A method for optimizing process parameters of copper tailings fly ash geopolymer based on cheminformatics

Abstract

This invention relates to the fields of cheminformatics and high-value utilization of solid waste, and particularly to a method for optimizing process parameters in the preparation of geopolymers from copper tailings and fly ash based on cheminformatics. The method involves constructing a heterogeneous reaction map of the geopolymer precursor and generating a silicon-aluminum-oxygen network topological chemical fingerprint vector. A dynamic graph convolution thermodynamic constraint collaborative prediction model is established, embedding physicochemical hard constraints to output the mapping relationship between process parameters and target performance. A chemical fingerprint similarity-weighted Pareto front adaptive exploration algorithm is used for global optimization, combined with in-situ vibrational spectroscopy to achieve online updating of the heterogeneous reaction map and iterative calibration of the model. This invention can effectively shorten the optimization cycle, comprehensively improve the target performance, and produce geopolymers with excellent mechanical properties and heavy metal solidification effects, providing efficient and precise technical support for the large-scale high-value utilization of copper tailings and fly ash.

Description

Technical Field

[0001] This invention relates to the field of cheminformatics and high-value utilization of solid waste, specifically to a method for optimizing process parameters in the preparation of geopolymers from copper tailings fly ash based on cheminformatics. Background Technology

[0002] Traditional methods for optimizing process parameters in geopolymer preparation mainly include orthogonal experimental design and response surface methodology. These methods, based on statistical principles, establish empirical relationships between process parameters and product performance through a finite number of experiments, thereby screening for optimal combinations of process parameters. However, these methods suffer from drawbacks such as a large number of experiments, long optimization cycles, and low optimization accuracy, and they cannot simultaneously address multiple conflicting performance objectives. Furthermore, traditional methods only consider the impact of macroscopic process parameters on product performance, completely ignoring the intrinsic relationship between precursor microstructure and reactivity, making it difficult to fundamentally reveal the mechanism by which process parameters affect product performance.

[0003] In recent years, machine learning technology has been increasingly applied to the optimization of process parameters in geopolymer preparation. By constructing data-driven predictive models, the number of experiments can be reduced and optimization efficiency improved to some extent. However, most existing machine learning models are purely data-driven black-box models, lacking consideration of the physicochemical laws governing the alkali-activated reaction of geopolymers. This easily leads to predictions that violate fundamental thermodynamic and kinetic principles, resulting in poor interpretability and weak generalization ability. Furthermore, most existing methods employ offline optimization, failing to perceive real-time changes in the reaction state during preparation. This makes it difficult to eliminate deviations in optimization results caused by batch fluctuations in raw materials, differences in the preparation environment, and equipment precision errors. Consequently, the process parameters obtained through offline optimization often fail to achieve the expected performance in actual production.

[0004] Furthermore, copper tailings contain various heavy metal elements, and their migration and transformation behavior during geopolymer preparation directly affects the heavy metal solidification effect of the product. Most existing process parameter optimization methods treat the heavy metal solidification rate as an independent performance indicator, without deeply analyzing the coupling relationship between the heavy metal-bearing phase structure and the evolution of the silicon-aluminum-oxygen network, making it difficult to achieve a synergistic improvement in mechanical properties and heavy metal solidification performance.

[0005] Therefore, there is an urgent need to develop a method for optimizing the preparation process parameters of copper tailings fly ash geopolymers that integrates cheminformatics and physicochemical principles. This method would enable precise characterization of the microscopic reactivity of precursors, construction of a physicochemically interpretable process performance prediction model, and establishment of an in-situ, real-time closed-loop optimization system. In this way, the optimal preparation process parameters that take into account multiple objectives can be obtained efficiently and accurately, thereby promoting the large-scale, high-value utilization of copper tailings and fly ash. Summary of the Invention

[0006] The purpose of this invention is to provide a method for optimizing process parameters in the preparation of copper tailings fly ash geopolymers based on cheminformatics, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for optimizing process parameters in the preparation of geopolymers from copper tailings fly ash based on cheminformatics includes the following steps:

[0009] S1: Collect data on the composition, phase composition and micromorphology of copper tailings and fly ash, construct a heterogeneous reaction spectrum of geopolymer precursors, and characterize the nodes of the spectrum as silicon-aluminum-oxygen tetrahedral clusters and heavy metal-bearing phases. The edge weights are calculated from the difference in electronegativity of elements, bond valence and coordination polyhedral sharing degree.

[0010] S2: Based on the principles of cheminformatics, topological dimensionality reduction and reaction path encoding are performed on heterogeneous reaction maps to generate silicon-aluminum-oxygen network topological chemical fingerprint vectors. The vectors include polyhedral connectivity index, spatial distribution entropy of base excitation sites, migration barrier of heavy metal ions, and chemical information entropy of reaction paths.

[0011] S3: Input the topological chemical fingerprint vector of the silicon-aluminum-oxygen network and the preset preparation process parameters into the dynamic graph convolution-thermodynamic constraint co-prediction model. The dynamic graph convolution-thermodynamic constraint co-prediction model embeds the Gibbs free energy change threshold of the geopolymer condensation reaction and the Arrhenius kinetic equation as hard constraints during the graph message passing process, and outputs the mapping relationship between process parameters and target performance.

[0012] S4: Based on the mapping relationship between process parameters and target performance, a chemical fingerprint similarity weighted Pareto front adaptive exploration algorithm is used to globally optimize process parameters. The algorithm constructs a hybrid kernel function prior with the Tanimoto similarity and Mahalanobis distance of the topological chemical fingerprint vector of the silicon aluminum oxide network, dynamically adjusts the exploration-utilization weight of the acquisition function, and outputs the optimal combination of process parameters.

[0013] S5: Input the optimal combination of process parameters into the preparation system, collect in-situ vibrational spectral data in real time and convert it into chemical coordinates of the reaction process, perform online reconnection of node weights and incremental update of graph structure on the heterogeneous reaction spectrum, drive the dynamic graph convolution-thermodynamic constraint collaborative prediction model for iterative calibration until the prediction residual converges to the preset threshold, and output the final optimized preparation process parameters.

[0014] As can be seen from the above technical solution provided by the present invention, the method for optimizing process parameters of copper tailings fly ash geopolymer preparation based on cheminformatics provided by the present invention has the following beneficial effects:

[0015] This invention breaks through the limitations of traditional methods that rely solely on macroscopic physicochemical indicators. By constructing a heterogeneous reaction map of geopolymer precursors, it uses silicon-aluminum-oxygen tetrahedral clusters and heavy metal-bearing phases as nodes. Edge weights are calculated based on elemental electronegativity differences, bond valence sums, and coordination polyhedral sharing degrees, thus comprehensively characterizing the interaction relationships between precursor microstructural units. The silicon-aluminum-oxygen network topological chemical fingerprint vector generated on this basis can simultaneously reflect the polyhedral connectivity, spatial distribution of alkali-excitation sites, heavy metal ion migration resistance, and reaction path diversity. This provides high-dimensional and low-redundancy structural feature inputs for process optimization, fundamentally improving the accuracy of the mapping relationship between process parameters and product performance.

[0016] The proposed dynamic graph convolution thermodynamic constraint collaborative prediction model embeds the Gibbs free energy change threshold and the Arrhenius kinetic equation as hard constraints during graph message transmission. It blocks the message flow of non-spontaneous reaction paths through thermodynamic feasibility gating units and dynamically adjusts the information transmission intensity using kinetic rate modulation units, forcing the model to perform feature learning and prediction within the physicochemically permissible parameter feasible domain. This model effectively solves the problem of traditional machine learning models easily producing prediction results that violate chemical laws, and can provide a reliable prediction basis for optimizing process parameters.

[0017] The chemical fingerprint similarity-weighted Pareto front adaptive exploration algorithm designed in this invention constructs a hybrid kernel function prior using the Tanimoto similarity and Mahalanobis distance of the topological chemical fingerprint vectors of the silicon-aluminum-oxygen network, which significantly improves the traversal efficiency of the process parameter search space. The algorithm can dynamically adjust the exploration and utilization weights of the acquisition function according to the distribution of the Pareto front, accelerating the convergence process of the global optimal solution. At the same time, the introduction of a chemical fingerprint similarity-weighted decision mechanism ensures that the optimal combination of process parameters not only takes into account the multi-objective performance requirements such as compressive strength, heavy metal solidification rate, and volume stability, but also has good reaction feasibility, avoiding theoretically optimal solutions that are practically unattainable.

[0018] This invention acquires in-situ vibrational spectral data during the preparation process in real time, converts it into chemical coordinates of the reaction process, realizes online reconnection of node weights and incremental update of graph structure in heterogeneous reaction spectra, and drives the prediction model to perform iterative calibration. This closed-loop mechanism effectively eliminates the deviation of offline optimization results caused by raw material batch fluctuations, differences in preparation environment and equipment precision errors, and greatly improves the applicability and stability of process parameters in actual production. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process parameters optimization method for preparing copper tailings fly ash geopolymer based on cheminformatics according to the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, this embodiment of the invention provides a method for optimizing process parameters in the preparation of copper tailings fly ash geopolymers based on cheminformatics, comprising the following steps:

[0023] S1: Collect data on the composition, phase composition and micromorphology of copper tailings and fly ash, construct a heterogeneous reaction spectrum of geopolymer precursors, and characterize the nodes of the spectrum as silicon-aluminum-oxygen tetrahedral clusters and heavy metal-bearing phases. The edge weights are calculated from the difference in electronegativity of elements, bond valence and coordination polyhedral sharing degree.

[0024] S2: Based on the principles of cheminformatics, topological dimensionality reduction and reaction path encoding are performed on heterogeneous reaction maps to generate silicon-aluminum-oxygen network topological chemical fingerprint vectors. The vectors include polyhedral connectivity index, spatial distribution entropy of base excitation sites, migration barrier of heavy metal ions, and chemical information entropy of reaction paths.

[0025] S3: Input the topological chemical fingerprint vector of the silicon-aluminum-oxygen network and the preset preparation process parameters into the dynamic graph convolution-thermodynamic constraint co-prediction model. The dynamic graph convolution-thermodynamic constraint co-prediction model embeds the Gibbs free energy change threshold of the geopolymer condensation reaction and the Arrhenius kinetic equation as hard constraints during the graph message passing process, and outputs the mapping relationship between process parameters and target performance.

