Method for performance prediction and proportion design of low-carbon cement
By constructing a low-carbon cement mechanical performance prediction model and proportion optimization design module, the problems of high carbon emissions in cement production and difficulty in developing new cement systems are solved, and efficient R&D and performance improvement of low-carbon cement are achieved.
Patent Information
- Application Number
- CN202411599290.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-11
- Publication Date
- 2025-02-18
AI Technical Summary
The carbon emissions in the existing cement production process are high, and the clinker coefficient is close to the limit, making it difficult to further reduce carbon emissions through traditional methods. The experimental volume of developing new cement systems is large, costly and low efficiency.
By analyzing the intrinsic characteristics of clinker and mixed materials, a high-quality data set was constructed, a low-carb cement mechanical performance prediction model was constructed using deep neural network and Bayesian optimization algorithm, and the Markov chain Monte Carlo method and Kerry were added. Gold agent model realizes the optimized design of low-carbon cement ratio.
The forward prediction and reverse design of low-carbon cement performance are achieved, the clinker coefficient is reduced, the number of tests and costs are reduced, and the development speed of low-carbon cement is improved.
Smart Images
Figure CN119476023A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of cement, and in particular to a method for predicting performance and designing mix ratio of low-carbon cement. Background Art
[0002] For more than 200 years, Portland cement has been an important component of many building materials, such as concrete. However, the production process of Portland cement clinker accounts for more than 13% of industrial CO2 emissions, and has always been at the forefront of carbon emissions. The main active mineral phases in Portland cement clinker are formed by the high-temperature reaction of CaO after limestone decarbonization with SiO2, Al2O3 and Fe2O3. In this case, adjusting the clinker composition is one of the solutions to reduce CO2 emissions. In addition, using admixtures to reduce the cement clinker coefficient is also an important technical approach to reduce CO2 emissions. At present, my country's clinker coefficient is close to the "limit", and there is "limited" room for significant carbon reduction. In 2020, the cement clinker coefficient is about 0.65. Admixtures in cement can not only significantly optimize its hydration process and microstructure, but also improve its working performance. The latest national standard GB175-2023 clearly gives 14 types of cement admixtures, such as slag, fly ash, and limestone, but in addition to these admixtures, there are also many "potential" admixtures, such as steel slag, copper slag, and lithium slag. Despite the potential environmental and performance benefits of the above-mentioned admixtures, it is important to note that their quality significantly affects the quality of cement and, therefore, their feasibility for use in cement production.
[0003] In addition to traditional fly ash, slag and pozzolana cement, a new type of ternary mixed cement, LC 3 Cement. This type of cement is made of Portland cement clinker supplemented with calcined clay and limestone. The calcium carbonate provided by limestone to the cement system and the additional alumina provided by calcined clay will further promote the cement hydration reaction. The amount of calcined clay is maintained at 35%. If the amount of limestone is maintained at about 0.5 times, the cement clinker coefficient can be reduced to 0.5. However, if you want to develop a new cement system and use other admixtures in the cement system to further reduce the clinker coefficient, a large number of tests are required; if you want to develop a four-component or even five-component cement system, the amount of testing is almost impossible to achieve and accept, it is difficult, the labor cost is high and the work efficiency is low. Summary of the invention
[0004] In view of the above problems, the present invention provides a method for predicting the performance of low-carbon cement and designing the proportions, which takes the intrinsic characteristics of clinker and admixture as the core, combines data-driven, deeply explores the adaptability of cement clinker and admixture, realizes forward prediction and reverse design of low-carbon cement, and improves the research and development speed of low-carbon cement.
[0005] The technical solution of the present invention is: a method for predicting the performance of low-carbon cement and designing the mix ratio, comprising the following steps: 1) Analyze the intrinsic properties of clinker and various mixed materials and build high-quality data sets; 2) Based on the data set in step 1), a low-carbon cement mechanical properties prediction model is constructed using a deep neural network and Bayesian optimization algorithm; 3) In step 2), the Markov chain Monte Carlo method and Kriging proxy model are added to the mechanical property prediction model of low carbon cement to achieve the optimal design of low carbon cement mix ratio.
