Concrete mix proportion design method and system based on neural network and gradient optimization

Abstract

The invention provides a concrete mix proportion design method and system based on a neural network and gradient optimization, and relates to the technical field of concrete mix proportion design, and the method comprises the steps: collecting actual production data and corresponding performance parameters of multiple groups of concrete mix proportions in the same region; establishing a variational auto-encoder model based on the data; extracting a latent variable space and a decoder part as a generative model; establishing a deep learning prediction model of concrete performance, and training the model by adopting the training set; a deep learning prediction model is obtained as an evaluation model; generating a plurality of groups of mix proportion samples through random sampling and decoding operation of latent variables; and based on multi-path gradient descent search, obtaining multiple groups of mix proportion schemes which meet performance requirements and have differentiated parameter distribution. The method has the advantages that different engineering requirements are met through multiple latent variable initialization and gradient descent search, limitation of a single solution is avoided, and the actual application effect of the design scheme in construction is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of concrete mix design technology, and more specifically, to a concrete mix design method and system based on neural networks and gradient optimization. Background Technology

[0002] Traditional concrete mix design relies primarily on empirical formulas (such as the water-cement ratio law and the ternary equilibrium method) and repeated experiments. It determines the optimal proportions of each component (cement, aggregates, mineral admixtures, water, etc.) through engineers' practical experience and extensive trial data. While this method is widely used in practical engineering, its limitations in design efficiency, adaptability, and accuracy are becoming increasingly apparent as engineering projects place higher performance demands on concrete materials (such as the synergistic effect of multiple objectives including strength, durability, workability, and cost) and as raw materials become increasingly diverse. Specifically, traditional methods struggle to effectively characterize the complex nonlinear mapping relationship between mix proportions and performance, and cannot adequately meet the demands for efficient and precise design.

[0003] In recent years, data-driven surrogate models combined with heuristic optimization algorithms have been widely applied to concrete mix design optimization. These methods typically employ regression models (such as support vector regression, random forest, gradient boosting tree, etc.) to establish a nonlinear mapping relationship between mix proportion and performance, and combine them with swarm search strategies such as genetic algorithms and particle swarm optimization to perform iterative optimization within the design space to obtain mix proportion solutions that satisfy multiple objectives or complex constraints. However, these methods still face many challenges, mainly in the following aspects: First, these algorithms lack gradient information and rely primarily on swarm intelligence for searching, making them prone to getting trapped in local optima in high-dimensional non-convex spaces, resulting in poor stability and convergence of the optimized solution; Second, heuristic optimization methods heavily rely on multiple evaluations of the objective function, requiring the use of computationally expensive performance prediction models in each iteration to select the optimal solution. As the population size and number of iterations increase, the computational cost rises significantly, thus extending the optimization time and making it difficult to meet the needs of engineering applications for efficient optimization; Furthermore, heuristic algorithms rely entirely on the objective function to drive optimization, lacking explicit modeling of the coupling relationship between concrete mix proportion parameters and engineering physical constraints, leading to insufficient coordination among variables and a tendency for compensatory imbalances. Summary of the Invention

[0004] The purpose of this invention is to provide a concrete mix design method and system based on neural networks and gradient optimization, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a concrete mix design method based on neural networks and gradient optimization, including: Step S100: Collect actual production data and corresponding performance parameters of multiple concrete mix proportions in the same region, including the compressive strength, workability and material cost of concrete; normalize the original data of the mix proportions, and the normalized data includes the performance data parameters of concrete strength, workability and cost corresponding to the raw material usage. Step S200: Based on the normalized concrete mix proportions from step S100, a variational autoencoder model is established. The variational autoencoder model uses the normalized data as input. The encoder extracts features layer by layer and maps them to a continuous, regularized low-dimensional latent space. The decoder generates new concrete mix proportion samples from the latent variables. The training framework of this model uses the PyTorch deep learning platform, and the model achieves optimal convergence by minimizing the reconstruction error and the KL divergence of the latent variable distribution. After training, the latent variable space and the decoder are extracted as the generative model. Step S300: Based on the normalized mix proportion data and corresponding performance data from step S100, a deep learning prediction model for concrete performance is established. Concrete performance includes strength, workability, and cost. The prediction model is constructed using PyTorch. The input layer takes the feature parameters of the normalized mix proportion data as input, and the output layer corresponds to the target performance index of the concrete. The model is trained using a training set, and the weights and biases of each layer are adjusted through the backpropagation algorithm, with the mean square error as the loss function. After training, the deep learning prediction model is obtained as the evaluation model. Step S400: Perform comprehensive performance verification on the generative model obtained in step S200 and the evaluation model obtained in step S300. In the generative model verification, multiple sets of mix proportion samples are generated through random sampling and decoding of latent variables, and it is verified whether their key ratios of water-cement ratio, sand ratio, admixture dosage ratio, fly ash dosage ratio, and mineral powder dosage ratio are within the reasonable range for engineering. In the evaluation model verification, the test set data is input into the trained evaluation model to predict its compressive strength, workability, and cost indicators, and the results are compared and analyzed with the measured values. Step S500: Connect the generated model obtained in step S200 with the evaluation model obtained in step S300 to form an integrated design framework; establish an optimization function based on engineering requirements with the goal of achieving target performance and minimizing cost, where the target performance includes strength and workability; obtain the gradient information of the optimization objective function with respect to latent variables through automatic differentiation technology, and iteratively solve the problem using the L-BFGS second-order approximate gradient optimization algorithm based on the gradient information to obtain the optimal concrete mix proportion solution in the latent variable space; perform multiple different initializations of the latent variables, and obtain multiple mix proportion schemes that meet the performance requirements and have different parameter distributions based on multi-path gradient descent search.

[0005] Preferably, the actual production data of the mix proportion includes the amount of cement, mineral powder, fly ash, water, water-reducing agent, fine aggregate, and coarse aggregate.

