A data-driven risk identification and decision optimization method for new material industry investment

Abstract

The application discloses a kind of data-driven risk identification and decision optimization methods for new material industry investment, comprising the following steps: by multi-source heterogeneous data acquisition module, the multidimensional data related to new material industry investment is collected in real time, multidimensional data includes technical research and development data, supply chain data, market and policy data, enterprise operation data and environmental social governance data;Based on the multidimensional data collected, an "industry element-geography-time" three-dimensional anchoring knowledge graph is constructed.In the application, by constructing an counterfactual reasoning engine containing positive causal reasoning, counterfactual reasoning and causal chain prediction, the transition from passive response to active prediction is realized, the risk factor is set to the non-occurrence state in the counterfactual reasoning module and the causal effect intensity is calculated, the causal chain prediction module generates the conduction path prediction based on the current graph state and the historical mode library before the risk event occurs, so that the investor can be warned at the risk budding stage, and the decision foresight is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of investment analysis and risk management technology in the new materials industry, and in particular to a data-driven risk identification and decision optimization method for investment in the new materials industry. Background Technology

[0002] The new materials industry is characterized by rapid technological iteration, long R&D cycles, long supply chains, and significant susceptibility to macroeconomic policies and geopolitical influences. Investment decisions in this industry face high uncertainty and multi-dimensional, complex risks. Traditional risk assessment methods for new materials industry investments largely rely on expert experience and static financial indicator analysis. These methods struggle to systematically capture the complex relationships and dynamic evolution between technological substitution risks, supply chain transmission risks, and policy change risks. Furthermore, they lack the ability to trace the causal paths of risks and the capacity for forward-looking prediction, resulting in insufficient scientific rigor and timeliness in investment decisions. The patent document with publication number CN119026919A discloses a supply chain risk assessment and early warning system and method based on multidimensional data maps. The method obtains the timeliness assessment data and accuracy assessment data of each link in the supply chain, calculates the timeliness assessment value and accuracy assessment value, and combines the reliability assessment data to obtain the reliability assessment value, so as to realize the assessment and early warning of supply chain risks.

[0003] While the aforementioned methods incorporate multidimensional data mapping for supply chain risk assessment, their assessment dimensions are limited to static indicators such as timeliness, accuracy, and reliability. They lack the ability to model the dynamic propagation of supply chain risks between upstream and downstream nodes, and fail to identify special nodes with risk buffering functions within the supply chain to guide investment decisions. This makes it difficult to meet the practical needs of forward-looking risk identification and optimization decisions in new materials industry investment. Therefore, we propose a data-driven risk identification and decision optimization method for new materials industry investment to address this problem. Summary of the Invention

[0004] The purpose of this invention is to provide a data-driven risk identification and decision optimization method for investment in the new materials industry, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A data-driven risk identification and decision optimization method for investment in the new materials industry includes the following steps: S100. Real-time collection of multi-dimensional data related to investment in the new materials industry through a multi-source heterogeneous data acquisition module. The multi-dimensional data includes technology research and development data, supply chain data, market and policy data, enterprise operation data, and environmental and social governance data. S200. Based on the collected multidimensional data, construct a three-dimensional anchored knowledge graph of "industrial elements-geography-time". Each industrial entity in the knowledge graph is simultaneously bound to a geographical anchor and a time anchor. The geographical anchor includes the policy and regulatory vector, resource endowment vector and geopolitical risk index of the region where the entity is located. The time anchor includes the state time series and relationship time window of the entity. S300. Based on the three-dimensional anchored knowledge graph, a counterfactual reasoning engine is constructed. The counterfactual reasoning engine includes a forward causal reasoning module, a counterfactual reasoning module, and a causal chain prediction module. The forward causal reasoning module identifies the causal structure between risk factors from the multidimensional data through a causal discovery algorithm and generates a risk causal graph. The counterfactual reasoning module sets the value of a specific risk factor to a "not occurred" state through intervention and drives the structural causal model to recalculate the counterfactual results. Based on the counterfactual results, the causal effect strength of each risk factor is calculated. The causal chain prediction module predicts the risk transmission path before the risk event actually occurs, based on the current knowledge graph state and the historical causal pattern library, through pattern matching and extrapolation. S400. Based on the three-dimensional anchored knowledge graph and the risk causal graph, construct a digital twin of the new material project or enterprise to be evaluated. The digital twin includes a financial model, a supply chain model, a technology evolution model, and a market response model. Data interaction between the sub-models is achieved through a two-way data interface. S500: Perform multi-scenario stress tests on the digital twin to generate a dynamic risk evolution trajectory; S600. Construct a multi-layer network model for the contagion of risks in the industrial chain. The multi-layer network model includes a resource layer, a processing layer, a material layer, and an application layer, which are connected by directed weighted edges. Based on the dynamic risk evolution trajectory, the multi-layer network model is used to simulate the propagation process of risks between the layers of the industrial chain, calculate the risk absorption coefficient of each node, and identify key hub nodes and risk buffer nodes. S700: Input the output results of the risk causal diagram, the dynamic risk evolution trajectory, and the multi-layer network model of industrial chain risk contagion into the interpretable artificial intelligence decision engine to generate an interpretable investment decision report containing risk identification basis, transmission path description, and investment advice; S800: Based on user interaction feedback, the decision parameters of the explainable artificial intelligence decision engine are dynamically adjusted to form a closed-loop decision optimization mechanism for human-machine collaboration.

