and construct a temperature measurement compensation loss. Simultaneously extract the initial spatial collocation set. The network predicts the initial temperature at the location. Calculate the initial temperature of concrete poured into the formwork in known Kelvin units. The deviation is used to construct the initial condition loss. The calculation formulas are as follows: ; .

[0065] S315, in A set of boundary spatiotemporal coordinate points Fourier's law of heat conduction and Newton's law of cooling are applied to obtain the predicted temperature of the network at the boundary nodes. and its spatial gradient's outward normal projection Obtain the convective heat transfer coefficient in constant units of W / (m²·K). The mixed thermal conductivity is a constant in W / (m·K). And the ambient temperature in known Kelvin units. Calculate the continuity residual between the outward thermal conduction flux and the convective heat dissipation flux at the boundary, and construct the boundary condition loss. : .

[0066] S316. Introduce adaptive gradient adjustment coefficient. The above sub-loss terms are aggregated to construct the total loss function that drives the optimization of the entire network parameters. : ; The adjustment coefficients evolve dynamically in the early stages of training to balance the backpropagation magnitude of the gradients of various losses, preventing single boundary conditions or data-driven losses from dominating parameter updates on the numerical range, and ensuring that the model falls into the globally stable convergence range of multi-objective optimization.

[0067] As one possible implementation, an adaptive adjustment method based on the shared hidden layer gradient norm ratio is provided to dynamically update the adaptive gradient adjustment coefficients during network training. The specific implementation steps and calculation rules are as follows: S316-1. After extracting the material mapping subnetwork input layer, the cascaded shared fully connected hidden layer defines its trainable weight matrix as the anchor parameter tensor. After each forward propagation computation during network training, the sub-loss functions are independently calculated with respect to the anchor parameter tensor using an automatic differentiation framework. The partial derivative matrix is ​​used to obtain the data-driven loss gradient matrix. and the physical and boundary constraint loss gradient matrix ,index Corresponding to physical constraint losses Temperature measurement compensation loss Initial condition loss Boundary condition loss .

[0068] S316-2. Solve for the L2 norm of each extracted gradient matrix to quantify the magnitude of the backward gradient update triggered by each loss function in the shared feature extraction layer. Calculate the... The next iteration step Instantaneous gradient balance coefficients corresponding to the term constraint loss The calculation formula is as follows: ; In the formula, The gradient L2 norm of the data-driven loss. For the first The gradient 2 norm of the constraint loss. To prevent infinitesimal floating-point constants with a denominator of zero (taking values ​​of...) This operation forces the gradient magnitudes of the various physical and boundary constraint terms to numerically align with the gradient magnitudes of the data-driven loss.

[0069] S316-3. To prevent numerical divergence caused by the gradient minima in a single step, a truncation operator is introduced to generate truncation instantaneous coefficients. The calculation formula is: In the formula, The lower threshold is... The upper limit threshold is determined by constructing an absolute numerical cutoff boundary to forcibly constrain the range of the L2 norm ratio of the partial derivative matrices between multi-source loss terms. This prevents the adaptive gradient adjustment coefficient from distorting and surging due to the gradient of a certain physical constraint term approaching zero during a single iteration, thus avoiding divergence of backpropagation network parameters. For example, setting... ,set up Based on the adjustment coefficient of the previous iteration step, calculate the... The final adaptive gradient adjustment coefficient in the next iteration step The calculation formula is: In the formula, This is the momentum smoothing decay constant. Its value is determined by applying a continuous exponential moving average filter to the instantaneous local gradient balance coefficients using a large proportion of historical iteration coefficient information. This suppresses high-frequency transient fluctuations in a multi-objective gradient game environment and maintains the smooth convergence of the global joint training gradient. For example, setting... .

[0070] S32. Perform data-driven warm-up and full-chain alternating joint training, specifically divided into the following sub-steps: S321. Extract the training set of multi-source material data. Mixing ratio parameters Obtain cement usage fly ash dosage Slag powder dosage and water consumption Calculate the dimensionless water-cement ratio per unit volume. The proportion of mineral admixtures The calculation formulas are as follows: and .

[0071] A physical empirical estimation model for the intrinsic parameters of hydration is constructed, and the baseline initial values ​​of the shape parameters for the heat release during hydration are calculated. Reference initial values ​​of hydration characteristic time parameters The calculation formula is as follows: ; In the formula, and It is an empirical constant pre-calibrated based on historical standard hydration tests, and is dimensionless.

[0072] For the second branch of the material mapping subnetwork, a penalty term based on physical empirical initial values ​​is constructed. Preheating loss function The calculation formula is as follows: In the formula, and The empirical penalty coefficient is dimensionless.

[0073] Using the preheating loss function Backpropagation and parameter updates are performed independently on the second branch of the material mapping subnetwork until the preheating loss function is reached. Once the value drops below the preheating convergence threshold, stop the preheating training and forcibly constrain the second branch of the network to output initial values ​​of hydration intrinsic parameters with reasonable physical benchmarks.

[0074] S322. Freeze all trainable parameters of the preheated material mapping subnetwork, blocking the back propagation of the gradient vector to the material mapping subnetwork in the automatic differentiation mechanism. Explicitly eliminate the data-driven loss, which exhibits static constant error properties due to parameter freezing. Constructing a loss function based on physical constraints Temperature measurement compensation loss Initial condition loss With boundary condition loss Composition of local training total loss The calculation formula is: ; Extracting the weight tensor of the first fully connected hidden layer of the spatiotemporal temperature prediction subnetwork As a local calculation anchor point parameter, the measured temperature compensation loss is selected. Set the corresponding adjustment coefficient as the target alignment reference. .