[0006] Preferably, the training process for establishing a variational autoencoder model based on this data includes: Using all normalized concrete mix proportions from actual production data as input, unsupervised learning is initiated to train the variational autoencoder model. Based on input data from actual production data, the encoder of the variational autoencoder maps the fully connected layer of the input concrete mix proportion data into the mean vector and standard deviation vector in the latent space, and then transforms them into latent variables through reparameterization techniques. The decoder, based on these latent variables, maps back from the low-dimensional latent space to the original data space and reconstructs the mix proportion close to the original input data through nonlinear transformation. The optimization process of the decoder is carried out by maximizing the KL divergence between the reconstruction error and the latent variable distribution. Based on the working process of the encoder and decoder, the training objective of the variational autoencoder is determined to be minimizing the loss function; Based on a defined loss function, the Adam optimization algorithm is used for backpropagation during training to update the weights and biases in the encoder and decoder. The model iterates in multiple training epochs until the loss function converges. In each training epoch, the entire training set data is passed once, and the network parameters are updated based on the optimization. After the loss function converges and the model is trained, the decoder part of the variational autoencoder is extracted and its parameters are fixed. The decoder part is used for the generation of subsequent matching samples. Preferably, the process of training the evaluation model using the obtained deep learning prediction model as the evaluation model includes the following steps: The normalized mix ratio data and its corresponding performance data in step S100 are divided into training set and test set according to the proportion. The training set is used for model training and the test set is used for subsequent model validation. Based on the partitioned dataset, an evaluation model is constructed using deep learning methods, with a fully connected neural network as the basic architecture. The input of this neural network is the feature vector of the concrete mix proportion, which includes the dosage information of cement, aggregate and admixtures. The output is the target performance index of the concrete, which includes compressive strength, workability and cost. Based on the established model architecture, the training objective of the evaluation model is set as minimizing the difference between the predicted and actual values, and the mean squared error is used as the loss function. Based on a defined loss function, the evaluation model is trained using the training set data after step partitioning. The Adam optimization algorithm is used for backpropagation to optimize the model's weights and biases. The training process involves multiple training cycles, with each iteration traversing the entire training set and updating the network parameters based on the optimization, ultimately resulting in a fully trained evaluation model.

[0007] Preferably, step S400, which involves performance verification and rationality evaluation of the generated model, includes the following steps: In the generative model trained in step S200, multiple latent variables are generated by random sampling of latent variables, and then the latent variables are mapped to the corresponding concrete mix proportion data by the decoder in the generative model. The mix proportion data is evaluated to determine whether the key ratios of water-cement ratio, sand ratio, admixture dosage ratio, fly ash dosage ratio, and mineral powder dosage ratio are within the reasonable range for the project. If so, the evaluation is output; otherwise, the evaluation continues.

[0008] Preferably, step S400, which involves performance verification and rationality assessment of the evaluation model, includes the following steps: The data in the test set divided in step S500 is input into the evaluation model trained in step S300. The evaluation model is used to predict the corresponding target performance, where the target performance includes compressive strength, workability and cost. Based on the prediction results obtained in step S100, the root mean square error and the coefficient of determination are used as performance evaluation indicators.

[0009] Preferably, in step S500, the gradient optimization solution and solution set generation include the following steps: Based on the constructed optimization objective function, the gradient of the optimization objective function is calculated using automatic differentiation technology to obtain gradient information about the design variables, and then the gradient of the objective function with respect to the latent variables is calculated. Based on the gradient information of the objective function with respect to the latent variables obtained by calculation, the gradient optimization algorithm of L-BFGS second-order approximation is used to optimize the latent variables. Multiple iterations are performed in the latent variable space until the optimization process converges. According to the optimization process, multiple starting points of the latent variables are initialized multiple times, and multiple optimization solutions are found through gradient descent along different paths. During each initialization, the latent variables are optimized from random positions, and the L-BFGS algorithm is used for iteration. Based on the iteration results, multiple different mix proportion solutions are obtained, and finally multiple sets of concrete mix proportion solutions are generated.

[0010] Secondly, this application also provides a concrete mix design system based on neural networks and gradient optimization, comprising: Step S100, Data Acquisition Module: Used to collect actual production data and corresponding performance parameters of multiple concrete mix proportions in the same region, including the compressive strength, workability and material cost of concrete; and to normalize the raw data of the mix proportions, including the performance data parameters of concrete strength, workability and cost corresponding to the raw material usage. Step S200: Establishing a module: This module is used to establish a variational autoencoder model based on the normalized concrete mix proportions from step S100, using the actual production data. The variational autoencoder model takes the normalized data as input, and the encoder part extracts features layer by layer and maps them to a continuous, regularized low-dimensional latent space. The decoder part is used to generate new concrete mix proportion samples from the latent variables. The training framework of this model uses the PyTorch deep learning platform, and the optimization convergence of the model is achieved by minimizing the reconstruction error and the KL divergence of the latent variable distribution. After training, the latent variable space and the decoder part are extracted as the generative model. Step S300, Training Module: This module is used to establish a deep learning prediction model for concrete performance based on the normalized mix proportion data and corresponding performance data from step S100. Concrete performance includes strength, workability, and cost. The prediction model is built using PyTorch. The input layer takes the feature parameters of the normalized mix proportion data as input, and the output layer corresponds to the target performance index of the concrete. The model is trained using a training set, and the weights and biases of each layer are adjusted through the backpropagation algorithm, with the mean square error as the loss function. After training, the deep learning prediction model is obtained as the evaluation model. Step S400, Verification and Evaluation Module: This module is used to comprehensively verify the performance of the generative model obtained in step S200 and the evaluation model obtained in step S300. In the generative model verification, multiple sets of mix proportion samples are generated through random sampling and decoding of latent variables, and it is verified whether their key ratios of water-cement ratio, sand ratio, admixture dosage ratio, fly ash dosage ratio, and mineral powder dosage ratio are within the reasonable range for engineering. In the evaluation model verification, the test set data is input into the trained evaluation model to predict its compressive strength, workability, and cost indicators, and the results are compared and analyzed with the measured values. Step S500, Solving Module: This module connects the generated model obtained in step S200 with the evaluation model obtained in step S300 to form an integrated design framework; it establishes an optimization function based on engineering requirements, aiming to achieve target performance while minimizing cost, where target performance includes strength and workability; it obtains the gradient information of the optimization objective function with respect to latent variables through automatic differentiation technology, and iteratively solves the problem using the L-BFGS second-order approximate gradient optimization algorithm based on this gradient information to obtain the optimal concrete mix proportion solution in the latent variable space; it performs multiple different initializations of the latent variables, and obtains multiple mix proportion schemes that meet performance requirements and have different parameter distributions based on multi-path gradient descent search.