[0006] In step 1), the mixed material includes calcium raw material, silicon raw material, aluminum raw material and gypsum raw material; The calcium raw materials include slag, limestone, steel slag, waste concrete, etc. The siliceous raw materials include fly ash, sandstone, volcanic ash, tuff, zeolite, pumice, diatomaceous earth, burnt shale, coal slag, lithium slag, nickel slag, tailings powder, magnesium slag, copper slag, etc. The aluminum raw materials include calcined coal gangue, burnt clay, bauxite, etc. The gypsum raw materials include anhydrite, desulfurized gypsum, phosphogypsum, fluorinated gypsum and the like.
[0007] In step 1), the admixture also includes an active activator, specifically high benzene sulphoaluminate cement, aluminate cement, ferroaluminate cement, sulphoaluminate cement, sodium carbonate, quicklime, carbide slag, sodium sulfate and the like.
[0008] In step 1), the intrinsic properties of the mixed material include mineral composition, chemical composition, specific surface area, fineness distribution, density, 1-24h dissolution rate of silicon / aluminum / calcium ions under alkaline environment conditions, etc.
[0009] In step 2), during the training of the low carbon cement mechanical properties prediction model, the deep neural network includes: 2.1) Input feature extraction: clinker, mixed material components and their intrinsic properties are used as input values; 2.2) Neural network architecture design: The deep neural network designed according to the requirements includes input layer, hidden layer and output layer; 2.3) Initialization and optimization of neurons and their weights: Each neuron in a layer receives the feature data output by the previous layer, and after a linear combination of weights and biases, it is input into the activation function for nonlinear transformation to obtain the output value of the 3d / 28d cement mortar compressive strength; 2.4) Model training and prediction evaluation: The batch gradient descent method is used to divide the data set into small batches, and one small batch of data is trained each time to speed up the training process. In each training iteration, the error between the predicted output and the true value is calculated, and the back propagation algorithm is used to calculate the gradient of each layer parameter and update the weight. After the model training is completed, the test data is input into the network, the prediction results are obtained through forward propagation, and the model accuracy is evaluated by MSE.
[0010] The Bayesian optimization algorithm includes: Select the hyperparameters to be optimized in the neural network, determine the minimum MSE as the optimization indicator, and use the Gaussian distribution model as the backend model; Randomly select a set of initial hyperparameter combinations, perform model training, and record the MSE of the model at this time. Update the Gaussian process model, calculate the optimal point in the hyperparameter space based on the posterior distribution of the model, select the next experimental point, and repeat the above process until the MSE meets the expected conditions. Stop iteration and output the best hyperparameter combination as the hyperparameter setting of the final model.
[0011] Step 3) includes: 3.1) Initial setting: Set the initial value of low carbon cement mix ratio according to the applicable scope of each component; 3.2) Parameter setting: the step size is selected between 0.001-0.15, the number of iterations is usually between 1000-3000, the burn-in period is 5%-15%, and the acceptance rate is 20%-40%; 3.3) Sample generation: Based on the initial parameters, a large number of potential new low-carbon cement mixes are generated by random sampling; by exploring the complex input parameter space, a mix ratio scheme that meets the target mechanical properties is found; 3.4) Ratio verification: The generated new ratio will be fitted by the Kriging proxy model to obtain the output prediction value and uncertainty estimate of each ratio scheme; 3.5) The Kriging proxy model needs to be verified again with the low-carbon cement mechanical properties prediction model to verify the optimized mix ratio scheme to ensure that its predicted performance is consistent with the actual performance; If there are deviations in the results of the Kriging proxy model verification, the sampling and fitting parameters need to be adjusted based on the feedback to optimize the sample generation accuracy.
[0012] The kernel function in the Kriging proxy model is RBF kernel function or Matérn kernel function, the scale parameter is between 0.2-4, and the noise level is between 5*10 -3 -5*10 -2 The optimizer uses L-BFGS-B.
[0013] The present invention proposes a method for predicting the performance of low-carbon cement and designing its mix ratio. Based on the intrinsic characteristics of different regions, different processes, and different types of admixtures, a deep learning method is used to construct a prediction model for the mechanical properties of low-carbon cement, and a low-carbon cement mix ratio reverse design module is added to the model to achieve the optimization of the low-carbon cement mix ratio and the improvement of its mechanical properties. By exploring the compatibility of admixtures and cement in cement, the cement clinker coefficient of the clinker can be further reduced.
