Coal-based solid waste differential modification method and system, heat storage particles

Abstract

The application discloses a coal-based solid waste differential modification method and system and heat storage particles, and belongs to the technical field of solid waste resource utilization, and comprises the following steps: obtaining full-component chemical composition data; calculating a thermodynamic equilibrium fingerprint vector of the solid waste in a target sintering temperature range according to the full-component chemical composition data; performing JMAK non-equilibrium dynamics correction on the thermodynamic equilibrium fingerprint vector to generate a non-equilibrium thermodynamic fingerprint vector; obtaining a multi-dimensional performance prediction result through a graph neural network proxy model; embedding a multi-objective Bayesian optimization algorithm with thermodynamic and kinetic physical constraints in a process parameter space to obtain an optimal process parameter combination; and performing differential modification processing to obtain differential modification process parameters. In the prior art, the labeled data set is seriously insufficient, so that the MAE of the GNN model is about 12.3%. Compared with the prior art, the application can achieve a prediction MAE of 5.2% with only 30-50 groups of labeled data, and is convenient for engineering verification and trust establishment.

Description

Technical Field

[0001] This invention relates to the field of solid waste resource utilization technology, and in particular to a method and system for differentiated modification of coal-based solid waste and thermal storage particles. Background Technology

[0002] Labeled data in the field of solid waste ceramicization is extremely scarce. Each complete set of "composition-process-performance" data requires several days of experimentation and costs hundreds of yuan. Therefore, only 30 to 50 sets of labeled data are typically available. However, 30 to 50 sets of labeled data are severely insufficient for deep learning models, falling far short of the basic data requirements for deep learning models.

[0003] This data scarcity severely limits model performance. In the second-generation approach, researchers used graph neural networks (GNNs) to predict key performance indicators of ceramic products. However, the limited number of labeled samples cannot support effective fitting of GNNs in a broad feature space. The model is prone to overfitting to a few samples and exhibits highly unstable responses to sensitive factors such as compositional fine-tuning and temperature changes. Experimental results show that the mean absolute error (MAE) of this GNN model is approximately 12.3%. In practical engineering, this deviation means that the predicted value can differ from the true value by more than ten percentage points, making it difficult to meet the stringent requirements for quality stability in solid waste resource recovery products.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] This invention provides a method for differentiated modification of coal-based solid waste using a self-supervised graph neural network, comprising: S1: performing chemical composition analysis on the solid waste to be treated to obtain full-component chemical composition data; S2: Based on the chemical composition data of all components, calculate the thermodynamic equilibrium fingerprint vector F^eq of solid waste within the target sintering temperature range. The thermodynamic equilibrium fingerprint vector includes the equilibrium liquid phase fraction f_L^eq(T), equilibrium phase content, equilibrium distribution coefficient of alkali metal in each phase {K_Alk,i^eq(T)}, and equilibrium distribution of sulfur occurrence forms {S_j^eq(T)}. S3: Perform JMAK non-equilibrium kinetic correction on the thermodynamic equilibrium fingerprint vector to generate a non-equilibrium thermodynamic fingerprint vector F^neq; S4: Input the full-component chemical composition data and non-equilibrium thermodynamic fingerprint vector into the graph neural network surrogate model to obtain multi-dimensional performance prediction results; S5: Based on the multidimensional performance prediction results, a multi-objective Bayesian optimization algorithm with thermodynamic and kinetic physical constraints is embedded in the process parameter space to obtain the optimal combination of process parameters. S6: Based on the optimal combination of process parameters, perform differentiated modification treatment to obtain differentiated modification process parameters.

[0006] Furthermore, the full-component chemical composition data includes the mass fractions of SiO2, Al2O3, CaO, Fe2O3, Na2O, K2O, MgO, SO3, and TiO2.

[0007] Furthermore, S3 specifically includes: performing JMAK nonequilibrium dynamic correction on the thermodynamic equilibrium fingerprint vector F^eq; The equilibrium phase content is converted into the non-equilibrium phase content at the actual sintering time t to generate the non-equilibrium thermodynamic fingerprint vector F^neq; The conversion formula is:

[0008] The content of the non-equilibrium phase of the i-th phase at sintering temperature T and actual sintering time t; The equilibrium phase content of the i-th phase at sintering temperature T; Let JMAK be the rate constant of the i-th phase at sintering temperature T; denoted as the Avrami index of the i-th phase at sintering temperature T.

[0009] Furthermore, S4 specifically includes: constructing positive and negative sample pairs based on unlabeled solid waste composition data, performing comparative learning pre-training on the positive and negative sample pairs, and using NT-Xent loss as the pre-training loss function; Multi-task fine-tuning is performed on the labeled dataset to output multidimensional performance prediction results, which include multidimensional performance prediction vectors and calibrated prediction uncertainties.