[0011] Thirdly, this application also provides a concrete mix design device based on neural networks and gradient optimization, comprising: Memory, used to store computer programs; A processor is used to implement the steps of the concrete mix design method based on neural networks and gradient optimization when executing the computer program.

[0012] Fourthly, this application also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described concrete mix design method based on neural networks and gradient optimization.

[0013] The beneficial effects of this invention are as follows: This invention effectively overcomes the limitations of traditional methods by combining generative and evaluation models, thereby improving design efficiency, accuracy, and solution diversity.

[0014] This invention first learns the latent distribution of concrete mix proportion data through a variational autoencoder (VAE), mapping the six-dimensional mix proportion parameters to a continuous and regularized two-dimensional latent space. After training, the decoder portion of the VAE is extracted and fixed as the generative model; subsequently, a fully connected neural network is constructed as the evaluation model to achieve a precise mapping from the mix proportion vector to the target performance (such as compressive strength and cost). The two are combined to form an end-to-end design network. Based on this, automatic differentiation technology can be used to efficiently calculate the gradient of the objective function with respect to the latent variables, and combined with the second-order gradient optimization algorithm L-BFGS, rapid iterative back-optimization within the latent space is achieved, thereby efficiently generating multiple mix proportion schemes that meet the target performance requirements. When applied to the field of concrete mix proportion design, this invention effectively overcomes the technical bottlenecks of traditional methods, such as low efficiency, susceptibility to local optima, single solution set, and poor feasibility. It can generate physically feasible, stable, and diverse design schemes without explicit constraints, significantly improving design efficiency and optimization accuracy. This provides an efficient, stable, and engineering-scalable solution for the intelligent and high-performance design of concrete materials.

[0015] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the concrete mix design method based on neural networks and gradient optimization as described in an embodiment of the present invention. Figure 2 This is the model architecture of the VAE model in the concrete mix design method based on neural networks and gradient optimization described in the embodiments of the present invention; Figure 3 This is the model architecture in the concrete mix design method based on neural networks and gradient optimization described in the embodiments of the present invention; Figure 4 The following is an evaluation diagram of the generative model in the concrete mix design method based on neural network and gradient optimization described in the embodiments of the present invention. (a) shows the original dataset in the distribution of water-cement ratio and sand ratio in different sample sets, (b) shows 4000 sets of samples generated based on VAE in the distribution of water-cement ratio and sand ratio in different sample sets, and (c) shows 10000 sets of samples generated based on VAE in the distribution of water-cement ratio and sand ratio in different sample sets. The left figure in (a), (b) and (c) shows the distribution of water-cement ratio, and the right figure shows the distribution of sand ratio. Figure 5 The evaluation model in the concrete mix design method based on neural network and gradient optimization described in the embodiments of the present invention is evaluated, wherein (a) the figure shows the correlation between the predicted and actual values ​​of the training set and test set in the prediction performance of the FNN model for 28-day compressive strength; (b) the figure shows the predicted value and error distribution of the entire dataset in the prediction performance of the FNN model for 28-day compressive strength. Figure 6 This refers to the network architecture for model concatenation and optimization in the concrete mix design method based on neural networks and gradient optimization described in this embodiment of the invention. Figure 7This is a schematic diagram of the concrete mix design system based on neural networks and gradient optimization described in this embodiment of the invention. Figure 8 This is a schematic diagram of the concrete mix design equipment based on neural networks and gradient optimization as described in an embodiment of the present invention.

[0018] In the diagram: 701, Acquisition module; 702, Establishment module; 703, Training module; 704, Verification and evaluation module; 705, Solution module; 800, Concrete mix design equipment based on neural network and gradient optimization; 801, Processor; 802, Memory; 803, Multimedia component; 804, I / O interface; 805, Communication component. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] Example 1:

[0022] This embodiment provides a concrete mix design method based on neural networks and gradient optimization.

[0023] This invention addresses the problems of low design efficiency, poor optimization stability, insufficient solution set diversity, and inadequate modeling of physical constraints between parameters in existing concrete mix design technologies. It proposes a differentiable and scalable intelligent design framework. Currently, traditional concrete mix design methods mainly rely on empirical formulas or heuristic algorithms, which suffer from high computational costs, unstable optimization processes, and difficulty in generating diverse design solutions, failing to fully consider the inter-component coupling relationships in the concrete mix. While existing surrogate models and heuristic optimization methods can provide relatively flexible solution sets, they still face technical bottlenecks such as getting trapped in local optima, long computation times, and poor design solution rationality.

[0024] See Figure 1 The figure shows that the method includes steps S100, S200, S300, S400 and S500.

[0025] S100. Collect actual production data and corresponding performance parameters of multiple concrete mix proportions in the same region. The performance parameters include the compressive strength, workability and material cost of the concrete. Normalize the original data of the mix proportions. The normalized data includes the performance data parameters of concrete strength, workability and cost corresponding to the raw material usage.

[0026] Understandably, in this step, data acquisition and preprocessing involve collecting actual production data and corresponding performance parameters for multiple concrete mix proportions from the same region, including concrete compressive strength, workability, and material costs. The raw data for these mix proportions are then normalized to eliminate dimensional differences and improve the stability of model training. The normalized data includes parameters such as raw material usage corresponding to concrete strength, workability, and cost.

[0027] In this step, the computational model for standardization is as follows: (1) in, x ' This represents the normalized variable value. x min and x max The minimum and maximum values ​​of the variable should be respectively identified.

[0028] S200. Based on the normalized concrete mix proportions from step S100, a variational autoencoder model is established. The variational autoencoder model uses the normalized data as input. The encoder extracts features layer by layer and maps them to a continuous, regularized low-dimensional latent space. The decoder is used to generate new concrete mix proportion samples from the latent variables. The training framework of this model uses the PyTorch deep learning platform. The model achieves optimal convergence by minimizing the reconstruction error and the KL divergence of the latent variable distribution. After training, the latent variable space and the decoder are extracted as the generative model.

[0029] Understandably, in this step, a variational autoencoder (VAE) model based on actual production data of concrete mix proportions is established. The VAE model uses the actual production data of concrete mix proportions as input. The encoder part extracts features layer by layer and maps them to a continuous, regularized low-dimensional latent space; the decoder part is used to generate new concrete mix proportion samples from the latent variables. The training framework of this model adopts the PyTorch deep learning platform, and the model achieves optimal convergence by minimizing the reconstruction error and the KL divergence of the latent variable distribution. After training, the latent variable space and the decoder part are extracted as a generative model for subsequent rapid generation and reconstruction of mix proportions.

