An industrial solid waste carbon sequestration potential prediction and mechanism mining method

Abstract

The application belongs to the technical field of carbon sequestration, and particularly relates to an industrial solid waste carbon sequestration potential prediction and mechanism mining method, which comprises the following steps: S1: constructing a multi-dimensional parameter system and a field mechanism knowledge base, and realizing vector storage of industry standards and literature; S2: standardizing and semantically labeling multi-modal data to form a consistent multi-modal data set with semantic labels; S3: introducing a routing chain to hierarchically fuse the multi-modal data and dynamically allocate feature weights; S4: constructing a multi-modal large model based on an intelligent agent mechanism, and realizing parameter optimization through a retrieval enhancement module; and S5: using a reasoning chain to perform distributed logical reasoning to generate an interpretable report with physical and chemical mechanism support. The method effectively improves the accuracy and robustness of industrial solid waste carbon sequestration potential prediction, and enhances the interpretability of the model decision-making process. Through the dual driving of multi-modal data fusion and knowledge enhanced reasoning, not only high-precision prediction is realized, but also scientific discovery ability is possessed.

Description

Technical Field

[0001] This invention belongs to the field of carbon sequestration technology, specifically a method for predicting the carbon sequestration potential and mechanism of industrial solid waste. Background Technology

[0002] Industrial solid waste is widely used in steelmaking, cement, and heating industries to produce green building materials, artificial aggregates, and high-performance permeable bricks. Industrial solid waste carbon sequestration is a key technology that utilizes industrial solid waste such as red mud and fly ash to react with captured carbon dioxide through mineralization, adsorption, or co-sequestration reactions (preferably adapted to slurry carbonation processes) to generate stable carbonate minerals or stationary products. This achieves long-term carbon dioxide sequestration and resource utilization of solid waste, possessing both environmental value and promising engineering applications.

[0003] This process involves multiphase reaction coupling, multi-scale structural evolution (micropore-macro system), and dynamic changes under complex operating conditions. Accurate modeling of its core operating parameters (reaction rate, carbon conversion efficiency, pore structure evolution, etc.) is a prerequisite for achieving process control, efficiency optimization, and risk prevention, and is directly related to carbon sequestration effectiveness, product stability, and engineering economics.

[0004] In recent years, machine learning (ML) has become a powerful tool for modeling complex nonlinear relationships in large multidimensional datasets, thereby revealing key factors and mechanisms that traditional models may miss. It can also be used to uncover patterns and correlations that traditional models might overlook.

[0005] In addition, machine learning algorithms have the ability to handle complex nonlinear relationships, identify patterns and regularities in high-dimensional data, actively extract hidden key features from data, and have strong generalization and iterative optimization capabilities.

[0006] For example, using the maximum likelihood ratio test to examine the correlation between lignocellulose properties and the methane production process can effectively and accurately predict methane production, but specific quantitative analysis of the parameters could not be carried out. An extreme gradient boost (XGBoost) model was developed to predict the carbon dioxide emission response of thermal power plants using several key observable factors, but the specific mechanisms underlying the reaction could not be explored.

[0007] The aforementioned machine learning-based solutions have poor generalization ability for reaction parameters in actual industrial solid waste carbon sequestration scenarios. They cannot guarantee the accuracy and robustness of predictions, nor can they uncover the intrinsic mechanisms of the reactions, and they lack interpretability. Their "black box" characteristics cannot meet the stringent requirements of determinism and traceability in industrial scenarios. Furthermore, they lack the ability to autonomously discover scientific laws from multimodal data, thus making it difficult to provide clear and reliable optimal parameter combinations and unable to support the precise control of industrial solid waste carbon sequestration projects. Summary of the Invention

[0008] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a method for predicting the carbon sequestration potential and mechanism of industrial solid waste.

