A knowledge-data double-driven intelligent decision system and method for recycling waste lithium batteries

Abstract

The application discloses a kind of knowledge-data double-drive intelligent decision system and method of waste lithium battery recycling, belong to new energy solid waste resource and artificial intelligence cross technical field, to solve the industry pain points that existing lithium battery recycling intelligent tool mechanism atlas cannot be completely deduced, industrial label is incomplete, lack of economic carbon synchronous evaluation.The system includes: knowledge graph automatic construction module constructs directed attribute knowledge base;Graph topology multi-step jump connection search module carries out materialization constraint;Double-track machine learning prediction module distinguishes process and respectively completes performance prediction;Multi-path intelligent agent cognitive decision module shunts service path;Atlas linkage optimization module quantifies parameter residual error contribution;Process economic carbon integrated evaluation module synchronously accounts processing cost and carbon footprint.The present application greatly reduces laboratory trial and error cost, with industrial digital landing value.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of new energy solid waste resource utilization and artificial intelligence, and in particular relates to a knowledge-data dual-driven intelligent decision-making system and method for recycling waste lithium batteries. Background Technology

[0002] With the large-scale retirement of power batteries for new energy vehicles worldwide, three types of waste lithium battery recycling processes—wet acid leaching, pyrometallurgical roasting-water leaching, and direct regeneration—have become core components of the circular economy. In recent years, machine learning and large-scale model retrieval technologies have been gradually applied to lithium battery recycling process development. However, the lithium battery recycling system is complex, and existing intelligent auxiliary systems face six major technological bottlenecks in industrial implementation: Pure data-driven machine learning suffers from black-box limitations, lacking tangible prior constraints. Existing models rely solely on experimental data to fit parameters without introducing boundary conditions, easily outputting theoretically feasible but practically chemically ineffective process parameters, resulting in extremely high trial-and-error costs in the laboratory. Traditional vector-based RAG retrieval cannot perform multi-causal mechanism deduction, and conventional text-based vector retrieval struggles to complete cross-document multi-step causal link reasoning from process operation → reaction mechanism → recovery performance. Industrial datasets generally suffer from asymmetric labeling, and existing data augmentation methods are unsuitable for the physicochemical differences of lithium battery metals. In actual production, many samples only label partial metal recovery rates, and simple sample removal and mean imputation will lose synergistic reaction information. Existing intelligent systems can only predict performance unidirectionally, lacking autonomous iterative diagnostic closed-loop. Existing process decision-making tools only focus on recovery performance optimization, lacking simultaneous evaluation of economic and environmental dimensions. Existing solutions have limited functionality and cannot meet the differentiated needs of multiple scenarios.

[0003] Currently, there are a few lithium battery material recycling optimization solutions that combine knowledge graphs and machine learning. These are simple solutions that rely on knowledge graphs as static text databases. The industry urgently needs a complete intelligent system that integrates mechanism graphs, differentiated machine learning, multi-path intelligent decision-making, and integrated cost and carbon emission calculation. Summary of the Invention

[0004] The purpose of this invention is to address the problems of existing technologies that rely on graphs as static text databases, which can only achieve basic performance prediction, are difficult to complete high-precision, interpretable, and autonomous iterative process optimization within physical safety boundaries, and lack the ability to quantify economic and carbon footprint assessment capabilities. Therefore, this invention proposes a knowledge-data dual-driven intelligent decision-making system and method for waste lithium battery recycling.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The overall operation is divided into two main phases: offline initialization training and online intelligent decision-making. The offline phase simultaneously completes the knowledge graph construction and machine learning model training and consolidation. The online phase utilizes the constructed knowledge graph and trained model to perform process optimization calculations. This solution comprises six core functional modules: an automatic knowledge graph construction module, a graph topology multi-step jump-link retrieval module, a dual-track machine learning prediction module, a multi-path intelligent agent cognitive decision-making module, a multi-process parameter attribution and graph linkage optimization module, and an integrated process, economic, and carbon emission assessment module. These modules work together to form a complete closed loop: mechanism constraint input → data performance prediction → multi-path intelligent reasoning → physical self-consistency verification → parameter attribution optimization → comprehensive economic and carbon emission assessment. The specific definitions of each module are as follows: The knowledge graph automatic construction module is an offline initialization unit that only runs during system deployment or database update phases. The completed knowledge graph is stored in a graph database for online retrieval. This module employs a hybrid construction strategy combining top-down expert ontology design and bottom-up large model extraction. It extracts contextual knowledge from unstructured literature on lithium battery recycling using a large language model, and combines materialization rules with expert experience for joint quality control to construct a domain-specific knowledge graph in the form of a directed attribute graph. Graph nodes are divided into seven categories: literature, battery systems, experimental procedures, operational variables, reaction mechanisms, performance, and environmental indicators. Directed relationships such as subordination, procedures, mechanism triggering, and optimal parameter mapping are configured between nodes, and qualitative prior constraints across the entire domain are stored. The graph topology multi-step hop-connection retrieval module receives process optimization retrieval requests forwarded by multiple intelligent agents. It parses the user's business intent using a large language model and dynamically generates graph database retrieval instructions adapted to the target battery system, combining the independent battery tag rules (NCM, LFP, LCO, LMO) in the graph ontology. The retrieval executes 1-4 depth adaptive topology subgraph traversal, with a default 1-3 hop computing power configuration under normal operating conditions, expandable to 4 hops in scenarios with sufficient computing power. It fully extracts causal evidence across the entire chain from operational variables to reaction mechanisms to recovery performance. The module's generated retrieval statements forcibly disable sorting and quantity truncation instructions, achieving full recall of literature, process, parameter, and mechanism-related data without evidence omission. Simultaneously, it incorporates a dual-type retrieval switching logic: performing full data extraction for process optimization and performance prediction needs, and automatically adding keyword filtering for single mechanism or single reagent queries, reducing retrieval computing power overhead. The module is equipped with a hierarchical multi-hop question-and-answer evaluation system to quantify the logical integrity and result stability of multi-step causal reasoning. After the retrieval is completed, the thermodynamic and dynamic laws recorded in the literature are organized into physical boundary constraints and sent back to the multi-channel intelligent agent as exclusive preconditions for process generation and iterative optimization. At the same time, the retrieval and traceability structured data are automatically archived.

[0006] A dual-track machine learning prediction module. This module completes model training and solidification offline, and only performs inference online. The offline training data comes from experimental data in publicly available industry literature and complete or incomplete recycling datasets from actual factory production lines. Online inference only receives standardized numerical features transformed by multiple agents and does not receive raw natural language text. The dual-track machine learning prediction module is used to build differentiated prediction models for data features of different recycling processes. The module adopts a metal-differentiated track-based modeling strategy to automatically identify knowledge transfer boundaries between different metals and suppress the negative transfer effect of multi-tasks. The module has two built-in parallel learning frameworks: First Track Labeling Issue: Addressing the asymmetric data deficiency commonly found in industrial datasets for hydrometallurgical and pyrometallurgical recycling, where only some metals in a single sample are labeled while others are missing labels, this track constructs a two-layer learning architecture: differentiated track segmentation and teacher-student pseudo-label evolution. The core architecture consists of three complete stages: teacher model training, pseudo-label generation and filtering, and student fusion retraining. Throughout the process, the modeling branches for transition metals and lithium are differentiated to avoid prediction accuracy degradation caused by cross-elemental physicochemical conflicts. The definitions of each stage are as follows: 1) Teacher Model Training Phase. A complete training subset is formed by selecting fully labeled samples from the original dataset for lithium, cobalt, nickel, and manganese. Considering the high homology in crystal structure, acid leaching reduction thermodynamics, and ion migration mechanism of the ternary transition metals Ni, Co, and Mn, and the strong positive synergistic correlation in their leaching behavior, two parallel modeling schemes can be used: Scheme 1 uses MultiOutputRegressor to encapsulate the XGBoost basic regressor to complete multi-task synchronous fitting; Scheme 2 directly uses the native multi-output tree structure of XGB for joint training. Both methods can simultaneously mine the positive synergistic features among the three metals. However, the lattice insertion / extraction, dissolution kinetics, and redox mechanism of Li are fundamentally different from those of transition metals. If combined with the ternary metals for global multi-output training, it will produce severe negative feature transfer. Therefore, a separate independent XGBoost prediction branch is built.

[0007] Hyperparameters are optimized using a multi-layered grid traversal. After training the track-specific model with each set of candidate hyperparameters, the model's generalization ability is quantified by substituting them into the generalization gap penalty comprehensive scoring function (Formula 1). (1) Among them, Avg_R 2 _tr is the average coefficient of determination for the training set, Avg_R 2 _te is the average coefficient of determination on the test set, and λ is the overfitting penalty weight. In this embodiment, λ is set to 5. This scoring mechanism is used in the hyperparameter optimization stage to evaluate the generalization ability of the model corresponding to each group of candidate hyperparameters, and automatically filters out models with high training accuracy and low training or testing R. 2For overfitted parameter combinations with excessively large discrepancies, only hyperparameters with stable generalization are retained for constructing the teacher model, thereby reducing the overall risk of distortion from subsequent false labels from the source.