[0030] In this embodiment, training the variational autoencoder model specifically includes the following steps: ① Using all normalized concrete mix proportions from actual production data as input, unsupervised learning is performed to train a variational autoencoder (VAE) model. Through learning from the training data, the model can extract the latent structure of the data and generate latent variable representations that can efficiently reconstruct the input data.

[0031] ② A variational autoencoder consists of two main parts: an encoder and a decoder. The encoder maps the input concrete mix proportion data (standardized feature vectors) to distribution parameters in the latent space through a series of fully connected layers, typically the mean vector and standard deviation vector. The encoder's goal is to learn the latent structure of the data and then transform it into latent variables using a reparameterization trick. This approach allows the model to perform gradient descent optimization while maintaining the continuity and differentiability of the latent variable space.

[0032] Decoder: The decoder's task is to map latent variables from the low-dimensional latent space back to the original data space. Through the decoder's nonlinear transformation, the model can reconstruct a fit that closely approximates the original input data. The decoder's optimization process is achieved by maximizing the KL divergence between the reconstruction error and the latent variable distribution, thereby ensuring effective data reconstruction while maintaining a reasonable distribution of latent variables.

[0033] ③ The training objective of a variational autoencoder is to minimize the following loss function:

[0034] in, This represents the reconstruction error between the samples generated by the decoder and the input data, reflecting the model's ability to reconstruct the data; KL divergence represents the difference between the posterior and prior distributions of the latent variables, helping to optimize the latent variable space to approximate a standard normal distribution. By minimizing this loss function, the variational autoencoder can effectively learn the latent distribution of the data and generate high-quality samples.

[0035] ④ During training, the Adam optimization algorithm is used for backpropagation to update the weights and biases in the encoder and decoder to minimize the loss function. The Adam optimization algorithm is an adaptive learning rate optimization method based on gradient-based first and second moment estimation. It can efficiently handle large-scale datasets and adapt to sparse gradient problems during training. The model iterates through multiple training epochs until the loss function converges. In each training epoch, the entire training set is passed through once, and the network parameters are updated through optimization, thereby reducing reconstruction errors and KL divergence, and improving the model's accuracy and stability.

[0036] ⑤ After the model training is completed, extract the decoder part from the VAE and fix its parameters. The decoder part is used to map latent variables from the low-dimensional latent space back to the original data space, thereby generating new concrete mix proportion samples.

[0037] S300. Based on the normalized mix proportion data and corresponding performance data in step S100, a deep learning prediction model for concrete performance is established. Concrete performance includes strength, workability, and cost. The prediction model is constructed using PyTorch. The input layer takes the feature parameters of the normalized mix proportion data as input, and the output layer corresponds to the target performance index of concrete. The model is trained using a training set, and the weights and biases of each layer are adjusted through the backpropagation algorithm to minimize the mean square error as the loss function. After training, the deep learning prediction model is obtained as the evaluation model.

[0038] Understandably, this step involves establishing a deep learning prediction model for concrete performance (including strength, workability, and cost). This prediction model is built using PyTorch, with the input layer taking feature parameters such as actual production data of the concrete mix proportions as input, and the output layer corresponding to the target performance indicators of the concrete. The model is trained using a training set, and the weights and biases of each layer are continuously adjusted through backpropagation, with the mean squared error (MSE) used as the loss function. After training, a deep learning prediction model with high performance prediction accuracy is obtained, which serves as the evaluation model for this invention, used to evaluate the performance of the generated mix proportion schemes.

[0039] In this embodiment, training the evaluation model specifically includes the following steps: ① The normalized combination ratio data and its corresponding performance data are proportionally divided into training and test sets for model training and validation. By appropriately dividing the dataset, data diversity is ensured during training, and the test set is used to evaluate the model's generalization ability.

[0040] ② A deep learning approach is used to construct the evaluation model, specifically employing a fully connected neural network (FNN) as its basic architecture. The input to this neural network is the feature vector of the concrete mix proportion, which includes information on the dosage of raw materials such as cement, aggregates, and admixtures. The output is the target performance indicators of the concrete, including compressive strength, workability (such as fluidity and water retention), and cost. Through a multi-layer fully connected network, the evaluation model establishes a mapping relationship between the mix proportion and performance indicators, enabling accurate prediction of the concrete mix proportion's performance.

[0041] ③ The training objective of the evaluation model is to minimize the difference between the predicted and actual values, typically using the mean squared error (MSE) as the loss function. The formula for the loss function is as follows:

[0042] in, This represents the performance metric value predicted by the model. This represents the actual measured performance value, where n is the number of samples. By minimizing this loss function, the model weights are continuously adjusted during training, enabling the model to better fit the data and improve prediction accuracy.

[0043] ④ The evaluation model is trained using the training set data, and the Adam optimization algorithm is used for backpropagation to optimize the model's weights and biases. The Adam optimization algorithm dynamically adjusts the learning rate of each parameter based on the estimation of the first moment (mean) and second moment (variance), adapting to different gradient changes and thus improving training efficiency. The training process typically involves multiple training epochs. Each iteration traverses the entire training set and gradually updates the network parameters to minimize the loss function and improve the model's prediction accuracy. Through repeated training on the training set, an evaluation model with high generalization ability and accuracy is finally obtained.

[0044] S400. Perform comprehensive performance verification on the generative model obtained in step S200 and the evaluation model obtained in step S300. In the generative model verification, generate multiple sets of mix proportion samples through random sampling and decoding of latent variables, and verify whether their key ratios of water-cement ratio, sand ratio, admixture dosage ratio, fly ash dosage ratio and mineral powder dosage ratio are within the reasonable range for engineering. In the evaluation model verification, input the test set data into the trained evaluation model, predict its compressive strength, workability and cost indicators respectively, and compare and analyze them with the measured values.

[0045] Understandably, this step involves a comprehensive performance verification of both the generative and evaluation models. In the generative model, multiple mix proportion samples are generated through random sampling and decoding of latent variables. These samples are then examined to verify whether key ratios such as water-cement ratio (W / B), sand ratio (S / a), admixture ratio (Ad / B), fly ash ratio (FA / B), and mineral powder ratio (Ca / B) are within the engineeringly reasonable range, thus validating the effectiveness and physical feasibility of the generative model. In the evaluation model, test set data is input into the trained model to predict its compressive strength, workability, and cost indicators. These predictions are then compared with measured values ​​to evaluate the model's generalization ability and prediction accuracy.