[0009] To solve the above-mentioned technical problems, the specific solution provided by the present invention is as follows: A method for predicting and exploring the carbon sequestration potential of industrial solid waste includes the following steps: S1: Construct a multi-dimensional parameter system and domain mechanism knowledge base: collect general data related to carbon sequestration of industrial solid waste, clarify the carbon sequestration objects and methods, construct a multi-dimensional parameter system for the entire process of the reaction; and vectorize industry standards and academic literature related to carbon sequestration of industrial solid waste to construct a domain mechanism knowledge base; S2: Standardization and Semantic Labeling of Multimodal Data: Multimodal data is collected during the industrial solid waste storage process based on a multi-dimensional parameter system covering the entire process. The multimodal data is then classified, organized, and standardized to form a consistent multimodal dataset with semantic labels. S3: Hierarchical fusion of multimodal datasets: A progressive fusion strategy is adopted to perform weighted fusion and feature selection on multi-source data to generate multimodal fusion features for modeling; S4: Construction of a multimodal large model based on intelligent agent architecture: A multimodal large model is constructed based on multimodal fusion features. A retrieval enhancement module and intelligent agent mechanism are introduced to achieve dynamic alignment between experimental data and domain mechanism knowledge base, and output carbon sequestration performance prediction results and optimized parameter combinations. Specifically, S4 constructs a closed-loop optimization architecture based on LangChainAgent: (1) Retrieval enhancement module: Before generating predictions, the model retrieves chemical kinetic descriptions of similar working conditions from the vector database of S1 through the LangChain retrieval tool to assist the model in predicting and verifying complex nonlinear reactions; (2) Intelligent agent mechanism: The LLaVA model is encapsulated as the Agent core, and the Agent is driven to automatically call different tool plugins through the Prompt instruction template to realize the self-correction of prediction results and the automatic optimization of process parameters.

[0010] Therefore, a cross-modal correlation model is established between the mineral phase composition before reaction, the reaction process conditions, and the carbon conversion results after reaction, so as to predict and evaluate the carbon sequestration performance of industrial solid waste, and provide data-driven support for the screening of key mineral phase parameters and the determination of their quantitative analysis.

[0011] S5: Parameter Quantitative Analysis and Mechanism Mining Based on Logical Chains: Using inference chains, the visual recognition results and quantitative prediction values ​​of the multimodal large model are logically inferred step by step with the domain mechanism knowledge base to quantitatively analyze the causal relationship between parameter changes, mineral phase evolution and carbon dioxide sequestration behavior, and generate an interpretable report with physicochemical mechanism support.

[0012] Specifically, based on the multimodal fusion prediction model determined by S4, we will conduct research on carbon sequestration reaction mechanisms based on multimodal interpretable analysis. By fusing data-driven results with experimental characterization information, we will logically link the semantic and numerical importance of the LLaVA output images using the LangChain inference chain. The specific process is as follows: After receiving the Shap analysis results, the Agent automatically triggers the mechanism mining chain, semantically aligns high-contribution features with physicochemical laws in the vector database, reverse-analyzes the implicit reaction paths in the model, reveals the inherent causal logic between mineral phase evolution, microstructural changes, and CO2 fixation behavior, and constructs an interpretable and verifiable cognitive framework.

[0013] Finally, the results of multimodal data-driven analysis were cross-validated with the results of Shap analysis to extract reusable knowledge rules and parameter empirical models of carbon sequestration mechanism, providing interpretable and verifiable mechanistic support for the optimization of industrial solid waste carbon sequestration process, model updating and large-scale engineering application.

[0014] Furthermore, the high-contribution features obtained from the multimodal large-scale model analysis were grouped according to chemical composition parameters, mineral phase characteristics, microstructure characteristics, and process condition parameters. Interactive analysis was conducted on the joint contributions of different feature groups to identify synergistic or inhibitory relationships among various parameters. Combining the mineral phase transformation patterns characterized by XRD and the pore structure evolution characteristics revealed by SEM and CT, the implicit reaction pathways in the multimodal large-scale model were analyzed in reverse. By comparing the changes in the Shap contribution of mineral phase parameters before and after the reaction, the key controlling factors of the transformation of the dominant mineral phase from reactive oxides to carbonate minerals during carbonation were identified, revealing a deeper reaction mechanism and ultimately forming a causal chain in which parameter changes and mineral phase evolution promote the improvement of carbon sequestration efficiency.

[0015] Preferably, in S1, the multi-dimensional parameter system for the entire process covers pre-reaction parameters, in-reaction parameters, and post-reaction parameters. Specifically, pre-reaction parameters include images of physical properties such as chemical composition, mineral phase composition, and specific surface area of ​​industrial solid waste obtained through instruments such as XRF, XRD, and BET; in-reaction parameters include process parameters such as reaction temperature, carbon dioxide partial pressure, and reaction time collected in real time by online sensors, as well as derived kinetic parameters; post-reaction parameters include verification parameters such as carbon conversion rate, carbonate mineral formation, product mineral phase composition, and pore structure evolution.