[0010] Furthermore, the positive sample pair is: for the same solid waste, perform steps S2 and S3 at temperatures T and T+ΔT respectively to obtain the non-equilibrium thermodynamic fingerprint vector F^neq(T) at temperature T and the non-equilibrium thermodynamic fingerprint vector F^neq(T+ΔT) at temperature T+ΔT. Negative sample pairs are obtained by performing step S2 for different solid wastes to obtain the thermodynamic equilibrium fingerprint vectors of different solid wastes.

[0011] Furthermore, the physical constraints include at least one of the following: C1: f_L^eq(T_sinter) ∈ [0.15,0.35]; C2: K_Alk,liquid^eq(T_sinter) ≤0.30; C3: S_anhydrite^eq / S_total^eq>0.85; C4: ΔG_An(T_sinter) ≤ -50 kJ / mol; C5: α_An^{neq}(T_sinter,t_sinter) ≥ 0.90.

[0012] Furthermore, it also includes the following steps: comparing the full-component chemical composition data with the Mahalanobis distance in the CALPHAD database; when the comparison result exceeds a preset threshold, using general machine learning interatomic potential molecular dynamics simulation as a supplementary thermodynamic calculation branch. The Mahalanobis distance is: ; Let x be the Mahalanobis distance of the solid waste component vector. The solid waste component vector x is the 9-dimensional oxide component vector of the solid waste; A component space mean vector that is fully validated for the CALPHAD database; S is the covariance matrix of the validation ensemble; Let S be the inverse of the covariance matrix S; (·)^T is the transpose operation for a vector or matrix; The preset threshold is 2.5σ.

[0013] Furthermore, when N consecutive batches exceed 2σ, the following process is executed: The number of graph convolutional layers in the frozen graph neural network surrogate model; Re-execute steps S2 and S3; Repeat step S5.

[0014] A self-supervised graph neural network-based differential modification system for coal-based solid waste, used to execute the aforementioned method, includes: a component detection module, a CALPHAD thermodynamic calculation engine, a JMAK non-equilibrium correction engine, a graph neural network inference engine, and a physically constrained Bayesian optimization engine; the component detection module is used to acquire chemical composition data of all components; the CALPHAD thermodynamic calculation engine is used to calculate thermodynamic equilibrium fingerprint vectors based on the chemical composition data of all components; the JMAK non-equilibrium correction engine is used to calculate non-equilibrium thermodynamic fingerprint vectors based on the thermodynamic equilibrium fingerprint vectors; the graph neural network inference engine is used to output multi-dimensional performance prediction results based on the thermodynamic equilibrium fingerprint vectors and non-equilibrium thermodynamic fingerprint vectors; and the physically constrained Bayesian optimization engine is used to search for the Pareto optimal combination of process parameters in the process parameter space through multi-objective Bayesian optimization with embedded physical constraints based on the multi-dimensional performance prediction results.

[0015] A ceramicized thermal storage particle is prepared using the method described above.

[0016] Compared with the prior art, the present invention has the following advantages: This invention utilizes thermodynamic and physical principles to construct positive sample pairs, which is superior to purely data-driven enhancement strategies.

[0017] This invention requires only 30-50 sets of labeled data to achieve a predicted MAE of 5.2%, significantly reducing labeling costs.

[0018] The positive sample pairs used in this invention have clear physical meaning, which facilitates engineering verification and trust establishment.

[0019] The thermodynamic characterization learned through pre-training can be transferred to different types of solid waste.

[0020] The enhancement strategy of this invention has a clear physical meaning. That is, the thermodynamic behavior of the same solid waste at different temperatures is intrinsically related, and learning this relationship helps the model understand the thermodynamic nature of solid waste, rather than simply memorizing the statistical correlation between composition and performance.

[0021] In the learning process, this invention actually learns the thermodynamic nature of solid waste (such as the rate of change of liquid phase fraction with temperature, crystal phase transformation sequence, etc.), rather than simply memorizing the statistical distribution of components. This makes the representations learned through pre-training have better generalization ability. Attached Figure Description

[0022] Figure 1 Overall process diagram.

[0023] Figure 2 JMAK correction principle diagram.

[0024] Figure 3 : Diagram of the Genetic Neural Network Agent (GNN) architecture.

[0025] Figure 4 : Schematic diagram of CALPHAD conditional augmentation pre-training strategy.

[0026] Figure 5 Flowchart of Physically Constrained Bayesian Optimization.

[0027] Figure 6 Comparison of ablation experiment results.

[0028] Figure 7 Causal discovery DAG diagram.

[0029] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0030] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0031] Example 1 This invention provides a method for differentiated modification of coal-based solid waste using a self-supervised graph neural network, referring to... Figure 1 , Figure 1 This is a schematic diagram of the process of the present invention. It includes: S1: Performing a full-component chemical composition analysis on the coal-based solid waste raw material to be treated to obtain full-component chemical composition data. The full-component chemical composition data includes at least the following: mass fractions of SiO2, Al2O3, CaO, Fe2O3, Na2O, K2O, MgO, SO3, and TiO2.