[0026] S4: Based on the mapping relationship between process parameters and target performance, a chemical fingerprint similarity weighted Pareto front adaptive exploration algorithm is used to globally optimize process parameters. The algorithm constructs a hybrid kernel function prior with the Tanimoto similarity and Mahalanobis distance of the topological chemical fingerprint vector of the silicon aluminum oxide network, dynamically adjusts the exploration-utilization weight of the acquisition function, and outputs the optimal combination of process parameters.

[0027] S5: Input the optimal combination of process parameters into the preparation system, collect in-situ vibrational spectral data in real time and convert it into chemical coordinates of the reaction process, perform online reconnection of node weights and incremental update of graph structure on the heterogeneous reaction spectrum, drive the dynamic graph convolution-thermodynamic constraint collaborative prediction model for iterative calibration until the prediction residual converges to the preset threshold, and output the final optimized preparation process parameters.

[0028] In this embodiment, the core function of step S1 is to systematically collect multi-dimensional basic physicochemical data of copper tailings and fly ash, generate a high-quality precursor physicochemical feature dataset through standardized data preprocessing and key feature extraction, and construct a heterogeneous reaction spectrum based on graph theory principles that can accurately characterize the correlation between precursor microstructure units and reactivity, providing basic data structure and reactivity characterization support for subsequent topological chemical fingerprint generation, process performance prediction, and parameter optimization; the detailed steps are as follows:

[0029] Step S1-1: Basic physicochemical data collection and preprocessing of copper tailings and fly ash:

[0030] X-ray fluorescence spectrometry was used to collect the main elemental composition data of copper tailings and fly ash, X-ray diffraction was used to collect phase composition data, and scanning electron microscopy was used to collect microscopic morphology data. The data collection process must strictly follow the standard operating procedures of the corresponding instruments. Each sample was tested three times and the average value was taken as the final raw data to ensure the accuracy and repeatability of the data.

[0031] Normalization calibration and background noise filtering are performed on the acquired raw data. The normalization calibration adopts the minimum-maximum normalization method to map the raw data of different dimensions to the range of 0 to 1. The background noise filtering adopts the wavelet transform denoising algorithm to remove instrument noise and environmental interference signals.

[0032] Key topological feature parameters such as silicon-aluminum molar coordination ratio, mass fraction of each crystal phase, average particle size, specific surface area of ​​particles, and surface roughness of particles are extracted from the preprocessed data. All extracted feature parameters are associated and stored according to sample number to generate a standardized precursor physicochemical feature dataset.

[0033] Step S1-2: Generation of feature sequences of heterogeneous reaction map nodes:

[0034] Based on the precursor physicochemical characteristic dataset, the phase region analysis algorithm is used to spatially analyze the phase distribution of copper tailings and fly ash, and delineate the coordination environment boundaries of different silicon-aluminum-oxygen tetrahedral clusters. The phase region analysis algorithm determines the cell parameters and spatial distribution range of each silicon-aluminum-oxygen tetrahedral cluster by performing peak fitting and quantitative phase analysis on X-ray diffraction patterns.

[0035] Based on the analysis results of the occurrence state of heavy metal elements, a phase space mapping model of heavy metal elements was established. The occurrence state analysis of heavy metal elements adopted the continuous extraction method, which divided heavy metal elements into five occurrence states: exchangeable state, carbonate bound state, iron and manganese oxide bound state, organic bound state and residue state, and determined the spatial location and content distribution of each occurrence state.

[0036] Based on the spatial mapping results between the coordination environment boundary of the silicon-aluminum-oxygen tetrahedral cluster and the heavy metal-bearing phase, a node feature sequence carrying topological type labels and three-dimensional spatial coordinates is generated. Each node in the node feature sequence corresponds to a silicon-aluminum-oxygen tetrahedral cluster or a heavy metal-bearing phase. The node features include topological type, elemental composition, three-dimensional spatial coordinates and content information.

[0037] Step S1-3: Construction of the edge weight attribute matrix of the heterogeneous reaction map:

[0038] Based on the spatial adjacency relationship and elemental properties of adjacent nodes in the node feature sequence, the distribution of electronegativity difference between bonding elements is calculated; the electronegativity difference is the difference between the Pauling electronegativity values ​​of the two bonding elements, reflecting the polarity and reactivity of the chemical bond.

[0039] Bond valence and values ​​are derived by matching crystallographic parameters. Bond valence and values ​​are obtained by adding the bond valences of each chemical bond in the node. The bond valence is calculated based on the relationship between the bond length and the standard bond length, reflecting the strength and stability of the chemical bond.

[0040] The proportional weights of vertex sharing, edge sharing, and face sharing of coordination polyhedra are quantified based on the topological feature parameters of particle surfaces. By performing image segmentation and feature extraction on scanning electron microscope images, the frequency of occurrence of different types of coordination polyhedron sharing modes is counted, and the proportional weights of each sharing mode are calculated.

[0041] By fusing multi-source information and gradient quantization of electronegativity difference distribution, bond valence and numerical values, and shared proportion weights, the edge weights of the heterogeneous reaction spectrum are calculated; the formula for calculating the edge weights is as follows: ,in, For edge weights, It represents the electronegativity difference between bonding elements. For key valence and value, For coordination polyhedron sharing degree, , , These are the weighting coefficients for electronegativity difference, bond valence and value, and coordination polyhedron sharing degree, respectively, and satisfying the following conditions: ;

[0042] Based on the edge weights of all node pairs, construct the edge weight attribute matrix of the heterogeneous response graph; the rows and columns of the edge weight attribute matrix correspond to the nodes in the node feature sequence, and the values ​​of the matrix elements are the edge weights between the corresponding two nodes.

[0043] Steps S1-4: Topological assembly and optimization of heterogeneous reaction maps:

[0044] The node feature sequence and the edge weight attribute matrix are assembled using graph theory topology to generate an initial heterogeneous reaction map. The initial heterogeneous reaction map uses nodes to represent silicon aluminum oxygen tetrahedral clusters and heavy metal-bearing phases, edges to represent the interaction relationships between nodes, and edge weights to represent the strength of the interaction.

[0045] Thermodynamic stability criteria and spatial steric constraints are introduced to filter out virtual connections in the initial heterogeneous reaction spectrum; the thermodynamic stability criteria are based on the Gibbs free energy change principle to filter out thermodynamically unstable interaction edges; the spatial steric constraints are based on the three-dimensional spatial coordinates and dimensions of the nodes to filter out edges that cannot form effective interactions in space.

[0046] The edge weight attribute matrix after filtering out virtual connected edges is subjected to graph structure smoothing. The graph structure smoothing uses the Gaussian kernel smoothing algorithm to perform local weighted averaging of edge weights, eliminating abrupt changes and noise in edge weights, making the graph structure smoother and more continuous.

[0047] After smoothing the graph structure, the final heterogeneous reaction map of the geopolymer precursor is constructed. This map can comprehensively and accurately characterize the microstructure and reactivity of copper tailings and fly ash precursors, providing a reliable basic data structure for subsequent steps.

[0048] In this embodiment, the core function of step S2 is to perform topological dimensionality reduction and reaction path encoding on the heterogeneous reaction map of the geopolymer precursor constructed in step S1, based on the principles of cheminformatics. This extracts key topological features that can quantitatively characterize the precursor's reactivity and structural evolution potential, generating a standardized silicon-aluminum-oxygen network topological chemical fingerprint vector. This provides high-dimensional and low-redundancy input features for subsequent process performance prediction models. The detailed steps are as follows: ,in, The polyhedral connectivity index. The mean of the degree centrality of the nodes. The mean of the betweenness centrality of the nodes. This represents the mean degree of closure of the ring system. , , The weights for degree centrality, betweenness centrality, and cyclic closure are respectively, and satisfy the following conditions: ;

[0049] Step S2-2: Calculation of spatial distribution entropy of alkali excitation sites:

[0050] Based on the local topological constraints characterized by the polyhedral connectivity index, and combined with the three-dimensional spatial coordinates and element attribute labels in the node feature sequence, surface non-bridging oxygen nodes and charge-compensated cation nodes are selected as base-excited active sites. The precursor reaction space is divided into cubic grid cells of equal volume using a three-dimensional spatial grid partitioning algorithm. The number of base-excited active sites in each grid cell is counted to obtain the site density distribution frequency.

[0051] The electronegativity difference distribution of elements in the edge weight attribute matrix is ​​introduced as a site reactivity correction coefficient, and the site density distribution frequency of each grid cell is weighted by activity. Based on the information theory entropy calculation principle, the degree of aggregation and dispersion of active sites in space is solved to generate the spatial distribution entropy of alkali-excited sites.

[0052] The formula for calculating the spatial distribution entropy of base excitation sites is as follows: ,in, The entropy of the spatial distribution of base-excitation sites. This represents the total number of three-dimensional spatial grid cells. For the first The proportion of the activity-weighted site density within each grid cell to the total site density;

[0053] Step S2-3: Calculation of migration barrier for heavy metal ions:

[0054] Based on the spatial steric hindrance distribution characterized by the spatial distribution entropy of base-excited sites, a heavy metal ion migration network is constructed along the edge path of the heterogeneous reaction spectrum. The bond valence and value in the edge weight attribute matrix are inversely mapped to the proportion weight of coordination polyhedra sharing as migration resistance parameters. The larger the bond valence and value and the higher the degree of coordination polyhedra sharing, the greater the corresponding migration resistance.

[0055] The constrained random walk algorithm was used to simulate the diffusion trajectory of heavy metal ions in the voids of silicon-aluminum-oxygen polyhedra, and the residence time and transfer probability of ions at each node were recorded. Combined with the constraints of the local coordination environment of the nodes, the cumulative potential energy extreme value and the transition state energy drop on each migration path were calculated, and the minimum transition state energy drop among all possible migration paths was taken as the migration barrier of the heavy metal ion.

[0056] The formula for calculating the migration barrier of heavy metal ions is: ,in, This acts as a barrier for the migration of heavy metal ions. The total number of possible migration paths. For the first The maximum cumulative potential energy along the migration path For the first The initial potential energy of the migration path;

[0057] Step S2-4: Calculation of chemical information entropy of reaction pathway and generation of topological chemical fingerprint vector of silicon-aluminum-oxygen network:

[0058] The polyhedral connectivity index, the spatial distribution entropy of alkali excitation sites, and the migration barrier of heavy metal ions are fused to construct a state transition network for precursor condensation reaction. Based on the temporal characteristics of the reaction process, the nodes of the heterogeneous reaction map are divided into three subgraphs: dissolved state, migrating state, and condensation state. The topological features and edge weight attributes of each subgraph are extracted respectively.