[0016] Preferably, in S4, the intelligent agent mechanism includes: Prediction Agent: Used to invoke a multimodal large model to simulate and predict carbon conversion rate or unit carbon sequestration under specified parameters; Retrieval Agent: Used to retrieve mechanism descriptions under similar working conditions from the domain mechanism knowledge base to correct the prediction when abnormal fluctuations occur in the prediction results; Optimization Agent: Used to automatically iteratively search for the optimal parameter range for reaction temperature, carbon dioxide partial pressure, and reaction time based on predictive feedback.

[0017] Preferably, in S2, the multimodal data encompasses experimental characterization image data and microstructure image data. Specifically, the multimodal data is first categorized and organized according to four core types. Experimental characterization image data, including chemical composition, mineral phases, and pore characteristics before and after the reaction, are acquired using specialized instruments such as XRF and XRD. Real-time online monitoring time-series data, such as reaction temperature and carbon dioxide partial pressure, are captured using multiple sensors. Microstructure image data of microstructure evolution is obtained using SEM and CT. Historical operation and maintenance records and experimental records are integrated to form process operation correlation data. All data is archived according to "parameter category - data source - acquisition time" to ensure spatiotemporal and logical consistency.

[0018] To address the heterogeneity of multimodal data, outliers were removed using the 3σ criterion, missing data were filled in using linear interpolation, and Z-score standardization was applied to unify the dimensionality. Feature quantization was performed on XRD spectra and SEM images, and time synchronization and dimensionality matching of real-time and offline data, as well as micro and macro data, were achieved based on reaction time. Finally, a standardized dataset was generated, laying a solid foundation for subsequent hierarchical fusion and modeling.

[0019] Preferably, in S3, the weighted fusion and feature selection adopt a progressive fusion strategy of data layer, feature layer and decision layer.

[0020] The data fusion strategy at the data layer involves weighted fusion of similar data based on the measurement accuracy of different instruments and sensors. Specifically, weights are first assigned according to the accuracy of the instruments and sensors, and then a weighted average method is used to integrate multi-source data of the same type and dimension, thereby reducing errors.

[0021] The feature layer fusion strategy is as follows: key influencing features are selected using dimensionality reduction methods and attention weighting mechanisms. Specifically, the core features of various data types are extracted, and after PCA / KPCA dimensionality reduction and attention weighting, key influencing features are focused to generate fused features.

[0022] The decision-making layer's fusion strategy involves constructing fusion decision rules by combining the carbon sequestration reaction mechanism of industrial solid waste. Specifically, it involves constructing fusion decision rules by combining the carbon sequestration reaction mechanism with the microscopic laws revealed by the instruments, eliminating data redundancy, strengthening core information, and providing a high-quality fusion feature set for subsequent collaborative modeling.

[0023] Preferably, the feature layer fusion strategy specifically includes extracting numerical feature vectors. and image feature vectors By combining dimensionality reduction and attention weighting, a fused feature representation is obtained: in , Here, α and β are the feature mapping functions for numerical and image data, respectively, and are adaptive weighting coefficients. Preferably, in S4, the multimodal large model includes a feature encoding module and a feature extraction module. The feature encoding module processes numerical parameters, while the feature extraction module processes image data. This multimodal large model utilizes a cross-modal alignment mechanism to achieve collaborative representation and key modeling of features from different modalities.

[0024] Preferably, the training of a multimodal large model aims to obtain quantitative analysis of feature parameters and uncover the underlying mechanisms. The model parameters θ are optimized through maximum likelihood estimation, and the training objective function is expressed as: Where L is the sequence length, The model outputs the predicted carbon sequestration performance. To fuse feature inputs, The instruction information consists of numerical parameters. Based on the trained model, the unit fly ash carbon sequestration amount y is output and embedded with the Shap analysis to obtain the importance analysis of different mineral phases and reaction processes.

