Green geopolymer optimization design method based on molecular dynamics and machine learning
By using methods based on molecular dynamics and machine learning, a polymer molecular model of aluminum tailings base was established. By using dynamic weight adaptive generative adversarial networks and improved residual neural networks, the high cost and low precision problems of traditional geopolymer material research were solved, and efficient and accurate material optimization design was achieved.
Patent Information
- Application Number
- CN202510967497.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-10-17
AI Technical Summary
Traditional geopolymer material research methods have the problems of high experimental costs, long cycles, and the optimization results are prone to falling into local extremes. The molecular dynamics method is not highly intelligent, and the combination of machine learning and traditional experiments has the problem of insufficient accuracy.
A green geopolymer optimization design method based on molecular dynamics and machine learning is adopted. By establishing a polymer molecular model for an aluminum tailings base, a sample data set is constructed, and a dynamic weight adaptive generative adversarial network and an improved residual neural network are used in combination with a ternary phase diagram to predict material properties and optimize the design.
The optimization design efficiency and accuracy of polymer materials in aluminum tailings bases have been significantly improved, the test costs have been reduced, and more accurate mechanical property predictions and optimized ratios have been achieved.
Smart Images
Figure CN120808962A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of material analysis and design, and specifically relates to a green geopolymer optimization design method based on molecular dynamics and machine learning. BACKGROUND
[0002] From the perspective of raw materials, traditional geopolymer materials mainly use industrial by-products such as fly ash and metakaolin as main raw materials. Such materials have significant advantages in reducing carbon emissions and saving raw material costs due to their waste resource recycling characteristics. However, there are significant regional distribution differences in industrial solid waste resources in China - fly ash is concentrated in the thermal power base, and metakaolin depends on the distribution of kaolin veins, resulting in a cross-regional transportation cost of up to 40%-50% of the raw material cost for application enterprises in non-resource producing areas. Therefore, it is necessary to select geopolymer raw materials according to local conditions. Guangxi, as an important aluminum industrial base in China, produces a large amount of aluminum tailings waste during the bauxite mining and processing process. If aluminum tailings can be used for geopolymer production, the raw material cost can be greatly reduced. Unfortunately, systematic research on aluminum tailings-based geopolymer is still in its infancy, and key technical issues such as mix design need to be addressed.
[0003] In the field of performance optimization of geopolymer materials, traditional research methods mainly use experimental design methods such as trial-and-error method, orthogonal test method and response surface method. By preparing samples with different raw material ratios and conducting compression strength tests and other laboratory characterization methods, the mechanical properties and engineering characteristics of geopolymer materials are systematically evaluated. Based on the comparative analysis of multiple parallel test data, researchers can preliminarily determine the optimal raw material ratio scheme. However, this method has limitations such as insufficient mechanism revelation, long experimental period and high trial-and-error cost. Essentially, it is still an optimization mode based on experience accumulation. This method has high experimental time cost and the results obtained are prone to fall into local extreme value traps, resulting in an optimal ratio that may not be the best.
[0004] In recent years, with the breakthrough of computational materials science, optimization design methods based on molecular dynamics have been gradually applied in the field of geopolymer material research and development. This method can simulate the depolymerization-polymerization kinetics of silicate and aluminate groups under alkali activation environment by constructing a nanoscale silicate and aluminate molecular model. Compared with traditional experiments, a single MD simulation can obtain structural evolution data equivalent to 28 days of curing samples in 48-72 hours, reducing the research and development period by more than 80%. However, this method also relies on manual observation and determination by researchers, and when the data volume is large, researchers need to spend a lot of time on quantitative analysis of a single simulation, and the error rate is also high. and
[0005] In addition, researchers have developed a method for collaboratively optimizing design by combining machine learning with traditional laboratory experiments. This method addresses the shortcomings of traditional experiments, but still faces significant bottlenecks in predictive accuracy. This is primarily because machine learning relies on vast amounts of data, while traditional laboratory experiments can only obtain relatively limited amounts of data, resulting in poor generalization of machine learning models. Summary of the Invention
[0006] The present invention provides a green geopolymer optimization design method based on molecular dynamics and machine learning, which can solve the technical problems of high experimental costs of traditional experimental methods, low intelligence level of molecular dynamics methods, and low precision when combining machine learning with traditional experiments.