[0032] S2: Based on the chemical composition data of the whole components, the thermodynamic equilibrium fingerprint vector F^eq of the solid waste in the target sintering temperature range is calculated using the CALPHAD method. The thermodynamic equilibrium fingerprint vector includes the equilibrium liquid phase fraction f_L^eq(T): the mole fraction of the liquid phase at temperature T.

[0033] Equilibrium phase content: The equilibrium mole fraction of each crystalline phase (calcium feldspar, mullite, wollastonite, hematite, anhydrite, etc.).

[0034] The equilibrium distribution coefficient of alkali metals in each phase {K_Alk,i^eq(T)}: the distribution ratio of Na and K in each phase.

[0035] The equilibrium distribution of sulfur occurrence forms {S_j^eq(T)}: the proportion of sulfur existing in the form of anhydrite (CaSO4), free SO3, etc.

[0036] S3: Perform JMAK non-equilibrium kinetic correction on the thermodynamic equilibrium fingerprint vector F^eq, converting the equilibrium phase content into the non-equilibrium phase content at the actual sintering time t. (See attached...) Figure 2 As shown. The conversion formula is:

[0037] in, The content of the non-equilibrium phase of the i-th phase at sintering temperature T and actual sintering time t; The equilibrium phase content of the i-th phase at sintering temperature T; Let JMAK be the rate constant of the i-th phase at sintering temperature T.

[0038] denoted as the Avrami index of the i-th phase at sintering temperature T.

[0039] α = 1 - exp(-k·t^n) represents the degree of phase transition completion.

[0040] Thus, the non-equilibrium thermodynamic fingerprint vector F^neq is obtained through the aforementioned process.

[0041] in, The following formula is used to determine:

[0042] Specifically, is the pre-factor of the i-th phase.

[0043] is the activation energy of the i-th phase.

[0044] R is the ideal gas constant.

[0045] T represents the sintering temperature.

[0046] is the thermodynamic correction function driven by the Gibbs free energy of the i-th phase.

[0047] Let represent the Gibbs free energy change of the i-th phase at temperature T.

[0048] in,

[0049] Using the characteristic energy scale, we ensure that the rate constant also approaches zero when the driving force approaches zero, giving the JMAK parameters a thermodynamic basis rather than a purely empirical fit.

[0050] on the other hand, , , The initial value is derived from the experimental calibration value reported in the literature, and is calibrated and corrected through experimental feedback during system operation.

[0051] Specifically, the JMAK parameters of the key crystalline phases and their sources are as follows:

[0052] For crystal phases not listed in the table, their JMAK parameters are obtained by the following methods: (a) prioritizing the search for experimental calibration values ​​in the literature; (b) if no literature values ​​are available, using the parameters of crystal phases of the same structural type as the initial estimate, and then calibrating and correcting them through experimental feedback during system operation.

[0053] In summary, CALPHAD calculations provide the phase composition under thermodynamic equilibrium conditions. However, industrial rotary kiln sintering is a non-equilibrium process, and the actual sintering time (2-6 hours) is far insufficient to bring the system to complete thermodynamic equilibrium. This invention corrects the equilibrium phase content through the aforementioned process. S4: Input the full-component chemical composition data and non-equilibrium thermodynamic fingerprint vector into a two-stage trained graph neural network surrogate model (GNN) to obtain multi-dimensional performance prediction results. (See attached image) Figure 3 As shown. The graph neural network proxy model adopts the following architecture: (a) Node encoding layer—each oxide is treated as a node, and the initial feature vector of the node includes the mass fraction, mole fraction, ion field strength Z / r², Pauling electronegativity, oxygen ion polarizability, and the activity coefficient of the oxide at the target temperature calculated by CALPHAD, for a total of 6 dimensions. (b) Edge encoding layer—an edge exists between any two oxides, and the initial feature vector of the edge includes the binary eutectic temperature, liquidus temperature, enthalpy of mixing ΔH_mix, and the ratio of binary reaction rate constants k_i / k_j after non-equilibrium correction calculated by CALPHAD, for a total of 4 dimensions; (c) Graph convolutional layer—3 layers of EdgeConv convolution, each followed by BatchNorm and LeakyReLU activation. The EdgeConv operation is defined as:

[0054] Specifically, This is the updated feature vector of node i in the (l+1)th layer.

[0055] i is the index of the target node.

[0056] j is the index of a neighboring node, and j ∈ N(i) means that j belongs to the neighbor set N(i) of node i.

[0057] `max` is an aggregation operation that takes the maximum value among all neighboring nodes.

[0058] MLP is the nonlinear transformation function of a multilayer perceptron.