[0014] The present invention predicts the performance of low-carbon cement and then realizes the mix ratio design; it has the following advantages: 1) Explore the structure-activity relationship between cement admixtures and cement mechanical properties, and build an ideal model for predicting cement mechanical properties from cement admixtures; 2) Further develop new low-carbon cement with a clinker coefficient below 0.5, thereby reducing carbon emissions; 3) It is particularly suitable for small data sets, especially in the field of cement research and development where high-quality data is difficult to collect. By reducing the number of tests, the test cycle and cost can be reduced, and a large number of optimized mix ratio schemes can be quickly generated to guide the research and development of low-carbon cement. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 This is a schematic diagram of the principle structure of the present invention. DETAILED DESCRIPTION
[0016] like Figure 1 As shown, the present invention provides a method for predicting the performance of low-carbon cement and designing the mix ratio, comprising the following steps: 1) Analyze the intrinsic properties of clinker and various mixed materials and build high-quality data sets; 2) Based on the data set in step 1), a low-carbon cement mechanical properties prediction model is constructed using a deep neural network and Bayesian optimization algorithm; 3) In step 2), the Markov chain Monte Carlo method and Kriging proxy model are added to the mechanical property prediction model of low carbon cement to achieve the optimal design of low carbon cement mix ratio.
[0017] In the operation of the present invention, step 1) is used to collect data and construct a data set, the low-carbon cement mechanical property prediction model in step 2) is used to train the performance prediction model, the mix ratio is specifically adjusted in step 3), and the performance is predicted using step 2), and then the mix ratio that meets the requirements is screened out in step 3).
[0018] The present invention uses a method based on knowledge and data coupling to greatly improve the development efficiency of low-carbon cement, guide R&D personnel to purposefully develop high-performance low-carbon cement, and realize the development of multi-element low-carbon cement. This method takes the intrinsic characteristics of mixed materials as the core, combines data-driven, deeply explores the compatibility of cement clinker and mixed materials, realizes forward prediction and reverse design of low-carbon cement, thereby accelerating the development of low-carbon cement and promoting the marketization of low-carbon cement.
[0019] In step 1), the mixed material includes calcium raw material, silicon raw material, aluminum raw material and gypsum raw material; The calcium raw materials include slag, limestone, steel slag, waste concrete, etc. The siliceous raw materials include fly ash, sandstone, volcanic ash, tuff, zeolite, pumice, diatomaceous earth, burnt shale, coal slag, lithium slag, nickel slag, tailings powder, magnesium slag, copper slag, etc. The aluminum raw materials include calcined coal gangue, burnt clay, bauxite, etc. The gypsum raw materials include anhydrite, desulfurized gypsum, phosphogypsum, fluorinated gypsum and the like.
[0020] In step 1), the admixture also includes an active activator, specifically high benzene sulphoaluminate cement, aluminate cement, ferroaluminate cement, sulphoaluminate cement, sodium carbonate, quicklime, carbide slag, sodium sulfate and the like.
[0021] In step 1), the intrinsic properties of the mixed material include mineral composition (crystals and glass), chemical composition, specific surface area, fineness distribution, density, 1-24h dissolution rate of silicon / aluminum / calcium ions under alkaline environment conditions, etc.
[0022] In step 2), during the training of the low carbon cement mechanical properties prediction model, the deep neural network includes: 2.1) Input feature extraction: clinker, mixed material components and their intrinsic properties are used as input values; 2.2) Neural network architecture design: The deep neural network designed according to the requirements includes input layer, hidden layer and output layer. The requirement is to use MSE as the judgment; 2.3) Initialization and optimization of neurons and their weights: Each neuron in a layer receives the feature data output by the previous layer, and after a linear combination of weights and biases, it is input into the activation function for nonlinear transformation to obtain the output value of the 3d / 28d cement mortar compressive strength; each neuron in the network has a corresponding weight parameter 𝑊 and bias 𝑏. Before the training starts, the initialization method randomly generates the initial weights to ensure that the network has a good initial state and avoid the problem of gradient disappearance or explosion.
[0023] 2.4) Model training and prediction evaluation: The batch gradient descent method is used to divide the data set into small batches, and one small batch of data is trained each time to speed up the training process; in each training iteration, the error between the predicted output and the true value is calculated, and the gradient of each layer parameter is calculated using the back propagation algorithm and the weight is updated. During the training process, a regularization method (Dropout layer) is introduced to prevent overfitting. After the model training is completed, the test data is input into the network, the prediction result is obtained through forward propagation, and the model accuracy is evaluated by MSE.