[0025] Compared with the prior art, the present invention has at least the following beneficial effects: 1. By constructing a multi-dimensional parameter system covering the entire process, multimodal data from the industrial solid waste carbon sequestration process are collected to form a multimodal dataset. Subsequently, the heterogeneous data from multiple sources are preprocessed in a unified manner, and the multimodal dataset is fused in a hierarchical manner to finally construct a large multimodal model, which effectively improves the accuracy and robustness of predicting the carbon sequestration potential of industrial solid waste. At the same time, the interpretability of the model's decision-making process is enhanced to meet the stringent requirements of determinism and traceability in industrial scenarios. Through the dual-drive of "multimodal data fusion + knowledge-enhanced reasoning", not only is high-precision prediction achieved, but also scientific discovery capabilities are also developed.

[0026] 2. This invention adopts a collaborative modeling approach based on LangChainAgent, enabling the system to autonomously schedule prediction tools and knowledge bases for comparison and verification. Compared with a single model, the accuracy of carbon conversion rate prediction is higher, and it also has the ability to automatically optimize process parameters.

[0027] 3. This invention utilizes reasoning chains to logically couple LLaVA's visual understanding ability with Shap's quantitative contribution, upgrading from "correlation" mining to "causal" deduction, quantitatively identifying key mineral phase parameters and providing scientific mechanistic support.

[0028] 4. This invention combines multimodal modeling results with mineral facies evolution and microstructure characterization information to construct a causal relationship between parameter changes, mineral facies evolution, and carbon sequestration efficiency improvement, thereby enabling the interpretable mining of carbon sequestration reaction mechanisms and providing reliable support for process optimization and engineering applications. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart illustrating the method for predicting and exploring the carbon sequestration potential of industrial solid waste according to an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the process of hierarchical fusion of multimodal large model data in an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram of the instruction information generation process for the numerical parameters of a multimodal large model according to an embodiment of the present invention.

[0033] Figure 4 This is a quantitative analysis diagram based on a multimodal large model in an embodiment of the present invention. Detailed Implementation

[0034] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. These embodiments are implemented based on the technical solution of the present invention and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0035] This embodiment takes the prediction of a typical industrial solid waste blast furnace slag carbon sequestration scenario as an example, such as... Figure 1 The implementation process of the present invention is described in detail below: S1: Taking blast furnace slag as a typical industrial solid waste carbon sequestration object, a slurry carbonization process is adopted to construct a multi-dimensional parameter system covering the entire process from "before reaction to during reaction to after reaction".

[0036] Pre-reaction parameters: Obtain the chemical composition (mass fractions of CaO, SiO2, MgO, Al2O3, etc.), mineral phase composition (yellow feldspar, grayish yellow feldspar, dicalcium silicate, etc.) and specific surface area of ​​blast furnace slag. Collect XRD spectra and SEM images before and after the reaction to characterize the initial mineral phase and microstructure features.

[0037] Reaction parameters: Process parameters such as reaction temperature, carbon dioxide partial pressure, slurry pH, stirring speed, liquid-solid ratio and reaction time are collected in real time by online sensors. Kinetic parameters such as apparent reaction rate constant and apparent activation energy are obtained by fitting experimental data and used to characterize the rate control characteristics of the carbonization process.

[0038] Post-reaction parameters: Verification parameters such as carbon conversion rate, unit carbon fixation, residual alkalinity change, product mineral phase composition, pore structure evolution, and mechanical properties are measured to evaluate the carbon sequestration effect and product stability under different operating conditions.

[0039] A domain-specific mechanism knowledge base was simultaneously constructed: Academic literature and industry standards related to blast furnace slag mineralization kinetics, calcium magnesium silicate carbonization mechanisms, and multiphase mass transfer reactions were imported using a document loading module. After segmenting, denoising, and structurally annotating the literature content, the text was transformed into high-dimensional semantic vectors using a vector embedding engine and stored in a vector database, forming a domain-specific mechanism knowledge base with similarity retrieval capabilities. This knowledge base covers mechanistic descriptions such as "the influence of magnesium ions on calcium carbonate crystal form and nucleation pathways," "aluminum element enhancing CSH gel stability and its effect on the layer structure of carbonization products," and "blast furnace slag glass phase dissolution kinetics," providing external knowledge support for subsequent retrieval enhancement, model verification, and mechanism reasoning.

[0040] S2: Collected data from 219 sets of experimental samples of blast furnace slag slurry carbonization, covering typical operating conditions such as different CaO / SiO2 alkalinity, MgO content, liquid-solid ratio, temperature and CO2 partial pressure.