[0059] The probability distribution of path transitions from the dissolved state to the migrated state and from the migrated state to the condensation state is calculated. A reaction path diversity measurement model from cheminformatics is introduced. After normalizing the path transition probability distribution, the information entropy is solved to generate the chemical information entropy of the reaction path.

[0060] Manifold space projection dimensionality reduction and feature alignment operations are performed on the polyhedral connectivity index, the spatial distribution entropy of alkali excitation sites, the migration barrier of heavy metal ions, and the chemical information entropy of reaction pathways to eliminate the linear correlation between topological redundancy dimensions and features. The processed features are then spliced ​​and standardized to output a silicon-aluminum-oxygen network topological chemical fingerprint vector with fixed dimensions and standardization.

[0061] In this embodiment, the core function of step S3 is to construct a dynamic graph convolution prediction model that integrates physicochemical hard constraints. The silicon-aluminum-oxygen network topological chemical fingerprint vector generated in step S2 is coupled with preset fabrication process parameters as input. Through thermodynamic feasibility gating and kinetic rate modulation, the topological pruning of ineffective reaction paths and the precise control of the graph structure evolution rhythm are achieved. The model outputs a quantitative mapping relationship between process parameters and target performance, providing a reliable predictive basis for subsequent global optimization of process parameters. The detailed steps are as follows:

[0062] Step S3-1: Construction of initial node representation of process-structure coupling graph:

[0063] The silicon-aluminum-oxygen network topological chemical fingerprint vector output in step S2 is tensor aligned and multidimensional feature spliced ​​with the preset preparation process parameters. The preset preparation process parameters include the mass ratio of copper tailings to fly ash, the modulus of alkali activator, the concentration of alkali activator, curing temperature, curing humidity and curing time.

[0064] The concatenated feature vectors are standardized to eliminate the differences in dimensionality and numerical range of different feature dimensions. The standardized feature vectors are then mapped to the initial feature vectors of the graph nodes to construct the initial node representation of the process-structure coupling graph, which serves as the input layer state of the dynamic graph convolution-thermodynamic constraint co-prediction model.

[0065] Step S3-2: Topology pruning of ineffective reaction paths under thermodynamic hard constraints:

[0066] In the graph message passing process of the dynamic graph convolution-thermodynamic constraint co-prediction model, the thermodynamic feasibility gate unit is activated; this unit calculates the Gibbs free energy change corresponding to each possible reaction path based on the basic thermodynamic principle of geopolymer condensation reaction.

[0067] The formula for determining the Gibbs free energy change is: in, The Gibbs free energy change for condensation polymerization. The preset Gibbs free energy change threshold;

[0068] Using the Gibbs free energy change threshold as the criterion, all message transmission channels between adjacent nodes are screened; message flows corresponding to non-spontaneous condensation paths with Gibbs free energy changes higher than the threshold are blocked, realizing topological pruning of invalid reaction paths by thermodynamic hard constraints, ensuring that the model only performs feature learning and prediction within the thermodynamically feasible parameter space.

[0069] Step S3-3: Dynamic control of graph message passing rate under dynamic hard constraints:

[0070] Synchronously activate the kinetic rate modulation unit to transform the Arrhenius kinetic equation into a temperature-concentration sensitive weight scaling factor;

[0071] The Arrhenius kinetic equation is: ,in, The reaction rate constant is... Pre-exponential factor, The activation energy of the reaction. The gas constant is Thermodynamic temperature;

[0072] The kinetic rate scaling factor is calculated based on the reaction rate constant, using the following formula: ,in, This is the scaling factor for the dynamic rate. The reaction rate constant under standard reference conditions;

[0073] The information transmission intensity of each adjacent edge in the graph convolution aggregation operator is dynamically controlled by the kinetic rate scaling factor; so that the graph message transmission rate of the dynamic graph convolution-thermodynamic constraint co-prediction model matches the kinetic stages of dissolution, migration and condensation of the alkaline-activated reaction, and realizes the full-process intervention of the kinetic hard constraint on the evolution rhythm of the graph structure.

[0074] Step S3-4: Global graph feature reading and generation of process-performance mapping relationship:

[0075] After multiple rounds of dynamic graph message passing and node state updates controlled by both thermodynamic and kinetic hard constraints, a global graph feature reading operation is performed; a combination of global average pooling and global max pooling is used to transform the converged graph structure representation into a fixed-dimensional global feature vector.

[0076] The global feature vector is input into the performance mapping decoder, which adopts a fully connected neural network structure and outputs the predicted distribution of three target properties: compressive strength, heavy metal solidification rate, and volume stability. A quantitative mapping relationship between the preset preparation process parameters and the three target properties is established and output.

[0077] Step S3-5: Construction and training of the dynamic graph convolution-thermodynamic constraint co-prediction model:

[0078] Historical test batch data covering different precursor ratios, alkali activator concentrations, and curing regimes were obtained; the topological chemical fingerprint vectors, preparation process parameters, and measured target performance of the silicon-aluminum-oxygen network corresponding to each batch were extracted to construct a collaborative training dataset; data distribution balancing and abnormal test batch removal were performed on the dataset to ensure the quality and representativeness of the training data.

[0079] Initialize the network topology of the dynamic graph convolution-thermodynamic constraint co-prediction model; sequentially configure the graph feature embedding layer, multi-scale dynamic graph convolution layer, thermodynamic-dynamic dual-constraint projection layer and target performance output layer; wherein the thermodynamic-dynamic dual-constraint projection layer is pre-set with the initialization parameters of the Gibbs free energy change threshold decision matrix and the Arrhenius kinetic rate scaling matrix to complete the basic structure of the model;

[0080] A composite loss function that senses physicochemical constraints is constructed to drive model parameter optimization; the formula for calculating the composite loss function is as follows: ,in, For the total composite loss, The target performance prediction deviation term is calculated using the mean square error. Penalty item for violating thermodynamic feasibility. For the kinetic rate to deviate from the canonical term, and These are the weighting coefficients for the thermodynamic penalty term and the kinetic regularization term, respectively;

[0081] During the backpropagation iteration, orthogonal projection truncation is performed on the gradient components that touch the hard constraint boundary; the model parameter update trajectory is forcibly constrained within the physicochemically allowed parameter feasible region, completing the gradient descent iteration with hard constraint embedding;

[0082] A dynamic early stopping strategy and a five-fold cross-validation mechanism are used to monitor the model training process. When the hard constraint satisfaction rate on the validation set reaches the preset convergence standard, and the prediction error of the mapping relationship between process parameters and target performance no longer decreases significantly for ten consecutive rounds, the network layer weights and thermodynamic-kinetic dual-constraint projection layer hyperparameters of the model are solidified, and a training-ready dynamic graph convolution-thermodynamic constraint co-prediction model instance is output.

[0083] In this embodiment, the construction and training of the dynamic graph convolution-thermodynamic constraint joint prediction model includes:

[0084] I. Construction and Preprocessing of Co-training Datasets:

[0085] Obtain historical test batch data covering different precursor ratios, alkali activator concentrations, and curing regimes; the historical test batch data should cover the entire area of ​​the process parameter search space; ensure that the data has sufficient diversity and representativeness;

[0086] Three types of core information were extracted from the data of each historical test batch: the first type is the topological chemical fingerprint vector of the silicon-aluminum-oxygen network; the second type is the preparation process parameters; and the third type is the measured target performance. The preparation process parameters include the mass ratio of copper tailings to fly ash, the modulus of the alkali activator, the concentration of the alkali activator, the curing temperature, the curing humidity, and the curing time. The measured target performance includes compressive strength, heavy metal solidification rate, and volume stability.

[0087] The three types of core information extracted are associated and stored according to the test batches; an initial collaborative training dataset is constructed.

[0088] Perform data distribution balancing processing on the initial co-training dataset; use synthetic minority class oversampling technique to augment the data in the sparse process parameter range; generate synthetic samples to supplement the initial co-training dataset; make the distribution of the dataset tend to be uniform throughout the entire process parameter search space;

[0089] Perform outlier test batch removal on the dataset after data distribution equalization; identify outlier samples using the 3σ criterion; calculate the mean and standard deviation of each target performance index; remove outlier test batches whose target performance index exceeds the range of mean plus or minus three times the standard deviation;

[0090] After data preprocessing, the final co-training dataset is obtained. The final co-training dataset is then randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1, which are used for model training, model validation, and model performance testing, respectively.

[0091] II. Initialization of Model Network Topology:

[0092] The network topology architecture of the dynamic graph convolution-thermodynamic constraint co-prediction model is initialized. The model adopts an end-to-end graph neural network structure, consisting of a graph feature embedding layer, a multi-scale dynamic graph convolutional layer, a thermodynamic-dynamic dual-constraint projection layer, and a target performance output layer.

[0093] Graph feature embedding layer:

[0094] The graph feature embedding layer is responsible for mapping the initial node representations of the input process-structure coupling graph to a unified high-dimensional feature space. The dimension of the input node representation is the sum of the dimension of the silicon-aluminum-oxygen network topological chemical fingerprint vector and the dimension of the fabrication process parameters. The graph feature embedding layer adopts a single-layer fully connected neural network structure. The output dimension is set to 64. The activation function is a linear rectified function.

[0095] Multi-scale dynamic graph convolutional layers:

[0096] The multi-scale dynamic graph convolutional layer is responsible for extracting the multi-scale topological features of the process-structure coupling graph. This model has three multi-scale dynamic graph convolutional layers with output dimensions of 64, 128, and 128, respectively.

[0097] Each multi-scale dynamic graph convolutional layer contains three graph convolutional branches with different receptive fields; the three branches correspond to feature extraction at the vertex, edge, and ring scales, respectively; each branch uses an independent graph convolutional kernel for message passing and feature aggregation; the output features of the three branches are concatenated to obtain the multi-scale topological features of the layer;

[0098] Multi-scale dynamic graph convolutional layers use a message-passing mechanism for feature updates; the node feature update formula is: ,in, For the first Layer The feature vector of each node For the first Layer The feature vector of each node For the first The set of adjacent nodes of a node. For the first The node and the first Edge weights between nodes For the first The learnable weight matrix of the layer, For the first The learnable bias vector of the layer, It is a linear rectification activation function;

[0099] Thermodynamic-Kinetic Dual-Constraint Projection Layer:

[0100] The thermodynamic-kinetic dual-constraint projection layer is the core innovative layer of this model; this layer embeds thermodynamic and kinetic hard constraints during graph message passing; it enables topological pruning of ineffective reaction paths and precise control of the graph structure evolution rhythm;

[0101] The thermodynamic-kinetic dual-constraint projection layer is composed of a thermodynamic feasibility gated unit and a kinetic rate modulation unit connected in parallel; the two units simultaneously act on the feature map output by the multi-scale dynamic graph convolutional layer.