[0046] In this embodiment, the performance verification and rationality evaluation of the model specifically includes the following steps: ① Model Generation Validation: In the model generation section, multiple latent variables are generated through random sampling of latent variables, and these latent variables are mapped to corresponding concrete mix proportion data using a decoder. The generated mix proportion data is evaluated for reasonableness to ensure that the generated mix proportions are within the reasonable range for key ratios such as W / B (water-cement ratio), S / a (sand ratio), Ad / B (admixture ratio), FA / B (fly ash ratio), and Ca / B (mineral powder ratio). This step verifies whether the generated model can produce feasible mix proportions that meet the actual engineering requirements.

[0047] ② Model Validation Evaluation: In the model evaluation section, data from the test set is input into the trained evaluation model to predict its corresponding target performance (such as compressive strength, workability, and cost). Root mean square error (RMSE) and coefficient of determination (R²) are used as performance evaluation metrics to evaluate the model. RMSE: Calculates the root mean square error between predicted and actual values, measuring the accuracy of the model's predictions. A smaller RMSE value indicates a stronger predictive ability of the model. The formula is:

[0048] in, This represents the performance metric value predicted by the model. This represents the actual measured performance value, where n is the sample size. R² (coefficient of determination): used to measure the model's ability to explain the variance of the data, calculated using the following formula:

[0049] in, This represents the mean of the actual data. The closer the R² value is to 1, the better the model fits the data.

[0050] S500: Connect the generated model obtained in step S200 with the evaluation model obtained in step S300 to form an integrated design framework; establish an optimization function based on engineering requirements with the goal of achieving target performance and minimizing cost, where the target performance includes strength and workability; obtain the gradient information of the optimization objective function with respect to latent variables through automatic differentiation technology, and use the L-BFGS second-order approximate gradient optimization algorithm to iteratively solve the problem based on the gradient information to obtain the optimal concrete mix proportion solution in the latent variable space; perform multiple different initializations of the latent variables, and obtain multiple mix proportion schemes that meet the performance requirements and have different parameter distributions based on multi-path gradient descent search.

[0051] Understandably, in this step, the generative model and the evaluation model are linked to form a complete integrated "generation-evaluation" design framework. An optimization function is established with the goal of achieving target performance (including strength, workability, etc.) while minimizing cost. The gradient information of the objective function with respect to latent variables is directly obtained through automatic differentiation techniques, and a second-order approximate gradient optimization algorithm such as L-BFGS is used for rapid iterative solution, thereby obtaining the optimal concrete mix design in the latent variable space. To enhance the diversity of the solution set, the latent variables are initialized multiple times with different initializations. Through multi-path gradient descent search, multiple mix design schemes that meet performance requirements and have differentiated parameter distributions can be obtained, achieving the "one-to-many" optimization objective of concrete mix design.

[0052] In this embodiment, the concatenation and optimization solution of the model specifically includes the following steps: ① Connecting the Generative Model and the Evaluation Model: The generative model trained in step S2 and the evaluation model trained in step S3 are connected in series to form a complete design framework. In this framework, the generative model is used to generate concrete mix proportion schemes, while the evaluation model is used to evaluate the performance of the generated mix proportions and predict their target performance such as compressive strength, workability, and cost.

[0053] ② Establish the optimization objective function: Based on engineering requirements, construct the optimization objective function. The objective function aims to achieve multi-objective optimization, specifically minimizing the cost of the concrete mix design while ensuring that target performance (such as compressive strength, workability, etc.) meets the requirements. The objective function can be expressed as:

[0054] Where x is the design variable for the concrete mix proportion, Cost is the cost of the mix proportion, and Performance is the target performance of the mix proportion. and This is a weighting coefficient used to adjust the trade-off between target performance and cost.

[0055] ③ Automatic Differentiation and Gradient Calculation: Automatic differentiation is used to calculate the gradient of the objective function. Through automatic differentiation, gradient information about the design variables can be directly obtained from the objective function. This process avoids the tedious steps of manual differentiation, effectively improving computational efficiency. Based on this, the gradient of the objective function with respect to the latent variables can be calculated to guide subsequent optimization processes.

[0056] ④ Gradient Optimization Solution: Based on the gradient information obtained from automatic differentiation, a second-order approximation gradient optimization algorithm, such as L-BFGS, is used to quickly optimize the latent variables. The L-BFGS optimization algorithm is a quasi-Newton method that estimates the second derivative of the objective function using current gradient information and historical update information, thus achieving an efficient optimization process. Multiple iterative updates are performed in the latent variable space until the optimization process converges, obtaining the optimal mix ratio that meets the design requirements.

[0057] ⑤ Multiple Initializations and Solution Set Generation: To avoid getting trapped in local optima, latent variables are initialized multiple times with different starting points, and multiple optimal solutions are found through gradient descent along different paths. During each initialization, the latent variables are optimized from random positions, and the L-BFGS algorithm is used iteratively to obtain multiple different mix proportion solutions. Each optimization path is different, ultimately generating multiple diverse concrete mix proportion solution sets, ensuring that more possible optimal solutions are explored within the design space.