[0008] 2) Pseudo-label generation and two-layer filtering stage. The optimal teacher is used to perform batch inference on incomplete samples, sequentially filtering through two layers of rules: confidence threshold and material metal presence mask. Low-confidence predictions that violate basic chemical common sense are eliminated. After the two-layer screening, predictions that meet the confidence and physicochemical constraints are retained as high-quality pseudo-labels.

[0009] 3) Student Model Fusion and Retraining Stage. The original complete labeled subset and the filtered high-quality pseudo-labeled samples are merged to construct an expanded global hybrid training dataset. The student model reuses the differentiated track design of the teacher model. Ni / Co / Mn can be jointly fitted using either encapsulated multi-output regression or XGB native multi-output, while Li still independently builds a single-output branch, completely avoiding cross-element feature conflicts. The hybrid dataset simultaneously incorporates real experimental values ​​and high-confidence model predictions, fully exploring the hidden metal synergistic physicochemical characteristics in the incomplete industrial data. Finally, a prediction model adapted to the incomplete industrial data is trained, which can simultaneously output the leaching rates of lithium, cobalt, nickel, and manganese.

[0010] Second-track label integrity: Direct regeneration of the positive electrode. The second-track direct regeneration process with complete labels: For direct regeneration datasets with complete labels, a random forest algorithm is used for feature importance pre-screening. An importance threshold is set to remove redundant variables, retaining core heat treatment and chemical doping features. A Bayesian optimization framework is introduced during the data partitioning stage, using the test set determination coefficient as the objective function to adaptively optimize the training-test set partitioning ratio. Five-fold cross-validation is performed on the finalized prediction model, and residual distribution analysis is used to evaluate the model's physical fidelity under multi-dimensional nonlinear parameter interaction constraints. In the feature engineering stage, high-order physical descriptors representing thermodynamic and kinetic properties are derived online from basic process variables using pre-defined mathematical formulas. These descriptors serve as model input features, ensuring that the model's decision boundaries have interpretable physical meaning. The core derivation formula is as follows: Heating intensity of tubular furnace: (2) Annealing heat treatment strength: (3) Total heat treatment volume: (3) Oxygen release propensity index: (4) in (5) Temperature-time synergy coefficient: (6) Reducing agent interaction items: (7) The derived features are concatenated into the basic feature matrix and used as input to the machine learning model to ensure the physical interpretability of the model's decision boundary. The physical descriptor derivation formula is not limited to the specific mathematical expression mentioned above. Any equivalent mathematical transformation that uses the logarithmic or exponential relationship between tube furnace temperature and time, or annealing temperature and time to characterize the heat treatment intensity, including base substitution, linear scaling, and fine-tuning by introducing constant terms, falls within the scope of protection of this patent.

[0011] Multi-channel intelligent agent cognitive decision-making module. This module is the core unit for online global scheduling, cross-modal feature conversion, multi-scenario task routing, and iterative reflection closed loop of the system. It builds a standardized state transition framework based on conditional directed state graphs, and has a built-in unified state dictionary to store full-process information such as user input, process constraints, graph mechanism context, model prediction results, iteration counts, and process schemes. It achieves full-process traceability and control by relying on conditional jumps of state nodes. The module adopts a hierarchical collaborative scheduling architecture of large and small models, and divides the computing power according to the task reasoning difficulty: the lightweight large language model is responsible for lightweight basic tasks, such as user intent recognition, cross-modal alignment of natural process parameters to standardized physical values, task routing and distribution, and state information normalization; the deep reasoning large model undertakes high-order complex reasoning tasks, such as generating new process parameters, tracing the root cause of failures, correcting iterative parameters, and identifying conflicts in physical mechanisms. The module incorporates dedicated cross-modal parameter alignment rules. For three types of recycling processes—wet, pyrometallurgical, and direct regeneration—it unifies textual descriptions of acid, roasting atmosphere, and doping reducing agents into quantifiable values ​​such as dissociation constants, discrete state codes, and standard generation enthalpy, eliminating ambiguity in natural language descriptions and outputting a standardized feature matrix to the downstream prediction module. Based on the large model intent parsing results, the system dynamically divides business requirements into four independent workflows: process exploration and optimization, pure mechanism question answering, direct operating condition prediction, and failure experiment diagnosis. Each of these four paths has differentiated operational logic: only the process exploration and optimization path and the failure experiment diagnosis path have an iterative reflection closed loop with controllable upper limits; the pure mechanism question answering and direct operating condition prediction paths lack iterative verification logic. Specifically, the process exploration path automatically backtracks upstream nodes to iteratively correct process parameters when prediction performance is substandard or when operating conditions conflict with the physical mechanisms in the graph. The failure diagnosis path is configured with dedicated recursive processing logic, which reversely retrieves the knowledge graph to locate the failure mechanism by comparing the measured and predicted performance gaps, generates corrected parameters, and performs secondary prediction verification, forming a self-consistent closed loop of prediction—diagnosis—correction—re-prediction. The iterative reflection closed loop embeds game theory parameter attribution logic: After each round of process conditions completes the physical self-consistency verification of the graph, the multi-process parameter attribution and graph linkage optimization module is called within the closed loop to quantify and analyze the contribution of the current operating condition parameters and the cross-process coupling effect, and guide the correction of the next round of process parameters based on the attribution results; after the iteration converges and the final qualified process scheme is generated, the finalized operating condition data is uniformly distributed to the process economy and carbon emission integrated assessment module, and a comprehensive decision report that takes into account recycling performance, cost, and carbon emissions is output; the pure mechanism question-and-answer path has no predictive operating conditions, directly assembles the mechanism analysis report output, and does not enter the attribution and carbon emission assessment link.

[0012] The module for multi-process parameter attribution and graph-linked optimization is embedded within the reflection and evaluation closed loop of the multi-agent cognitive decision-making module. It is only activated during process exploration and failure diagnosis iteration processes. The module reuses the pre-stored global Shapley values ​​and feature importance from the offline TreeSHAP dual-track machine learning model. It performs local SHAP game-theoretic attribution with the measured-predicted residual deviation as the target, locating high-contribution process features causing performance deviations and analyzing the nonlinear coupling laws of parameters for wet, pyrometallurgical, and direct regeneration processes. The module uses the selected core features as retrieval anchors, and through a global verification backflow path, drives a knowledge graph-oriented retrieval of associated materialization mechanisms, optimal parameter ranges, and cross-process coupling boundaries, simultaneously updating process search constraints and failure matching rules. The updated mechanism constraints are fed back to the multi-agent to correct parameter generation strategies and initiate a new round of iterative prediction. The system iteratively executes the attribution-spectrum retrieval-constraint update-repreneurial prediction process until the deviation between the measured and predicted data reaches a preset threshold. It then exits the reflection loop and outputs process parameter tuning and optimization suggestions. The converged operating data is then transferred to the downstream economic carbon emission assessment module.

[0013] The system includes an integrated assessment module for process economy and carbon emissions. This module incorporates a database of raw material prices and process carbon emission factors, and pre-sets material loss calculation formulas. After receiving process plans, material parameters, and performance predictions, it automatically calculates the processing cost per batch and the total carbon emissions for the entire process, outputting a comprehensive assessment report that considers recycling efficiency, economics, and low-carbon characteristics. The system leverages knowledge graphs to provide white-box mechanism constraints, dual-track machine learning to achieve high-precision performance prediction, multi-channel intelligent agents to perform iterative optimization across multiple scenarios, an attribution module to ensure mechanism interpretability, and a carbon emission module to perform multi-dimensional comprehensive calculations, forming a complete knowledge-data dual-driven collaborative optimization architecture.