[0102] The thermodynamic feasibility gating unit is based on the Gibbs free energy change threshold of the geopolymer condensation reaction; it performs path screening on the message transmission channels between adjacent nodes; and blocks the message flow corresponding to non-spontaneous condensation paths with Gibbs free energy changes higher than the threshold.

[0103] The kinetic rate modulation unit is based on the Arrhenius kinetic equation; it transforms the reaction rate constant into a temperature-concentration sensitive weight scaling factor; and it dynamically controls the information transmission intensity of each adjacent edge in the graph convolution aggregation operator.

[0104] The initialization parameters of the pre-set Gibbs free energy change threshold determination matrix and Arrhenius kinetic rate scaling matrix for the thermodynamic-kinetic dual-constraint projection layer are: the Gibbs free energy change threshold is set based on the thermodynamic data of geopolymer condensation reaction; the pre-exponential factor and reaction activation energy of the Arrhenius kinetic equation are obtained by fitting based on previous experimental data.

[0105] Target performance output layer:

[0106] The target performance output layer is responsible for mapping the extracted global graph features to the predicted values ​​of the target performance. First, it performs a global graph feature readout operation. It uses a combination of global average pooling and global max pooling to transform the converged graph structure representation into a global feature vector of fixed dimensions.

[0107] The global feature vector is input to a performance mapping decoder consisting of two fully connected neural networks; the first fully connected layer has 64 neurons and uses a linear rectified function as the activation function; the second fully connected layer has 3 neurons and corresponds to three target performance parameters: compressive strength, heavy metal solidification rate, and volume stability; the output layer does not use an activation function and directly outputs the predicted values ​​of the target performance.

[0108] After completing the configuration of each layer structure; initialize all weight parameters and bias parameters of the model; use the Xavier uniform initialization method for the weight parameters; initialize the bias parameters to 0;

[0109] III. Construction of the composite loss function for physicochemical constraint sensing:

[0110] A composite loss function with physicochemical constraint awareness is constructed to drive model parameter optimization. The composite loss function is composed of three weighted components: a target performance prediction bias term, a thermodynamic feasibility violation penalty term, and a kinetic rate deviation regularization term. The calculation formula for the composite loss function is as follows: ,in, For the total composite loss, For the target performance prediction deviation term, Penalty item for violating thermodynamic feasibility. For the kinetic rate to deviate from the canonical term, and These are the weighting coefficients for the thermodynamic penalty term and the kinetic regularization term, respectively;

[0111] Target performance prediction bias term:

[0112] The target performance prediction deviation term measures the difference between the model's predicted values ​​and the measured values; it is calculated using the mean squared error; the calculation formula is: ,in, For batch size, For the first The first sample Predicted values ​​of target performance, For the first The first sample Measured values ​​of the target performance;

[0113] Penalty for violating thermodynamic feasibility:

[0114] Thermodynamic feasibility violation penalty term is used to penalize violations of fundamental thermodynamic laws in model predictions; the number of reaction pathways with Gibbs free energy changes exceeding a threshold in each batch is counted; the thermodynamic feasibility violation rate is calculated; the thermodynamic feasibility violation penalty term is proportional to the square of the violation rate; the calculation formula is: ,in, The number of reaction pathways whose Gibbs free energy change is higher than a threshold. This represents the total number of reaction pathways.

[0115] Dynamic rate deviation from the canonical term:

[0116] The kinetic rate deviation regularization term is used to constrain the consistency between the model-predicted reaction rate and the actual reaction kinetics; it calculates the relative error between the model-predicted reaction rate and the theoretical reaction rate calculated by the Arrhenius kinetic equation; the kinetic rate deviation regularization term is the sum of squares of the relative errors; the calculation formula is: ,in, For the first The model predicts the reaction rate for each sample. For the first The theoretical reaction rate for each sample;

[0117] IV. Gradient Descent Training with Hard Constraint Embedding:

[0118] The Adam optimizer is used to update model parameters; the initial learning rate is set to 0.001; the batch size is set to 32; the maximum number of iterations is set to 200 rounds; the learning rate adopts a cosine annealing decay strategy; the learning rate decays to 0.99 times that of the previous round in each iteration.

[0119] During the backpropagation iteration, an orthogonal projection truncation operation is performed on the gradient components that touch the hard constraint boundary. This operation decomposes the gradient vector into components within the feasible region and components outside the feasible region. Only the components within the feasible region are retained for parameter updates, thus forcibly constraining the model parameter update trajectory within the physicochemically allowed parameter feasible region.

[0120] The formula for calculating the orthogonal projection truncation operation is: ,in, The gradient vector after projection. This is the original gradient vector. The unit normal vector of the hard-constrained boundary;

[0121] After each iteration, the model performance is evaluated on the validation set; the target performance prediction error, hard constraint satisfaction rate, and total composite loss are calculated on the validation set.

[0122] The formula for calculating the satisfaction rate of hard constraints is: ,in, The hard constraint satisfaction rate;

[0123] V. Model Training Monitoring and Consolidation:

[0124] A dynamic early stopping strategy and a five-fold cross-validation mechanism are used to monitor the model training process. The training set is divided into five non-overlapping subsets. Four subsets are selected as the training set and one subset is selected as the validation set each time. The training process is repeated five times. The average of the five validation results is taken as the final validation performance of the model.

[0125] The model is considered to have converged when the following two conditions are met: the first condition is that the satisfaction rate of the hard constraints on the validation set reaches the preset convergence criterion; the second condition is that the prediction error of the target performance on the validation set no longer decreases significantly after ten consecutive rounds; the criterion for a significant decrease is that the decrease in prediction error is less than 0.1%.

[0126] After the model training converges, solidify the network layer weights and thermodynamic-dynamic dual-constraint projection layer hyperparameters of the dynamic graph convolution-thermodynamic constraint co-prediction model; save the model's structure file and weight parameter file; output a training-ready instance of the dynamic graph convolution-thermodynamic constraint co-prediction model.

[0127] The trained model was tested using a test set. The test results showed that the model's average prediction error for the target performance was less than 3%, the hard constraint satisfaction rate was greater than 98%, and it could meet the accuracy requirements for optimizing the process parameters of geopolymer preparation.

[0128] In this embodiment, the core function of step S4 is to use a multi-objective optimization algorithm that integrates chemical prior knowledge to globally optimize the process parameters based on the mapping relationship between the process parameters and target performance generated in step S3. The algorithm improves search efficiency by constructing a hybrid kernel function prior, dynamically adjusts the balance between exploration and utilization of the acquisition function, generates a Pareto front solution set covering the optimal performance range, and finally combines chemical fingerprint similarity to screen out the optimal combination of process parameters that takes into account both multi-objective performance and reaction feasibility. The detailed steps are as follows:

[0129] Step S4-1: Construction of hybrid kernel function priors based on chemical fingerprints:

[0130] Based on the mapping relationship between the silicon-aluminum-oxide network topological chemical fingerprint vector generated in step S2 and the process parameters and target performance established in step S3, the Tanimoto similarity between the fingerprint vectors corresponding to the historical process parameter samples is calculated; the formula for calculating the Tanimoto similarity is: ,in, For the first The and the first Tanimoto similarity of fingerprint vectors of each sample For the first The topological chemical fingerprint vector of a sample silicon-aluminum-oxygen network. For the first The topological chemical fingerprint vector of a silicon-aluminum-oxygen network for each sample;

[0131] Calculate the Mahalanobis distance between the fingerprint vectors corresponding to historical process parameter samples; the formula for calculating the Mahalanobis distance is: ,in, For the first The and the first Mahalanobis distance of fingerprint vectors of samples Let be the covariance matrix of all sample fingerprint vectors;

[0132] The chemical structure overlap features represented by Tanimoto similarity and the distribution discrete features represented by Mahalanobis distance are nonlinearly fused to construct a hybrid kernel function prior; the calculation formula of the hybrid kernel function is as follows: ,in, For the hybrid kernel function in the first... The and the first The values ​​between samples, The weighting coefficients for Tanimoto similarity. The bandwidth parameter of the Gaussian kernel;

[0133] Based on the hybrid kernel function, the kernel function values ​​among all historical samples are calculated to generate a kernel function covariance matrix covering the entire process parameter search space;

[0134] Step S4-2: Initialization of the desired improved acquisition function:

[0135] Input the kernel function covariance matrix into the acquisition function construction module of the chemical fingerprint similarity weighted Pareto front adaptive exploration algorithm; use the multi-objective performance prediction distribution output in step S3 as a benchmark to initialize the desired improved acquisition function;

[0136] Set the initial ratio of the exploration weight factor and the utilization weight factor; set the initial value of the exploration weight factor to 0.7 and the initial value of the utilization weight factor to 0.3; output the sample collection function configuration example to be iteratively optimized.

[0137] Step S4-3: Generation of the initial Pareto front solution set:

[0138] Based on the sample acquisition function configuration instance to be iteratively optimized, an initial candidate parameter combination sequence is generated in the process parameter search space using the Latin hypercube sampling method; the number of initial candidate parameter combinations is set to 10 times the dimension of the process parameters.

[0139] Input the initial candidate parameter combination sequence into the mapping relationship between process parameters and target performance established in step S3 to obtain the corresponding multi-target performance prediction value set of compressive strength, heavy metal solidification rate and volume stability;

[0140] Using the overall optimality of target performance as the criterion, a non-dominated sorting is performed on the multi-target performance prediction value set; the non-dominated sorting divides all solutions into different non-dominated levels according to the dominance relationship of the solutions, and the solutions of the first level constitute the initial non-dominated solution set;

[0141] Perform a crowding assessment on the initial non-dominated solution set, calculate the crowding distance of each solution in the target space; remove dense solutions with crowding distances less than a preset threshold, generate an initial Pareto front solution set; output a front distribution state matrix characterizing the spatial distribution of the solution set.