[0059] Let i be the feature vector of node i in the l-th layer.

[0060] Let be the feature vector of the l-th layer neighbor node j.

[0061] Let be the edge feature vector connecting node i and node j.

[0062] ⊕ represents the vector concatenation operation.

[0063] (d) Global pooling layer – a concatenation of global average pooling and global max pooling; (e) Multi-task prediction head – six parallel fully connected branches (two layers each, with hidden dimensions 64→32), which output the predicted values ​​and variances of compressive strength, porosity, free quartz content, residual unstable SO3 content, thermal cycling mass loss rate, and alkali metal volatilization rate (estimated via MC Dropout, T=30 forward propagation cycles).

[0064] As attached Figure 4 As shown, it specifically includes the following two stages: Phase 1 is CALPHAD conditional enhanced self-supervised pre-training: constructing positive and negative sample pairs based on unlabeled solid waste composition data. Positive sample pairs are formed by performing steps S2 and S3 at temperatures T and T+ΔT, respectively, to obtain the non-equilibrium thermodynamic fingerprint vector F^neq(T) at temperature T and the non-equilibrium thermodynamic fingerprint vector F^neq(T+ΔT) at temperature T+ΔT.

[0065] Negative sample pairs are obtained by performing step S2 for different solid wastes to obtain the thermodynamic equilibrium fingerprint vectors of different solid wastes.

[0066] Comparative learning is performed on positive and negative sample pairs for pre-training. The pre-training loss function uses NT-Xent (Normalized Temperature Scaled Cross Entropy) loss. The specific details of the NT-Xent loss are as follows:

[0067] in, The NT-Xent normalized temperature-scaled cross-entropy (NT-Xent) comparison learning loss function value is used during the pre-training phase.

[0068] Let be the natural logarithm function with base e.

[0069] It is an exponential function with base e.

[0070] For two graph-level representation vectors in a positive sample pair and Cosine similarity between them.

[0071] Let be the graph-level representation vector of the i-th sample after being encoded by a graph neural network.

[0072] To and Another graph-level representation vector that constitutes a positive sample pair.

[0073] The temperature parameter controls the sharpness of the distribution, with a value of 0.1.

[0074] This represents the number of original samples in a batch.

[0075] 2N is the total number of samples including both positive and negative samples after augmentation.

[0076] k is the summation index variable, which iterates through all samples in the batch except for i itself.

[0077] This is an indicator function; it takes the value 1 when k ≠ i, and 0 otherwise.

[0078] Let be the cosine similarity between the representation of the i-th sample and the representation of the k-th sample.

[0079] Preferably, ΔT = 50℃. The sources of unlabeled solid waste composition data are as follows: Part 1: Public solid waste composition databases—including a subset of coal ash composition data from the USGS World Coal Quality Inventory (≥5000 sets), the World Coal Ash Database (≥3000 sets), and composition data from publicly published coal-based solid waste research literature (≥2000 sets), totaling ≥10 4 Part I: Unlabeled solid waste XRF detection data accumulated during the historical operation of this method. Part II: Synthetic data generated by Latin hypercube sampling in a known coal-based solid waste composition space (the mass fraction range of each oxide was determined by statistical analysis of historical data).

[0080] For records in the public database that lack trace components such as Na2O, K2O, and SO3, the following processing strategies are adopted: (a) If a certain component is explicitly marked as "not detected" in the database record, the content of that component is set to zero; (b) If the component is not measured / reported, the mean value of that component in all records of the database is used to fill it, and a higher Dropout rate (0.3 vs 0.1) is applied to the filled value during pre-training to reduce the model's dependence on the filled value.

[0081] Phase 2 is multi-task supervised fine-tuning: multi-task fine-tuning is performed on the labeled dataset, which contains 30-50 sets of known solid waste chemical composition, process parameters and corresponding measured performance data; after fine-tuning, multi-dimensional performance prediction results are output, which include multi-dimensional performance prediction vectors and calibrated prediction uncertainties.

[0082] Specifically, multi-task fine-tuning is performed on 30-50 sets of labeled data, with the fine-tuning loss function being the multi-task weighted MSE. The formula is as follows:

[0083] in, The total loss function value during the multi-task supervised fine-tuning phase. This is a task index variable, with values ​​from 1 to 6, corresponding to six performance metrics. is the learnable homoscedasticity uncertainty parameter for the m-th task, used to automatically balance the weights of each task.

[0084] Let be the mean squared error loss for the m-th task.

[0085] Let be the natural logarithm function with base e.

[0086] On the other hand, the prediction uncertainty is estimated using MC Dropout (T=30 forward propagations), as shown in the following formula:

[0087] in, This is the average prediction after multiple forward propagations of MC Dropout.

[0088] T represents the number of forward propagations in MC Dropout, with a value of 30.