[0007] The purpose of the present invention is achieved through the following technologies:
[0008] The present invention provides a green geopolymer optimization design method based on molecular dynamics and machine learning, comprising the following steps:
[0009] (1) Establish a molecular model of aluminum tailings base polymers that can be used for molecular dynamics simulation;
[0010] (2) Based on the molecular dynamics simulation results of the aluminum tailings base polymer molecular model in step (1), a sample data set is constructed: a molecular dynamics tensile simulation is performed on the aluminum tailings base polymer molecular model in step (1), and the stress-strain curve of the aluminum tailings base polymer under different characteristic parameters is output to obtain the mechanical performance index; the sample data set is constructed with the specific values of the characteristic parameters of the aluminum tailings base polymer as input variables and the specific values of the mechanical performance index as output variables;
[0011] (3) Propose a dynamic weight adaptive generative adversarial network Data enhancement method to generate high-quality data to effectively expand the dataset in step (2);
[0012] (4) Establish a prediction system for the mechanical properties of polymers in aluminum tailings bases Residual neural network is a Based on the residual neural network, a multi-branch residual neural network is improved for low-dimensional data sets; The improved part of the residual neural network for low-dimensional datasets includes two modules: a dynamic sparse branch selection module and a low-rank feature intersection module;
[0013] (5) Construct different feature parameter combinations to input the trained step (4) The residual neural network outputs the predicted value of mechanical properties; the coupling effect of characteristic data on mechanical performance indicators is displayed through the ternary phase diagram, the optimal parameter range is marked, and the optimized design of aluminum tailings base polymer is achieved.
[0014] Further, step (1) is based on the complex chemical environment of the aluminum tailings multi-component system, combined with the microstructure characteristics of the geopolymer hydration product, to construct a molecular model of the iron-doped silico-aluminate of the aluminum tailings geopolymer;
[0015] The aluminum tailings multi-component system presents typical non-equilibrium mineral phase distribution characteristics, which is mainly composed of a complex multi-phase system of high-reactivity amorphous silico-alumina oxide and crystalline transition metal oxide phase;
[0016] The microstructure of the geopolymer hydration product is essentially a silico-alumina oxide network system with multi-level ordered characteristics, and its structural characteristics can be deconstructed into a highly cross-linked three-dimensional network skeleton and a discrete branched oligomer;
[0017] The formation of the highly cross-linked three-dimensional network skeleton cross-linked structure is due to the structural reconstruction of the aluminum oxide tetrahedron to the silico-oxide network, which specifically manifests as: To the topological chemical substitution of the tetrahedron, resulting in structural units;
[0018] The structural units are each tetrahedron bridged by three adjacent tetrahedrons through bonds, and the positive correlation between network connectivity and polymerization index has a significant enhancing effect on macroscopic mechanical properties;
[0019] The discrete branched oligomer refers to oligomeric silico-aluminate precursors in the form of Q1 or Q2 coordination, including linear dimers (Q1) and branched structures (Q2), which dominate the early hydration kinetics and rheological behavior of geopolymer;
[0020] The component characteristics of the aluminum tailings are that the chemical components contain a certain amount of iron oxide composition.
[0021] Further, the characteristic parameters of step (2) include the silicon-aluminum ratio and the sodium-calcium ratio, and the mechanical property indicators include the ultimate tensile strength and the Young's modulus; the Young's modulus and the ultimate tensile strength are taken from the linear region slope and the peak stress intensity of the stress-strain curve, respectively.
[0022] Further, the dynamic weight self-adaptive generation of step (3) mainly includes three parts: a dynamic mode division unit, a non-linear constraint module, and a weight self-adaptive distributor, which significantly improves the reliability of small sample data and prediction accuracy through causal enhancement generation and dynamic balance training;
[0023] The dynamic mode division unit is realized based on an improved density clustering algorithm DBSCAN.