[0142] Step S4-4: Exploring the Acquisition Function - Adaptive Adjustment Using Weights:

[0143] The local clustering density gradient and global expansion rate of the solution in the frontier region are analyzed in real time based on the frontier distribution state matrix. The local clustering density gradient is obtained by calculating the local density change rate of the frontier solution in the target space, and the global expansion rate is obtained by calculating the change rate of the hypervolume of the frontier solution set.

[0144] The local cluster density gradient is converted into a decay adjustment amount for the exploration weight factor; the larger the local cluster density gradient, the greater the decay of the exploration weight factor; the global expansion rate is converted into a gain adjustment amount for utilizing the weight factor; the smaller the global expansion rate, the greater the gain of utilizing the weight factor.

[0145] The exploration and utilization weight ratios in the acquisition function configuration instance to be iteratively optimized are adaptively redistributed; after redistribution, the sum of the exploration weight factor and the utilization weight factor remains at 1; and a dynamically updated acquisition function driver engine is generated.

[0146] Step S4-5: Pareto Front Iterative Update and Final State Locking:

[0147] The acquisition function with dynamically updated weights drives the engine to perform iterative sampling and prediction evaluation loops within the process parameter search space; each iteration generates 5 new candidate parameter combinations, and their performance prediction results are incorporated into the historical evaluation set;

[0148] Based on the updated historical evaluation set, re-execute the non-dominated sorting and solution set crowding assessment, update the initial Pareto front solution set; repeat the front distribution situation analysis and exploration using weighted adaptive redistribution operation;

[0149] When the rate of change of the hypervolume index of the Pareto front solution set is lower than the preset convergence threshold for three consecutive iterations and no new nondominated solution is generated for five consecutive iterations, the iteration process is terminated; the current Pareto front solution set is locked as the final Pareto front solution set.

[0150] Step S4-6: Optimal process parameter selection based on chemical fingerprint similarity:

[0151] Based on the final state Pareto front solution set, a chemical fingerprint similarity weighted decision mechanism is introduced; an ideal reaction state reference fingerprint vector is preset, which corresponds to the theoretical state where the geopolymer condensation reaction is complete and the heavy metal solidification effect is optimal;

[0152] Calculate the Tanimoto similarity between the topological chemical fingerprint vector of the silicon-aluminum-oxygen network corresponding to each solution in the final Pareto front solution set and the reference fingerprint vector of the ideal reaction state; use the Tanimoto similarity as a weighting coefficient and multiply it by the multi-objective performance normalization score of the solution to obtain the comprehensive score;

[0153] Select the process parameter combination with the highest comprehensive score that meets the safety margin requirements of the thermodynamic and kinetic dual-constraint boundary; output this combination as the optimal process parameter combination.

[0154] In this embodiment, the core function of step S5 is to input the optimal combination of process parameters obtained through offline optimization into the actual preparation system, monitor the reaction process in real time through in-situ vibrational spectroscopy, convert the spectral signal into chemical coordinates of the reaction process, realize online dynamic updating of heterogeneous reaction spectra and iterative calibration of the prediction model, form a data-driven closed-loop optimization loop, and finally output the final preparation process parameters with the highest accuracy and best stability under actual preparation conditions; the detailed steps are as follows:

[0155] Step S5-1: In-situ vibrational spectral data acquisition and preprocessing:

[0156] The optimal combination of process parameters output in step S4 is sent to the preparation control system to start the geopolymer preparation process; in-situ Raman spectra or in-situ Fourier transform infrared spectra in the reaction system are continuously collected at a preset sampling period; the sampling period is set according to the reaction rate, with the sampling period set to 1 minute in the early stage of the reaction and extended to 5 minutes in the later stage of the reaction.

[0157] The acquired in-situ vibrational spectrum raw data were sequentially processed by baseline drift correction, fluorescence background subtraction, and characteristic peak deconvolution. Baseline drift correction adopted the adaptive iterative reweighted penalized least squares method. Fluorescence background subtraction adopted the polynomial fitting method. Characteristic peak deconvolution adopted the Gaussian function fitting method to separate overlapping characteristic peaks into independent single peaks.

[0158] Extract the peak position shift of the silicon-oxygen framework asymmetric stretching vibration, the half-width evolution curve, and the characteristic peak area integral ratio from the preprocessed spectral data; arrange the above characteristic parameters in chronological order to generate a time-series spectral characteristic sequence.

[0159] Step S5-2: Generation of chemical coordinates for the reaction process:

[0160] Based on the time-series spectral feature sequence, a joint phase space projection model of peak position shift and vibration intensity is constructed; the ratio of the peak position shift of the silicon-oxygen skeleton asymmetric stretching vibration to the integral of the characteristic peak area at each sampling time is used as the two coordinate dimensions of the phase space to construct a two-dimensional phase space.

[0161] Principal component orthogonal decomposition is used to eliminate thermal noise and solvation interference signals in the phase space; the first two principal components are extracted as the reaction progress axis, and the spectral evolution trajectory is mapped to the multidimensional reaction progress axis.

[0162] Based on the variation of spectral characteristics, the reaction process is divided into the precursor dissolution stage, the ion migration stage, and the network condensation stage; the chemical coordinate sequence of the reaction process is output to characterize the continuous evolution of the three stages; each element in the chemical coordinate sequence of the reaction process corresponds to a reaction progress value at a sampling time, with a value range of 0 to 1; 0 corresponds to the reaction start time, and 1 corresponds to the reaction completion time;

[0163] Step S5-3: Online reconnection of node weights in the heterogeneous reaction map:

[0164] Based on the current reaction phase identified by the chemical coordinate sequence of the reaction process, the heterogeneous reaction map of the geopolymer precursor constructed in step S1 is called; the bond order evolution index in the chemical coordinate sequence of the reaction process is extracted; the bond order evolution index is calculated by the peak position shift of the asymmetric stretching vibration of the silicon-oxygen framework, reflecting the change of chemical bond strength with the reaction process;

[0165] The bond-level evolution index is mapped to the node activity weight correction coefficient; the formula for calculating the node activity weight correction coefficient is as follows: ,in, This is the node activity weight correction coefficient. This is the proportionality coefficient. The offset of the peak position of the asymmetric stretching vibration of the silicon-oxygen framework relative to the initial moment.

[0166] For silicon-aluminum-oxygen tetrahedral cluster nodes and heavy metal-bearing phase nodes in the condensation critical state in the spectrum, the adjacency probability is recalculated and the connection channel is dynamically switched. The adjacency probability recalculation is based on the product of the node activity weight correction coefficient and the original edge weight. The connection channel dynamic switching means that when the adjacency probability is greater than the preset threshold, a new connection edge is established; when the adjacency probability is less than the preset threshold, the original connection edge is disconnected.

[0167] After completing the weight update and connection channel switching of all critical state nodes, a local topology rewiring graph is generated;

[0168] Step S5-4: Incremental update of heterogeneous reaction spectrum structure:

[0169] The local topological rewire map was compared with the original geopolymer precursor heterogeneous reaction map by sub-graph isomorphism comparison and topological overlay; the newly generated bridging oxygen connection edges due to condensation reaction and the coordination polyhedral distortion regions caused by ion solidification were identified.

[0170] A graph embedding alignment algorithm is used to align the node features of the local topology rewire graph with the node features of the original graph. The graph embedding alignment algorithm achieves feature space unification by minimizing the Euclidean distance between the corresponding node feature vectors of the two graphs.

[0171] The adjacency matrix incremental splicing technique is used to seamlessly integrate the reconnection weights into the edge weight attribute matrix of the original graph; the continuity of the graph Laplacian feature spectrum distribution is maintained, and the model prediction instability caused by abrupt changes in graph structure is avoided.

[0172] A dynamic incremental heterogeneous reaction map was constructed; this map reflects the current microstructure state and reactivity distribution of the reaction system in real time.

[0173] Step S5-5: Iterative calibration of the dynamic graph convolution-thermodynamic constraint co-prediction model:

[0174] The node characterization and edge weight attributes of the dynamic incremental heterogeneous reaction spectrum are fed into the dynamic graph convolution-thermodynamic constraint co-prediction model trained in step S3 in real time; the model outputs the predicted distribution of compressive strength, heavy metal solidification rate and volume stability under the current process parameters.

[0175] Meanwhile, the measured performance surrogate value is obtained by inversion based on the in-situ spectral data; the measured performance surrogate value is calculated by a pre-established quantitative relationship model between spectral characteristics and target performance;

[0176] Calculate the predicted residual vector of the process parameter mapping path; the formula for calculating the predicted residual vector is: ,in, To predict the residual vector, The target performance prediction vector output by the model. This is the vector of measured performance surrogate values ​​for in-situ spectral inversion;

[0177] Based on the residual gradient direction, constrained partial derivative corrections are performed on the message passing weight matrix and kinetic rate scaling factor of the dynamic graph convolution-thermodynamic constraint co-prediction model; during the correction process, it is necessary to ensure that the model parameters are always within the physicochemically permissible parameter feasible region; complete the iterative calibration of model parameters and output the calibration status;

[0178] Steps S5-6: Iteration Termination and Final Process Parameter Output:

[0179] The closed-loop process from in-situ spectral acquisition to model calibration is executed iteratively; the norm decay rate of the predicted residual vector is monitored in real time; the norm of the predicted residual vector is calculated using the following formula: ,in, For the first The square of the prediction residuals for each target performance;

[0180] The iteration loop terminates when the norm of the predicted residual vector is consistently below the preset threshold for five consecutive sampling periods, and the fluctuation amplitude of the multi-objective performance prediction output by the dynamic graph convolution-thermodynamic constraint co-prediction model falls within the allowable tolerance band.

[0181] Extract the combination of process parameters corresponding to the current round; after dual verification of thermodynamic feasibility boundary and kinetic safety margin, output the final optimized preparation process parameters.

[0182] The present invention will be further described below with reference to specific embodiments:

[0183] Example

[0184] I. Experimental Materials and Instruments:

[0185] 1. Experimental materials:

[0186] Copper tailings: Taken from flotation tailings of a copper mine in Fujian Province, ball-milled to a particle size ≤74μm. Main chemical composition (mass fraction): SiO2 58.2%, Al2O3 12.6%, Fe2O3 8.7%, CuO 0.32%, PbO 0.18%, ZnO 0.25%, CaO 3.1%, MgO 2.4%, loss on ignition 4.85%; the phase composition is mainly quartz, feldspar, and kaolinite, with small amounts of chalcopyrite and galena.