[0014] Based on the above system execution, the following steps are included: S1. Knowledge Graph Construction: Based on the large language model, extract the triplet of lithium battery recycling literature, and combine expert ontology and physical rules for joint quality inspection to construct a knowledge graph of the battery recycling field with directed attributes. S2, Graph Multi-hop Constraint Retrieval: Parse user process requirements, dynamically compile graph database retrieval statements, perform multi-depth topological traversal, and extract reaction mechanisms and optimal process intervals as prerequisite physical constraints; S3. Differentiated machine learning prediction: For wet and pyrometallurgical processes with missing labels, a teacher-student pseudo-label track-sharing model is used to predict the multi-metal leaching rate; for direct regeneration processes with complete labels, physical and chemical derivative features are embedded, and a fully supervised ensemble model is used to predict the regeneration cycle capacity. S4, Multi-path Intent Routing: Parses user business needs and dynamically distributes them to four types of workflows: process exploration, pure mechanism Q&A, direct operating condition prediction, and failure diagnosis. Only the exploration and diagnosis paths initiate an iterative reflection closed loop. S5. Iterative Game Theory Parameter Attribution and Mechanism Verification: The physical self-consistency comparison of the prediction working conditions in a single round is completed. Within the reflection closed loop, the parameter contribution and multi-dimensional coupling effect are quantified through game theory to lock the high-contribution process feature linkage graph and generate parameter tuning scheme. If the working condition violates the physical mechanism or the performance does not meet the standard, the state machine is triggered to backtrack and correct the parameters and return to S3 for re-prediction until the iteration converges. S6. Integrated assessment of process economy and carbon emissions: After iterative convergence to obtain the optimal process scheme, the system automatically calculates material consumption, processing costs, and the carbon footprint of the entire process, and outputs a comprehensive process assessment report that takes into account recycling performance, economic efficiency, and low-carbon indicators.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) Applicable to the optimization of process parameters for the entire process of wet leaching, pyrometallurgical roasting, and direct cathode regeneration: In wet leaching scenarios, predict the leaching rate of multiple metals and optimize acid concentration, temperature, solid-liquid ratio, and reducing agent ratio based on spectral and attribution results; In pyrometallurgical scenarios, predict the metal recovery rate after roasting, analyze the temperature-time interaction across processes, and optimize the entire set of roasting and leaching process parameters; In direct regeneration scenarios, predict the cycle discharge capacity and optimize heat treatment, lithium replenishment, and doping schemes based on the lattice repair mechanism; In experimental diagnostic scenarios, automatically trace the root cause of failure conditions and iteratively reconstruct the process; Simultaneously calculate costs and carbon emissions to support the selection of green and low-carbon processes.

[0016] 2) Leveraging domain-specific knowledge graphs and multi-hop causal retrieval, pre-process design incorporates material constraints, significantly reducing ineffective process schemes that violate thermodynamics and kinetics; an innovative dual-track, student-teacher pseudo-label framework distinguishes the physicochemical differences between lithium and transition metals, resolving the multi-task negative transfer problem in industrial datasets lacking labels, and simultaneously improving multi-metal prediction accuracy; a four-way split intelligent agent and state machine reflection loop are built to achieve automatic diagnosis of experimental failures and iterative parameter optimization, significantly shortening the manual trial-and-error cycle and reducing raw material and energy costs; game theory-based quantitative attribution is introduced to bridge the gap between machine learning black boxes and metallurgical mechanisms, providing theoretical support for global process optimization; a two-way anti-illusion link and hierarchical scheduling of large and small models are configured to balance inference accuracy and computational power consumption, improving the overall operational stability of the system; an integrated cost and carbon emission assessment unit is set up to standardize the representation of the three types of recycling processes. Attached Figure Description

[0017] Figure 1 is a schematic diagram of the ontology topology and entity relationship mapping of the battery recycling-specific dynamic knowledge graph (BR-KG) in an embodiment of the present invention; Figure 2 is a logic diagram of pseudo-label evolution and fully supervised learning data flow of the dual-track machine learning prediction module in an embodiment of the present invention. Figure 3 is a flowchart of the cognitive routing workflow and state machine reflection closed loop of the multi-agent cognitive routing in an embodiment of the present invention; Detailed Implementation

[0018] To fully elucidate the technical solution and practical effects of the present invention, the core concept, implementation process, and resulting technical benefits of the present invention will be systematically described below in conjunction with specific embodiments. These embodiments are intended to provide detailed evidence for understanding the purpose, innovative features, and technical advantages of the present invention; however, it should be understood that the scope of protection of the present invention is not limited to the listed embodiments. The above solution will be further described below with reference to specific embodiments, and preferred embodiments of the present invention are detailed below: The system is divided into two main operating phases: the offline initialization phase and the online decision-making and operation phase.

[0019] Offline initialization phase: Start the knowledge graph automatic construction module to complete the construction of a knowledge graph specific to the lithium battery recycling field; collect public industry literature and complete or incomplete recycling datasets from factory production lines, input them into the dual-track machine learning prediction module to complete model training and solidify weights; the completed knowledge graph and the trained machine learning model are stored persistently in a unified manner, and the graph construction and model retraining are not repeated during the online operation phase.

[0020] During the online operation phase: the industrial control server only loads the pre-built graphs and the solidified prediction models; the system's standardized port receives two types of input information: structured process parameters uploaded from the laboratory or production line, failure experiment feedback data, and natural language query commands submitted by users.

[0021] Natural language instructions are only distributed to the graph topology multi-step jump connection retrieval module and the multi-agent cognitive decision-making module for processing; structured process data must first undergo cross-modal standardized numerical conversion by multi-agents before being sent to the solidified dual-track machine learning prediction module to perform performance inference.

[0022] For the three business paths that generate predicted operating conditions—process exploration, failure diagnosis, and direct operating condition prediction—the single-round prediction output first enters the prediction-graph bidirectional anti-illusion verification; if it is in an iterative reflection loop, then after verification, game theory parameter attribution analysis is embedded, and the process parameters are corrected and re-predicted based on the attribution conclusions until the operating condition meets the physical constraints and the performance indicators meet the standards. After iterative convergence, a finalized process scheme is obtained; the finalized scheme is then uniformly sent to the process economy and carbon emission integrated assessment module to complete cost and carbon footprint accounting.

[0023] For the pure mechanism question-and-answer path: there is no process performance prediction process, and the mechanism explanation report is only output based on the spectrum retrieval, without participating in the verification, attribution and carbon emission assessment links.

[0024] The system ultimately integrates and outputs a comprehensive decision-making solution, which includes: optimal process parameter prescriptions, predicted values ​​of metal recovery rate or regeneration capacity, evidence for physicochemical mechanism tracing, optimization suggestions for root cause diagnosis of failure experiments, calculation results of single-batch processing costs, and quantitative assessment data of the entire process carbon footprint. The following sections, with reference to the accompanying diagrams, explain the implementation details of each module.

[0025] For the optimal implementation of the knowledge graph automatic construction module, refer to Figure 1 ,include: This module only runs during the system's offline initialization phase. During initialization, it collects a database of non-structured chemical terms from papers, experimental reports, and patents related to wet acid leaching, pyrometallurgical roasting-water leaching, and direct regeneration. A small number of high-quality entity-relation triples are pre-annotated by domain experts as few-sample prompt examples to construct context-aware extraction prompts.

[0026] For example, for the document text "NCM layered structure can be repaired by annealing at 850℃ for 5 hours", the large language model extracts triples: (Step: Annealing, R_Temp_HAS_OPTIMAL, Variable: 850℃), (Step: Annealing, R_Time_HAS_OPTIMAL, Variable: 5h), (NCM, INCLUDES, Step: Annealing), (Annealing, ACHIEVES, Effect: Layered structure repair). The extraction results are uniformly output in JSON format (head entity, relation, tail entity). After extraction, manual review and materialization rules are used for dual filtering to remove isolated nodes, conflicting, and semantically ambiguous triples. Validated triples are imported into a graph database and instantiated to generate a directed attribute knowledge graph for battery recycling. As shown in Figure 1, the graph has seven preset node types: Paper, Battery Material NCM / LFP / LCO / LMO, Experiment Type, Step, Variable, Effect, and Mechanism; and is equipped with directed relation edges for INCLUDES, PROCESS, ACHIEVES, various optimal parameter mappings, and mechanism explanations. In the online phase, only this graph is read to perform the search.

[0027] A preferred implementation of the graph topology multi-step jump link retrieval module includes: During the online phase, the module receives retrieval requests forwarded by multiple intelligent agents, analyzes user retrieval intent using a large language model, and dynamically generates dedicated retrieval statements for the graph database based on pre-defined battery recycling knowledge graph ontology rules. The module incorporates standardized graph matching rules: different cathode materials (NCM / LFP / LCO / LMO) are precisely matched using independent node labels, and the retrieval path traversal depth supports 1-4 hops, with a default topology traversal of 1-3 hops under normal operating conditions. The retrieval generation logic forcibly disables result sorting and quantity truncation instructions to ensure full recall of all causal relationship subgraphs. Simultaneously, it differentiates between two retrieval modes based on user needs: for scenarios involving process parameter recommendations and performance prediction, it extracts full process, variable, and mechanism data; for scenarios only querying a single reaction mechanism or a single reagent, it automatically adds keyword filtering logic to narrow the retrieval scope. Example query requirement: Optimal annealing temperature for direct regeneration of waste lithium iron phosphate. The system automatically matches LFP battery node tags, links regeneration and annealing related process nodes for multi-hop traversal, and fully associates literature, operational variables, and reaction mechanisms. After the search is completed, thermodynamic and kinetic qualitative constraints such as "annealing temperature above 900℃ will cause grain coarsening and increased impedance" are extracted, and the complete set of physical boundary conditions is fed back to the multi-channel intelligent agent as a prerequisite constraint for process generation and iterative optimization. The original structured data is automatically and persistently stored, and a complete literature traceability record is retained.