[0024] Further, the step (3) comprises the following steps:
[0025] S1. Dynamic mode recognition and weight smoothing based on DBSCAN core point clustering and noise adaptive suppression; core points and noise points in the data are dynamically recognized through a neighborhood radius and a minimum number of points, and a mode label is generated, the classification rule of which is shown in formula (1); at the same time, for each effective mode with a sample number of n, the weight is calculated by using formula (2), and a smoothing term is introduced to avoid infinite weight;
[0026] (1);
[0027] (2);
[0028] wherein, ε is the neighborhood radius, minPts is the minimum neighborhood point number required by the core point, k is the number of clusters automatically discovered, and noise represents the noise class data. The generated label is added to the data set D, which is subsequently input into the generator and the discriminator as a constraint condition, which can guide the generator to perform;
[0029] S2. The embedding layer Embedding is introduced to map the dynamic mode (k+1)-dimensional label into a fixed-dimensional conditional vector to adapt to the generator input, as shown in formula (3), and the generated process variable and the result variable can be represented as:
[0030] (3);
[0031] S3. Multi-mode discriminator optimization based on gradient penalty and weighted confrontation; the constructed discriminator loss function is shown in formula (4), which includes three parts: the expected score of real data, the expected score of generated data, and the gradient penalty term. In the training process, the generator maximizes the expected score of real data by mode condition guidance and introduction of weighted loss, so that the discriminator can learn the distribution characteristics of real data and accurately generate multi-mode data; the discriminator minimizes the expected score of generated data by weighted confrontation training, so that the generated data is difficult to be distinguished by the discriminator; the gradient penalty term forces the gradient norm of the discriminator to be close to 1, ensuring that the generator can cover all data modes and avoid generating only single-mode data;
[0032] (4);
[0033] wherein, is the linear interpolation of the real result variable and the pseudo result variable;
[0034] S4. Construct a multi-modal conditional generative adversarial-regression joint framework based on physical correlation, optimize the generator by dynamically weighting the adversarial loss and double regression constraints, and jointly optimize the generator and regression model using an alternating co-training strategy. The specific implementation steps are as follows:
[0035] S41. Regression relationship embedding and conditional generation: based on the physical correlation between process variables and result variables, a regression model is constructed, as shown in equation (5);
[0036] (5);
[0037] In the formula, is the mode label obtained by DBSCAN clustering, and the discrete mode is encoded as a continuous vector as the joint conditional input of the generator and the regression model;
[0038] S42. Generator joint optimization and adaptive weighting: the loss function of the generator G integrates the adversarial loss and the regression error, and dynamically balances the multi-modal contribution through the mode weight, as shown in equation (6);
[0039] (6);
[0040] In the formula, the first term is the adversarial loss, which forces the generator G to deceive the discriminator D; the second term is the double regression constraint, which forces the generated quality variable to simultaneously approach the regression predicted value and the true value, is the weight of the regression term, used to balance the adversarial training and regression accuracy;
[0041] S43. Regression model co-training and stability enhancement: the regression model utilizes generated data and real data for joint training, as shown in equation (7), whose objective function is the weighted mean square error. The generator and the regression model are trained alternately through the freeze / thaw strategy, which not only avoids mode collapse, but also enhances the model's generalization ability to noise and new working conditions;
[0042] (7).
[0043] Further, step (4) realizes the training and prediction of low-dimensional data sets through the following technical combination:
[0044] (1) Dynamic sparse branch selection module: first, a gating network is designed at the branch entrance, which generates four branch normalized weights, i.e., branch probabilities, which represent the relative importance of each branch at the current input; second, approximate discrete weights, i.e., branch masks, are generated by sampling the branch probabilities through Gumbel-Softmax, which directly control the activation or shutdown of the branch; finally, the corresponding branch is activated according to the branch mask and weighted and merged;
[0045] (2) Low-rank feature cross module:
[0046] S1. Low-rank projection: Compress the 32-dimensional features from the dynamic sparse branch selection module into a 2-dimensional space, reducing the number of parameters while retaining the main information, as shown in Equation (8):
[0047] (8);
[0048] S2. Implicit crossover: Perform element-wise product (Hadamard product) of the projected features in the low-dimensional space to generate implicit crossover terms. Assume that the two-dimensional vector is as shown in formula (9):
[0049] (9);
[0050] S3. Low-rank reconstruction: Map the cross-dimensional features back to the original dimension, as shown in formula (10):
[0051] (10).
[0052] Compared with the prior art, the present invention has the following beneficial effects:
[0053] (1) In view of the fact that aluminum tailings contain some iron oxide and considering the structural characteristics of geopolymers, a molecular model of aluminum tailings base polymers is proposed to provide a model basis for material performance optimization.
[0054] (2) In DRW-GAN, the DBSCAN clustering algorithm is used to dynamically identify data patterns and noise. Combined with the pattern weight adaptive mechanism, the generation weight is dynamically adjusted according to the sample size. The physical regression model and the generative adversarial network are innovatively integrated. Through the joint optimization of dual regression constraints and gradient penalty terms, the quality of small sample pattern generation is significantly improved.
[0055] (3) Aiming at the characteristics of low-dimensional data sets that are prone to overfitting and gradient disappearance, this paper proposes a Residual neural networks can further improve the accuracy of material property prediction models in scenarios with limited data feature dimensions or complex feature interactions.