[0041] For numerical data, the 3σ criterion is used to remove significant outliers, and missing data is filled by linear interpolation or neighbor sample imputation. The Z-score standardization is used to unify the dimensions of each parameter. Intensity normalization was performed on the XRD spectrum, and characteristic parameters such as peak position, peak area, and peak width were extracted to reflect changes in mineral phase composition and crystal structure. SEM images are converted to grayscale, noise filtered, and texture features extracted to quantify microstructure indicators such as porosity, pore size distribution, and crystal morphology. Noise reduction and resampling are performed on the online monitoring time series data to ensure alignment with the offline test data on the time axis; By utilizing metadata management functions, each set of experimental data is associated with corresponding literature indexes, operating condition tags, and sample numbers ("high alkalinity and high magnesium operating condition", "low temperature and high pressure operating condition"). While ensuring spatiotemporal consistency, semantic tags are added to the multimodal datasets to make them traceable and interpretable, providing an index basis for subsequent RAG retrieval and mechanism mining.

[0042] S3: Based on standardized multimodal datasets, such as Figure 2 As shown, a multimodal fusion feature suitable for blast furnace slag carbon sequestration is constructed by using a hierarchical fusion strategy of "data layer - feature layer - decision layer" through a visual encoder and a trainable mapping matrix.

[0043] Data layer: Based on the calibration accuracy and repeatability of different instruments and sensors, weighted fusion of similar numerical data is performed; for data from multiple repeated tests, a weighted average method is used for integration to reduce measurement noise and improve data reliability; Feature layer: Extracting numerical feature vectors respectively , and image feature vector yi, .

[0044] The feature dimension is compressed using dimensionality reduction methods such as PCA / KPCA, and different feature components are weighted using an attention mechanism to obtain a fused feature representation. in, and These are numerical and image feature mapping functions, respectively, with α and β being adaptive weighting coefficients used to automatically balance the contributions of physical parameters and microstructural features to the prediction task, thereby focusing on key features highly correlated with carbon sequestration behavior. Decision layer: A routing chain is introduced to identify the type and operating characteristics of the input solid waste. When it is identified as "blast furnace slag" and meets the specific alkalinity and MgO content range, the fusion weight allocation scheme and decision rules optimized for the highly amorphous characteristics of calcium magnesium silicate are automatically invoked to enhance the characterization of the active glass phase, CSH gel and early carbonate phase, further eliminate redundant information, highlight key mineral phase and pore structure information, and provide a high-quality fusion feature set for subsequent model training.

[0045] S4: As Figure 3 As shown, a multimodal large model is constructed based on the fused feature Z, and a closed-loop prediction of blast furnace slag carbon sequestration performance and optimization of process parameters are achieved through an intelligent agent architecture.

[0046] (1) Model structure and training: The multimodal large model includes a feature encoding module, a feature extraction module, and a cross-modal alignment module. The feature encoding module is used to process physical and chemical parameters and process conditions before and after the reaction. The feature extraction module is used to process image data such as XRD and SEM. The cross-modal alignment module achieves collaborative representation of features of different modalities through attention mechanism and shared latent space.

[0047] The model training primarily aims to predict the carbon sequestration rate y per unit blast furnace slag. The model parameters θ are optimized using maximum likelihood estimation. The training objective function is: Wherein, Xa is the carbon sequestration performance prediction sequence. The structured instruction information consists of numerical parameters such as CaO / SiO2 alkalinity ratio, MgO content, liquid-solid ratio, temperature, and carbon dioxide partial pressure, with Z representing the fused feature input. During training, a training set, validation set, and test set partitioning strategy is employed, and the model's prediction accuracy and generalization ability are evaluated using metrics such as R², MAE, and RMSE.

[0048] (2) Enhanced instruction generation and retrieval: Key process parameters such as CaO / SiO2 basicity ratio, MgO content, liquid-solid ratio, reaction temperature, and CO2 partial pressure are encapsulated as structured instructions (Xinstruct) and input into the intelligent agent. The intelligent agent first calls the retrieval enhancement generation module to retrieve mechanistic literature fragments similar to the current working condition from the vector database, such as "the role of Mg element in blast furnace slag carbonization" and "glass phase dissolution behavior under high basicity conditions." Background knowledge such as "Mg²⁺ promotes calcite crystal growth and induces aragonite formation" and "high basicity accelerates the dissolution of calcium silicate phase" is obtained and injected as additional conditions into the multimodal large model to achieve knowledge-enhanced prediction of complex multiphase reactions.