[0089] This is the index variable for forward propagation, with values ​​ranging from 1 to T.

[0090] This is the predicted value for the t-th forward propagation.

[0091] The variance of cognitive uncertainty reflects the degree to which the model lacks sufficient understanding of the distribution of the input data.

[0092] (·)² represents the squaring operation.

[0093] Total prediction uncertainty = cognitive uncertainty (σ²_epistemic) + accidental uncertainty (σ²_aleatoric, directly output by the model).

[0094] S5: Based on the multidimensional performance prediction results, the optimal Pareto process parameter combination is searched within the process parameter space consisting of CaO addition amount, sintering temperature, sintering time, pre-washing temperature and number of washes, and pre-calcination temperature and time, using a multi-objective Bayesian optimization algorithm with embedded thermodynamic and kinetic physical constraints. (See attached...) Figure 5 As shown.

[0095] The physical constraints employed include at least one of the following: C1: f_L^eq(T_sinter) ∈ [0.15, 0.35]. The physical meaning is: too low a liquid fraction results in insufficient densification, while too high a liquid fraction leads to particle deformation.

[0096] C2: K_Alk,liquid^eq(T_sinter) ≤0.30; Physical meaning: Alkali metals are enriched in the liquid phase at no more than 30% to avoid volatilization.

[0097] C3: S_anhydrite^eq / S_total^eq>0.85; Physical meaning: Sulfur exists in the form of stable anhydrite in greater than 85% of the form.

[0098] C4: ΔG_An(T_sinter) ≤ -50 kJ / mol; physical meaning: the calcium feldspar reaction is thermodynamically feasible.

[0099] C5: α_An^{neq}(T_sinter,t_sinter) ≥ 0.90. Physically, this means the kinetic completion rate of the anorthosite formation reaction is ≥90%.

[0100] Specifically, the formula used for multi-objective optimization is as follows:

[0101] in, For multi-objective minimization optimization operators.

[0102] The decision vector for the process parameters to be optimized includes CaO addition amount, sintering temperature, sintering time, pre-washing temperature and number of washes, and pre-calcination temperature and time.

[0103] It is the negative value of the compressive strength σ_c(x).

[0104] This is the predicted value for porosity.

[0105] This is the predicted value for the free quartz content.

[0106] This is the predicted value for the alkali metal volatilization rate.

[0107] This is the predicted value for the residual unstable SO3 content.

[0108] The predicted mass loss rate for 2000 thermal cycles.

[0109] The acquisition function employs Expected Hypervolume Improvement (EHVI) to search for the Pareto front in a 6-dimensional target space.

[0110] S6: Based on the optimal combination of process parameters, perform differentiated modification treatment, including at least one combination of the following sub-steps: pre-washing treatment, pre-calcination treatment, CaO addition and mixing, granulation, and sintering treatment, to obtain differentiated modification process parameters.

[0111] It also includes the following steps: By comparing the Mahalanobis distance between the full-component chemical composition data and the coverage of the CALPHAD database, when the comparison result exceeds a preset threshold, Universal MLIP molecular dynamics simulation is triggered as a supplementary thermodynamic calculation branch.

[0112] The Mahalanobis distance is: ; Let x be the Mahalanobis distance of the solid waste component vector. The solid waste component vector x is the 9-dimensional oxide component vector of the solid waste; A component space mean vector that is fully validated for the CALPHAD database; S is the covariance matrix of the validation ensemble; Let S be the inverse of the covariance matrix S; (·)^T is the transpose operation for a vector or matrix; The preset threshold is 2.5σ. That is, the MLIP branch is triggered when D_M>2.5. The MLIP simulation results and the CALPHAD+JMAK correction results are weighted and fused using the Bayesian model averaging (BMA) method.

[0113] It also includes a causal structure discovery step: based on the historically accumulated measured process-performance data (non-proxy model prediction data), when the accumulated data volume is ≥100 sets, the causal structure discovery algorithm is executed to construct a directed acyclic graph (DAG) between component variables, process parameters and performance indicators; the causal structure discovery algorithm adopts the PC (Peter-Clark) algorithm, uses Fisher's Z-test as the conditional independence test method (significance level α=0.05), and is supplemented by the NOTEARS algorithm for cross-validation - only the causal edges jointly confirmed by the two algorithms are retained; the causal DAG is used to identify the key regulatory variables with the greatest causal effect on the target performance and to detect spurious correlations.

[0114] It also includes a solid waste composition drift adaptive mechanism: when the chemical composition analysis results of N consecutive batches (N≥3) of solid waste show that the moving average of a certain key oxide deviates from the original calibration value by more than 2σ, the following adaptive process is automatically triggered: First: Freeze the graph convolutional layer parameters of the graph neural network surrogate model and only fine-tune the multi-task prediction head. Second: Re-execute steps S2 and S3 to update the thermodynamic equilibrium fingerprint vector and the non-equilibrium thermodynamic fingerprint vector. Third: Re-execute step S5 to update the optimal combination of process parameters.