[0056] (4) The present invention is based on The residual neural network can predict the mechanical properties of multi-component aluminum tailings matrix, and the ternary phase diagram can be used to more comprehensively analyze the impact of three variables on the performance of geopolymers, thereby achieving the optimal design of the mix ratio.
[0057] Figure 1 Flowchart of the green geopolymer optimization design method based on molecular dynamics and machine learning of the present invention.
[0058] Figure 2 Molecular diagram of the polymer composition of aluminum tailings base.
[0059] Figure 3 Alumina tailings-based polymer initial model graph.
[0060] Figure 4 Residual neural network architecture graph.
[0061] Figure 5 Ternary phase diagram of neural network predicting ultimate tensile strength.
[0062] Figure 6 Ternary phase diagram of neural network predicting Young's modulus. DETAILED DESCRIPTION
[0063] In order to make the objects, technical solutions and advantages of the present application clearer, the specific embodiments of the present application are further described in detail below with reference to the drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.
[0064] The green geopolymer optimization design method based on molecular dynamics and machine learning provided by the specific embodiments includes the following steps:
[0065] Step 1: Establishing an aluminum tailings-based polymer molecular model that can be used for molecular dynamics simulation.
[0066] Based on the complex chemical environment of the aluminum tailings multi-component system, combined with the microstructure characteristics of the geopolymer hydration product, an iron-doped silico-aluminate aluminum tailings-based polymer molecular model is constructed;
[0067] The aluminum tailings multi-component system presents typical non-equilibrium mineral phase distribution characteristics, which mainly consists of amorphous silicon-aluminum oxide with high reactivity and crystalline transition metal oxide phase (represented by ) to form a complex multi-phase system;
[0068] The microstructure of the geopolymer hydration product is essentially a silicon-aluminum oxide network system with multi-level ordered characteristics, and its structural characteristics can be decomposed into a highly cross-linked three-dimensional network skeleton and a discrete distribution of branched oligomers;
[0069] The formation of the cross-linked structure of the highly cross-linked three-dimensional network skeleton is due to the structural reconstruction of the aluminum oxide tetrahedron to the silicon oxide network, which specifically manifests as: Topochemical substitution of tetrahedron, thereby generating structural units;
[0070] The structural unit is that each tetrahedron is connected to three adjacent tetrahedrons through The positive correlation between the bridging of the bonds, the network connectivity and the aggregation index has a significant reinforcing effect on the macro-mechanical properties.
[0071] The discrete distribution of branched oligomers refers to oligomeric silicate-aluminate precursors in Q1 or Q2 coordination forms, including linear dimers (Q1) and branched structures (Q2), which dominate the early hydration kinetics and rheological behavior of the geopolymer;
[0072] The component characteristics of the aluminum tailings are that the chemical components contain a certain amount of iron oxide components.
[0073] The aluminum tailings geopolymer molecular model that can be used for molecular dynamics simulation is established, which comprises the following steps:
[0074] S1. The iron oxide content of the aluminum tailings samples from different places is characterized by X-ray fluorescence spectroscopy (XRF), and the mass fraction distribution interval is determined to be 15%-33% by range statistical method.
[0075] S2. The aluminum tailings from different places are sampled, and the same type and concentration of alkali activator is used to prepare geopolymer cement test blocks. The silicon-oxygen coordination structure in the aluminum tailings geopolymer solidified body is resolved by using nuclear magnetic resonance technology, and the Q3(1Al) characteristic peak and the Q2(1Al) characteristic peak are identified, and the chemical shifts are between -85ppm and -90ppm and -75ppm and -80ppm, respectively. Then, the integral areas of the Q3(1Al) characteristic peak and the Q2(1Al) characteristic peak are calculated, and the relative proportions of the Q3(1Al) and Q2(1Al) structures are in the interval [0.4, 22].
[0076] S3. According to the structural characteristics of Q3(1Al) and Q2(1Al) in the silicate-aluminate geopolymer network, MaterialsStudio is used to construct the corresponding cross-linked and chain-like structure skeleton. For the iron oxide doping parameters in the system, the iron oxide content gradient is selected within the range of the actual content of the S1 sample [15%, 23%]. Based on the measured data of the S2 aluminum tailings solidified body (the relative proportion of Q3 / Q2 is in the range [0.4, 22]), the proportion gradient of the cross-linked and chain-like structures is set within this interval. Based on the total charge of the silicate-aluminate anion skeleton, the number of sodium and calcium cations is dynamically adjusted to achieve charge balance, while the Na / Ca molar ratio is limited to the range (0, 6] to set multiple compensation schemes. The above structure types, iron oxide content parameters and cation ratio are fully combined to establish a parameterization construction system for the aluminum tailings geopolymer molecular model.