[0058] It is understood that, in this embodiment, the method for designing concrete mix proportions using generative neural networks and gradient optimization specifically includes the following steps: Step 1: Collect actual production data and corresponding performance parameters for 1201 sets of concrete mix proportions, including compressive strength, workability, and cost of the concrete. To improve the training stability of the model, all collected raw data were normalized to eliminate dimensional differences and ensure the stability of the model training process. Step 2: Construct a generative model using a Variational Autoencoder (VAE). The input to the generative model is normalized concrete mix proportion data (cement C, fly ash FA, water W, water-reducing agent WR, coarse aggregate CA, manufactured sand MS). The encoder extracts features layer by layer and maps them to a continuous, regularized low-dimensional latent space. The decoder generates new concrete mix proportion samples (cement C', fly ash FA', water W', water-reducing agent WR', coarse aggregate CA', manufactured sand MS') from the latent variables. The VAE model architecture is as follows: Figure 2 As shown; Step 3: Establish a deep learning prediction model for evaluating concrete performance (such as strength, workability, and cost). The input layer consists of feature parameters such as actual production data of concrete mix proportions, and the output layer consists of the target performance indicators. The evaluation model architecture is as follows: Figure 3 As shown; Step 4: Conduct comprehensive validation of the generative model and the evaluation model. By randomly sampling latent variables from the generative model and decoding them to generate 4000 and 10000 sets of mix proportion data, the rationality of its key ratios such as water-cement ratio (W / B) and sand ratio (S / a) is verified. The evaluation results of the generative model are as follows: Figure 4 As shown in the figure. Next, the trained evaluation model is input using the test set data to predict its compressive strength, and the prediction is compared with the actual data to evaluate the model's accuracy and generalization ability. The evaluation results of the model are shown in the figure. Figure 5 As shown; Step 5: Connect the generated model and the evaluation model in series. The model architecture is as follows: Figure 6 As shown, the optimization function designed to achieve the target performance (such as strength, workability, etc.) while minimizing cost is constructed as follows:

[0059] Wherein, UCS and COST represent the predicted compressive strength and material cost corresponding to the current mix proportion, respectively, and UCS* and COST* represent the target strength and cost required by the user, respectively. λ is a Lagrange multiplier used to coordinate the weight relationship between performance error and cost constraint in the objective function, and its value is set to 10 in this study.

[0060] Step 6: Directly obtain the gradient information of the objective function with respect to the latent variables using automatic differentiation technology, and perform rapid iterative solution using second-order approximation gradient optimization algorithms such as L-BFGS to obtain the optimal concrete mix design. To enhance the diversity of the solution set, the latent variables are initialized multiple times, and through multi-path gradient descent search, multiple mix design schemes that meet performance requirements and have differentiated parameter distributions are finally obtained, thereby achieving the "one-to-many" optimization objective of concrete mix design.

[0061] Example 2:

[0062] like Figure 7 As shown, this embodiment provides a concrete mix design system based on neural networks and gradient optimization. See [link to documentation]. Figure 7 The system includes: Step S100, Data Acquisition Module: Used to collect actual production data and corresponding performance parameters of multiple concrete mix proportions in the same region, including the compressive strength, workability and material cost of concrete; and to normalize the raw data of the mix proportions, including the performance data parameters of concrete strength, workability and cost corresponding to the raw material usage. Step S200: Establishing a module: This module is used to establish a variational autoencoder model based on the normalized concrete mix proportions from step S100, using the actual production data. The variational autoencoder model takes the normalized data as input, and the encoder part extracts features layer by layer and maps them to a continuous, regularized low-dimensional latent space. The decoder part is used to generate new concrete mix proportion samples from the latent variables. The training framework of this model uses the PyTorch deep learning platform, and the optimization convergence of the model is achieved by minimizing the reconstruction error and the KL divergence of the latent variable distribution. After training, the latent variable space and the decoder part are extracted as the generative model. Step S300, Training Module: This module is used to establish a deep learning prediction model for concrete performance based on the normalized mix proportion data and corresponding performance data from step S100. Concrete performance includes strength, workability, and cost. The prediction model is built using PyTorch. The input layer takes the feature parameters of the normalized mix proportion data as input, and the output layer corresponds to the target performance index of the concrete. The model is trained using a training set, and the weights and biases of each layer are adjusted through the backpropagation algorithm, with the mean square error as the loss function. After training, the deep learning prediction model is obtained as the evaluation model. Step S400, Verification and Evaluation Module: This module is used to comprehensively verify the performance of the generative model obtained in step S200 and the evaluation model obtained in step S300. In the generative model verification, multiple sets of mix proportion samples are generated through random sampling and decoding of latent variables, and it is verified whether their key ratios of water-cement ratio, sand ratio, admixture dosage ratio, fly ash dosage ratio, and mineral powder dosage ratio are within the reasonable range for engineering. In the evaluation model verification, the test set data is input into the trained evaluation model to predict its compressive strength, workability, and cost indicators, and the results are compared and analyzed with the measured values. Step S500, Solving Module: This module connects the generated model obtained in step S200 with the evaluation model obtained in step S300 to form an integrated design framework; it establishes an optimization function based on engineering requirements, aiming to achieve target performance while minimizing cost, where target performance includes strength and workability; it obtains the gradient information of the optimization objective function with respect to latent variables through automatic differentiation technology, and iteratively solves the problem using the L-BFGS second-order approximate gradient optimization algorithm based on this gradient information to obtain the optimal concrete mix proportion solution in the latent variable space; it performs multiple different initializations of the latent variables, and obtains multiple mix proportion schemes that meet performance requirements and have different parameter distributions based on multi-path gradient descent search.

[0063] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0064] Example 3:

[0065] Corresponding to the above method embodiments, this embodiment also provides a concrete mix design device based on neural networks and gradient optimization. The concrete mix design device based on neural networks and gradient optimization described below and the concrete mix design method based on neural networks and gradient optimization described above can be referred to in correspondence.

[0066] Figure 8 This is a block diagram illustrating a concrete mix design device 800 based on neural networks and gradient optimization, according to an exemplary embodiment. Figure 8 As shown, the concrete mix design device 800 based on neural networks and gradient optimization includes a processor 801 and a memory 802. The device also includes one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0067] The processor 801 controls the overall operation of the neural network and gradient optimization-based concrete mix design device 800 to complete all or part of the steps in the aforementioned neural network and gradient optimization-based concrete mix design method. The memory 802 stores various types of data to support the operation of the neural network and gradient optimization-based concrete mix design device 800. This data may include, for example, instructions for any application or method operating on the neural network and gradient optimization-based concrete mix design device 800, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as a keyboard, mouse, or buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the neural network and gradient optimization-based concrete mix design device 800 and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or one or more combinations thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0068] In an exemplary embodiment, the concrete mix design device 800 based on neural networks and gradient optimization can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the aforementioned concrete mix design method based on neural networks and gradient optimization.

[0069] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the above-described concrete mix design method based on neural networks and gradient optimization. For example, the computer-readable storage medium may be the memory 802 including the program instructions, which may be executed by the processor 801 of the concrete mix design device 800 based on neural networks and gradient optimization to complete the above-described concrete mix design method based on neural networks and gradient optimization.

[0070] Example 4:

[0071] Corresponding to the above method embodiments, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the concrete mix design method based on neural networks and gradient optimization described above.