[0075] After the initial forward propagation calculation of local training is completed (i.e., the initial moment before update), the true scaling is directly mapped using the L2 norm of the extracted initial gradient matrix. The calculation formula is as follows: ; After calculation, assign , , Real physical initial values.

[0076] After entering the regular training iteration, calculate the instantaneous local gradient balance coefficient. Application of extreme value range Perform a numerical truncation operation to obtain The output local coefficients are updated by applying the momentum smoothing decay formula. Input the expanded multi-source material data training set. The forward inference of the spatiotemporal temperature prediction subnetwork is driven. A dual convergence evaluation system is established to simultaneously extract the original physical constraint loss values ​​without dynamic weighting. and the original temperature measured compensation loss value As an objective monitoring indicator, the backpropagation parameters are continuously updated until... and All iterations exhibit a global convergence trend and consistently fall below the set upper limit of the corresponding error, at which point the local phase training stops. The instantaneous local gradient balance coefficients that were not truncated during the last ten iterations before convergence are extracted. Calculate the arithmetic mean. Construct and output the relative proportion prior vector. The extraction rules are as follows , , , .

[0077] S323. Unfreeze all parameters of the material mapping sub-network to restore the globally joint trainable state of the overall dual-branch decoupled physical information neural network. At the initial moment of global joint training, receive the relative proportion prior vector output in step S322. To eliminate the global gradient magnitude jump caused by anchor point transformation, temperature measurement is used to compensate for the loss. As a transitional benchmark, calculate the absolute benchmark multiplier. And perform the same boundary truncation on it to prevent surge, calculated as follows: ; All adaptive gradient adjustment coefficients are initialized using the proportional migration rule, and final boundary constraints are applied. The calculation formula is as follows: ; Inject the aforementioned initialized adaptive gradient adjustment coefficients into the total loss function defined in step S316. In the middle. For buffering. The sudden recovery causes weight contention and gradient shocks, so a low initial learning rate is configured in the early stages of global training for a warm-up transition. In subsequent iterations, the learning rate defined in step S316 is activated and strictly executed. An adaptive update mechanism for anchor point parameters to quickly correct coefficient deviations. Connectivity from mix proportion parameters. An end-to-end gradient update link mapped to the macroscopic spatiotemporal temperature field and apparent quality indicators. Global parameter joint training is performed, awaiting the total loss function. Once the global convergence condition is met, the iteration stops, and the trained surrogate evaluation model is saved and output.

[0078] As one possible implementation, pre-collection of... Isothermal calorimetric datasets of historical standard hydration tests with different mix proportions were collected, and historical samples were extracted from each group. Actual water-cement ratio per unit volume The proportion of mineral admixtures The corresponding true hydration heat release rate curve was obtained. The true hydration heat release rate curve was fitted using the nonlinear least squares method, and the corresponding true hydration heat release shape parameters were extracted through inversion. Compared with the actual hydration characteristic time parameters . Build contains Linear regression objective loss function for a set of historical data The calculation formula is as follows: Let the target loss function be... To each , , The partial derivatives are zero. Solve the normal equations and calculate the constants to determine the output. , , Numerical values. Logarithmic linearization combined with a univariate linear regression model was used to determine empirical constants. , The calculation formula is as follows: The linearized equation was fitted using ordinary least squares to obtain the regression coefficients. With slope The parameters are obtained by inverse solving through exponential operations. Output a definite constant. Numerical values. Calculation of the set of true hydration heat release shape parameters. Statistical variance With the set of real hydration characteristic time parameters Statistical variance The penalty coefficient is set to the reciprocal of the statistical variance of the corresponding parameter to eliminate the absolute difference in numerical magnitude between different physical intrinsic parameters and prevent the loss function from becoming gradient-dominated during the warm-up phase due to the excessive magnitude of a certain parameter. The calculation formula is as follows: , .

[0079] As one possible implementation, the training hyperparameters involved in the joint training of network parameters are set as follows: For the preheating training of the second branch of the material mapping subnetwork, the preheating convergence threshold is set to... The initial learning rate is set to [value] for the local training phase in step S322, which involves training the parameters of the frozen material mapping subnetwork and performing local forward propagation calculations on the spatiotemporal temperature prediction subnetwork. The warm-up iteration strategy is not enabled; for the joint training of global parameters of the decoupled physical information neural network, the initial learning rate is set to... The number of preheating iteration steps is set to Total loss function The global convergence condition is set to continuous. The decrease in the loss function value within each iteration step is less than .

[0080] In step S3 of this embodiment, an adaptive gradient adjustment coefficient is introduced to aggregate data-driven loss, physical constraint loss, and temperature measurement compensation loss to construct a multi-source composite loss function. The instantaneous gradient balance coefficient is obtained by independently calculating the ratio of the gradient norm 2 of the partial derivative matrix of each sub-loss function with respect to the shared hidden layer. This adjustment coefficient is dynamically updated using an extreme value truncation and exponential moving average filtering mechanism to balance the backpropagation update amplitude of each loss. Specifically, for example: assuming that in the... In the next iteration, after extracting the anchor parameter tensor of the shared fully connected hidden layer, the gradient norm 2 of the data-driven loss is calculated using an automatic differentiation framework. Gradient L2 norm of physical constraint loss And take an infinitesimal floating-point constant to prevent the denominator from being zero. Calculate the instantaneous gradient balance coefficient corresponding to the first constraint loss, and substitute it into the formula: Set a lower threshold for extreme value cutoff. With upper limit threshold Since the calculation results are within the threshold range, the process of generating truncated instantaneous coefficients using the extreme value truncation operator is as follows: Further assume the adjustment coefficient of the previous iteration step. Momentum smoothing decay constant The final adaptive gradient adjustment coefficient is updated using an exponential moving average filtering mechanism. The calculation process is as follows: This computational mechanism can suppress the divergence of adjustment coefficients caused by excessively small physical constraint gradients in a single iteration, suppress high-frequency transient fluctuations in a multi-objective gradient game environment, and maintain the smooth convergence of the global joint training gradient.