[0115] In summary, this invention differs from traditional self-supervised learning enhancement strategies based on data perturbation (such as adding noise to components). This invention utilizes the physical implications of CALPHAD thermodynamic calculations: CALPHAD calculations and JMAK non-equilibrium corrections are performed on the same solid waste component at different temperatures (T and T+ΔT) to obtain non-equilibrium thermodynamic fingerprint vectors at two different temperatures, which serve as positive sample pairs for comparative learning. This enhancement strategy has clear physical significance. That is, the thermodynamic behavior of the same solid waste at different temperatures is intrinsically related; learning this relationship helps the model understand the thermodynamic essence of solid waste, rather than merely memorizing the statistical correlation between components and performance.

[0116] On the other hand, positive sample pairs constructed by pure compositional perturbation enhancement (noise addition) only contain statistical information—"composition A and composition A + noise should be similar." In contrast, positive sample pairs constructed by CALPHAD conditional enhancement contain physical information—"the thermodynamic behaviors of the same solid waste at 1180℃ and 1230℃, such as phase composition, liquid phase fraction, and alkali metal distribution, are intrinsically related." In learning this relationship, this invention is actually learning the thermodynamic essence of the solid waste (such as the rate of change of liquid phase fraction with temperature, crystal phase transformation sequence, etc.), rather than merely memorizing the statistical distribution of components. This allows the pre-trained representation to have better generalization ability.

[0117] To accurately understand the aforementioned process, the following running example is provided: Raw material: Guizhou coal gangue (GZ-01), chemical composition data of all components: SiO2 52.3%, Al2O3 24.1%, CaO 2.8%, Fe2O3 6.5%, Na2O 0.4%, K2O 1.8%, MgO 1.2%, SO3 1.5%, TiO2 4.2%.

[0118] CALPHAD equilibrium calculation:

[0119] JMAK unbalanced correction (t=4h):

[0120] GNN performance prediction:

[0121] Physically constrained Bayesian optimization determines process parameters: The constraint check results are as follows: C1: f_L^eq(1180℃)=0.18 ∈ [0.15, 0.35] satisfies this condition.

[0122] C2: K_Na,liquid^eq=0.22 ≤ 0.30 is satisfied.

[0123] C3: S_anhydrite / S_total=0.91>0.85 is satisfied.

[0124] C4: ΔG_An(1180℃)=-62 kJ / mol ≤ -50 kJ / mol, which is satisfied.

[0125] C5: α_An^neq(1180℃,4h)=0.82<0.90 is not satisfied.

[0126] Constraint C5 is not satisfied – the kinetics of the calcium feldspar conversion reaction are insufficient. BO is searched under constraints, and the recommended process parameters are: CaO addition = 14%, sintering temperature = 1200℃, sintering time = 5h.

[0127] After adjustment, C5: α_An^neq(1200℃,5h)=0.93 ≥ 0.90, all constraints are satisfied.

[0128] Final formula: CaO 14%, 1200℃×5h, pre-washing at 80℃×2 times, pre-calcination at 600℃×1h.

[0129] The experimental results for the final formula are as follows: compressive strength 16.2 MPa, porosity 8.5%, free quartz 0.7%, and all indicators are qualified.

[0130] Causal discovery detection for spurious correlations, details are attached. Figure 7 As shown.

[0131] After the system accumulates 120 sets of measured data, the causal discovery engine is triggered.

[0132] Cross-validation results of PC algorithm (α=0.05) and NOTEARS algorithm (λ=0.1): The jointly confirmed causal edges are as follows: CaO addition amount → anorthite formation amount → compressive strength (cause and effect +0.52).

[0133] Sintering temperature → Liquid phase fraction → Porosity (cause and effect -0.58).

[0134] Na2O content → Alkali metal volatilization rate (causal effect +0.85).

[0135] False correlation warning (high Pearson correlation but neither algorithm confirms causal edges): Fe2O3 content vs. compressive strength: Pearson r = 0.65, no causal edge (confounding factor: SiO2 content).

[0136] MgO content vs. porosity: Pearson r = -0.48, no causal edge (confounding factor: sintering temperature).

[0137] Engineering Decisions: The focus of process optimization should be on the amount of CaO added and the sintering temperature (the two variables with the greatest causal effect), rather than being misled by spurious correlations to adjust Fe2O3 or MgO.

[0138] On the other hand, to further demonstrate the technical effects of the present invention, a control experiment was set up as follows: Annotated dataset: 30 groups of coal-based solid waste (15 groups of fly ash + 15 groups of coal gangue), each group containing complete composition-process-performance data.

[0139] Test set: 10 independent solid waste samples (not used in training).