[0187] Fly ash: Taken from Class I fly ash of a coal-fired power plant, with a particle size ≤45μm. Main chemical composition (mass fraction): SiO2 52.3%, Al2O3 28.7%, Fe2O3 6.2%, CaO 4.5%, MgO 1.8%, loss on ignition 3.2%; the phase composition is mainly glassy phase, containing a small amount of mullite and quartz.

[0188] Alkali activator: It is a compound of industrial-grade sodium hydroxide (purity 96%) and sodium silicate solution (modulus 3.3, SiO2 28.5%, Na2O 8.8%).

[0189] 2. Main instruments:

[0190] X-ray fluorescence spectrometer, X-ray diffractometer, scanning electron microscope, in-situ Raman spectrometer, microcomputer-controlled electronic universal testing machine, inductively coupled plasma atomic emission spectrometer, constant temperature and humidity curing chamber;

[0191] II. Specific Implementation Steps:

[0192] Step S1: Constructing the heterogeneous reaction map of geopolymer precursors:

[0193] Basic data collection: Three parallel samples of copper tailings and fly ash were prepared respectively. The elemental composition was determined by XRF, the phase composition was determined by XRD (scanning range 5°-80°, step size 0.02°), and the microscopic morphology images were acquired by SEM (magnification of 5000x and 20000x). The average value of the three tests was taken as the raw data.

[0194] Data preprocessing: The original data was mapped to the [0, 1] interval using min-max normalization, and background noise was filtered out using db4 wavelet (decomposition level 5). Key parameters such as silicon-aluminum molar coordination ratio (copper tailings 1.56, fly ash 1.23), crystal phase mass fraction, average particle size (copper tailings 32μm, fly ash 18μm), and specific surface area were extracted to generate a standardized precursor physicochemical characteristic dataset.

[0195] Node feature generation: The coordination environment boundary of the silicon-aluminum-oxygen tetrahedral clusters is delineated using the phase region analysis algorithm. Combined with the heavy metal occurrence state data obtained by the continuous extraction method (Cu mainly exists in the iron-manganese oxide bound state, accounting for 62.3%; Pb mainly exists in the carbonate bound state, accounting for 58.7%), a feature sequence containing 128 silicon-aluminum-oxygen tetrahedral cluster nodes and 16 heavy metal occurrence phase nodes is generated. Each node carries the topology type, elemental composition, and three-dimensional spatial coordinate label.

[0196] Edge weight calculation and graph assembly: setting weight coefficients , , The edge weights are calculated by combining the electronegativity difference of bonding elements, bond valence, and coordination polyhedron sharing degree. The edge weight calculation formula is as follows: ,in, For edge weights, The electronegativity difference between bonding elements is represented by BVS, where BVS is the bond valence and numerical value. For coordination polyhedron sharing degree;

[0197] Introducing the Gibbs free energy change threshold kJ / mol and steric hindrance constraints are used to filter out virtual connections, followed by Gaussian kernel smoothing. The final heterogeneous reaction map of the geopolymer precursor was then constructed.

[0198] Step S2: Generate the topological chemical fingerprint vector of the silicon-aluminum-oxygen network:

[0199] Calculation of the connectivity index of a polyhedron: setting weighting coefficients , , By combining the mean of node degree centrality (2.37), the mean of betweenness centrality (0.082), and the mean of cycle closure (0.64), the polyhedral connectivity index is obtained: Where CI is the polyhedral connectivity index. The mean of the degree centrality of the nodes. The mean of the betweenness centrality of the nodes. The mean of the closure degree of the ring system is calculated as follows: =1.26;

[0200] Calculation of the spatial distribution entropy of base activation sites: The reaction space is divided into... A cubic mesh was used to select 89 surface non-bridging oxygen nodes and 27 charge-compensated cation nodes as active sites. An electronegativity difference correction coefficient was introduced to weight the site density, and the spatial distribution entropy of the alkali-excited sites was calculated. ,in, The entropy of the spatial distribution of base-excitation sites. This represents the total number of three-dimensional spatial grid cells. For the first The proportion of the activity-weighted site density within each grid cell to the total site density; calculated ;

[0201] Calculation of migration barriers for heavy metal ions: Construction along the edge paths of heterogeneous reaction maps , The migration network was simulated using a restricted random walk algorithm to calculate the migration barrier of heavy metal ions. ,in, This acts as a barrier for the migration of heavy metal ions. The total number of possible migration paths. For the first The maximum cumulative potential energy along the migration path For the first The initial potential energy of the migration path is calculated. The migration barrier is 18.7 kJ / mol. The migration barrier is 21.3 kJ / mol.

[0202] Calculation of chemical information entropy of reaction pathways: The heterogeneous reaction map is divided into sub-graphs of dissolved state, migrating state, and condensation state according to time sequence. The path transition probability distribution between sub-graphs is calculated to obtain the chemical information entropy of the reaction pathways. =2.45;

[0203] Feature dimensionality reduction and encoding: t-SNE manifold dimensionality reduction (target dimension 16) and Z-score normalization are performed on the above 4 features to generate a 16-dimensional normalized silicon-aluminum-oxygen network topological chemical fingerprint vector;

[0204] Step S3: Construct and train a dynamic graph convolutional-thermodynamically constrained co-prediction model:

[0205] Training dataset construction: 120 sets of historical experimental data were collected, covering the mass ratio of copper tailings to fly ash (3:7-7:3), modulus of alkali activator (1.0-1.8), concentration of alkali activator (8-16 mol / L), curing temperature (20-80℃), and curing time (1-28 days). Corresponding topological chemical fingerprint vectors, process parameters, and measured performance (compressive strength, Cu curing rate, Pb curing rate, volume shrinkage rate) were extracted. The SMOTE algorithm was used to balance the data distribution, and 6 sets of outliers were removed using the 3σ criterion. The dataset was then divided into training, validation, and test sets in a 7:2:1 ratio.

[0206] Model initialization: Configuration graph feature embedding layer (input dimension 22, output dimension 64), 3-layer multi-scale dynamic graph convolutional layer (output dimensions 64, 128, 128), thermodynamic-kinetic dual-constraint projection layer (preset). kJ / mol, pre-exponential factor ,activation energy kJ / mol), target performance output layer (2 fully connected layers, output dimension 4); weights are initialized uniformly using Xavier, and biases are initialized to 0; model training: Adam optimizer is used, initial learning rate 0.001, batch size 32, maximum iterations 200 epochs, cosine annealing learning rate decay coefficient 0.99; construct a physicochemical constraint-aware composite loss function: ,in, For the total composite loss, The target performance prediction bias term (mean square error) is... Penalty item for violating thermodynamic feasibility. For the kinetic rate to deviate from the canonical term, and These are the weighting coefficients for the thermodynamic penalty term and the kinetic regularization term, respectively;

[0207] set up , Perform orthogonal projection truncation on gradient components that touch hard constraint boundaries: ,in, The gradient vector after projection. This is the original gradient vector. The unit normal vector of the hard-constrained boundary;

[0208] Model convergence verification: Training was stopped when the hard constraint satisfaction rate on the validation set reached 98.7% and the target performance prediction error decreased by less than 0.1% for 10 consecutive rounds; the test set results showed that the model's average prediction error for the target performance was 2.68%, which met the accuracy requirements for process optimization.

[0209] Step S4: Optimization using a chemical fingerprint similarity weighted Pareto front adaptive exploration algorithm:

[0210] Hybrid kernel function construction: Calculate the Tanimoto similarity between historical sample fingerprint vectors: ,in, For the first The and the first Tanimoto similarity of fingerprint vectors of each sample For the first The topological chemical fingerprint vector of a sample silicon-aluminum-oxygen network. For the first The topological chemical fingerprint vector of a silicon-aluminum-oxygen network for each sample;

[0211] Calculate the Mahalanobis distance between historical sample fingerprint vectors: ,in, For the first The and the first Mahalanobis distance of fingerprint vectors of samples Let be the covariance matrix of all sample fingerprint vectors;

[0212] By nonlinearly fusing Tanimoto similarity and Mahalanobis distance, a hybrid kernel function is constructed: ,in, For the hybrid kernel function in the first... The and the first The values ​​between samples, The weighting coefficients for Tanimoto similarity. The bandwidth parameter of the Gaussian kernel;

[0213] set up , Generate a kernel function covariance matrix that covers the entire process parameter search space;

[0214] Acquisition function initialization: The desired improved acquisition function is adopted, with an initial exploration weight factor of 0.7 and a utilization weight factor of 0.3.

[0215] Initial Pareto front generation: 60 initial candidate parameter combinations are generated using Latin hypercube sampling. After model prediction, non-dominated sorting and solution set crowding evaluation are performed. Dense solutions with crowding distance <0.05 are removed, generating an initial Pareto front solution set containing 12 solutions.

[0216] Adaptive iterative optimization: Each iteration samples 5 new parameters and dynamically adjusts the exploration-utilization weights based on the local clustering density gradient and global expansion rate of the solution in the front region; when the rate of change of the Pareto front hypervolume index is <0.05% for 3 consecutive rounds and no new non-dominated solutions are generated for 5 consecutive rounds, the final Pareto front solution set (a total of 18 solutions) is locked.

[0217] Optimal parameter selection: A reference fingerprint vector of the ideal reaction state is preset (corresponding to the theoretical state where the polycondensation reaction is complete and the heavy metal solidification effect is optimal). The Tanimoto similarity between each frontier solution and the reference vector is calculated. Combined with the multi-objective performance normalization score, the initial optimal process parameter combination with the highest comprehensive score and meeting the dual-constraint safety margin requirements is selected: copper tailings to fly ash mass ratio 5:5, alkali activator modulus 1.4, alkali activator concentration 12mol / L, curing temperature 60℃, curing humidity 90%, curing time 7d;

[0218] Step S5: In-situ spectral closed-loop calibration and final process parameter output:

[0219] In-situ vibrational spectroscopy acquisition: Initial optimal parameters are sent to the preparation control system to start the geopolymer preparation process; reaction data are continuously acquired using an in-situ Raman spectrometer, with sampling periods of 1 min for 0-2 h, 5 min for 2-24 h, and 30 min after 24 h; baseline drift correction, fluorescence background subtraction, and Gaussian function characteristic peak deconvolution are sequentially performed on the original spectra to extract the asymmetric stretching vibration peak of the silicon-oxygen framework (-1050 cm⁻¹). -1 The peak position shift, half-width evolution curve and characteristic peak area integral ratio of the peak position are used to generate a time-series spectral characteristic sequence.