[0028] The preferred implementation of the dual-track machine learning prediction module is as follows: Figure 2 ,include: The module is divided into two parts: offline model training and online inference. The offline stage uses literature and complete or incomplete datasets from production lines to complete the training. The online stage only receives standardized numerical features converted by the agent and does not receive the original natural language, so it does not retrain.

[0029] Track 1: Offline training process for wet and pyrometallurgical pseudo-label evolution. For samples with asymmetric and incomplete labels, a subset of all four metals with complete annotations is first selected to train the teacher model. For Ni / Co / Mn, XGBoost can be encapsulated using MultiOutputRegressor, or XGB's native multi_output_tree can be used for joint training. Li uses a single independent output branch to avoid negative transfer. The underlying algorithm can be XGBoost, Random Forest, or LightGBM. Hyperparameter mesh optimization uses a generalization gap penalty formula; in this implementation, λ=5. After the teacher infers on the incomplete samples, a 0.9 confidence threshold and a material-specific metal mask (zeroing the predicted values ​​of Ni / Co / Mn in the LFP system) are set to filter pseudo-labels. Real labels and high-quality pseudo-labels are combined to train the student model, and the weights are solidified for online use after training.

[0030] The second step: Directly regenerating the fully supervised offline training process. The complete labeled dataset uses random forest for initial feature screening with an importance threshold of 0.1; Bayesian optimization is used to partition the training set, achieving an optimal test set ratio of 13.7% in this implementation; five-fold cross-validation combined with residual analysis is used to evaluate the physical fidelity of the model.

[0031] Based on the process temperature and time variables, the higher-order characteristics of heat treatment are derived, and the exemplary logarithmic calculation formula is only a preferred form.

[0032] The optimal implementation of the multi-agent cognitive decision-making module is described in reference to... Figure 3 ,include: This module is the core of the system's online phase global scheduling and cross-modal transition. Internally, it builds a directed graph framework for state transitions with conditional decision-making, accompanied by a globally unified state dictionary. The state dictionary includes key fields: original user input, parsed process target constraints, physicochemical characteristics of the feed material, mechanistic context subgraph returned by the graph retrieval, performance prediction results output by the machine learning module, reflection iteration count, candidate process parameter set, mechanistic source evidence, and temporary economic carbon calculation data. The module relies on directed conditional edges to achieve transitions between business state nodes. After a lightweight large language model parses the user intent, tasks are automatically routed to four types of business paths.

[0033] The module adopts a layered collaborative reasoning mechanism with lightweight and heavy models, dividing the reasoning carriers according to task complexity: the lightweight large language model undertakes lightweight processing work, including user intent classification, conversion of natural process parameters into standardized physical values, task distribution for each business path, and regularization of state dictionary information; the large model with deep reasoning capabilities of long-chain thinking is responsible for high-order complex operations, including autonomous generation of new process parameters, root cause localization of failure experiments, iterative reflection and correction, and judgment of material constraint conflicts.

[0034] 1) Exploratory Path. Triggering Scenario: The user only provides the material composition and process constraints of the waste battery, without any ready-made complete process parameters. Execution Flow: The deep reasoning model retrieves the pre-built offline battery recycling knowledge graph and derives multiple sets of initial process operation parameters by combining the physical and chemical characteristics of the material; after the lightweight model completes the parameter standardization encoding, it is sent to the offline-trained dual-track machine learning prediction module to output performance predictions; the single-round prediction results are sent to the internal reflection and evaluation closed-loop points of this module to complete the physical consistency verification between the working condition and the graph mechanism; if the prediction index does not reach the user's preset qualified threshold or the working condition violates the physical and chemical mechanism of the graph, the iteration count inside the state dictionary is automatically incremented; within this round of closed loop, the multi-process parameter attribution and graph linkage optimization module is called to carry out game theory feature analysis, and the parameter shortcomings are located based on the coupling law of high contribution parameters, generating a corrected process and sending it back to the prediction module for recalculation; this path allows a maximum of three rounds of reflection and backtracking, and parameter iteration stops when the iteration limit is reached or the working condition meets the dual constraints of performance and mechanism; the converged and finalized working condition is stored in the state dictionary, waiting to be uniformly distributed to the evaluation module.

[0035] 2) Mechanism-based path. Triggering scenario: Users only need theoretical analysis of process reaction mechanisms, parameter interaction laws, and material phase transition boundaries, without requiring performance numerical prediction. Execution process: Only the graph topology multi-step jump connection retrieval module is called to extract the complete causal subgraph. The deep model sorts out the thermodynamic and kinetic interaction logic based on graph entities and relationships, and directly outputs a pure mechanism analysis report. This path does not call the machine learning prediction module, does not enter the reflection and evaluation closed loop, and does not trigger parameter attribution and carbon emission assessment processes.

[0036] 3) Predictive Path. Triggering Scenario: The user has provided a complete set of process parameters and only needs to quickly obtain the predicted results of metal leaching rate or regeneration cycle capacity. Execution Flow: The lightweight model completes the standardized feature splicing of the text parameters and directly sends them to the solidified dual-track machine learning prediction module to output performance values, while simultaneously generating a simplified single-condition parameter contribution analysis; no parameter generation, multi-round iterative optimization logic; the prediction results enter the internal reflection and evaluation closed loop of the multi-path intelligent agent to complete the physical consistency verification. After the verification is completed, it directly enters the process economy carbon emission integrated evaluation module, without activating the complete game theory analysis of the multi-process parameter attribution and graph linkage optimization module.

[0037] 4) Diagnostic Path. Triggering Scenario: Failure test data is uploaded from the laboratory or production line, requiring the identification of failure causes and the provision of improved processes. Execution Flow: The system first calculates the performance gap between the model's theoretical prediction results and the actual experimental indicators. Using the gap as a search keyword, it reversely retrieves the knowledge graph to match the corresponding reaction mechanism, identifying the type of failure originating from exceeding temperature limits, insufficient reaction time, or missing key dopant materials. The deep reasoning model, combined with the optimal parameter range of the graph, generates a correction scheme. After standardization, it is sent back to the prediction module for secondary performance verification, and then sent to this module for reflection and evaluation of closed-loop points to conduct physical consistency verification. If the verification finds that there are still mechanism conflicts or performance gaps in the operating conditions, the multi-process parameter attribution and graph linkage optimization module is called to analyze the parameter interaction effect. The graph retrieval constraints are updated through the global verification backflow path, and the updated constraints are sent back to regenerate the correction parameters for recursive calculation, forming a fixed recursive closed loop of prediction output → performance comparison → mechanism tracing → parameter correction → secondary prediction → closed-loop verification. The number of iterations is limited by a counter in the state dictionary, and only qualified operating conditions after iteration convergence are retained in the state dictionary.

[0038] After all four business paths have been processed, they will be distributed and processed uniformly. 1. Explore the path, diagnose the path after iterative convergence, and predict the path after internal closed-loop verification. These are then simultaneously sent to the integrated assessment module for process economy and carbon emissions. Combined with the game theory attribution analysis results already completed within the closed loop, a complete comprehensive decision report is generated. 2. The mechanism path has no predicted operating conditions, and the mechanism explanation report can be directly assembled and output without going through the verification, attribution, and carbon emission assessment links.

[0039] A preferred implementation of a game-theoretic attribution explanation module for multi-process parameter interactions includes: This module interfaces with the prediction-graph bidirectional anti-hallucination verification module, reusing the offline pre-computation and caching of global SHAP feature importance and parameter interaction patterns using a dual-track machine learning model. The computational logic is divided into two layers: 1. Offline global computation: Based on the complete training dataset of the model, the TreeSHAP algorithm is used to traverse all decision trees, calculate the overall contribution of each parameter, and store the global parameter importance and interaction heatmap results in advance; 2. Online single-sample calculation: After the user inputs a set of standardized process features, the same tree model is reused to quickly solve the local SHAP contribution value of each parameter under the current working conditions. The values ​​are compared to select the core parameters that have the greatest impact on the prediction performance.

[0040] Using this high-contribution parameter as a search keyword, a targeted search is initiated to the graph topology multi-step jump-link retrieval module. This retrieves related literature records, thermodynamic mechanisms, and process feasibility boundaries, forming a complete causal subgraph centered on this feature. The module performs parameter coupling analysis for three types of recycling processes: wet acid leaching analyzes the nonlinear compensatory relationship between slurry concentration and leaching temperature; pyrometallurgical roasting quantifies the temperature and time compensation effect and the critical range of grain sintering degradation; and direct regeneration clarifies the synergistic control window of annealing and lithium doping. The optimization direction of the selected high-contribution parameter is simultaneously pushed to the multi-agent cognitive decision-making module, updating the priority adjustment logic for subsequent parameter generation and iterative correction, forming a complete closed loop of performance prediction → single-condition key feature identification → targeted mechanism retrieval → intelligent parameter tuning strategy update.