[0140] GNN architecture: uniformly uses 3-layer EdgeConv + 6-head multi-task prediction.

[0141] Evaluation metrics: Normalized MAE of 6 performance metrics (each metric is averaged after being divided by its target value).

[0142] The experimental group is as follows:

[0143] The experimental results are as follows:

[0144] Combining experimental results with appendix Figure 6 The content shown was analyzed, and the results are as follows: Compared to the baseline without pre-training, pure component perturbation enhanced pre-training reduced the average MAE by 34.1% (12.3% → 8.1%), demonstrating the effectiveness of self-supervised pre-training itself. Compared with pure component perturbation enhancement, CALPHAD conditional enhancement pre-training further reduced the average MAE by 35.8% (8.1%→5.2%), demonstrating the unique incremental technical effect of the present invention; Compared to the baseline without pre-training, the CALPHAD conditional enhancement reduced the average MAE by 57.7% (12.3% → 5.2%), demonstrating the effectiveness of the overall technical solution of this invention.

[0145] On the other hand, experiments have shown that the present invention achieves the following performance improvements:

[0146] At the same time, the following performance advantages were achieved: This invention utilizes thermodynamic and physical principles to construct positive sample pairs, which is superior to purely data-driven enhancement strategies.

[0147] This invention requires only 30-50 sets of labeled data to achieve a prediction MAE of 5.2%, significantly reducing labeling costs.

[0148] The positive sample pairs used in this invention have clear physical meaning, which facilitates engineering verification and trust establishment.

[0149] The thermodynamic characterization learned through pre-training can be transferred to different types of solid waste.

[0150] Example 2: A self-supervised graph neural network-based differential modification system for coal-based solid waste is provided for executing the method described in Example 1. Specifically, it includes: a component detection module, a CALPHAD thermodynamic calculation engine, a JMAK non-equilibrium correction engine, a graph neural network inference engine, and a physically constrained Bayesian optimization engine.

[0151] The component detection module includes an X-ray fluorescence spectrometer (XRF) and an X-ray diffractometer (XRD) to acquire full-component chemical composition data and mineral phase composition.

[0152] The CALPHAD thermodynamic calculation engine deploys the FactSage software and the FT Oxide thermodynamic database to calculate thermodynamic equilibrium fingerprint vectors based on the chemical composition data of all components.

[0153] The JMAK non-equilibrium correction engine deploys the JMAK kinetic parameter database and correction calculation module to calculate non-equilibrium thermodynamic fingerprint vectors based on thermodynamic equilibrium fingerprint vectors.

[0154] The graph neural network inference engine deploys a graph neural network model trained in two stages (CALPHAD conditional augmentation self-supervised pre-training + multi-task supervised fine-tuning) to output multi-dimensional performance prediction results based on thermodynamic equilibrium fingerprint vectors and non-equilibrium thermodynamic fingerprint vectors.

[0155] The physical constraint Bayesian optimization engine is used to search for the Pareto optimal combination of process parameters in the process parameter space by embedding physical constraints, based on multidimensional performance prediction results and multi-objective Bayesian optimization.

[0156] Example 3: A ceramicized thermal storage particle, prepared using the method described in Example 1. Specifically, it meets the following performance indicators (n≥15 batches, n≥5 parallel samples per batch, data expressed as mean ± SD):

[0157] Example 4: A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps described in Embodiment 1.

[0158] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0159] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0160] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0161] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

[0162] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

Claims

1. A method for differentiated modification of coal-based solid waste using a self-supervised graph neural network, characterized in that, include: S1: Perform chemical composition analysis on the solid waste to be treated to obtain full-component chemical composition data; S2: Based on the chemical composition data of the whole components, calculate the thermodynamic equilibrium fingerprint vector F^eq of the solid waste within the target sintering temperature range. The thermodynamic equilibrium fingerprint vector includes the equilibrium liquid phase fraction f_L^eq(T), equilibrium phase content, equilibrium distribution coefficient of alkali metal in each phase {K_Alk,i^eq(T)}, and equilibrium distribution of sulfur occurrence forms {S_j^eq(T)}. S3: Perform JMAK non-equilibrium kinetic correction on the thermodynamic equilibrium fingerprint vector to generate a non-equilibrium thermodynamic fingerprint vector F^neq; S4: Input the full-component chemical composition data and the non-equilibrium thermodynamic fingerprint vector into the graph neural network surrogate model to obtain multi-dimensional performance prediction results; S5: Based on the multidimensional performance prediction results, a multi-objective Bayesian optimization algorithm with thermodynamic and kinetic physical constraints is embedded in the process parameter space to obtain the optimal combination of process parameters. S6: Based on the optimal combination of process parameters, perform differentiated modification treatment to obtain differentiated modification process parameters.