[0049] (3) Closed-loop prediction and parameter optimization: Predictive Agent: Under the constraints of fused feature Z and retrieved mechanistic knowledge, the multimodal large model predicts the unit carbon sequestration amount y and the corresponding carbon conversion rate under different operating conditions. Retrieval Agent: When the prediction result deviates from the validation data or exceeds the set error threshold, the intelligent agent automatically triggers the logic verification tool to perform local search and correction on the input parameter space. Optimization Agent: Gradually adjusts variables such as reaction temperature, carbon dioxide partial pressure, reaction time, and liquid-solid ratio, forming a closed-loop optimization process of "prediction-verification-correction" until the optimal or near-optimal parameter combination that satisfies multiple constraints such as carbon conversion rate, energy consumption, and operational safety is obtained, thus achieving intelligent optimization of the blast furnace slag carbon sequestration process.

[0050] S5: After completing model training and performance verification, based on multimodal interpretable analysis and knowledge-enhanced reasoning, a systematic exploration of the carbon sequestration reaction mechanism in blast furnace slag is conducted, such as... Figure 4 As shown.

[0051] (1) Contribution conversion: The Shap method is used to perform feature importance analysis on the trained multimodal large model to obtain the contribution of each input feature to the unit carbon fixation amount y. The numerical results are mapped into readable semantic descriptions, such as "high MgO content", "large specific surface area", "high pore structure openness", "moderate reaction temperature", "high CO2 partial pressure", etc., and a mapping relationship from the internal weight space of the model to the physicochemical semantic space is constructed.

[0052] (2) Reasoning Link Construction: After receiving the Shap analysis results, the intelligent agent starts the reasoning chain and performs semantic alignment and reasoning with the multimodal information such as "significant Shap value of Mg element", "SEM image shows needle-like aragonite crystals" and "XRD characterizes the transformation of calcite into aragonite and magnesite" and the text knowledge in the vector library such as "magnesium ion interference of calcium carbonate nucleation and growth mechanism" and "structural reconstruction of CSH gel in carbonization process". Through multi-hop retrieval and logical link construction, the data-driven conclusions and mechanism knowledge are linked together to form a causal chain from "parameter change - micro-evolution - macro-performance".

[0053] (3) Mechanism generation and quantitative conclusions: Based on the reasoning chain, the system automatically deduced the following mechanism: Magnesium in blast furnace slag gradually dissolves during carbonization and forms an adsorption layer on the surface of calcite crystals, increasing the interfacial energy, inhibiting the coarsening of calcite crystals, and inducing the formation of needle-like or platy aragonite with a larger specific surface area, thereby maintaining a high effective reaction interface and improving the overall carbon fixation efficiency; at the same time, the CSH / CASH gel generated by the reconstruction of Si and Al components has a spatial restriction and template effect on the nucleation and growth of carbonate crystals, further affecting the pore structure and mass transfer pathway. By comparing the changes in the contribution of the mineral phase parameter Shap under different working conditions, the key control factors of the transformation of the dominant mineral phase from active oxides to carbonate minerals were identified, and the marginal contributions of factors such as alkalinity, MgO content, and specific surface area to the unit carbon fixation amount were quantified.

[0054] Based on multi-round reasoning chain analysis, this embodiment shows that the maximum carbonization potential of blast furnace slag can exceed 250 kg / t. The optimal carbon sequestration effect is achieved when the CaO content is 38%–42%, the MgO content is higher than 8%, and the liquid-to-solid ratio, reaction temperature, and CO2 partial pressure are within their optimized ranges. At this point, the synergistic effect of magnesium and aluminum significantly optimizes the pore structure of the carbonization product layer, preventing premature reaction termination due to excessively rapid densification of the calcium carbonate product layer. These conclusions were further organized into a structured knowledge graph, forming a causal network of "composition and operating parameters—mineral phase and pore evolution—improved carbon sequestration efficiency," providing explainable and verifiable mechanistic support for the scale-up, engineering design, and cross-operating-condition migration of blast furnace slag carbon sequestration technology.