[0041] The optimal implementation of the integrated assessment module for process economy and carbon emissions includes: This module simultaneously receives two input data streams: one is the performance indicators such as metal recovery rate and material recycling capacity output by the dual-track machine learning prediction module; the other is the complete set of standardized process operation parameters and material addition ratios output by the multi-channel intelligent agent cognitive decision-making module. The module has a locally built-in database containing the unit prices of various lithium battery recycling raw materials, acid-base reducing agents, dopant consumables, equipment unit power consumption, and carbon emission basic factors corresponding to various chemical raw materials and heat treatment processes. It also pre-configures calculation formulas for material loss, reagent consumption, and power consumption for each process route. During the online calculation phase, the module matches the corresponding calculation formula based on the current process's wet, pyrometallurgical, or direct regeneration route, and automatically calculates step-by-step: raw material procurement cost, process power consumption cost, and consumable reagent usage cost, summarizing to obtain the total processing cost for a single batch. Simultaneously, it calculates the total carbon dioxide emissions for the entire process by cumulatively adding carbon emission factors at each stage, and can also output the carbon footprint per unit of recycled metal and carbon footprint per unit of recycled material for comparison. The module integrates cost figures, carbon emissions, high-cost process labeling, and comprehensive evaluation conclusions on economic efficiency and low carbon emissions. These are incorporated into the final comprehensive decision report along with process prescriptions, performance predictions, and mechanism traceability information, allowing users to compare the economic and environmental advantages and disadvantages of different process solutions from multiple dimensions.

[0042] Example 1: Exploratory Approach: Environmental Optimization of Direct Recycling Process for Waste Ternary Black Powder Implementation Scenario: A retired new energy vehicle battery recycling company processes a batch of waste ternary lithium-ion battery powder. ICP-OES analysis revealed the following feed composition: Li=4.77 wt%, Ni=23.28 wt%, Co=9.2 wt%, Mn=12.88 wt%, Al=2.86 wt%, with a particle size r=44 μm. The factory is subject to strict environmental regulations, requiring the recycling process to achieve environmental optimization, i.e., reducing acid mist emissions and lowering wastewater treatment pressure.

[0043] System execution process: Users input natural language commands through the system's interactive interface: "There is a batch of waste ternary black powder, with ICP data as above. Minimizing environmental impact is required. Please recommend the optimal combination of operands and predict the effects." Step 1: Intent Routing. Based on the routing workflow in Figure 3, the system's multimodal LLM intent routing engine identifies that the query contains exploratory terms such as "recommendation" and "minimal environmental impact," clearly determining it as an "exploratory path (Path 1)," thus triggering the knowledge graph retrieval and parameter creation process.

[0044] Step 2: Graph Retrieval and Process Selection. As shown in Figure 1, the graph topology multi-step hop-connection retrieval module dynamically compiles Cypher query statements, accurately matches NCM node labels based on battery system type, and performs a multi-hop traversal with a depth of 2 to retrieve operational variables, chemical mechanisms, and environmental indicators related to the direct regeneration process. The system extracts the following qualitative assessments from the graph literature: molten salt-assisted direct regeneration can avoid the strong acid waste liquid problem of wet processes, molten salt can be recycled (recovery rate >95%), and the low-temperature process reduces energy consumption by 35%-50% compared to traditional pyrometallurgical processes. Based on this, the system selects the direct regeneration process as the optimal path.

[0045] Step 3: Parameter Creation and Feature Engineering. The deep thinking model in the multi-agent cognitive decision-making module, combined with the user's feeding conditions and spectral literature, independently derives the initial operating parameters: The tube furnace temperature is set to 320°C: According to spectral literature, in a LiOH:LiNO3=2:3 molten salt system (melting point 433°C), operating in a sub-molten state can enhance Li... + It diffuses, and at 113°C lower than the melting point, it reduces the risk of high-temperature decomposition; The tube furnace time is set to 4 hours: based on reaction kinetic calculations, to ensure that Li + It diffuses fully within the 44μm particles; The annealing temperature is set at 850°C: based on the thermodynamics of lattice reconstruction, this temperature provides sufficient energy margin to complete defect repair, while saving 15% energy compared to the traditional 900°C sintering and suppressing Al impurity segregation; The annealing time is set to 5 hours: based on the Ostwald ripening control principle, this duration allows for appropriate grain growth and repair of grain cracks. The lithium replenishment amount is set to 1.05: Based on the material balance equation, for a lithium loss rate of 17.7%, replenishment is carried out at a stoichiometric ratio of 1.05, matching the optimal practice of the spectrum; The reducing agent type was set to LiOH:LiNO3=2:3 (mapping coefficient -5.909282379), and the impurity control adopted the "no addition" mode.

[0046] The system further calculates higher-order features online using built-in physical descriptor-derived formulas: Tube_heating_intensity = 320 × ln (4 + 1) ≈ 514.7; Annealing_intensity = 850 × ln (5 + 1) ≈ 1523.0; Total_heat_treatment ≈ 2037.7; Oxygen_release_tendency is calculated based on the Ni / Co / Mn ratio; Temp_time_coordination, Reducing_agent_interaction, etc. are generated synchronously.

[0047] Step 4: Machine Learning Prediction. As shown in the fully supervised orbit in Figure 2, the above feature vectors are input into the direct regeneration fully supervised orbit (random forest model) in the dual-track machine learning prediction module. The model outputs a predicted discharge specific capacity of 140.39 mAh / g after the 100th cycle of the recycled material.

[0048] Step 5: Internal reflection of the agent to verify physical consistency and game theory attribution analysis.

[0049] The predicted operating conditions were fed into the built-in reflection and evaluation closed-loop point of the multi-channel intelligent agent to carry out the physical consistency verification of the graph; the graph literature records that the optimal capacity range under similar operating conditions is 135-145 mAh / g, the predicted value falls within this range, there is no mechanism conflict, and the iterative backtracking logic is not triggered.

[0050] The multi-process parameter attribution and graph linkage optimization module is called synchronously, and the offline pre-stored TreeSHAP global feature contribution value is reused to carry out local game theory SHAP attribution calculation for the direct regeneration conditions of this group. The positive and negative contributions of four core parameters, namely tubular furnace temperature, annealing temperature and time, lithium replenishment ratio, and molten salt reducing agent, to the cycle capacity are quantified. The parameter coupling effect of temperature and time synergy and lithium replenishment doping is analyzed, the key control parameters affecting the regeneration capacity are identified, and the parameters are stored in the state dictionary to form interpretable analysis basis.

[0051] Since the operating conditions and physical constraints meet the preset standards, the state machine determines that the convergence termination condition has been met and exits the reflection loop.

[0052] Step 6: Calculation of the integrated assessment module for process economy and carbon emissions.

[0053] The converged and finalized process parameters, material composition, and the predicted capacity of 140.39 mAh / g are input into the integrated process economy and carbon emission assessment module. The module retrieves the built-in raw material unit price and carbon emission factor library for the heat treatment process, automatically calculates the cost of consumables and electricity consumption for processing ternary black powder per batch, and accumulates the carbon dioxide emissions of the entire process for each process to quantify the carbon footprint of a unit of recycled material. The assessment results show that the comprehensive energy consumption is about 1.2 GJ / ton, the carbon emission is reduced by about 57% compared with the traditional high-temperature pyrometallurgical process, and there are no additional costs for wet acid mist and wastewater treatment.

[0054] Step 7: Generate an interpretability comprehensive report.

[0055] The system integrates spectral mechanism tracing, SHAP parameter attribution analysis, performance prediction values, and economic and carbon emission accounting results to assemble a complete comprehensive decision-making report. The core content of the report includes: Process selection criteria. Direct regeneration avoids strong acid waste liquid, and molten salt is recyclable; parameter creation and derivation. The derivation logic and literature review for each core value; game theory attribution analysis. The contribution weight and coupling control law of each heat treatment and lithium replenishment parameter to cycle capacity; performance prediction results. The discharge capacity at the 100th cycle is 140.39 mAh / g; economic and environmental dual-dimensional assessment. Single-batch processing cost calculation, zero acid mist emission, near-zero wastewater load, comprehensive energy consumption, and quantitative data on carbon emission reduction.

[0056] Implementation Results: The company conducted experiments according to the SOP recommended by the system, and the actual measured discharge capacity of the recycled material after 100 cycles was 138.5 mAh / g, with an error of only 1.3% compared to the predicted value. At the same time, combined with the key parameter control logic of attribution analysis, the company further improved the stability of the recycled capacity after slightly optimizing the annealing and heat preservation range. This verified the complete function and reliability of the system in process optimization, parameter interpretability analysis, and integrated economic and low-carbon assessment under environmental constraints.

[0057] Example 2: Diagnostic Approach: Failure Diagnosis and Optimization of Cobalt Recovery in Wet Leaching Process Implementation Scenario: A recycling company used wet leaching to treat a batch of ternary lithium black powder. ICP analysis data showed: Ni=28.02 wt%, Co=8.04 wt%, Mn=13.1 wt%, Li=6.16 wt%, Al=2.48 wt%, particle size r=44 μm. The user-defined operating parameters were: 1.5 mol / L sulfuric acid, 2 vol% hydrogen peroxide, solid-liquid ratio 50 g / L, reaction temperature 60°C, and reaction time 2 hours. However, actual experimental results showed that the cobalt leaching rate was only 68%, while the nickel leaching rate reached 92%, indicating extremely poor cobalt recovery.