[0081] S4. After dimensionality reduction of the mix proportion parameters, embed them into a non-dominated sorting genetic algorithm. With the dual objectives of maximizing the apparent quality index and minimizing material cost, a collaborative constraint of apparent quality-temperature gradient calibrated using historical data is introduced on top of the absolute constraints. Evolution is driven by a feasibility-first strategy, and finally, the optimal mix proportion scheme is automatically selected from the Pareto front using a sorting method that approximates the ideal solution. Specifically, the following sub-steps are included: S41. Perform dimensionality reduction transformation of multi-objective optimization variables and establish the optimization boundary, which is divided into the following sub-steps: S411, Mixing ratio parameters These are set as control variables for multi-objective optimization. To avoid the curse of dimensionality reduction and algorithmic infeasibility caused by absolute volume equality constraints, the absolute volume equality is applied to reconstruct the optimization variables to eliminate redundant degrees of freedom. From the control variables... Selecting the amount of fine aggregate Cement usage is retained as the dependent variable. Water consumption fly ash dosage Slag powder dosage , Coarse aggregate dosage Dosage of water-reducing agent A total of 6 parameters are independent optimization variables, and a set of dimensionality reduction control variables is constructed. Dependent variable: Fine aggregate dosage Uniquely determined by the absolute volume equation, and after logical corrections based on dimensional consistency and physical conservation, the calculation formula is as follows: ; In the formula, The volume fraction of air content in the design of dimensionless concrete is given. , , The true densities of cement, fly ash, and slag powder are respectively determined using the anhydrous kerosene discharge method, and the units are 1. ; and The apparent densities of fine and coarse aggregates are respectively determined using the volumetric flask method combined with the saturated surface-dry condition. The units are... ; and The densities of water and water-reducing agent were determined using a precision liquid density meter, and the units are... The reconstructed set of dimensionality-reduced control variables Embedded in chromosome encoding of non-dominated sorting genetic algorithms.

[0082] As one possible implementation method, the design air content by volume fraction of concrete The determination method is as follows: Based on the freeze-thaw resistance requirements of the natural environment where the target project is located and the maximum nominal particle size of the coarse aggregate used, refer to the "Standard Limit Table for Air Content in Concrete" in the "Specification for Mix Proportion Design of Ordinary Concrete" (such as JGJ55) to obtain the baseline percentage value. (e.g., 4.0%). Convert the unit of measurement to a dimensionless volume fraction using the following formula: .

[0083] S412. Determine the dual objective function for multi-objective optimization: maximize the predicted appearance quality index. (fitness) Minimize concrete material costs (fitness) The material cost calculation equation is as follows: , This corresponds to the market purchase price of the basic raw materials. Before each forward inference, the complete mix proportion parameters are reconstructed using the dependent variable equation. Input the data into the trained proxy evaluation model in step S323 to obtain the standardized compressive strength prediction value. Compared with the predicted value of the apparent quality index .

[0084] S413. Establish the physical boundary and constraint set for the optimization process: For independent optimization variables, apply hard boundary constraints of upper and lower quality limits given by engineering experience. .

[0085] Regarding the dependent variable, the amount of fine aggregate used... To prevent the dependent variable from yielding a negative value or exceeding the reasonable aggregate ratio, a reasonable quality range constraint based on historical data distribution is applied. .

[0086] Establish water-cement ratio inequality constraints .

[0087] Predicted values ​​of standardized compressive strength using the extended extreme value range set in step S131 Perform de-standardization operations to restore actual physical dimensions and establish mandatory lower limit constraints for engineering strength. The inverse standardization equation is as follows: In the formula, The minimum engineering target compressive strength is defined.

[0088] As one possible implementation method, the minimum engineering target compressive strength The method for determining this is as follows: extract the training set of multi-source material data. Actual compressive strength Sample statistical standard deviation The target compressive strength is derived using the quantiles of the normal distribution with a 95% probability guarantee rate. The calculation formula is as follows: In the formula, This is the guarantee rate coefficient; The value represents the characteristic value of the concrete strength grade based on structural design requirements, expressed in MPa. This formula ensures that the generated mix proportions accurately meet the load safety redundancy requirements of the engineering structure from a mathematical and statistical perspective.

[0089] As one possible implementation method, the upper and lower limits of the water-cement ratio and The method for determining it is as follows: Upper limit of water-to-glue ratio Constrained by both structural durability standards and mechanical strength requirements, the maximum water-cement ratio extreme value specified in the "Code for Design of Durability of Concrete Structures" for the current environmental impact level of the project is extracted. Based on the mechanical water-cement ratio limit derived from the Bolomey strength formula, the minimum value between the two is taken, and the calculation formula is as follows: In the formula, and These are regression coefficients (e.g., 0.53 and 0.20 for crushed stone), dimensionless; The measured compressive strength of the cementitious material after 28 days is expressed in MPa.