[0024] The Bayesian optimization algorithm includes: Select the hyperparameters to be optimized in the neural network (such as learning rate, batch size, number of layers, number of neurons, etc.), and determine the minimum MSE as the optimization indicator, and use the Gaussian distribution model as the backend model; Randomly select a set of initial hyperparameter combinations, perform model training, and record the MSE of the model at this time. Update the Gaussian process model, calculate the optimal point in the hyperparameter space based on the posterior distribution of the model, select the next experimental point, and repeat the above process until the MSE meets the expected conditions. Stop iteration and output the best hyperparameter combination as the hyperparameter setting of the final model.
[0025] In the process of training the prediction model of mechanical properties of low-carbon cement, clinker, admixture components and their intrinsic properties are used as input values, 3d / 28d cement mortar compressive strength is used as output value (prediction result), 80% of the data is used as training set, and 20% of the data is used as test set to train the model. MSE and R 2 as an evaluation indicator for the model.
[0026] The Bayesian optimization algorithm is used to optimize the main parameters in the above-mentioned neural network to prevent the model from overfitting. By continuously updating the model weights, the hyperparameter combination that can optimize the objective function is found, making the prediction accuracy of the low-carbon cement mechanical properties prediction model higher.
[0027] In step (2), the main parameters of the deep neural network are the number of neural layers layers, the number of neurons units, the learning rate Ir, the activation function AF, the amount of single training data batch_size, the number of training rounds epochs, the optimizer optimizer, and the residual connection module RC.
[0028] Among them, the number of neural layers is generally set to 5-20; the number of neurons is generally set to 32-256; the activation function is used to enhance the model's ability to handle nonlinear problems, and generally one or more of sigmoid, ReLu, Tanh, Softmax, and LeakyReLU are selected; the amount of single training data and the number of training rounds are used to update the model parameters, which are controlled by the amount of data, and are generally set to 16-128 and 500-1000 respectively; the optimizer is used to improve the performance and generalization ability of the model, and generally one of Adam, SGD, and AdamW is selected. The residual connection module is mainly to solve the problem that the neural network becomes difficult to train as the number of layers increases. Its placement and jump connection method need to be determined according to actual conditions.
[0029] Step 3) includes: 3.1) Initial setting: Set the initial value of low carbon cement mix ratio based on expert knowledge (applicable range of each component); 3.2) Parameter setting: The step size is selected between 0.001-0.15, the number of iterations is usually between 1000-3000, the burn-in period is generally 5%-15%, and the acceptance rate is generally 20%-40%; 3.3) Sample generation: Based on the initial parameters, the MCMC method generates a large number of potential new low-carbon cement mixes by random sampling; by exploring the complex input parameter space, MCMC can find a mix that meets the target mechanical properties. This process helps to effectively search the input space and obtain the possibility of different combinations; 3.4) Ratio Verification: The generated new ratios will be fitted with the Kriging proxy model to obtain the output prediction value and uncertainty estimate of each ratio scheme; the Kriging proxy model is based on Gaussian process modeling and can efficiently interpolate these ratios to ensure prediction accuracy.
[0030] 3.5) The Kriging proxy model needs to be verified again with the low-carbon cement mechanical properties prediction model to verify the optimized mix ratio scheme to ensure that its predicted performance is consistent with the actual performance; If there are deviations in the results of the Kriging proxy model verification, the sampling and fitting parameters need to be adjusted based on the feedback to further optimize the sample generation accuracy.
[0031] The kernel function in the Kriging proxy model is RBF kernel function or Matérn kernel function, the scale parameter is between 0.2-4, and the noise level is between 5*10 -3 -5*10 -2 The optimizer uses L-BFGS-B.
[0032] The Kriging proxy model includes: 1) Model training: Use the data from the previous neural network model training to train the Kriging proxy model. The kernel function hyperparameters, such as scale parameters and noise level, are determined by maximum likelihood estimation (MLE). During the training process, the Kriging proxy model constructs a Gaussian process. 2) Prediction and uncertainty assessment: predict new input parameters and generate predicted values of output responses and their uncertainties (confidence intervals). This means that the expected value and variability of each mix can be obtained.