[0077] S4. The basic parameters of the aluminum tailings-based polymer molecular model are combined in the AC module of Materials Studio to assemble different numbers of iron oxide molecules, cross-linked and chain-like silicate molecular chains, and sodium and calcium atoms, while maintaining the total number of system atoms at about 10,000. The geometry is initialized using the Monte Carlo algorithm to obtain a variety of aluminum tailings-based polymer molecular models with different parameters.
[0078] Step two: Based on the results of the molecular dynamics simulation of the aluminum tailings-based polymer molecular model, a sample data set is constructed.
[0079] In this embodiment, energy minimization is required to reduce the energy of the system to a lower level. Specifically, the conjugate gradient algorithm is selected, and four convergence conditions are set: the maximum component tolerance of force is 0, the energy change tolerance is , the maximum iteration limit is 50,000 steps, and the maximum function evaluation times is 10,000. In the minimization process, the minimization stops running when any of the four conditions is met.
[0080] Alternatively, the commonly used algorithms for energy minimization include the steepest descent method and the conjugate gradient method.
[0081] In this embodiment, after energy minimization, the system needs to be dynamically relaxed to eliminate residual stress and achieve a more stable configuration. Specifically, initial velocities conforming to 300K are generated for all atoms, and the NPT ensemble is set to gradually increase the temperature of the entire system from 300K to 900K within 10ps, and then gradually decrease the temperature from 900K to 300K within 20ps, and finally balance at 300K for 10ps.
[0082] In this embodiment, after structure minimization and dynamic relaxation, the aluminum tailings-based polymer molecular model is subjected to molecular dynamics tensile simulation, and the stress-strain curve under different characteristic parameters is output to obtain the mechanical property index. Specifically, the linear region slope is solved to obtain the Young's modulus, and the peak stress intensity is used as the ultimate tensile strength.
[0083] In this embodiment, the sample data set is the data for training the aluminum tailings-based polymer prediction model, as shown in Table 1. Specifically, the sample data set includes an input matrix of specific values of geopolymer characteristic parameters (silicon aluminum ratio, sodium calcium ratio), and an output matrix representing specific values of geopolymer mechanical property indexes (ultimate tensile strength, Young's modulus).
[0084] Table 1
[0085]
[0086]
[0087]
[0088] Step 3: Data enhancement based on generative adversarial network (DRW-GAN).
[0089] In this embodiment, before building a generative adversarial network framework, the DBSCAN clustering algorithm is used to divide the original data into dynamic patterns and noise data, and the weights of the divided patterns are calculated. Specifically, the dynamic pattern is a high-density data area with relationship connections, which represents a subgroup with similar characteristics or behavioral patterns in the data set. The noise data is a low-density discrete point in the data set that cannot be classified into any dynamic pattern. Each sample in the data set includes two process variables. and two outcome variables.
[0090] The specific implementation steps of this embodiment are as follows:
[0091] S1. Dynamic pattern recognition and weight smoothing based on DBSCAN core point clustering and noise adaptive suppression. and minimum points Dynamically identify the core points and noise points in the data and generate pattern labels. The classification rules are shown in formula (1). At the same time, for each sample number Effective mode , use formula (2) to calculate the weight and introduce the smoothing term Avoid infinite weights.
[0092] (1);
[0093] (2);
[0094] in, is the neighborhood radius, is the minimum number of neighborhood points required for the core point, k is the number of automatically discovered clusters, and noise represents noise-like data. The generated labels are added to the dataset , which is then input into the generator and discriminator as a constraint, and the constraint can guide the execution of the generator.
[0095] S2. The embedding layer is introduced to map the dynamic pattern (k+1)-dimensional label into a fixed-dimensional conditional vector to adapt the generator input. As shown in Equation (3), the generated process variables and result variables can be expressed as:
[0096] (3);
[0097] S3. Multi-modal discriminator optimization based on gradient penalty and weighted adversarial. The constructed discriminator loss function is shown in equation (4), which includes three parts: expected score of real data, expected score of generated data, and gradient penalty term. During the training process, the generator maximizes the expected score of real data by mode-conditioned guidance and introduces a weighted loss, so that the discriminator can learn the characteristics of the real data distribution and accurately generate multi-modal data; the discriminator minimizes the expected score of generated data through weighted adversarial training, so that the generated data is difficult to be distinguished by the discriminator; the gradient penalty term forces the gradient norm of the discriminator to be close to 1, ensuring that the generator can cover all data modes and avoid generating only single-mode data.