[0058] System execution process: The user uploads complete operating condition parameters and measured failure test data, and the system starts the full-link diagnostic closed loop.

[0059] Step 1: Intent Routing. Based on the routing workflow in Figure 3, the system recognizes that the user input contains clear real experimental anomaly feedback data and a set of fixed working conditions, and determines it as "diagnostic path (Path 4)". The system then initiates the multi-agent built-in prediction-diagnosis-correction-reprediction recursive reflection closed loop.

[0060] Step 2: Initial performance prediction and gap calculation. First, the user's original wet process condition standardized code is sent to the wet pseudo-label multi-output track of the dual-track machine learning prediction module in Figure 2 to obtain the initial multi-metal leaching rate prediction value; compared with the measured cobalt leaching rate of 68%, the quantified cobalt metal performance gap reaches 22%.

[0061] Step 3: Graph Reverse Mechanism Retrieval Alignment. Using the 22% performance gap as the retrieval anchor point, the graph topology multi-step jump-connection retrieval module in Figure 1 is invoked to traverse the BR-KG knowledge graph and match the core reaction mechanism: 60°C is not within the activation temperature range of the cobalt reduction reaction, and the hydrogen peroxide reduction activity cannot be fully released, resulting in a significantly low cobalt dissolution efficiency.

[0062] Step 4: First round of reflection on game theory attribution analysis within the closed loop.

[0063] The module calls upon the multi-process parameter attribution and graph linkage optimization module to conduct TreeSHAP game theory attribution with the deviation between measured and predicted leaching rate residuals as the target. It quantifies the local contribution of all wet process parameters and determines that reaction temperature and hydrogen peroxide addition are the two core parameters with high contribution to cobalt leaching failure. The module relies on the global verification backflow path to transmit the two types of key features back to the knowledge graph, and searches for the optimal operating window and parameter coupling boundary corresponding to temperature and reducing agent, and updates the process parameter search constraints.

[0064] Step 5: Targeted prescription generation. The multi-agent deep inference model generates an optimized process scheme based on the updated physical constraints of the graph. The reaction temperature was increased from 60°C to 70°C: Spectroscopic literature confirms that 70°C is the optimal temperature for H2O2-assisted leaching of cobalt, balancing the reaction rate and the stability of the reducing agent; the amount of hydrogen peroxide was reduced from 2 vol% to 1 vol%: Excess H2O2 is easily decomposed ineffectively at high temperatures, reducing the amount reduces reagent loss; the sulfuric acid concentration was maintained at 1.5 mol / L, the solid-liquid ratio at 50 g / L, and the reaction time at 2 hours.

[0065] Step 6: Secondary performance prediction and agent-embedded physical consistency verification.

[0066] The corrected standardized feature vector (Temperature=70°C, H2O2_vol=1.0%, 1_pKa1=-2 corresponding to sulfuric acid, Ni_in_Feed=28.02, etc.) is input into the dual-track machine learning prediction module wet pseudo-label multi-output orbital, and the leaching rates of four types of metals are predicted in a secondary manner: Li>98%, Ni>95%, Co>95%, Mn>96%; The prediction results are fed into a multi-channel intelligent agent for reflection and evaluation. The closed-loop process completes the physical consistency check of the spectral data, verifying that the 70°C condition perfectly matches the optimal range for cobalt leaching recorded in the spectral data, with no mechanistic conflicts, and that the measured-prediction residuals meet the preset deviation threshold. The state machine determines that the iteration has converged and terminates the recursive loop.

[0067] Step 7: Comprehensive calculation of the integrated assessment module for process economy and carbon emission.

[0068] The original failure condition and the two optimized process schemes are simultaneously input into the integrated process economy and carbon emission assessment module. The module retrieves the built-in database of acid and alkali reagents, electricity carbon emission factors and raw material prices, and calculates the single batch reagent consumption, energy consumption and processing cost, and total carbon emissions of the two schemes respectively. The optimized scheme has reduced hydrogen peroxide consumption, reduced low-temperature reaction energy consumption, and reduced reagent costs and process carbon footprint.

[0069] Step 8: Assemble and output the comprehensive diagnostic report.

[0070] The report integrates the following: numerical values ​​of failure performance gaps, knowledge graph-based tracing of reaction mechanisms, high-contribution failure parameters obtained from SHAP game theory attribution, optimized process operation prescriptions, comparison of metal leaching rate predictions under new and old operating conditions, quantitative comparison of the economics and carbon footprint of the two schemes, and experimental improvement suggestions.

[0071] Implementation Results: Enterprises conducted verification experiments according to the system optimization plan, achieving a measured cobalt leaching rate of 94.2% and a nickel leaching rate of 95.8%, which highly matched the system's secondary prediction results. This embodiment fully verifies the recursive closed-loop diagnostic path, game-theoretic attribution failure root cause localization, graph-linked parameter correction, and integrated economic carbon emission assessment across the entire chain. It compresses the traditional manual trial-and-error debugging cycle of several weeks to several hours, significantly reducing the R&D experimental costs and carbon emission trial-and-error losses in wet process technology.

[0072] Example 3: Prediction and Diagnosis Path: Prediction, Reflection, and Optimization of the Combined Pyrometallurgical and Hydrometallurgical Process for Lithium Iron Phosphate Black Powder Implementation Scenario: A company is processing a batch of waste lithium iron phosphate powder. The feed characteristics are: Li=2.85 wt%, Fe=23.89 wt%, Al=0.88 wt%, other metals not detected, particle size 44 μm. The user has decided to use a combined pyrometallurgical and hydrometallurgical process and has mandated the following operating conditions: graphite as additive (mapping value 12), additive-to-waste mass ratio 0.10, inert gas roasting atmosphere (mapping value 2), roasting temperature 700°C, roasting time 120 minutes, CO2 flow rate 0; subsequent leaching uses 2.0 mol / L sulfuric acid with 1.0 vol% hydrogen peroxide as a reduction aid, slurry density 50 g / L, leaching temperature 75°C, leaching time 60 minutes. The user requires direct prediction of lithium extraction rate and evaluation of the scientific validity of the parameters.

[0073] System execution process: The user uploads a fixed set of process parameters and material testing data. The system simultaneously performs performance prediction and operating condition mechanism evaluation. Unreasonable operating conditions automatically trigger a diagnostic iterative closed loop.

[0074] Step 1: Intent routing and cross-modal feature conversion. Based on the routing workflow in Figure 3, the system determines that the request is a composite request of predictive path and diagnostic reflection; the lightweight large model within the multi-channel agent starts the parameter standardization mapping rule, converting the natural language description of the agent and calcination conditions into a unified numerical feature matrix.

[0075] Step 2: Initial Performance Prediction. The system extracts standardized features: Particle size=44, Li_in_Feed=2.85, Fe_in_Feed=23.89, Al_in_Feed=0.88, Additive=12, Mass_ratio_to_waste=0.10, Roasting_temperature=700, Roasting_time=120, Atmosphere=2, CO2_flow_rate=0, I_AcidC=2.0 (sulfuric acid), H2O2=1.0, Pulp_density=50, Leaching_temperature=75, Leaching_time=60; these features are fed into the pyrometallurgical and hydrometallurgical joint pseudo-label multi-output orbital of the dual-track machine learning prediction module. The model outputs an initial predicted value of 53.79% for lithium leaching rate.

[0076] Step 3: Multi-agent built-in reflection closed-loop physical consistency verification.

[0077] The initial predicted operating conditions are sent to the internal reflection and evaluation loop of the multi-channel agent to complete the physical consistency verification of the graph; the graph topology multi-step jump connection retrieval module in Figure 1 is called to retrieve the relevant mechanism constraints of LFP carbothermal reduction and acid leaching: the optimal temperature window for LFP carbothermal reduction is 500~550℃, and 700℃ high temperature will cause the Fe-PO4 framework to melt and encapsulate lithium elements; the graphite addition ratio of 0.10 is lower than the complete reduction threshold; the hydrogen peroxide acid system easily oxidizes Fe. 2+ FePO4 passivation film is generated.

[0078] The verification results showed that the predicted values ​​conformed to the physical and chemical mechanism, but the operating parameters had serious design flaws. The system automatically triggered a reflection and backtracking loop and switched the diagnostic logic to intervene in the optimization.

[0079] Step 4: Reflect on the attribution of game theory parameters and the iterative linkage of the graph within the closed loop.