[0055] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the carbon sequestration potential and exploring the mechanism of industrial solid waste, characterized in that, Includes the following steps: S1: Construct a multi-dimensional parameter system and domain mechanism knowledge base: collect general data related to carbon sequestration of industrial solid waste, clarify the carbon sequestration objects and methods, construct a multi-dimensional parameter system for the entire process of the reaction; and vectorize industry standards and academic literature related to carbon sequestration of industrial solid waste to construct a domain mechanism knowledge base; S2: Standardization and Semantic Labeling of Multimodal Data: Multimodal data is collected during the industrial solid waste storage process based on a multi-dimensional parameter system covering the entire process. The multimodal data is then classified, organized, and standardized to form a consistent multimodal dataset with semantic labels. S3: Hierarchical fusion of multimodal datasets: A progressive fusion strategy is adopted to perform weighted fusion and feature selection on multi-source data to generate multimodal fusion features for modeling; S4: Construction of a multimodal large model based on intelligent agent architecture: A multimodal large model is constructed based on multimodal fusion features. A retrieval enhancement module and intelligent agent mechanism are introduced to achieve dynamic alignment between experimental data and domain mechanism knowledge base, and output carbon sequestration performance prediction results and optimized parameter combinations. S5: Parameter Quantitative Analysis and Mechanism Mining Based on Logical Chains: Using inference chains, the visual recognition results and quantitative prediction values ​​of the multimodal large model are logically inferred step by step with the domain mechanism knowledge base to quantitatively analyze the causal relationship between parameter changes, mineral phase evolution and carbon dioxide sequestration behavior, and generate an interpretable report with physicochemical mechanism support.

2. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 1, characterized in that, In S1, the multi-dimensional parameter system of the whole process includes pre-reaction parameters, in-reaction parameters, and post-reaction parameters. The pre-reaction parameters include the chemical composition of industrial solid waste, mineral phase composition, and specific surface area. The in-reaction parameters include reaction temperature, carbon dioxide partial pressure, and reaction time. The post-reaction parameters include carbon conversion rate, carbonate mineral formation, product mineral phase composition, and pore structure evolution parameters.

3. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 1, characterized in that, In S4, the intelligent agent mechanism includes: Prediction Agent: Used to invoke a multimodal large model to simulate and predict carbon conversion rate or unit carbon sequestration under specified parameters; Retrieval Agent: Used to retrieve mechanism descriptions under similar working conditions from the domain mechanism knowledge base to correct the prediction when abnormal fluctuations occur in the prediction results; Optimization Agent: Used to automatically iteratively search for the optimal parameter range for reaction temperature, carbon dioxide partial pressure, and reaction time based on predictive feedback.

4. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 1, characterized in that, In S2, the multimodal data includes experimental characterization image data and microstructure image data.

5. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 4, characterized in that, The experimental characterization image data includes image data of chemical composition, mineral phase, and pore characteristics before and after the reaction, and the microstructure image data includes images of the obtained pore structure and crystal morphology evolution.

6. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 1, characterized in that, In S3, weighted fusion and feature selection adopt a progressive fusion strategy of data layer, feature layer and decision layer.

7. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 6, characterized in that, The data layer fusion strategy includes weighted fusion of similar data based on the measurement accuracy of different instruments and sensors. The feature layer fusion strategy includes screening key influencing features through dimensionality reduction methods and attention weighting mechanisms. The decision layer fusion strategy includes constructing fusion decision rules by combining the reaction mechanism of industrial solid waste carbon sequestration.

8. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 7, characterized in that, The feature layer fusion strategy specifically includes extracting numerical feature vectors. and image feature vectors By combining dimensionality reduction and attention weighting, a fused feature representation is obtained: ； in , These are the numerical and image feature mapping functions, respectively, and α and β are adaptive weighting coefficients.

9. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 1, characterized in that, In S4, the multimodal large model includes a feature encoding module and a feature extraction module. The feature encoding module is used to process numerical parameters, and the feature extraction module is used to process image data.

10. The method for predicting and exploring the carbon sequestration potential of industrial solid waste according to claim 1 or 9, characterized in that, The training of the multimodal large model aims to obtain quantitative analysis of feature parameters and uncover the underlying mechanisms. It optimizes the model parameters θ through maximum likelihood estimation, and its training objective function is expressed as: ； Where L is the sequence length, The model outputs the predicted carbon sequestration performance. To fuse feature inputs, The instruction information consists of numerical parameters. Based on the trained model, the unit fly ash carbon sequestration amount y is output and embedded with the Shap analysis to obtain the importance analysis of different mineral phases and reaction processes.

Images