[0090] Water-to-gel ratio lower limit Constrained by the theoretical water requirement for complete hydration of cementitious materials and the limitations of historical pumping processes, a training set of multi-source material data was extracted. The lowest water-cement ratio ever recorded in China The dimensionless minimum water requirement limit constant for the complete hydration of silicate cement system. The two values ​​are compared, and the maximum value is taken. The calculation formula is as follows: .

[0091] As one possible implementation method, the lower limit of quality for independent control variables and upper limit The method for determining this is as follows: To ensure that the search space of the non-dominated sorting genetic algorithm is strictly within the reliable interpolation interval of the surrogate evaluation model trained in step 3, and to prevent unpredictable numerical divergence during distribution extrapolation, the optimization boundary is bound to the physical distribution domain of the training data. The expanded multi-source material data training set is then extracted through a traversal process. The Middle Historical minimum value of raw material usage Compared to historical maximum The values ​​are directly assigned to the boundary constraints of the optimization algorithm, i.e.: , This operation uses the historical prior distribution of data to force the blocking algorithm to generate physically absurd solutions such as water-cement ratio distortion or aggregate ratio imbalance.

[0092] As one possible implementation method, the dependent variable is the lower limit of the fine aggregate dosage. The upper limit of fine aggregate usage as the dependent variable The determination method is as follows: Traverse the multi-source material data training set expanded in step S1. Obtain all historical samples. This is for the training set of multi-source material data. The first in Extract the corresponding historical fine aggregate usage from each historical sample. Historical coarse aggregate usage Calculate the dimensionless sand ratio parameter of this historical sample. The calculation formula is as follows: Extract the sand ratio parameter set of all historical samples, obtain the historical minimum value in this set, and assign it as the benchmark minimum sand ratio. Obtain the historical maximum value in this set and assign it as the benchmark maximum sand ratio. In each optimization iteration of the non-dominated sorting genetic algorithm, the coarse aggregate dosage, an independent optimization variable already generated in the current chromosome encoding, is extracted. Introducing a benchmark minimum sand ratio Compared with the benchmark maximum sand ratio Based on the current coarse aggregate usage Perform inverse algebraic transformations to calculate the lower limit of fine aggregate content as the dependent variable under the current mix proportion. The upper limit of fine aggregate usage as the dependent variable The calculation formula is as follows: , In the formula, This represents the amount of coarse aggregate generated independently by the algorithm in the current optimization iteration step, in units of... .

[0093] S42. Perform sparsification temperature gradient assessment and collaborative crack prevention constraint determination, which is divided into the following sub-steps: S421. Extract the geometric center coordinates, surface key point coordinates, and discrete spatial coordinate set of the embedded temperature sensor determined in step S113 from the specimen. This constitutes a set of discrete nodes in the reduced-order space. The timeline covering the period of intense hydration was extracted, defining the time domain as 1 to 3 days after concrete pouring. The number of days was converted to seconds, and discretization sampling was performed at a set time step to generate a set of discrete time nodes. Discretize the set of nodes in space. With time discrete node set Perform a Cartesian product combination to generate a fixed-dimensional sparse reduced-order gradient evaluation domain matrix. This avoids algorithm redundancy and inference time expansion caused by global dense grid computing.

[0094] S422. Extract the complete mix ratio parameters restored after dimensionality reduction and reconstruction. Import the second branch of the material mapping subnetwork to obtain the hydration intrinsic parameters. and . Parameter and Together with the sparsified reduced-order gradient evaluation domain matrix The node coordinates are fed into the spatiotemporal temperature prediction subnetwork. Local forward inference and automatic differentiation are performed within the deep learning framework to directly obtain the spatiotemporal temperature prediction subnetwork output with respect to spatial coordinate variables. The partial derivative vector is calculated, and the L2 norm of the three-dimensional temperature gradient at each spatial discrete node is calculated. Extract the global maximum three-dimensional temperature gradient L2 norm within the evaluation domain. Establish an absolute constraint on the critical upper limit of crack prevention: In the formula, The critical temperature gradient threshold constant used to prevent thermal stress cracks in concrete structures, expressed in K / m.

[0095] As one possible implementation method, the critical temperature gradient threshold for crack prevention. The ultimate temperature gradient that concrete can withstand without cracking under internal thermal stress is characterized. It is determined by extracting the standard value of axial tensile strength from the standard physical test report of the concrete in the target project design. (Unit: MPa) Static compressive modulus of elasticity (Unit: MPa) and the coefficient of linear expansion of concrete (Unit: 1 / K). Extract the geometric model of the target component and calculate the effective heat dissipation feature length. (Unit: m), defined as the shortest straight-line distance from the geometric center of the component to the outermost heat dissipation surface. Using the constitutive equations of thermal stress and ultimate tensile strain in mechanics of materials, the critical temperature gradient threshold for crack prevention is quantitatively calculated. The calculation formula is as follows: .

[0096] S423. Call the predicted apparent quality index value output by the first branch of the material mapping subnetwork in step S412. Extract the maximum three-dimensional temperature gradient L2 norm calculated in step S422, establish a synergistic empirical inequality constraint equation between apparent quality and normalized temperature gradient, and use it as a physical defense against apparent microcracks and color difference degradation. The calculation formula is as follows: ; In the formula, the intercept term Characterizing the background apparent quality level without the influence of a temperature gradient This is a dimensionless temperature gradient sensitivity coefficient. This constraint operation ensures that, in pursuit of high apparent quality (objective function...),... At the same time, the model spontaneously suppresses the physical heating limit, and due to the forced definition This ensures that the upper limit of the calculation of the exponential decay term is always less than or equal to 1, and ensures that the cooperative inequality constraint has a non-empty physical feasible solution space.