[0072] A computer program is stored on a readable storage medium, and when the computer program is executed by a processor, it implements the steps of the concrete mix design method based on neural networks and gradient optimization described in the above method embodiments.

[0073] Specifically, the readable storage medium can be a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or any other readable storage medium capable of storing program code.

[0074] In summary, this invention effectively improves the efficiency and accuracy of concrete mix design by combining generative neural networks and gradient optimization techniques. The generative model can quickly generate multiple mix proportions that meet engineering requirements, and the generated mix proportions can be accurately predicted by the evaluation model, ensuring that multiple objectives such as concrete strength, workability, and cost are optimized.

[0075] Compared with traditional methods, this invention can provide a diverse set of solutions, and through multiple latent variable initializations and gradient descent searches, it can meet different engineering needs and avoid the limitations of a single solution.

[0076] Furthermore, by employing second-order approximate gradient optimization algorithms such as L-BFGS, the optimization solution is completed in a shorter time, significantly reducing computational overhead and meeting the requirements of efficient design in practical engineering. Through automatic differentiation technology, the optimization process is more accurate and efficient, improving the stability and convergence of the model.

[0077] Through rationality assessment, it was ensured that the generated mix proportions not only met mathematical optimization standards but also possessed practical engineering feasibility, thereby guaranteeing the actual application effect of the design scheme in construction. Most importantly, this scheme exhibits strong adaptability and generalization ability in multi-objective optimization, capable of coping with changes in different materials and construction conditions, generating high-performance mix proportions that meet engineering requirements, improving the overall quality of concrete mix design, and possessing broad application prospects and practical engineering value.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0079] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A concrete mix design method based on neural networks and gradient optimization, characterized in that, include: Step S100: Collect actual production data and corresponding performance parameters of multiple concrete mix proportions in the same region, including the compressive strength, workability and material cost of concrete; normalize the original data of the mix proportions, and the normalized data includes the performance data parameters of concrete strength, workability and cost corresponding to the raw material usage. Step S200: Based on the normalized concrete mix proportion actual production data in step S100, establish a variational autoencoder model based on the data. The variational autoencoder model takes the normalized data as input. The encoder part extracts features layer by layer and maps them to a continuous, regularized low-dimensional latent space. The decoder part is used to generate new concrete mix proportion samples from the latent variables. The training framework of this model adopts the PyTorch deep learning platform and achieves the optimization convergence of the model by minimizing the reconstruction error and the KL divergence of the latent variable distribution. After training, the latent variable space and decoder part are extracted as the generative model; Step S300: Based on the normalized mix proportion data and corresponding performance data from step S100, a deep learning prediction model for concrete performance is established. Concrete performance includes strength, workability, and cost. The prediction model is constructed using PyTorch. The input layer takes the feature parameters of the normalized mix proportion data as input, and the output layer corresponds to the target performance index of the concrete. The model is trained using a training set, and the weights and biases of each layer are adjusted through the backpropagation algorithm, with the mean square error as the loss function. After training, the deep learning prediction model is obtained as the evaluation model. Step S400: Perform comprehensive performance verification on the generative model obtained in step S200 and the evaluation model obtained in step S300. In the generative model verification, multiple sets of mix proportion samples are generated through random sampling and decoding of latent variables, and it is verified whether their key ratios of water-cement ratio, sand ratio, admixture dosage ratio, fly ash dosage ratio, and mineral powder dosage ratio are within the reasonable range for engineering. In the evaluation model verification, the test set data is input into the trained evaluation model to predict its compressive strength, workability, and cost indicators, and the results are compared and analyzed with the measured values. Step S500: Connect the generated model obtained in step S200 with the evaluation model obtained in step S300 to form an integrated design framework; establish an optimization function based on engineering requirements with the goal of achieving target performance and minimizing cost, where the target performance includes strength and workability; obtain the gradient information of the optimization objective function with respect to latent variables through automatic differentiation technology, and use the L-BFGS second-order approximate gradient optimization algorithm based on the gradient information to iteratively solve the problem and obtain the optimal concrete mix design in the latent variable space; By performing multiple different initializations of the latent variables and searching based on multi-path gradient descent, multiple sets of tack ratio schemes that meet performance requirements and have different parameter distributions are obtained.

2. The concrete mix design method based on neural networks and gradient optimization according to claim 1, characterized in that, The actual production data for the mix proportions include the amounts of cement, mineral powder, fly ash, water, water-reducing agent, fine aggregate, and coarse aggregate.

3. The concrete mix design method based on neural networks and gradient optimization according to claim 1, characterized in that, The calculation model for the standardization process in the normalization of the original data of the mix proportions is as follows: In the formula, x is the input variable for the concrete mix proportion. x ' This represents the normalized variable value. x min and x max The minimum and maximum values ​​of the variable should be respectively identified.

4. The concrete mix design method based on neural networks and gradient optimization according to claim 1, characterized in that, The training process for establishing a variational autoencoder model based on this data includes: All normalized concrete mix proportions from actual production data were used as input for unsupervised learning, and the training of the variational autoencoder model was initiated. Based on input data from actual production data, the encoder of the variational autoencoder maps the fully connected layer of the input concrete mix proportion data into the mean vector and standard deviation vector in the latent space, and then transforms them into latent variables through reparameterization techniques. The decoder, based on these latent variables, maps back from the low-dimensional latent space to the original data space and reconstructs the mix proportion close to the original input data through nonlinear transformation. The optimization process of the decoder is carried out by maximizing the KL divergence between the reconstruction error and the latent variable distribution. Based on the working process of the encoder and decoder, the training objective of the variational autoencoder is determined to be minimizing the loss function, and the expression of the loss function is as follows: in, This represents the reconstruction error between the sample generated by the decoder and the input data. Let KL divergence be the difference between the posterior and prior distributions of the latent variable. Based on a defined loss function, the Adam optimization algorithm is used for backpropagation during training to update the weights and biases in the encoder and decoder. The model iterates in multiple training epochs until the loss function converges. In each training epoch, the entire training set data is passed once, and the network parameters are updated based on the optimization. After the loss function converges and the model is trained, the decoder part of the variational autoencoder is extracted and its parameters are fixed. The decoder part is used for the generation of subsequent matching samples.