[0033] 3) Optimization and iteration: Based on the prediction results and uncertainty information, optimize the input parameters and find the optimal cement ratio.
[0034] In step 3), the Markov chain Monte Carlo (MCMC) method and the Kriging proxy model are used to solve the problem of high data similarity and the need to explore more latent spaces (clinker ratio), and at the same time achieve the purpose of inferring the optimal input combination from the output target (clinker mechanical properties).
[0035] The MCMC method generates a large number of potential new low-carbon cement mixes from the existing low-carbon cement data. The algorithm can effectively explore the complex input parameter space and generate mixes that meet the target mechanical properties.
[0036] The Kriging proxy model is used to fit the low-carbon cement mix generated by the MCMC method. The Kriging proxy model is based on Gaussian process modeling and can provide output predictions and uncertainty estimates for each mix. This enables the Kriging proxy model to efficiently interpolate large-scale data sets and provide accurate performance predictions. The Kriging proxy model enables the optimization of more low-carbon cement systems.
[0037] The Kriging proxy model needs to be verified again through the above-mentioned low-carbon cement mechanical properties prediction model to verify the optimized mix ratio scheme to ensure that its predicted performance is consistent with the actual performance. If there is a deviation, the fitting parameters of the MCMC sampling and Kriging proxy model can be adjusted according to the feedback to further improve the optimization accuracy and realize the intelligent reverse design of low-carbon cement.
[0038] The low carbon cement mix ratio and mechanical property range given by the above Kriging proxy model need to be verified by the low carbon cement mechanical property prediction model in step (2). When the output value of the prediction model is within the confidence interval of the prediction value given by the Kriging proxy model, the Kriging proxy model is deemed to be credible.
[0039] Embodiment 1: This embodiment provides a method for predicting the performance of low-carbon cement and designing the mix ratio. Customer B proposed that according to its actual needs, it is necessary to use three or more of clinker, gypsum (2 types), mineral powder (4 types), limestone (3 types), waste cement (2 types), fly ash (4 types), calcined coal gangue (4 types), red mud (4 types), and nickel slag (4 types) to prepare a cement product. The brackets after the components represent raw materials from different sources (the same region or different regions) and different manufacturers; The specific proportion range is 15-50% clinker, 0.5-3% gypsum, 0-60% mineral powder, 0-15% limestone, 0-10% waste cement, 0-50% fly ash, 0-40% calcined coal gangue, 0-20% red mud, and 0-20% nickel slag. The compressive strength of cement mortar is required to be greater than 48.5MPa, and the fineness of the mixed materials is selected according to the strength requirements.
[0040] According to customer needs, it is necessary to obtain the intrinsic properties of all raw materials used, such as cement admixtures and cement clinker. The specific intrinsic properties are shown in Table 1, and Table 2 gives the mineral phase composition of all admixtures. Using the intrinsic properties of the admixtures in Table 1 as input values, a deep neural network is built, and its optimal parameter combination is determined by the Bayesian optimization algorithm. The specific values are the number of neural layers: 10, the number of neurons: 48, the learning rate Ir: 0.0020124, the activation function AF: Relu, the amount of single training data batch_size: 64, the number of training rounds epochs: 1000, the optimizer optimizer: AdamW, and two residual connection modules RC are set, which are located from the second hidden layer to the fourth hidden layer and from the seventh hidden layer to the ninth hidden layer. The MSE of the low-carbon cement mechanical properties prediction model is 3.67, and R 2 is 0.9341. The initial values of the MCMC method are clinker 18%, gypsum 1% (C2), mineral powder 30% (D1), limestone 5% (C2), waste cement 6% (A1), fly ash 10% (D4), calcined coal gangue 10% (D4), red mud 10% (B2), and nickel slag 10% (A3), where ABCD is the code for different origins, 1234 is the different specific surface areas of the mixed materials, the step size is 0.01, the number of iterations is 2000, the burn-in period is 12.87%, the acceptance rate is 29.35%, the kernel function in the Kriging proxy model is the RBF kernel function, the size is 0.2204, the noise is 0.0084104, and the optimizer uses L-BFGS-B. At this time, the confidence interval of the output mechanical properties of the Kriging proxy model just includes the output value of the prediction model. Table 3 lists the recommended proportions and predicted strengths given by this method.