[0097] As one possible implementation, the intercept term Sensitivity coefficient to temperature gradient The method for determining it is as follows: traverse the expanded multi-source material data training set constructed in step S1. Extract all The first historical sample. (Regarding the first...) For each sample, extract its calculated True Apparent Quality Index. Take its corresponding mix proportion parameters The input is fed into the spatiotemporal temperature prediction subnetwork trained in step S3, and forward inference and automatic differentiation are performed to obtain the global maximum three-dimensional temperature gradient norm 2 of the sample in the reduced-order gradient evaluation domain. The threshold calculated above. Calculate the first Normalized temperature gradient of each sample The calculation formula is as follows: Based on the extracted data above... We construct the basic linear partial least squares regression equation under unconstrained conditions. First, we calculate the sample mean of the normalized temperature gradient. Sample mean of the logarithm of the apparent quality index : , ; Calculate the foundation coefficients under unconstrained conditions and : , ;because Characterizes the background apparent quality level without the influence of a temperature gradient, and the maximum theoretical value of the apparent quality index is constrained to be... Therefore The upper limit of the physical extreme value is strictly 1 To ensure the existence of a non-empty physically feasible region for the constraint equations, a forced projection operation based on the Karush-Kuhn-Tucker (KKT) conditions is performed: like This indicates that the basic fit satisfies the physical and logical constraints, and the parameter values ​​can be directly assigned: , ; like This indicates that the fitting result has exceeded the physical limit, and a forced activation of the effective set strategy is adopted to strictly anchor the intercept term to the upper limit of the physical extreme value (i.e., let...). The temperature gradient sensitivity coefficient, which minimizes the sum of squared residuals, is then recalculated under this hard boundary constraint. The formula is as follows: , .

[0098] S43. Based on feasibility priority, the Pareto front optimization method (TOPSIS) is used to determine the final concrete mix design, which is divided into the following sub-steps: S431. Extracting the strength constraint function Temperature gradient constraint function Collaborative experience constraint function and the fine aggregate dependent variable out-of-bounds constraint function Calculate the overall constraint violation degree. The calculation formula is as follows: ; In the non-dominated sorting genetic algorithm, a feasibility-first constraint handling strategy is adopted. The following decision criteria are followed in population selection and evolutionary operations: When individual of And individuals of At that time, determine the feasible solution for each individual. Individuals superior to infeasible solutions And prioritize retaining feasible solutions; When individual With individuals All are infeasible solutions (i.e.) When comparing the total degree of constraint violation, priority is given to retaining the violation. Smaller individuals; When individual With individuals When all solutions are feasible, based on the objective function (predicting the apparent quality index)... Cost of concrete materials The algorithm performs non-dominated sorting and crowding distance calculation. After a set number of generations of crossover and mutation optimization operations, the evolutionary algorithm outputs a set of non-dominated fits that are at the Pareto front.

[0099] S432. After the algorithm stops iterating and generates the final static Pareto front set, extract all solutions from the front solution set. The objective function values ​​for each individual. This will be used to predict the apparent quality index. Set as a benefit-oriented objective function The cost of concrete materials Set as a cost-type objective function To eliminate the discrepancy between the dimensions and the direction of objective optimization, the extracted objective function values ​​are normalized using range and extremum normalization, and all objectives are uniformly converted into benefit-type dimensionless parameters. ,in Represents the index of the objective function. This represents the index of an individual in the Pareto front set. The normalization calculation formula is as follows: ; ; In the formula, To prevent the occurrence of a minimal shift constant for meaningless poles in subsequent logarithmic operations, a value of is taken. .

[0100] Calculate the first The individual in the first Feature proportions under each objective and the Information entropy of a target The calculation formula is as follows: ; ; Based on information entropy Calculate the objective weight parameters of the two objective functions. The calculation formula is: This ensures that the target weights are entirely driven by the objective data dispersion at the optimization frontier.

[0101] S433. Decision-making using the Top-Optimal Solution Ranking Method (TOPSIS): Utilizing objective weight parameters. Multiply by normalization parameter Constructing a weighted normalized decision matrix In the weighted normalized decision matrix, the absolute positive ideal solution for each objective is extracted. With absolute negative ideal solution The calculation formula is: , Calculate the Euclidean distance between each individual in the Pareto front set and the positive ideal solution. and the Euclidean distance to the negative ideal solution : ; ; Calculate the relative proximity coefficients between each solution and the positive and negative ideal solutions. The calculation formula is as follows: ; for all Relative proximity coefficient of individuals The mix proportions were sorted in descending order, and the individual mix proportion with the highest relative similarity coefficient was selected as the final mix proportion scheme for the mirror concrete construction.