[0080] The multi-process parameter attribution and graph linkage optimization module was launched. TreeSHAP game theory attribution was carried out with the residual bias of the current low prediction performance as the target. The local contribution of all pyrometallurgical and hydrometallurgical parameters was quantified, and four types of parameters with high contribution and negative influence were identified: roasting temperature, graphite addition ratio, leaching reducing agent, and slurry density. The module uses four key features as retrieval anchors and pushes them to the knowledge graph retrieval node through a global verification backflow path. It then searches for the optimal range of each parameter and the coupling constraints of multiple processes, updates the internal process parameter search and failure judgment rules of the system, and corrects the intelligent agent parameter generation constraints.

[0081] Step 5: Optimize process formulation generation. The multi-agent deep inference model generates improved operating parameters based on the updated graph materialization constraints. Lowering the calcination temperature from 700°C to 500°C: 500°C allows for the realization of Fe 3+ Complete reduction, eliminating the problem of grain sintering and encapsulation; increasing the proportion of graphite additives from 0.10 to 0.15: satisfying the thermodynamic conditions for complete carbothermic reduction reaction, ΔG < -40 kJ / mol; replacing the leaching aid from 1.0 vol% hydrogen peroxide with 0.5 mol / L citric acid, relying on complexation to inhibit the formation of FePO4 passivation film; reducing the slurry density from 50 g / L to 30 g / L, improving hydrogen ion mass transfer efficiency by 20%~35%.

[0082] Step 6: Secondary performance prediction and convergence verification. The corrected standardized feature vector is fed back into the dual-track machine learning prediction module. The model outputs a predicted lithium leaching rate of >90% under optimized operating conditions. The secondary prediction result re-enters the intelligent agent reflection loop to complete the physical consistency verification. All improved parameters fall within the optimal process range recorded in the graph. The deviation between measured and predicted theories meets the preset threshold. The state machine determines that the iteration has converged and terminates the reflection loop.

[0083] Step 7: Visualization and Analysis of Game Theory Parameter Interaction. The global SHAP value cached by the reuse module is used to generate analytical charts: the output roasting temperature SHAP dependency curve visually demonstrates the strong negative inhibitory effect of temperature on lithium leaching rate at 700℃; simultaneously, multi-parameter interactive thermograms of roasting temperature, slurry concentration, and reducing agent type are output, clearly showing the positive gain control range of low temperature + citric acid + low slurry density, providing a complete explanation of parameter coupling and positive / negative driving mechanisms.

[0084] Step 8: Comparison and calculation of two schemes in the integrated assessment module of process economy and carbon emission.

[0085] The parameters of the original unreasonable operating condition and the optimized operating condition after iterative convergence are simultaneously input into the integrated process economy and carbon emission assessment module. The module retrieves the built-in database of consumable unit price, electricity and chemical carbon emission factors for each process, and calculates the single batch raw material and reagent consumption, comprehensive operating cost and full process carbon footprint of the two schemes respectively. The original process has high energy consumption of high temperature roasting and large hydrogen peroxide reagent loss. The comprehensive cost and carbon emission are significantly higher than the improved low temperature citric acid leaching scheme. The module quantitatively outputs the economic and environmental comparison indicators of the two schemes.

[0086] Step 9: Assemble and output the comprehensive evaluation report. The report integrates: material feed parameters, predicted lithium leaching values ​​under the original operating conditions, spectral mechanism verification conclusions, SHAP game theory parameter attribution analysis, iterative optimization of the process prescription, comparison of leaching rate predictions under the new and old operating conditions, quantitative comparison of processing costs and carbon footprints of the two schemes, and operational suggestions for process improvement.

[0087] Implementation Results: The company conducted comparative verification experiments using the two schemes in the system. The measured lithium leaching rate of the original parameter group was 52.1%, which highly matched the system's initial prediction of 53.79%; the measured lithium leaching rate of the optimized parameter group was 91.3%, matching the conclusion of the second prediction >90%. This embodiment fully verifies the system's closed-loop function, which includes prediction of user-defined unreasonable operating conditions, mechanism verification, game theory attribution to locate defects, graph-linked parameter correction, iterative re-prediction, and comprehensive comparison of economic carbon emissions. It can autonomously identify process parameters that violate thermodynamics and kinetics and provide improvement schemes that take into account recycling efficiency, cost, and low carbon emissions, significantly reducing the manual trial and error cycle of lithium battery recycling processes.

Claims

1. A knowledge-data dual-driven intelligent decision-making system for the recycling of waste lithium batteries, characterized in that, include: The knowledge graph automatic construction module collects various literatures in the field of waste lithium battery recycling, extracts the causal relationships of materials, processes, and reactions, constructs a directed attribute knowledge graph in the field of battery recycling, and outputs a standardized physical mechanism knowledge base. The graph topology multi-step jump connection retrieval module connects to the knowledge graph automatic construction module, receives user process optimization requirements forwarded by multiple intelligent agents, automatically generates graph retrieval statements and performs multi-step topology traversal, extracts thermodynamic and kinetic rules recorded in literature, outputs physical constraints of process feasible intervals and sends them back to the multiple intelligent agents. The dual-track machine learning prediction module receives process operation parameters and material physicochemical properties, and uses two differentiated training frameworks to operate on different recycling processes: for wet and pyrometallurgical processes with missing labels, a teacher-student pseudo-label track-based model is used to predict the multi-metal leaching rate; for direct regeneration with complete labels, an integrated model is used to predict the cycle capacity, and outputs the performance prediction results corresponding to the recycling process. The multi-channel intelligent agent cognitive decision-making module is connected to the graph topology multi-step jump connection retrieval module and the dual-track machine learning prediction module, respectively, to receive the process feasibility interval constraints and the performance prediction results of the retired battery recycling process; according to the user's business request, the task is dynamically routed to the process exploration and optimization path, the pure mechanism question answering path, the direct operating condition prediction path, or the failure experiment diagnosis path. The process exploration and optimization path and the failure experiment diagnosis path have built-in knowledge graph mechanism verification logic. The process exploration and optimization path outputs candidate recycling process solutions, the pure mechanism question answering path outputs material mechanism explanation information, the direct operating condition prediction path outputs the performance prediction results of the existing operating conditions, and the failure experiment diagnosis path outputs failure root cause analysis and diagnosis optimization solutions. The multi-process parameter attribution and graph linkage optimization module receives the prediction results, original operating condition input features, and residual deviations between measured and predicted data output by the dual-track machine learning prediction module. It then filters out the core process parameters that contribute significantly to the residuals and uses these parameters as anchors to drive the graph topology multi-step jump-link retrieval module to search for related mechanisms and parameter boundaries. The retrieved mechanism constraints are then fed back to the multi-channel intelligent agent to update the parameter tuning strategy. The module outputs process parameter tuning optimization suggestions, and the optimization scheme is globally verified and fed back to the retrieval module for iterative correction until the measured and predicted data meet the preset deviation threshold. The integrated process economy and carbon emission assessment module connects to the dual-track machine learning prediction module and the multi-channel intelligent agent cognitive decision-making module. It receives candidate recycling process schemes, material physicochemical properties, and process performance prediction results, calculates material consumption, processing costs, and carbon emissions throughout the process, and outputs an economic and carbon footprint assessment report.

2. The system according to claim 1, characterized in that, The knowledge graph automatic construction module adopts a hybrid construction strategy that combines top-down expert ontology design with bottom-up large language model instance extraction. The node types of the knowledge graph include document nodes, battery tag nodes, experiment type nodes, experiment step nodes, operation variable nodes, effect index nodes, and mechanism nodes. The relationship edge types include subordinate relationships, calling relationships, process relationships, result achievement relationships, and operation variable optimal value mapping edges.

3. The system according to claim 1, characterized in that, The graph topology multi-step hop retrieval module relies on a large language model to dynamically match the exclusive node labels of the graph database according to the battery system type. Under normal working conditions, it performs 1 to 3 hop topology traversal, and can be extended to 4 hop adaptive retrieval in scenarios with sufficient computing power. The generated retrieval statement disables sorting and quantity limit instructions, and adopts two retrieval modes, namely full extraction or keyword filtering, depending on the scenario, to achieve full retrieval of complete causal relationship data.

4. The system according to claim 1, characterized in that, The two-stage pseudo-label evolution framework of the first track in the dual-track machine learning prediction module includes: In the teacher model training phase, a subset with complete multi-metal annotations is extracted from the original dataset. Baseline training is completed using the differentiated track-based modeling strategy: Ni, Co, and Mn are grouped together and either a MultiOutputRegressor encapsulation scheme or XGBoost's native multi_output_tree vector leaf single-model joint fitting is used. Li is built as a separate independent prediction branch. In the hyperparameter grid search, the weighted difference between the average determination coefficient of the test set and the performance deviation between the training and test sets is used as a comprehensive score, automatically selecting the model with the best generalization ability as the teacher model. In the pseudo-label generation phase, the teacher model is used to perform multi-metal synchronous inference on samples with missing labels, retaining samples with prediction confidence higher than a preset threshold. A metal presence mask is forcibly applied according to the specific battery material system to remove low-quality pseudo-labels that violate physicochemical principles. In the student model training phase, the real labeled data and the filtered pseudo-label data are merged to construct a global augmentation dataset. The differentiated track-based training strategy is used to complete transition metal joint fitting and independent lithium modeling respectively, outputting the multi-metal recovery rate.