[0041] Table 1 Intrinsic properties of the mixed materials used - chemical composition (wt%) Table 2 Main mineral composition of all mixed materials (wt%) Table 3 Low carbon cement mix ratios with cement mortar compressive strength greater than 48.5 MPa (taking three groups of mix ratios as examples) From the clinker content in Table 3, it can be seen that the clinker coefficient is low, meeting the low carbon requirements.
[0042] Embodiment 2: This embodiment provides a method for predicting the performance of low-carbon cement and designing the proportion. Customer B proposed that according to its actual needs, it is necessary to use three or more of clinker, sulphoaluminate cement, gypsum, mineral powder (4 types), fly ash (4 types), calcined coal gangue (4 types), steel slag (6 types), and lithium slag (4 types) to configure a cement product. The specific proportion range is 15-50% clinker, 0-10% sulphoaluminate cement, 0.5-5% gypsum, 0-60% mineral powder, 0-50% fly ash, 0-40% calcined coal gangue, 0-20% steel slag, and 0-20% lithium slag. The compressive strength of cement mortar is required to be greater than 48.5MPa, and the fineness of the mixed materials is selected according to the strength requirements.
[0043] According to customer needs, it is necessary to obtain the intrinsic properties of all raw materials used, such as cement admixtures and cement clinker. The specific intrinsic properties are shown in Table 4, and Table 5 gives the mineral phase composition of all admixtures. Using the intrinsic properties of the admixtures in Table 4 as input values, a deep neural network is built, and the optimal parameter combination is determined by the Bayesian optimization algorithm. The specific values are the number of neural layers: 12, the number of neurons: 32, the learning rate Ir: 0.00047328, the activation function AF: Relu, the amount of single training data batch_size: 32, the number of training epochs: 1000, the optimizer optimizer: Adam, and the residual connection module RC is set to 1, which is located after the fourth hidden layer to the eighth hidden layer. The MSE of the low-carbon cement mechanical properties prediction model is 2.82, and R 2is 0.9408. The initial values of the MCMC method are clinker 23%, sulphoaluminate cement 5%, gypsum 2% (A3), mineral powder 25% (B2), fly ash 10% (C1), calcined coal gangue 5% (A4), steel slag 20% (E3), lithium slag 10% (D4), where ABCD are codes for different origins, 1234 are different specific surface areas of the mixed materials, the step size is 0.005, the number of iterations is 1500, the burn-in period is 9.47%, the acceptance rate is 24.31%, the kernel function in the Kriging proxy model is the Matérn kernel function, the size is 0.2730, the noise is 0.015397, and the optimizer uses L-BFGS-B. At this time, the confidence interval of the output mechanical properties of the Kriging proxy model just includes the output value of the prediction model. Table 6 lists the recommended proportions and predicted strengths given by this method.
[0044] Table 4 Intrinsic properties of the mixed materials used Table 5 Main mineral composition of all mixed materials (wt%) Table 6 Low carbon cement mix ratios with cement mortar compressive strength greater than 48.5 MPa (taking three groups of mix ratios as examples) The present invention is not limited to the types of cement admixtures specified in the national standard, expands the possibility of low-carbon cement preparation, further reduces the clinker coefficient of existing cement, and analyzes the intrinsic characteristics of the admixtures. Based on the mineral composition, chemical composition, physicochemical properties and other factors of the admixtures, a low-carbon cement mechanical properties prediction model is built through machine learning methods, and the reverse design method is used to achieve the mix design of low-carbon cement.
[0045] The technical means disclosed in the scheme of the present invention are not limited to the technical means disclosed in the above-mentioned implementation mode, but also include technical schemes composed of any combination of the above-mentioned technical features. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.
Claims
1. A method for predicting performance and designing mix ratio of low carbon cement, characterized in that: The following steps are involved: 1) Analyze the intrinsic properties of clinker and various mixed materials and build high-quality data sets; 2) Based on the data set in step 1), a low-carbon cement mechanical properties prediction model is constructed using a deep neural network and Bayesian optimization algorithm; 3) In step 2), the Markov chain Monte Carlo method and Kriging proxy model are added to the mechanical property prediction model of low carbon cement to achieve the optimal design of low carbon cement mix ratio.