[0220] Chemical coordinate generation of reaction process: Construct a joint phase space of peak position shift and vibration intensity, eliminate thermal noise and solvation interference by principal component orthogonal decomposition, and extract the first two principal components as the reaction progress axis; divide the reaction into dissolution stage (0-4h), migration stage (4-12h), and condensation stage (12-168h), and output the chemical coordinate sequence of reaction process with values ​​in the range of [0, 1].

[0221] Online updates of heterogeneous reaction spectra: mapping the peak shift of the silicon-oxygen framework to a node activity weighting correction coefficient (proportional coefficient). =0.008cm), the adjacency probability recalculation and connection channel dynamic switching are performed on the silicon-aluminum-oxygen tetrahedral cluster nodes and heavy metal-bearing phase nodes in the critical state of condensation; after subgraph isomorphism comparison and incremental splicing of adjacency matrix, a dynamic incremental heterogeneous reaction spectrum is generated.

[0222] Model Iterative Calibration: The node representations and edge weight attributes of the dynamic incremental heterogeneous response map are fed into the prediction model in real time. The predicted performance values ​​output by the model are compared with the measured performance surrogate values ​​obtained from in-situ spectral inversion, and the predicted residual vector is calculated. Based on the residual gradient direction, the message passing weight matrix and dynamic rate scaling factor of the model are corrected by constrained partial derivatives.

[0223] Iteration Termination and Final Parameter Output: The iteration terminates when the norm of the predicted residual vector is consistently below 0.02 for 5 consecutive sampling periods and the fluctuation amplitude of the multi-objective performance prediction is <±1%. After dual verification of thermodynamic feasibility boundary and kinetic safety margin, the final optimized preparation process parameters are output: copper tailings to fly ash mass ratio 5.2:4.8, alkali activator modulus 1.38, alkali activator concentration 11.8 mol / L, curing temperature 58℃, curing humidity 90%, and curing time 6.5 days.

[0224] III. Performance Test Results:

[0225] Geopolymer samples prepared according to the final optimized process parameters achieved a 7-day compressive strength of 42.6 MPa and a 28-day compressive strength of 58.3 MPa; Cu curing rate was 99.2%, and Pb curing rate was 99.5%; the 28-day volume shrinkage rate was 0.32%, meeting the requirements of GB / T14684-2011 "Construction Sand" and GB5085.3-2007 "Identification Standard for Hazardous Waste - Leaching Toxicity Identification"; compared with the traditional orthogonal experimental optimization method, the optimization cycle of the method of this invention is shortened by 65%, the overall target performance is improved by 18.7%, and the long-term curing stability of heavy metals is significantly improved.

[0226] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing process parameters in the preparation of geopolymers from copper tailings fly ash based on cheminformatics, characterized in that: Includes the following steps: S1: Collect data on the composition, phase composition and micromorphology of copper tailings and fly ash, construct a heterogeneous reaction spectrum of geopolymer precursors, and characterize the nodes of the spectrum as silicon-aluminum-oxygen tetrahedral clusters and heavy metal-bearing phases. The edge weights are calculated from the difference in electronegativity of elements, bond valence and coordination polyhedral sharing degree. S2: Based on the principles of cheminformatics, topological dimensionality reduction and reaction path encoding are performed on heterogeneous reaction maps to generate silicon-aluminum-oxygen network topological chemical fingerprint vectors. The vectors include polyhedral connectivity index, spatial distribution entropy of base excitation sites, migration barrier of heavy metal ions, and chemical information entropy of reaction paths. S3: Input the topological chemical fingerprint vector of the silicon-aluminum-oxygen network and the preset preparation process parameters into the dynamic graph convolution-thermodynamic constraint co-prediction model. The dynamic graph convolution-thermodynamic constraint co-prediction model embeds the Gibbs free energy change threshold of the geopolymer condensation reaction and the Arrhenius kinetic equation as hard constraints during the graph message passing process, and outputs the mapping relationship between process parameters and target performance. S4: Based on the mapping relationship between process parameters and target performance, a chemical fingerprint similarity weighted Pareto front adaptive exploration algorithm is used to globally optimize process parameters. The algorithm constructs a hybrid kernel function prior with the Tanimoto similarity and Mahalanobis distance of the topological chemical fingerprint vector of the silicon aluminum oxide network, dynamically adjusts the exploration-utilization weight of the acquisition function, and outputs the optimal combination of process parameters. S5: Input the optimal combination of process parameters into the preparation system, collect in-situ vibrational spectral data in real time and convert it into chemical coordinates of the reaction process, perform online reconnection of node weights and incremental update of graph structure on the heterogeneous reaction spectrum, drive the dynamic graph convolution-thermodynamic constraint collaborative prediction model for iterative calibration until the prediction residual converges to the preset threshold, and output the final optimized preparation process parameters.

2. The method for optimizing process parameters of copper tailings fly ash geopolymer preparation based on cheminformatics according to claim 1, characterized in that: Data on the component content, phase composition, and microstructure of copper tailings and fly ash were collected to construct a heterogeneous reaction map of geopolymer precursors. The nodes of the map represent silicon-aluminum-oxygen tetrahedral clusters and heavy metal-bearing phases. Edge weights were calculated based on elemental electronegativity differences, bond valences, and coordination polyhedral sharing. Specifically, the map includes: The original data of component content, phase composition and micromorphology of copper tailings and fly ash were obtained. The original data were normalized and calibrated and background noise was filtered out. The silicon-aluminum molar coordination ratio, crystal phase mass fraction and particle surface topological characteristic parameters were extracted to generate a precursor physicochemical characteristic dataset. Based on the precursor physicochemical feature dataset, the coordination environment boundary of silicon-aluminum-oxygen tetrahedral clusters is delineated using the phase region analysis algorithm. Combined with the analysis results of the occurrence state of heavy metal elements, an occurrence phase space mapping model is established to generate a node feature sequence carrying topological type labels and three-dimensional spatial coordinates. Based on the spatial adjacency relationship and elemental properties of adjacent nodes in the node feature sequence, the electronegativity difference distribution between bonding elements is calculated. The bond valence and value are deduced by matching crystallographic parameters. The proportional weights of corner sharing, edge sharing and face sharing of coordination polyhedra are quantized based on the topological feature parameters of the particle surface. The electronegativity difference distribution, bond valence and value and sharing proportional weights are fused with multi-source information and gradient quantized to generate an edge weight attribute matrix. By performing graph-theoretic topological assembly of node feature sequences and edge weight attribute matrices, introducing thermodynamic stability criteria and spatial steric constraints to filter out virtual edges, and performing graph structure smoothing on the edge weight attribute matrix, a heterogeneous reaction spectrum of geopolymer precursors is constructed.

3. The method for optimizing process parameters of copper tailings fly ash geopolymer preparation based on cheminformatics according to claim 1, characterized in that: Based on the principles of cheminformatics, topological dimensionality reduction and reaction path encoding are performed on heterogeneous reaction maps to generate a silicon-aluminum-oxygen network topological chemical fingerprint vector. This vector includes the polyhedral connectivity index, spatial distribution entropy of base excitation sites, heavy metal ion migration barriers, and chemical information entropy of reaction paths. Specifically, it includes: Graph traversal and substructure mining are performed on the heterogeneous reaction map of geopolymer precursors to identify the nodes of silicon-aluminum-oxygen tetrahedral clusters and the distribution of their adjacent coordination polyhedral rings. A multi-scale topological adjacency matrix is ​​constructed by combining the shared proportion weight of coordination polyhedra in the edge weight attribute matrix. Local connectivity features of nodes are extracted based on graph Laplacian feature spectrum decomposition of the map. The degree centrality, betweenness centrality and ring closure are weighted and fused to generate a polyhedral connectivity index. Based on the local topological constraints characterized by the polyhedral connectivity index, and combined with the three-dimensional spatial coordinates and element attribute labels in the node feature sequence, surface non-bridging oxygen and charge-compensated cation nodes are selected as alkali-excited active sites. A spatial grid partitioning algorithm is used to divide the neighborhood of the active sites, and the density distribution frequency of sites within each grid is calculated. The electronegativity difference distribution of elements in the edge weight attribute matrix is ​​introduced as a site reactivity correction coefficient to perform activity weighting on the density distribution frequency. Based on the information theory entropy calculation principle, the spatial aggregation and dispersion of sites are solved to generate the spatial distribution entropy of alkali-excited sites. Based on the spatial steric hindrance distribution characterized by the spatial distribution entropy of alkali-excited sites, a heavy metal ion migration network is constructed along the edge path of the heterogeneous reaction spectrum. The bond valence and numerical values ​​in the edge weight attribute matrix are inversely mapped to the proportional weights shared by the coordination polyhedra as migration resistance parameters. The constrained random walk algorithm is used to simulate the diffusion trajectory of heavy metal ions in the polyhedral voids. Combined with the constraints of the local coordination environment of the nodes, the cumulative potential energy extremum and the transition state energy drop on the migration path are calculated to quantify the heavy metal ion migration barrier. By fusing the polyhedral connectivity index, the spatial distribution entropy of alkali-excitation sites, and the migration barrier of heavy metal ions, a state transition network for precursor condensation polymerization is constructed. The graph nodes are divided into subgraphs of dissolved state, migration state, and condensation state according to the reaction process sequence. The path transition probability distribution between each state subgraph is calculated, and a reaction path diversity measurement model from cheminformatics is introduced to normalize the path transition probability distribution and solve for the information entropy, generating the chemical information entropy of the reaction path. Manifold space projection dimensionality reduction and feature alignment are performed on the polyhedral connectivity index, the spatial distribution entropy of alkali-excitation sites, the migration barrier of heavy metal ions, and the chemical information entropy of the reaction path to eliminate topological redundancy. After splicing and standardizing the encoding, the topological chemical fingerprint vector of the silicon-aluminum-oxygen network is output.