[0102] In step S4 of this embodiment, the mix proportion parameters are dimensionality-reduced and embedded into a multi-objective optimization algorithm. A surrogate evaluation model is called to predict the global maximum three-dimensional temperature gradient norm 2. The critical temperature gradient threshold for crack prevention is calculated based on the material constitutive mechanical parameters to establish absolute constraints. Based on this, a synergistic empirical inequality constraint equation between apparent quality and normalized temperature gradient is constructed, thereby selecting feasible solutions that meet the actual physical environment and appearance requirements. Specifically, for example: assuming the standard value of axial tensile strength is extracted... Static compressive modulus of elasticity and the coefficient of linear expansion of concrete Simultaneously, the geometric model of the target component is extracted, and the effective heat dissipation characteristic length is calculated. Using the constitutive equations for thermal stress and ultimate tensile strain in mechanics of materials, the critical temperature gradient threshold for crack prevention is quantitatively calculated. Substituting these values ​​into the formula, the calculation is as follows: Furthermore, assuming that for a given mix design, the surrogate evaluation model predicts the global maximum three-dimensional temperature gradient L2 norm. The intercept term obtained by calibrating historical data and temperature gradient sensitivity coefficient And the individual's predicted apparent quality index Calculate the right-hand side attenuation limit of the cooperative empirical inequality constraint equation between apparent quality and normalized temperature gradient, and substitute it into the formula for calculation: Because of the individual's predicted apparent quality index Greater than the above-mentioned constraint attenuation extreme value Based on the constraint violation determination logic, the individual is determined to be a feasible solution that satisfies the cooperative empirical inequality constraint equation. This operation introduces rigorous thermodynamic and physical law verification into the multi-objective optimization process, forcibly filtering out deteriorated mix proportions that have a very high theoretical probability but will cause surface cracking due to excessive temperature stress during actual pouring, thereby ensuring the structural integrity and apparent reflective effect of the mirror concrete.

[0103] Example 2 As a second embodiment of the present invention, based on embodiment 1, this embodiment also discloses a machine learning-based intelligent optimization system for the mix proportion of mirror concrete, specifically including: The tag building module is used to collect multi-source feature data to construct the appearance quality index. And the apparent quality index and actual compressive strength As an optimization target label, it is injected into the multi-source material data training set. ; The network building module is used to construct decoupled physical information neural networks, utilizing the coordination ratio parameter. Mapping hydration intrinsic parameters and To output the predicted spatiotemporal temperature field And construct the physical residual function Extracting temperature statistical feature scalar vector With mix proportion parameters Fusion, output standardized predicted compressive strength value and the predicted value of the apparent quality index ; The joint training module is used to train multi-source material data sets. Decoupled physical information neural network with spatiotemporal locus data input, combined with measured temperature field data. Physical boundary conditions and physical residual functions Constructing a multi-source composite loss function Train and generate an agent evaluation model; The multi-objective optimization module is used to reduce the set of control variables. An embedded multi-objective optimization algorithm is used for iterative optimization, and a proxy evaluation model is invoked to predict the global maximum three-dimensional temperature gradient L2 norm. Establish temperature-coordinated crack prevention constraints, generate a non-inferior mix design solution set, and select and output the mix design scheme for mirror concrete construction.

[0104] In the specific implementation of Embodiment 2 above, the label construction module obtains the total light reflectance of the historical specimens of mirror concrete. Surface pore distribution entropy and root mean square color difference This is transformed into an apparent quality index with quantitative representation. This module will include the apparent quality index. Together with the measured actual compressive strength They were integrated together as absolute labels into the multi-source material data training set. The network construction module receives the mix ratio parameters. The hydration exothermic shape parameter with positive definite properties is output through the kinetic parameter mapping branch. With hydration characteristic time parameters The spatiotemporal temperature prediction subnetwork constructs an explicit hydration heat release rate function based on the aforementioned parameters. Forward inference generates and predicts the spatiotemporal temperature field. The physical residual function is output by combining automatic differentiation technology. This module further extracts the predicted spatiotemporal temperature field. The static surface temperature sequence is processed and dimensionality is reduced by a global attention pooling layer to generate a scalar vector of temperature statistical features. Combine it with the mix proportion parameters After splicing, input the performance prediction branch and output the normalized compressive strength prediction value. Compared with the predicted value of the apparent quality index The joint training module extracts the set of internal physical spatiotemporal coordinates within a defined spatiotemporal domain. Boundary spacetime coordinate set and the initial spatial collocation set This module will train a multi-source material data set. Data-driven error, physical residual function Physical constraint error, measured temperature field data Error fusion compensation is achieved by using an adaptive gradient adjustment mechanism to update weight coefficients and construct a multi-source composite loss function. This completes the global joint training of network parameters and saves the converged surrogate evaluation model. The multi-objective optimization module applies the absolute volume equation to the mix proportion parameters. Dimensionality reduction is performed to obtain a set of dimensionality-reduced control variables. This is embedded into the chromosome encoding of a non-dominated sorting genetic algorithm, and a surrogate evaluation model is invoked during the bi-objective optimization iteration to predict the global maximum three-dimensional temperature gradient L2 norm. Combined with the critical temperature gradient threshold for crack prevention Establish constraint equations based on absolute constraints and collaborative empirical inequalities, and base them on the overall constraint violation degree. Prioritize retaining feasible solutions, generate a set of non-dominated solutions at the Pareto front, and calculate the relative proximity coefficient of the individuals using the approximation of ideal solutions ranking method. The individual ranked first will be selected to produce the final concrete mix design for the mirror concrete construction.