[0098] (4);
[0099] wherein, is the linear interpolation of the real result variable and the pseudo result variable.
[0100] S4. Construct a multi-modal conditional generative adversarial-regression joint framework based on physical correlation, optimize the generator through dynamic weighted adversarial loss and double regression constraints, and adopt an alternating co-training strategy to jointly optimize the generator and the regression model. The specific implementation steps are as follows:
[0101] S41. Regression relationship embedding and conditional generation: based on the physical correlation between process variables and result variables , construct a regression model , as shown in equation (5).
[0102] (5);
[0103] wherein, is the mode label obtained by DBSCAN clustering, discrete modes are encoded into continuous vectors as joint conditional inputs for the generator and the regression model.
[0104] S42. Generator joint optimization and adaptive weighting: the loss function of the generator G integrates adversarial loss and regression error, and dynamically balances multi-modal contributions through mode weight , as shown in equation (6).
[0105] (6);
[0106] wherein, the first term is the adversarial loss, which forces the generator G to deceive the discriminator D; the second term is the double regression constraint, which forces the generated quality variable to simultaneously approximate the regression predicted value and the real value , is the weight of the regression term, used to balance the adversarial training and regression accuracy.
[0107] S43. Regression model collaborative training and stability enhancement: regression model The generated data and real data are jointly trained, as shown in equation (7), and the objective function is the weighted mean square error. The generator and the regression model are alternately trained by freezing / thawing strategy, which not only avoids mode collapse, but also enhances the generalization ability of the model to noise and new working conditions.
[0108] (7);
[0109] Step four: establishment of residual neural network architecture for prediction of mechanical properties of aluminum tailings-based polymer Residual neural network architecture.
[0110] In this embodiment, The residual neural network architecture is improved based on the Resnet neural network, and the specific process is shown in Figure 4 The following technical combination is used to train and predict low-dimensional data sets:
[0111] (1) Dynamic sparse branch selection module: first, a gating network is designed at the branch entrance, and the gating network generates four branch normalized original weights, i.e., branch probability, which represents the relative importance of each branch at the current input; second, approximate discrete weights, i.e., branch mask, are generated by sampling the branch probability through Gumbel-Softmax, which directly controls the activation or shutdown of the branch; finally, the corresponding branch is activated according to the branch mask and weighted combined.
[0112] (2) Low-rank feature cross module:
[0113] S1. Low-rank projection: compress the 32-dimensional features from the dynamic sparse branch selection module to a 2-dimensional space, reduce the parameter amount while retaining the main information, as shown in equation (8).
[0114] (8);
[0115] S2. Implicit cross: element-wise product (Hadamard product) is performed on the projected features in the low-dimensional space to generate an implicit cross term, assuming that the two-dimensional vector is As shown in equation (9).
[0116] (9);
[0117] S3. Low-rank reconstruction: map the cross-processed low-dimensional features back to the original dimension, as shown in equation (10).
[0118] (10).
[0119] In this embodiment, each branch first receives the same 3D input, and then each branch contains a fully connected layer with a residual structure, and the activation function is , converting the 3-dimensional input into a 32-dimensional output, which is used to represent the extraction or generation of 32 different features from the input data.
[0120] In this embodiment, a residual connection is set in each part of the low-rank feature cross module. Even if the low-rank feature cross module fails to effectively capture the complex relationship between features, the original input information can still be retained through the residual connection, ensuring that the model performance will not degrade due to module failure.
[0121] In this embodiment, after the dynamic sparse selection branch module and low-rank feature cross processing, the two fully connected layers with residual connections are gradually reduced to two-dimensional feature outputs, namely Young's modulus and ultimate tensile strength. Here, the activation function of the fully connected layer is also .
[0122] Step 5: Optimization design of polymer ratio in aluminum tailings base.
[0123] In this embodiment, characteristic parameters such as silicon-aluminum ratio, sodium-calcium ratio, and iron oxide volume fraction are changed, and the trained step 4 is used to Residual neural network is used to predict the mechanical properties of aluminum tailings-based polymers under different characteristic parameters. In this way, the Young's modulus and ultimate tensile strength of aluminum tailings-based polymers with thousands of different components can be obtained, which can be used in the formation of (such as Figure 5 、 Figure 6 ) The ternary phase diagram is used to intuitively analyze the effects of the three variables on Young's modulus and ultimate tensile strength, so as to obtain the optimal variable value range and achieve the optimal design of the aluminum tailings base polymer.