2. The method for differentiated modification of coal-based solid waste using a self-supervised graph neural network as described in claim 1, characterized in that, The complete chemical composition data includes the mass fractions of SiO2, Al2O3, CaO, Fe2O3, Na2O, K2O, MgO, SO3, and TiO2.

3. The method for differentiated modification of coal-based solid waste using a self-supervised graph neural network as described in claim 1, characterized in that, S3 specifically includes: performing the JMAK non-equilibrium dynamics correction on the thermodynamic equilibrium fingerprint vector F^eq; The equilibrium phase content is converted into the non-equilibrium phase content at the actual sintering time t to generate the non-equilibrium thermodynamic fingerprint vector F^neq; The conversion formula is: The The content of the non-equilibrium phase of the i-th phase at sintering temperature T and actual sintering time t; The The equilibrium phase content of the i-th phase at sintering temperature T; The Let JMAK be the rate constant of the i-th phase at sintering temperature T; The denoted as the Avrami index of the i-th phase at sintering temperature T.

4. The method for differentiated modification of coal-based solid waste using a self-supervised graph neural network as described in claim 1, characterized in that, S4 specifically includes: constructing positive sample pairs and negative sample pairs based on unlabeled solid waste composition data, performing comparative learning pre-training on the positive sample pairs and negative sample pairs, and using NT-Xent loss as the pre-training loss function; Multi-task fine-tuning is performed on the labeled dataset to output multidimensional performance prediction results, which include multidimensional performance prediction vectors and calibrated prediction uncertainties.

5. The method for differentiated modification of coal-based solid waste using a self-supervised graph neural network as described in claim 4, characterized in that, The positive sample pair is: for the same solid waste, the S2 step and the S3 step are performed at temperature T and temperature T+ΔT respectively to obtain the non-equilibrium thermodynamic fingerprint vector F^neq(T) at temperature T and the non-equilibrium thermodynamic fingerprint vector F^neq(T+ΔT) at temperature T+ΔT. The negative sample pair is: for different solid wastes, the S2 step is performed to obtain the thermodynamic equilibrium fingerprint vector of different solid wastes.

6. The method for differentiated modification of coal-based solid waste using a self-supervised graph neural network as described in claim 1, characterized in that, The physical constraints include at least one of the following: C1: f_L^eq(T_sinter) ∈ [0.15,0.35]; C2: K_Alk,liquid^eq(T_sinter) ≤0.30; C3: S_anhydrite^eq / S_total^eq>0.85; C4: ΔG_An(T_sinter) ≤ -50 kJ / mol; C5: α_An^{neq}(T_sinter,t_sinter) ≥ 0.

90.

7. The method for differentiated modification of coal-based solid waste using a self-supervised graph neural network as described in claim 1, characterized in that, It also includes the following steps: The Mahalanobis distance between the full-component chemical composition data and the CALPHAD database is compared. When the comparison result exceeds a preset threshold, a general machine learning interatomic potential molecular dynamics simulation is used as a supplementary thermodynamic calculation branch. The Mahalanobis distance is: ; The Let x be the Mahalanobis distance of the solid waste component vector. The solid waste component vector x is a 9-dimensional oxide component vector of the solid waste; The A component space mean vector that is fully validated for the CALPHAD database; S is the covariance matrix of the validation integration. The Let S be the inverse of the covariance matrix S; The (·)^T is the transpose operation of a vector or matrix; The preset threshold is 2.5σ.

8. The method for differentiated modification of coal-based solid waste using a self-supervised graph neural network as described in claim 7, characterized in that, When N consecutive batches exceed 2σ, execute the following procedure: Freeze the number of graph convolutional layers in the graph neural network proxy model; Repeat steps S2 and S3; Repeat step S5.

9. A self-supervised graph neural network-based differential modification system for coal-based solid waste, used to execute the method according to any one of claims 1-8, comprising: Component detection module, CALPHAD thermodynamic calculation engine, JMAK non-equilibrium correction engine, graph neural network inference engine, physical constraint Bayesian optimization engine; The component detection module is used to acquire chemical composition data of all components; The CALPHAD thermodynamic calculation engine is used to calculate the thermodynamic equilibrium fingerprint vector based on the full component chemical composition data. The JMAK non-equilibrium correction engine is used to calculate the non-equilibrium thermodynamic fingerprint vector based on the thermodynamic equilibrium fingerprint vector. The graph neural network inference engine is used to output multi-dimensional performance prediction results based on the thermodynamic equilibrium fingerprint vector and the non-equilibrium thermodynamic fingerprint vector. The physical constraint Bayesian optimization engine is used to search for the Pareto optimal combination of process parameters in the process parameter space by embedding physical constraints through multi-objective Bayesian optimization based on the multi-dimensional performance prediction results.

10. A ceramicized thermal storage particle, characterized in that, Prepared by the method described in any one of claims 1 to 8.

Images