5. The system according to claim 1, characterized in that, The dual-track machine learning prediction module uses a random forest algorithm for feature importance pre-screening and redundant variable removal on a fully labeled direct regeneration dataset in its second track. During the data partitioning stage, a Bayesian optimization framework is introduced to adaptively optimize the training-test set partitioning ratio. Five-fold cross-validation is then performed on the finalized prediction model. The physical descriptor derivation formulas include: tubular furnace heating intensity, annealing heat treatment intensity, and total heat treatment amount based on the thermodynamic intensity integral principle; an oxygen release tendency index based on the element mole fraction weighting principle; a temperature-time synergy coefficient based on the temperature-time synergy ratio principle; and a reduction coefficient based on the reactant stoichiometric interaction principle. The reducing agent interaction term; wherein, the thermodynamic intensity integral formula adopts the product of temperature and logarithmic time, T×ln(t+1) or its equivalent mathematical transformation; the oxygen release tendency index adopts the linear weighted form of transition metal mole fraction or its equivalent transformation; the temperature-time synergy coefficient adopts the product of temperature ratio and time ratio or its equivalent transformation; and the reducing agent interaction term adopts the product of reactant content and category code or its equivalent transformation. Any equivalent physical descriptor that uses the logarithmic, exponential, power function, or polynomial relationship between temperature and time to characterize heat treatment intensity, or that uses elemental composition ratio, reactant interaction product, or chemical potential gradient, falls within the scope of protection of this patent.

6. The system according to claim 1, characterized in that, The multi-channel intelligent agent cognitive decision-making module has multiple functions, including parameter standardization, multi-path routing, large and small model collaboration, and iterative diagnosis. The built-in feature numerical dictionary mapping rules complete the cross-modal alignment of natural process parameters to model input values. Specifically, the acid type in wet acid leaching is mapped to the corresponding first dissociation constant pKa1 value, the roasting atmosphere is mapped to a specific discrete state code, and the core reducing agent, impurity and dopant compound types in the direct regeneration process are mapped to the corresponding standard generation enthalpy or reaction activity coefficient. The module operates using a directed state graph framework. The framework definition includes a state dictionary containing user input, intent path, target process, extracted features, graph context, prediction results, reflection count, and final answer. It relies on conditional edges to connect each state node to achieve dynamic diversion of four types of business paths: process exploration optimization path, pure mechanism question and answer path, direct operating condition prediction path, and failure experiment diagnosis path. External mechanism reflection and backtracking loops are supported only in the process exploration optimization path and the failure experiment diagnosis path. The pure mechanism question and answer path and the direct operating condition prediction path are not configured with iterative verification logic. The module is configured with a large and small model collaborative scheduling mechanism, which allocates a lightweight large language model to be responsible for simple tasks such as intent classification, parameter extraction, and routing decision, and allocates a deep thinking model with the ability of thought chain reasoning to complete complex reasoning tasks such as parameter creation, experimental diagnosis, and mechanism tracing. The failure experiment diagnosis path is configured with recursive processing trigger logic: after receiving real experimental anomaly feedback data, it is output to the experiment diagnosis optimization node, automatically calculates the performance gap, reversely retrieves the knowledge graph to complete the mechanism alignment, generates parameter correction prescriptions and re-sent them into the prediction link to carry out secondary performance verification, forming a closed loop process of "prediction-diagnosis-correction-re-prediction".

7. The system according to claim 1, characterized in that, The integrated process economy and carbon emission assessment module has a built-in raw material price database and a carbon emission factor database for electricity and chemicals in each process. It also pre-stores material loss calculation formulas for various recycling processes. After receiving candidate process schemes, material physicochemical properties, and process performance prediction results, the module automatically calculates the raw material cost and comprehensive operating cost of a single batch of waste battery processing, and simultaneously outputs the full-process carbon dioxide emissions and the quantitative assessment results of the carbon footprint per unit product.

8. The system according to claim 1, characterized in that, The multi-process parameter attribution and graph linkage optimization module reuses the global feature weights pre-stored offline by the dual-track machine learning model, and uses at least one of the game theory Shapley and SHAP values ​​to quantify the nonlinear interaction of parameters and the importance of global features in the three types of processes: wet process, pyrometallurgical process, and direct regeneration. The module has a built-in attribution-graph linkage subunit, which maps the selected high-contribution features to knowledge graph operation variables and mechanism nodes, performs targeted deep retrieval of corresponding reaction mechanisms, optimal intervals and multi-process coupling constraints, and corrects the intelligent agent's iterative optimization strategy based on the retrieval results.

9. A knowledge-data dual-driven intelligent decision-making method for waste lithium battery recycling, characterized in that, The system according to any one of claims 1-8 is executed, comprising the following steps: S1. Based on a large language model, context-aware knowledge extraction is performed on unstructured literature in the field of battery recycling to construct a knowledge graph of the field of battery recycling containing directed attribute graphs. S2. Receive the user's query intent, parse the intent by the large language model and dynamically compile the graph database query statement, perform a multi-step jump graph traversal retrieval with a preset depth in the knowledge graph, fully recall the matching subgraph structures, and extract prior physical constraints. S3. Input process operation parameters and material physicochemical properties, and call the dual-track prediction model for different recycling processes: For wet leaching and pyrometallurgical roasting-water leaching processes with asymmetric missing labels, a teacher-student pseudo-label track-splitting model is adopted. Through a two-stage pseudo-label evolution framework and a generalization gap penalty scoring function, metal-differentiated track-splitting modeling is performed to train the teacher model and student model, and output the multi-metal recovery rate; For direct regeneration processes with complete labels, a feature space is established by combining physical descriptor derivation formulas, and a fully supervised ensemble learning mechanism is used to train the prediction model and output the regeneration capacity performance index. S4. Based on the directed state graph framework, perform intent parsing and dynamically route tasks to the process exploration and optimization path, the pure mechanism question-and-answer path, the direct operating condition prediction path, or the failure experiment diagnosis path; only the process exploration and optimization path and the failure experiment diagnosis path perform mechanism verification, which can trigger the iterative correction process; the process exploration and optimization path outputs candidate recycling process schemes, the pure mechanism question-and-answer path outputs physical mechanism explanation information, the direct operating condition prediction path outputs operating condition performance prediction results, and the failure experiment diagnosis path outputs failure root cause analysis and diagnosis optimization schemes, respectively completing parameter recommendation, theoretical source tracing, operating condition prediction, or failure reconstruction; S5. When the deviation between the measured process data and the model prediction results exceeds a preset threshold, the attribution analysis step is executed: Attribution explanation is carried out based on the SHAP value corresponding to Shapley game theory, feature permutation importance, biased dependency analysis or cumulative local effect analysis, and feature global importance ranking and multi-dimensional parameter interaction heatmap are generated; the selected high contribution features are back-mapped to the knowledge graph, targeted mechanism retrieval and parameter optimization are carried out, revealing the positive and negative driving logic and coupling boundary of parameters on recovery efficiency, and outputting process parameter tuning optimization suggestions, and iterative correction is carried out according to the global verification return path until the deviation meets the threshold; S6. Based on candidate recycling process schemes, physical and chemical parameters of materials, and process performance prediction results, calculate material consumption, processing costs and carbon emissions of the whole process, and output a comprehensive assessment report on economic efficiency and carbon footprint.

10. The application of the method in optimizing the recycling process of waste lithium batteries, characterized in that, The uses include: The system or method is used to optimize the process parameters of the wet leaching, pyrometallurgical roasting-water leaching, and direct regeneration full recycling routes for waste lithium batteries. In the wet leaching process, the leaching rates of lithium, cobalt, nickel, and manganese are predicted, and based on the physicochemical constraints and characteristic attribution explanations of the knowledge graph, the acid concentration, leaching temperature, slurry concentration, solid-liquid ratio, and reducing agent ratio parameters are optimized. In the pyrometallurgical roasting process, the multi-metal recovery rate after high-temperature roasting is predicted, and the roasting temperature, roasting time, additive ratio, atmosphere control and water immersion liquid-solid ratio parameters are optimized based on the cross-process parameter interaction effect analysis. In the direct regeneration process, the discharge specific capacity of the regenerated cathode material after cycling is predicted, and the parameters of tube furnace temperature, tube furnace time, annealing temperature, annealing time, lithium replenishment agent type and chemical dopant ratio are optimized based on the lattice repair mechanism tracing and thermodynamic-kinetic synergistic constraint analysis. In the full recovery route, root cause diagnosis and parameter reconstruction are performed based on real experimental failure results. The state machine recursive closed loop of the diagnostic path realizes autonomous iterative optimization of the process until the experimental results are highly aligned with the spectral mechanism and model predictions.

Images