[0124] The test of the present invention selected several aluminum tailings with different composition ratios as raw materials to prepare geopolymer slurry test blocks, and simultaneously utilized the method proposed by the present invention to predict the component with the best performance, and the results are shown in Table 2. Since the molecular dynamics model is a completely ideal state, the atoms are arranged tightly and without defects, and the actual test blocks are often defective, and there are unreacted complete products, so the tensile strength accuracy is not high, but it can still reflect the variation trend of each component performance with the true tensile strength. For the elastic modulus, the true response value of each component can be better predicted, and the error with the true result is small. As can be seen from Table 2, the measured value of the elastic modulus and the predicted result error are all within 5%, and the predicted result of the tensile strength is consistent with the variation trend of the measured value. It can be considered that the test result is in good agreement with the predicted result.
[0125] Table 2
[0126]
Claims
1. A green geopolymer optimization design method based on molecular dynamics and machine learning, characterized in that: The following steps are involved: (1) Establish a molecular model of aluminum tailings base polymers that can be used for molecular dynamics simulation; (2) Based on the molecular dynamics simulation results of the aluminum tailings base polymer molecular model in step (1), a sample data set is constructed: a molecular dynamics tensile simulation is performed on the aluminum tailings base polymer molecular model in step (1), and the stress-strain curve of the aluminum tailings base polymer under different characteristic parameters is output to obtain the mechanical performance index; the sample data set is constructed with the specific values of the characteristic parameters of the aluminum tailings base polymer as input variables and the specific values of the mechanical performance index as output variables; (3) Propose a dynamic weight adaptive generative adversarial network Data enhancement method to generate high-quality data to effectively expand the dataset in step (2); (4) Establish a prediction system for the mechanical properties of polymers in aluminum tailings bases Residual neural network is a Based on the residual neural network, a multi-branch residual neural network is improved for low-dimensional data sets; The improved part of the residual neural network for low-dimensional datasets includes two modules: a dynamic sparse branch selection module and a low-rank feature intersection module; (5) Construct different feature parameter combinations to input the trained step (4) The residual neural network outputs the predicted value of mechanical properties; the coupling effect of characteristic data on mechanical performance indicators is displayed through the ternary phase diagram, the optimal parameter range is marked, and the optimized design of aluminum tailings base polymer is achieved.
2. The green geopolymer optimization design method based on molecular dynamics and machine learning according to claim 1, characterized in that: Step (1) Based on the complex chemical environment of the aluminum tailings multi-component system and the microstructural characteristics of the geopolymer hydration products, a molecular model of the aluminum tailings base polymer of iron-doped aluminosilicate was constructed; The aluminum tailings multi-component system presents a typical non-equilibrium mineral phase distribution characteristic, which is mainly composed of a complex multi-phase system consisting of highly reactive amorphous silicon aluminum oxide and crystalline transition metal oxide phases; The microstructure of the geopolymer hydration product is essentially a silicon-aluminum-oxygen network system with multi-level ordered characteristics, and its structural characteristics can be deconstructed into a highly cross-linked three-dimensional network skeleton and discretely distributed branched oligomers; The formation of the highly cross-linked three-dimensional network skeleton cross-linked structure is due to the structural reconstruction of the silicon-oxygen network by aluminum-oxygen tetrahedrons, which is specifically manifested as follows: right The resulting topological chemical substitution of tetrahedrons Structural unit; described The structural unit is each A tetrahedron with three adjacent Tetrahedron through The positive correlation between bond bridging, network connectivity and polymerization index has a significant enhancing effect on macroscopic mechanical properties; The discretely distributed branched oligomers refer to oligomeric aluminosilicate precursors in the form of Q1 or Q2 coordination, including linear dimers (Q1) and branched structures (Q2). These structural units dominate the early hydration kinetics and rheological behavior of geopolymers. The component characteristic of the aluminum tailings is that its chemical composition contains a certain amount of iron oxide.
3. The green geopolymer optimization design method based on molecular dynamics and machine learning according to claim 1, characterized in that: The characteristic parameters of step (2) include the silicon-aluminum ratio and the sodium-calcium ratio, and the mechanical performance indicators include the ultimate tensile strength and Young's modulus; the Young's modulus and ultimate tensile strength are respectively taken from the linear region slope and peak stress intensity of the stress-strain curve.