4. The method for optimizing process parameters of copper tailings fly ash geopolymer preparation based on cheminformatics according to claim 1, characterized in that: The topological chemical fingerprint vector of the silicon-aluminum-oxygen network and the preset preparation process parameters are input into a dynamic graph convolution-thermodynamic constraint co-prediction model. This model embeds the Gibbs free energy threshold and the Arrhenius kinetic equation for the geopolymer condensation reaction as hard constraints during graph message passing, outputting the mapping relationship between process parameters and target performance, specifically including: The topological chemical fingerprint vector of the silicon-aluminum-oxygen network is tensor aligned and multidimensional features are spliced ​​with the preset fabrication process parameters to construct the initial node representation of the process-structure coupling graph, which serves as the input layer state of the dynamic graph convolution-thermodynamic constraint collaborative prediction model. In the graph message passing process of the dynamic graph convolution-thermodynamic constraint co-prediction model, the thermodynamic feasibility gate unit is activated. Using the Gibbs free energy change threshold as the criterion, the message passing channel between adjacent nodes is screened to block the message flow corresponding to the non-spontaneous condensation path with a free energy change higher than the Gibbs free energy change threshold, thereby realizing the topology pruning of invalid reaction paths by thermodynamic hard constraints. Synchronously activate the kinetic rate modulation unit to transform the Arrhenius kinetic equation into a temperature-concentration sensitive weight scaling factor, dynamically adjust the information transmission intensity of each adjacent edge in the graph convolution aggregation operator, so that the graph message transmission rate of the dynamic graph convolution-thermodynamic constraint co-prediction model matches the kinetic stage of the alkali-induced reaction, and realize the full intervention of the kinetic hard constraint on the evolution rhythm of the graph structure. After multiple rounds of dynamic graph message passing and node state updates controlled by hard constraints, a global graph feature readout operation is performed. The converged graph structure representation is input into the performance mapping decoder, and the predicted distribution of target performance such as compressive strength, heavy metal solidification rate and volume stability is output. The mapping relationship between preset preparation process parameters and target performance is established and output.

5. The method for optimizing process parameters of copper tailings fly ash geopolymer preparation based on cheminformatics according to claim 4, characterized in that: The construction and training of the dynamic graph convolution-thermodynamic constraint co-prediction model includes: Historical test batch data covering different precursor ratios, alkali activator concentrations and curing regimes were obtained. The corresponding silicon-aluminum-oxygen network topological chemical fingerprint vectors, preset preparation process parameters and measured target performance were extracted. A collaborative training dataset was constructed, and data distribution balancing and abnormal test batch removal were performed. The network topology of the dynamic graph convolution-thermodynamic constraint co-prediction model is initialized by configuring the graph feature embedding layer, multi-scale dynamic graph convolution layer, thermodynamic-kinetic dual-constraint projection layer and target performance output layer in sequence. The thermodynamic-kinetic dual-constraint projection layer is pre-set with initialization parameters for the Gibbs free energy change threshold decision matrix and the Arrhenius kinetic rate scaling matrix, thus completing the basic structure of the dynamic graph convolution-thermodynamic constraint co-prediction model. A composite loss function with physicochemical constraint awareness is constructed to drive parameter optimization of the dynamic graph convolution-thermodynamic constraint co-prediction model. The composite loss function consists of a target performance prediction deviation term, a thermodynamic feasibility violation penalty term, and a kinetic rate deviation regularization term. During the backpropagation iteration, an orthogonal projection truncation operation is performed on the gradient components that touch the hard constraint boundary to force the model parameter update trajectory to be constrained within the physicochemically allowed parameter feasible region, thus completing the gradient descent iteration with hard constraint embedding. A dynamic early stopping strategy and cross-validation mechanism are used to monitor the training process of the dynamic graph convolution-thermodynamic constraint co-prediction model. When the hard constraint satisfaction rate on the validation set reaches the preset convergence criterion and the prediction error of the mapping relationship between the preset fabrication process parameters and the target performance no longer decreases significantly for multiple consecutive rounds, the network layer weights and thermodynamic-kinetic dual-constraint projection layer hyperparameters of the dynamic graph convolution-thermodynamic constraint co-prediction model are solidified, and a training-ready instance of the dynamic graph convolution-thermodynamic constraint co-prediction model is output.

6. The method for optimizing process parameters of copper tailings fly ash geopolymer preparation based on cheminformatics according to claim 1, characterized in that: Based on the mapping relationship between process parameters and target performance, a chemical fingerprint similarity-weighted Pareto front adaptive exploration algorithm is used for global optimization of process parameters. The algorithm constructs a hybrid kernel function prior using the Tanimoto similarity and Mahalanobis distance of the topological chemical fingerprint vectors of the silicon-aluminum-oxide network, dynamically adjusts the exploration of the acquisition function, utilizes weights, and outputs the optimal combination of process parameters, specifically including: Based on the mapping relationship between the topological chemical fingerprint vector of silicon-aluminum-oxygen network and process parameters and target performance, the Tanimoto similarity and Mahalanobis distance between the fingerprint vectors corresponding to historical process parameter samples are calculated. The chemical structure overlap feature represented by Tanimoto similarity and the distribution discrete feature represented by Mahalanobis distance are nonlinearly fused to construct a hybrid kernel function prior and output the kernel function covariance matrix covering the process parameter search space. The kernel function covariance matrix is ​​input into the acquisition function construction module of the chemical fingerprint similarity weighted-Pareto front adaptive exploration algorithm. Based on the multi-objective performance prediction distribution provided by the mapping relationship, the desired improved acquisition function is initialized, and the initial ratio of the exploration weight factor and the utilization weight factor is set. The acquisition function configuration instance to be iteratively optimized is output. Based on the configuration instance of the acquisition function to be iteratively optimized, an initial candidate parameter combination sequence is generated by sampling in the process parameter search space. The initial candidate parameter combination sequence is input into the mapping relationship for performance prediction mapping to obtain the corresponding multi-objective performance prediction value set. The non-dominated sorting and solution set crowding evaluation are performed with the comprehensive optimality of the target performance as the judgment criterion. Dominated inferior solutions are eliminated and an initial Pareto front solution set is generated. The front distribution state matrix representing the spatial distribution of the solution set is output. Based on the frontier distribution state matrix, the local clustering density gradient and global expansion rate of the frontier region solution are analyzed in real time. The local clustering density gradient is converted into the decay adjustment amount of the exploration weight factor, and the global expansion rate is converted into the gain adjustment amount using the weight factor. The exploration-using weight ratio in the acquisition function configuration instance to be iteratively optimized is adaptively redistributed to generate the acquisition function driving engine with dynamically updated weights. The acquisition function with dynamically updated weights drives the engine to perform iterative sampling and prediction evaluation loops in the process parameter search space. The performance prediction results of the new sampled parameter combination are incorporated into the historical evaluation set and the initial Pareto front solution set is updated. The front distribution situation analysis and exploration - weight adaptive redistribution operation is repeated until the rate of change of the hypervolume index of the initial Pareto front solution set is lower than the preset convergence threshold and no new non-dominated solutions are generated in multiple consecutive iterations, and the final Pareto front solution set is locked. Based on the final state Pareto front solution set, a chemical fingerprint similarity weighted decision mechanism is introduced to calculate the Tanimoto similarity weighted comprehensive score between the topological chemical fingerprint vector of the silicon-aluminum-oxygen network corresponding to each front solution and the preset ideal reaction state reference fingerprint vector. The process parameter combination with the highest comprehensive score and that meets the thermodynamic-kinetic dual-constraint boundary safety margin requirements is selected, and the optimal process parameter combination is output.

7. The method for optimizing process parameters of copper tailings fly ash geopolymer preparation based on cheminformatics according to claim 1, characterized in that: The optimal combination of process parameters is input into the preparation system. In-situ vibrational spectral data is acquired in real time and converted into chemical coordinates of the reaction process. The node weights of the heterogeneous reaction spectrum are reconnected online and the graph structure is updated incrementally. This drives the dynamic graph convolution-thermodynamic constraint co-prediction model to iteratively calibrate until the prediction residual converges to a preset threshold. The final optimized preparation process parameters are then output, including: The optimal combination of process parameters is sent to the preparation control system. In-situ vibrational spectral data in the reaction system are continuously collected at a preset sampling period. Baseline drift correction, fluorescence background subtraction and characteristic peak deconvolution processing are performed on the in-situ vibrational spectral data. The asymmetric stretching vibrational peak position shift, half-width evolution curve and characteristic peak area integral ratio of the silicon-oxygen framework are extracted to generate a time-series spectral characteristic sequence. Based on the time-series spectral feature sequence, a joint phase space projection model of peak position shift and vibration intensity is constructed. Thermal noise and solvation interference are eliminated by principal component orthogonal decomposition. The spectral evolution trajectory is mapped to a multidimensional reaction progress axis, and the chemical coordinate sequence of the reaction process characterizing the continuous evolution of precursor dissolution, ion migration and network condensation stages is output. Based on the current reaction phase state identified by the chemical coordinate sequence of the reaction process, the heterogeneous reaction map of the geopolymer precursor is called, and the bond order evolution index in the chemical coordinate sequence of the reaction process is mapped to the node activity weight correction coefficient. The adjacency probability recalculation and connection channel dynamic switching are performed on the silicon-aluminum-oxygen tetrahedral cluster nodes and heavy metal-bearing phase nodes in the condensation critical state in the map, and the node weights are reconnected online to generate a local topology rewire map. The local topological rewire map and the heterogeneous reaction map of geopolymer precursors are compared by subgraph isomorphism and topological overlay. The newly generated bridging oxygen connection edges due to condensation reaction and the coordination polyhedral distortion regions caused by ion solidification are identified. The graph embedding alignment algorithm and the adjacency matrix incremental splicing technology are used to seamlessly integrate the reconnection weights into the original graph structure, maintain the continuity of the graph Laplacian feature spectrum distribution, and construct a dynamic incremental heterogeneous reaction map. The node representations and edge weight attributes of the dynamic incremental heterogeneous reaction spectrum are fed into the dynamic graph convolution-thermodynamic constraint co-prediction model in real time. By comparing the target performance prediction distribution output by the model with the measured performance surrogate value obtained from in-situ spectral inversion, the prediction residual vector of the process parameter mapping path is calculated. Based on the residual gradient direction, the message passing weight matrix and kinetic rate scaling factor of the dynamic graph convolution-thermodynamic constraint co-prediction model are corrected under constraints. The model parameters are iteratively calibrated and the calibration status is output. The closed-loop process from in-situ spectral acquisition to model calibration is executed cyclically. The norm decay rate of the predicted residual vector is monitored in real time. When the norm of the predicted residual vector is stably lower than the preset threshold for multiple consecutive sampling periods, and the fluctuation amplitude of the multi-objective performance prediction output by the dynamic graph convolution-thermodynamic constraint co-prediction model falls within the allowable tolerance zone, the iteration loop is terminated. The process parameter combination corresponding to the current round is extracted. After dual verification of thermodynamic feasibility boundary and kinetic safety margin, the final optimized preparation process parameters are output.

Images