5. The concrete mix design method based on neural networks and gradient optimization according to claim 1, characterized in that, The process of training the evaluation model using the obtained deep learning prediction model as the evaluation model includes the following steps: The normalized mix proportion data and its corresponding performance data in step S100 are divided into training set and test set according to the proportion. The training set is used for model training and the test set is used for subsequent model validation. Based on the partitioned dataset, an evaluation model is constructed using deep learning methods, with a fully connected neural network as the basic architecture. The input of this neural network is the feature vector of the concrete mix proportion, which includes the dosage information of cement, aggregate and admixtures. The output is the target performance index of the concrete, which includes compressive strength, workability and cost. Based on the established model architecture, the training objective of the evaluation model is set as minimizing the difference between the predicted and actual values, and the mean squared error is used as the loss function. Based on a defined loss function, the evaluation model is trained using the training set data after step partitioning. The Adam optimization algorithm is used for backpropagation to optimize the model's weights and biases. The training process involves multiple training cycles, with each iteration traversing the entire training set and updating the network parameters based on the optimization, ultimately resulting in a fully trained evaluation model.

6. The concrete mix design method based on neural networks and gradient optimization according to claim 1, characterized in that, In step S400, the performance verification and rationality evaluation of the generated model includes the following steps: In the generative model trained in step S200, multiple latent variables are generated by random sampling of latent variables, and then the latent variables are mapped to the corresponding concrete mix proportion data by the decoder in the generative model. The mix proportion data is evaluated to determine whether the key ratios of water-cement ratio, sand ratio, admixture dosage ratio, fly ash dosage ratio, and mineral powder dosage ratio are within the reasonable range for the project. If so, the evaluation is output; otherwise, the evaluation continues.

7. The concrete mix design method based on neural networks and gradient optimization according to claim 1, characterized in that, In step S400, the performance verification and rationality assessment of the evaluation model includes the following steps: The data in the test set divided in step S500 is input into the evaluation model trained in step S300. The evaluation model is used to predict the corresponding target performance, where the target performance includes compressive strength, workability and cost. Based on the prediction results obtained in step S100, the root mean square error and the coefficient of determination are used as performance evaluation indicators; the formula for calculating the root mean square error is as follows: in, This represents the performance metric value predicted by the model. This represents the actual measured performance value, where n is the number of samples.

8. The concrete mix design method based on neural networks and gradient optimization according to claim 1, characterized in that, In step S500, the optimization objective function is established as follows: Based on engineering requirements, an optimization objective function is constructed, and the calculation formula for the optimization objective function is as follows: Where x is the design variable for the concrete mix proportion, Cost is the cost of the mix proportion, and Performance is the target performance of the mix proportion. and These are the weighting coefficients.

9. The concrete mix design method based on neural networks and gradient optimization according to claim 8, characterized in that, In step S500, gradient optimization and solution set generation include the following steps: Based on the constructed optimization objective function, the gradient of the optimization objective function is calculated using automatic differentiation technology to obtain gradient information about the design variables, and then the gradient of the objective function with respect to the latent variables is calculated. Based on the gradient information of the objective function with respect to the latent variables obtained by calculation, the gradient optimization algorithm of L-BFGS second-order approximation is used to optimize the latent variables. Multiple iterations are performed in the latent variable space until the optimization process converges. According to the optimization process, multiple starting points of the latent variables are initialized multiple times, and multiple optimization solutions are found through gradient descent along different paths. During each initialization, the latent variables are optimized from random positions, and the L-BFGS algorithm is used for iteration. Based on the iteration results, multiple different mix proportion solutions are obtained, and finally multiple sets of concrete mix proportion solutions are generated.

10. A concrete mix design system based on neural networks and gradient optimization, based on the concrete mix design method based on neural networks and gradient optimization as described in claim 1, characterized in that, include: Step S100, Data Acquisition Module: Used to collect actual production data and corresponding performance parameters of multiple concrete mix proportions in the same region, including the compressive strength, workability and material cost of concrete; and to normalize the raw data of the mix proportions, including the performance data parameters of concrete strength, workability and cost corresponding to the raw material usage. Step S200, Module Establishment: This module is used to establish a variational autoencoder model based on the normalized concrete mix proportions from step S100. The variational autoencoder model takes the normalized data as input. The encoder part extracts features layer by layer and maps them to a continuous, regularized low-dimensional latent space. The decoder part is used to generate new concrete mix proportion samples from the latent variables. The training framework of this model adopts the PyTorch deep learning platform and achieves the optimization convergence of the model by minimizing the reconstruction error and the KL divergence of the latent variable distribution. After training, the latent variable space and decoder part are extracted as the generative model; Step S300, Training Module: This module is used to establish a deep learning prediction model for concrete performance based on the normalized mix proportion data and corresponding performance data from step S100. Concrete performance includes strength, workability, and cost. The prediction model is built using PyTorch. The input layer takes the feature parameters of the normalized mix proportion data as input, and the output layer corresponds to the target performance index of the concrete. The model is trained using a training set, and the weights and biases of each layer are adjusted through the backpropagation algorithm, with the mean square error as the loss function. After training, the deep learning prediction model is obtained as the evaluation model; Step S400, Verification and Evaluation Module: This module is used to comprehensively verify the performance of the generative model obtained in step S200 and the evaluation model obtained in step S300. In the generative model verification, multiple sets of mix proportion samples are generated through random sampling and decoding of latent variables, and it is verified whether their key ratios of water-cement ratio, sand ratio, admixture dosage ratio, fly ash dosage ratio, and mineral powder dosage ratio are within the reasonable range for engineering. In the evaluation model verification, the test set data is input into the trained evaluation model to predict its compressive strength, workability, and cost indicators, and the results are compared and analyzed with the measured values. Step S500, Solving Module: This module connects the generated model obtained in step S200 with the evaluation model obtained in step S300 to form an integrated design framework; it establishes an optimization function based on engineering requirements, aiming to achieve target performance and minimize cost, where the target performance includes strength and workability; it obtains the gradient information of the optimization objective function with respect to latent variables through automatic differentiation technology, and iteratively solves the problem using the L-BFGS second-order approximate gradient optimization algorithm based on this gradient information to obtain the optimal concrete mix design in the latent variable space; By performing multiple different initializations of the latent variables and searching based on multi-path gradient descent, multiple sets of tack ratio schemes that meet performance requirements and have different parameter distributions are obtained.

Images