4. The green geopolymer optimization design method based on molecular dynamics and machine learning according to claim 1, characterized in that: Dynamic weight adaptive generative adversarial network in step (3) It mainly consists of three parts: a dynamic pattern division unit, a nonlinear constraint module, and a weight adaptive allocator. Through causal enhancement generation and dynamic balance training, it significantly improves the reliability and prediction accuracy of small sample data. The dynamic pattern division unit is implemented based on the improved density clustering algorithm DBSCAN.
5. The green geopolymer optimization design method based on molecular dynamics and machine learning according to claim 1, characterized in that: Step (3) includes the following steps: S1. Dynamic pattern recognition and weight smoothing based on DBSCAN core point clustering and noise adaptive suppression; the core points and noise points in the data are dynamically identified by the neighborhood radius and the minimum number of points, and the pattern labels are generated. The classification rules are shown in formula (1); at the same time, for each valid pattern with a sample number of , the weight is calculated using formula (2), and a smoothing term is introduced to avoid infinite weights; （1）； （2）； Where, is the neighborhood radius, is the minimum number of neighborhood points required for a core point, k is the number of automatically discovered clusters, and noise represents noisy data. The generated labels are added to the dataset and subsequently input into the generator and discriminator as constraints to guide the execution of the generator. S2. The embedding layer is introduced to map the dynamic pattern (k+1)-dimensional label into a fixed-dimensional conditional vector to adapt the generator input. As shown in Equation (3), the generated process variables and result variables can be expressed as: （3）； S3. Multimodal discriminator optimization based on gradient penalty and weighted adversarial training; The discriminator loss function is constructed as shown in formula (4), which consists of three parts: the expected score of real data, the expected score of generated data, and the gradient penalty term. During the training process, the generator maximizes the expected score of real data by guiding the mode condition and introducing weighted loss, so that the discriminator can learn the distribution characteristics of real data and accurately generate multimodal data; the discriminator minimizes the expected score of generated data through weighted adversarial training, making the generated data difficult to distinguish by the discriminator; the gradient penalty term forces the gradient norm of the discriminator to be close to 1, ensuring that the generator can cover all data modes and avoid generating data with only a single mode; （4）； where is the linear interpolation of the real outcome variable and the pseudo outcome variable; S4. Construct a multi-modal conditional generative adversarial-regression joint framework based on physical association. This framework optimizes the generator using a dynamic weighted adversarial loss and dual regression constraints. It also employs an alternating collaborative training strategy to jointly optimize the generator and regression models. The specific implementation steps are as follows: S41. Regression relationship embedding and condition generation: Based on the physical relationship between process variables and outcome variables, a regression model is constructed, as shown in formula (5); （5）； Where, is the pattern label obtained by DBSCAN clustering, and the discrete pattern is encoded into a continuous vector as the joint conditional input of the generator and regression model; S42. Joint optimization and adaptive weighting of the generator: The loss function of the generator G integrates the adversarial loss and the regression error, and dynamically balances the multimodal contributions through the mode weights, as shown in Equation (6): （6）； Where, the first term is the adversarial loss, which forces the generator G to deceive the discriminator D; the second term is the dual regression constraint, which forces the generated quality variable to approach both the regression prediction value and the true value. is the weight of the regression term, which is used to balance adversarial training and regression accuracy. S43. Co-training and stability enhancement of regression models: The regression model is jointly trained using generated data and real data, as shown in Equation (7). Its objective function is the weighted mean square error. The generator and regression model are alternately trained using a freeze / thaw strategy to avoid mode collapse and enhance the model's generalization ability to noise and new working conditions. （7）。 6. The green geopolymer optimization design method based on molecular dynamics and machine learning according to claim 1, characterized in that: Step (4) implements training and prediction of low-dimensional datasets through the following combination of techniques: (1) Dynamic sparse branch selection module: First, a gating network is designed at the branch entrance. The gating network generates the normalized original weights of the four branches, namely the branch probability, which represents the relative importance of each branch in the current input. Secondly, the branch probabilities are sampled by Gumbel-Softmax to generate approximate discrete weights, namely branch masks, which directly control the activation or closing of branches. Finally, the corresponding branches are activated according to the branch masks and weightedly merged. (2) Low-rank feature cross module: S1. Low-rank projection: Compress the 32-dimensional features from the dynamic sparse branch selection module into a 2-dimensional space, reducing the number of parameters while retaining the main information, as shown in Equation (8): （8）； S2. Implicit crossover: Perform element-wise product (Hadamard product) of the projected features in the low-dimensional space to generate implicit crossover terms. Assume that the two-dimensional vector is as shown in formula (9): （9）； S3. Low-rank reconstruction: Map the cross-dimensional features back to the original dimension, as shown in formula (10): （10）。