[0105] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A machine learning-based intelligent optimization method for the mix proportion of mirror concrete, characterized in that, include: The total light reflectance, surface pore distribution entropy, and root mean square color difference of historical mirror concrete specimens were collected to construct an apparent quality index. The apparent quality index and actual compressive strength were then injected into the multi-source material data training set as optimization target labels. First, a decoupled physical information neural network containing a material mapping subnetwork and a spatiotemporal temperature prediction subnetwork is constructed. Then, the material mapping subnetwork is used to map the mix proportion parameters into hydration intrinsic parameters. Based on the mix proportion parameters and the hydration intrinsic parameters, a hydration heat release rate function and a physical residual function are constructed. Then, the spatiotemporal temperature prediction subnetwork is used to output a predicted spatiotemporal temperature field based on the hydration heat release rate function. The temperature statistical features of the predicted spatiotemporal temperature field are extracted and concatenated with the mix proportion parameters, and then input back into the material mapping subnetwork to output the normalized compressive strength prediction value and the apparent quality index prediction value. The multi-source material data training set and spatiotemporal point data are input into the decoupled physical information neural network. A multi-source composite loss function is constructed by combining the measured temperature field data, physical boundary conditions and the physical residual function. The decoupled physical information neural network is jointly trained based on the multi-source composite loss function to generate a proxy evaluation model that can reflect the spatiotemporal evolution law of hydration heat release. The mix proportion parameters are reconstructed by dimensionality reduction. The reconstructed mix proportion parameters are embedded as optimization variables in a multi-objective optimization algorithm. The algorithm iterates optimization with the objective functions of maximizing the predicted value of the apparent quality index and minimizing the material cost. The proxy evaluation model is called to predict the maximum temperature gradient. Based on the maximum temperature gradient, temperature-coordinated crack prevention constraints are established and a non-inferior mix proportion solution set is generated. A multi-attribute decision algorithm is used to select the target mix proportion individual that meets the critical temperature gradient threshold for crack prevention and has the highest apparent quality index from the non-inferior mix proportion solution set, and the mirror concrete construction mix proportion scheme is output.

2. The intelligent optimization method for the mix proportion of mirror concrete based on machine learning according to claim 1, characterized in that: The construction of the apparent quality index includes: Extract the individual pixel color difference of the original color image and calculate the root mean square of the color difference of the entire image; The original color image is subjected to grayscale and binarization processing, and the area ratio of pore pixels within the grid region is extracted to calculate the surface pore distribution entropy. The extracted total light reflectance and the root mean square of color difference are simultaneously injected into a uniform random perturbation to generate preliminary perturbation feature values. Based on historical extreme values, a physical boundary truncation operator is applied to the preliminary perturbation feature values ​​to generate physical consistency enhancement feature values. The physical consistency enhancement feature value, the surface pore distribution entropy, the actual compressive strength, the mix proportion parameters, and the measured temperature field data are tensor-stitched to expand and generate a multi-source material data training set. The actual compressive strength and apparent characteristic parameters in the training set of the multi-source material data are subjected to range standardization. The feature weight coefficients are solved based on the individual information utility values. The standardized apparent characteristic parameters are then weighted and superimposed according to the feature weight coefficients to construct the apparent quality index.

3. The intelligent optimization method for the mix proportion of mirror concrete based on machine learning according to claim 1, characterized in that, The material mapping subnetwork is used to map the input mix proportion parameters to hydration intrinsic parameters, including: The mixing ratio parameters are input into the kinetic parameter mapping branch of the material mapping subnetwork to perform feature extraction. The activation function is used to output the hydration exothermic shape parameter and the hydration characteristic time parameter with positive definite properties as the hydration intrinsic parameters. Based on the amount of each raw material in the mix proportion parameters, the multiphase medium mixing law is introduced to establish the thermal property parameter equation of concrete. The mass specific heat capacity constant is extracted to calculate the equivalent specific heat capacity, the absolute apparent density is extracted to calculate the aggregate volume fraction, and the mixed thermal conductivity constant is calculated by combining the single-phase thermal conductivity. Based on the equivalent specific heat capacity and the mixed thermal conductivity constant, the equivalent density and thermal diffusivity are calculated.

4. The intelligent optimization method for the mix proportion of mirror concrete based on machine learning according to claim 3, characterized in that, Based on the mix proportion parameters and the intrinsic hydration parameters, an explicit hydration heat release rate function and a physical residual function are constructed, including: Extract the historical sequence of the predicted spatiotemporal temperature field, and construct an integral mapping to calculate the equivalent age of spatial nodes; The cement content, fly ash content, and slag powder content in the mix proportion parameters are extracted, and an explicit hydration heat release rate function is constructed by combining the equivalent age with the intrinsic hydration parameters. The first-order partial derivatives and Laplace operator of the predicted spatiotemporal temperature field are obtained with respect to time and space, respectively. The physical residual function is constructed by combining the equivalent specific heat capacity, the equivalent density, the thermal diffusivity, and the explicit hydration heat release rate function.

5. The intelligent optimization method for the mix proportion of mirror concrete based on machine learning according to claim 1, characterized in that, The output of the standardized compressive strength prediction value and apparent quality index prediction value includes: The surface temperature field data at the moment of maximum temperature rise is extracted from the predicted spatiotemporal temperature field, and a gradient truncation operator is introduced to obtain the static surface temperature sequence. The static surface temperature sequence is input into a global attention pooling layer to extract the surface temperature standard deviation and the global maximum surface temperature difference, and then merged to generate a surface temperature statistical feature scalar vector. The mixing ratio parameter and the surface temperature statistical feature scalar vector are tensor-concatenated to generate a fused auxiliary feature tensor. The fusion auxiliary feature tensor is input into the performance prediction branch of the material mapping subnetwork for forward computation, and the normalized compressive strength prediction value and the apparent quality index prediction value are output using the activation function.