2. A method for predicting performance and designing mix ratio of low carbon cement according to claim 1, characterized in that: In step 1), the mixed material includes calcium raw material, silicon raw material, aluminum raw material and gypsum raw material; The calcium raw materials include slag, limestone, steel slag, and waste concrete; The siliceous raw materials include fly ash, sandstone, volcanic ash, tuff, zeolite, pumice, diatomaceous earth, burnt shale, coal slag, lithium slag, nickel slag, tailings powder, magnesium slag, copper slag; The aluminum raw materials include calcined coal gangue, burnt clay, and bauxite; The gypsum raw materials include anhydrite, desulfurized gypsum, phosphogypsum and fluorinated gypsum.
3. A method for predicting performance and designing mix ratio of low carbon cement according to claim 2, characterized in that: In step 1), the admixture also includes an active activator, specifically high benzene sulphoaluminate cement, aluminate cement, ferroaluminate cement, sulphoaluminate cement, sodium carbonate, quicklime, carbide slag, and sodium sulfate.
4. A method for predicting performance and designing mix ratio of low carbon cement according to claim 1, characterized in that: In step 1), the intrinsic properties of the mixed material include mineral composition, chemical composition, specific surface area, fineness distribution, density, and 1-24h dissolution rate of silicon / aluminum / calcium ions under alkaline environment conditions.
5. A method for predicting performance and designing mix ratio of low carbon cement according to claim 1, characterized in that: In step 2), during the training of the low carbon cement mechanical properties prediction model, the deep neural network includes: 2.1) Input feature extraction: clinker, mixed material components and their intrinsic properties are used as input values; 2.2) Neural network architecture design: The deep neural network designed according to the requirements includes input layer, hidden layer and output layer; 2.3) Initialization and optimization of neurons and their weights: Each neuron in a layer receives the feature data output by the previous layer, and after a linear combination of weights and biases, it is input into the activation function for nonlinear transformation to obtain the output value of the 3d / 28d cement mortar compressive strength; 2.4) Model training and prediction evaluation: The batch gradient descent method is used to divide the data set into small batches, and one small batch of data is trained each time to speed up the training process. In each training iteration, the error between the predicted output and the true value is calculated, and the back propagation algorithm is used to calculate the gradient of each layer parameter and update the weight. After the model training is completed, the test data is input into the network, the prediction results are obtained through forward propagation, and the model accuracy is evaluated by MSE.
6. A method for predicting performance and designing mix ratio of low carbon cement according to claim 5, characterized in that: The Bayesian optimization algorithm includes: Select the hyperparameters to be optimized in the neural network, determine the minimum MSE as the optimization indicator, and use the Gaussian distribution model as the backend model; Randomly select a set of initial hyperparameter combinations, perform model training, and record the MSE of the model at this time. Update the Gaussian process model, calculate the optimal point in the hyperparameter space based on the posterior distribution of the model, select the next experimental point, and repeat the above process until the MSE meets the expected conditions. Stop iteration and output the best hyperparameter combination as the hyperparameter setting of the final model.
7. A method for predicting performance and designing mix ratio of low carbon cement according to claim 1, characterized in that: Step 3) includes: 3.1) Initial setting: Set the initial value of low carbon cement mix ratio according to the applicable scope of each component; 3.2) Parameter setting: the step size is selected between 0.001-0.15, the number of iterations is usually between 1000-3000, the burn-in period is 5%-15%, and the acceptance rate is 20%-40%; 3.3) Sample generation: Based on the initial parameters, a large number of potential new low-carbon cement mixes are generated by random sampling; by exploring the complex input parameter space, a mix ratio scheme that meets the target mechanical properties is found; 3.4) Ratio verification: The generated new ratio will be fitted by the Kriging proxy model to obtain the output prediction value and uncertainty estimate of each ratio scheme; 3.5) The Kriging proxy model needs to be verified again with the low-carbon cement mechanical properties prediction model to verify the optimized mix ratio scheme to ensure that its predicted performance is consistent with the actual performance; If there are deviations in the results of the Kriging proxy model verification, the sampling and fitting parameters need to be adjusted based on the feedback to optimize the sample generation accuracy.
8. A method for predicting performance and designing mix ratio of low carbon cement according to claim 1, characterized in that: The kernel function in the Kriging proxy model is RBF kernel function or Matérn kernel function, the scale parameter is between 0.2-4, and the noise level is 5*10 -3 -5*10 -2 The optimizer uses L-BFGS-B.