6. The intelligent optimization method for the mix proportion of mirror concrete based on machine learning according to claim 1, characterized in that, The construction of the multi-source composite loss function includes: Extract the internal physical spatiotemporal coordinate point set, the boundary spatiotemporal coordinate point set, and the initial spatial coordinate point set; Calculate the mean square error between the apparent quality label value and the network prediction value corresponding to the multi-source material data training set, and construct the data-driven loss. Extract the physical residual values ​​corresponding to the internal physical spatiotemporal coordinate point set to construct the physical constraint loss; Calculate the mean square error between the predicted temperature and the measured temperature field data, and construct a temperature measurement compensation loss; The deviation between the network-predicted initial temperature and the initial temperature of the mold pouring is calculated to construct the initial condition loss. Based on the laws of heat conduction and cooling, the continuity residuals of the outward heat conduction flux and convective heat dissipation flux of the boundary are calculated on the set of spatiotemporal loci at the boundary to construct the boundary condition loss. An adaptive gradient adjustment coefficient is introduced to aggregate the data-driven loss, the physical constraint loss, the temperature measurement compensation loss, the initial condition loss, and the boundary condition loss to construct a multi-source composite loss function.

7. The intelligent optimization method for the mix proportion of mirror concrete based on machine learning according to claim 6, characterized in that, The joint training of the parameters of the decoupled physical information neural network includes: Calculate the partial derivative matrix of each sub-loss function with respect to the shared fully connected hidden layer in the material mapping sub-network independently, and calculate the instantaneous gradient balance coefficient based on the gradient 2 norm ratio of the partial derivative matrix; The extreme value truncation operator is applied to generate truncation instantaneous coefficients, and the adaptive gradient adjustment coefficients are updated in combination with the exponential moving average filtering mechanism. The parameters of the training set of the multi-source material data are extracted to construct a physical empirical estimation model of the intrinsic hydration parameters. A preheating loss function is constructed using the physical empirical initial value penalty term, and preheating training is performed on the material mapping sub-network. Freeze the trainable parameters of the material mapping subnetwork, perform local forward propagation calculation using the target alignment benchmark corresponding to the temperature measurement compensation loss, extract the average value of the adaptive gradient adjustment coefficient before convergence, and construct a relative proportion prior vector. Unfreeze the trainable parameters of the material mapping subnetwork, initialize the adaptive gradient adjustment coefficients using the relative proportion prior vector, initiate the global parameter joint training of the decoupled physical information neural network, and output the surrogate evaluation model.

8. The intelligent optimization method for the mix proportion of mirror concrete based on machine learning according to claim 1, characterized in that, The dimensionality reduction reconstruction of the mix proportion parameters includes: The mix proportion parameters are reconstructed using the absolute volume equation. The amount of fine aggregate is selected as the dependent variable determined by the absolute volume equation, and the remaining mix proportion parameters are retained as independent optimization variables. Together, they form a set of dimensionality-reduced control variables that are embedded in a non-dominated sorting genetic algorithm. Apply hard boundary constraints on the upper and lower limits of quality, apply reasonable range constraints on the dependent variable quality based on historical data distribution, and apply water-cement ratio inequality constraints in combination with hydration conditions to establish physical optimization boundaries; The target compressive strength is derived based on the statistical standard deviation of the sample, and a mandatory lower limit constraint on the engineering strength is applied.

9. The intelligent optimization method for the mix proportion of mirror concrete based on machine learning according to claim 1, characterized in that, The output of the concrete mix design plan includes: The spatial discrete node set and the temporal discrete node set of the combined specimen are used to generate a sparsed reduced-order gradient evaluation domain matrix. The three-dimensional temperature gradient norm 2 is calculated based on the partial derivative vector of the predicted spatiotemporal temperature field with respect to spatial coordinate variables, and the global maximum three-dimensional temperature gradient norm 2 is extracted within the sparsified reduced-order gradient evaluation domain matrix. The critical temperature gradient threshold for crack prevention is calculated based on the constitutive equation, and an absolute constraint on the upper limit of the critical temperature gradient for crack prevention is constructed. Based on the background apparent quality level and the temperature gradient sensitivity coefficient, an exponential mathematical relationship between the predicted apparent quality index and the normalized temperature gradient is constructed as a cooperative empirical inequality constraint equation. The overall constraint violation degree of the optimization individual is calculated to prioritize the retention of feasible solution individuals, and a static Pareto front set is generated based on the non-dominated ranking and crowding distance as the non-dominated solution set; The objective function values ​​of individuals in the non-inferior mix ratio solution set are normalized, and objective weight parameters are generated by combining the objective function dispersion to construct a weighted normalized decision matrix. Based on the weighted normalized decision matrix, the relative proximity coefficient between each feasible solution individual and the absolute positive and negative ideal solutions is calculated. The individual ranked first is determined as the target mix proportion individual, and the mirror concrete construction mix proportion scheme is output.

10. A machine learning-based intelligent optimization system for the mix proportion of mirror concrete, employing the intelligent optimization method for the mix proportion of mirror concrete as described in any one of claims 1-9, characterized in that, include: The label building module is used to construct the apparent quality index and inject the apparent quality index and actual compressive strength as optimization target labels into the multi-source material data training set; The network construction module is used to build a decoupled physical information neural network. It uses the mix proportion parameters to map the hydration intrinsic parameters to output the predicted spatiotemporal temperature field and constructs the physical residual function. It extracts temperature statistical features and fuses them with the mix proportion parameters to output the standardized predicted values ​​of compressive strength and apparent quality index. The joint training module is used to input multi-source material data training set and spatiotemporal locus data into a decoupled physical information neural network, and combine measured temperature field data, physical boundary conditions and physical residual functions to construct a multi-source composite loss function, and train to generate a proxy evaluation model; The multi-objective optimization module is used to embed the reconstructed mix proportion parameters into a multi-objective optimization algorithm for optimization iteration, call the proxy evaluation model to predict the maximum temperature gradient to establish temperature-coordinated crack prevention constraints, generate a non-inferior mix proportion solution set, and select and output the mirror concrete construction mix proportion scheme.

Images