[0006] Preferably, the construction of the "industrial element-geography-time" three-dimensional anchored knowledge graph (100) in step S200 specifically includes: S210. Perform entity recognition and relationship extraction on the collected multidimensional data to identify enterprise entities, material entities, technology entities, supply chain node entities, policy and regulatory entities, and geographical area entities in the new materials industry; S220. Construct geographic anchor points for each entity, and connect to the regional policy database, mineral resource distribution database and geopolitical risk assessment database through the geographic information system data interface to generate policy and regulatory vectors, resource endowment vectors and geopolitical risk indices for the regions where each entity is located. S230. Construct time anchors for each entity, store the state time series of each entity based on the time series database, and set relationship validity windows for each entity relationship. The relationship validity windows are used to define the effective interval of the relationship in the time dimension. S240. The geographic anchor point and the time anchor point are embedded as independent dimensions into the entity-relationship-attribute triple structure of the knowledge graph to form a three-dimensional anchoring graph.

[0007] Preferably, the counterfactual reasoning engine described in step S300 is constructed in the following manner: S310. A structural causal model based on a directed acyclic graph is adopted as the mathematical framework of the positive causal reasoning module to learn the causal structure between variables from the multidimensional data and generate a risk causal graph. S320. The counterfactual reasoning module is based on the Pearl causal hierarchy model and performs the intervention operation through Do-calculus to change the value of the target risk factor to a "not occurred" state. S330. A counterfactual sample space is generated using Monte Carlo simulation, and the structural causal model is driven to recalculate the counterfactual results on the counterfactual sample space. S340. Compare the counterfactual results with the original results, and calculate the causal effect strength of each risk factor, wherein the causal effect strength characterizes the difference in results caused by whether or not the risk factor occurs; S350. The causal chain prediction module performs pattern matching between the current knowledge graph state and the historical causal pattern library, identifies historical causal patterns similar to the current state, extrapolates subsequent causal chains in the historical patterns to the current state, and generates forward-looking risk transmission path prediction results.

[0008] Preferably, the sub-models of the digital twin in step S400 are specifically: The financial model is built based on the company's historical financial data, industry benchmark data, and macroeconomic data, and supports cash flow forecasting, balance sheet simulation, and profitability assessment. The supply chain model is built based on raw material supply data, logistics data, inventory data, and geopolitical risk data, and supports supply disruption simulation and alternative supply solution deduction. The technology evolution model is constructed based on patent data, R&D investment data, technology life cycle data, and technology substitution data, and supports prediction of technology roadmap evolution and assessment of technology obsolescence risk. The market response model is constructed based on downstream market demand data, competitive landscape data, and price elasticity data, and supports market demand fluctuation simulation and competitive situation deduction. The sub-models interact with each other through a standardized two-way data interface, enabling changes in the state of any sub-model to trigger a coordinated response from other sub-models.

[0009] Preferably, the multi-scenario stress test described in step S500 specifically includes: S510. Based on historical extreme event data and expert experience, generate a stress test scenario library that includes supply disruption scenarios, demand plunge scenarios, policy change scenarios, technology substitution scenarios, and compound shock scenarios. S520. Input each stress scenario into the digital twin, drive each sub-model to run the simulation, and generate the risk evolution trajectory of the project or enterprise under each scenario. S530, The risk evolution trajectory includes a risk probability change curve, a risk loss distribution curve, and a threshold breakthrough time sequence of key risk indicators in the time dimension. S540. Based on the simulation results, calculate the value at risk and value at risk for each scenario, and generate a dynamic risk heat map.

[0010] Preferably, step S600, which involves constructing a multi-layered network model for the contagion of supply chain risks and identifying key hub nodes and risk buffer nodes, specifically includes: S610. Based on the three-dimensional anchored knowledge graph, construct a multi-layer network topology structure for the industrial chain. The resource layer of the multi-layer network model includes mineral resource enterprise nodes, the processing layer includes raw material processing enterprise nodes, the material layer includes new material manufacturing enterprise nodes, and the application layer includes downstream application enterprise nodes. S620. An embedding learning is performed on each node in the multi-layer network model using a graph neural network, and the influence score of each node within and between layers is calculated by coupling the PageRank algorithm. S630. Simulate single-point failure scenarios and multi-point cascading scenarios using a cascading failure model, calculate the total downstream risk propagation caused by the failure of each node, and use it as an indicator of the risk destructive power of that node. S640. By comparing the difference in the total amount of risk propagation under the two network states of "including the node" and "removing the node", the risk absorption coefficient of each node is calculated. The risk absorption coefficient represents the node's buffering capacity against upstream risks. S650. Nodes with risk destructive power indicators higher than the first threshold and risk absorption coefficient higher than the second threshold are marked as critical hub nodes, and nodes with risk destructive power indicators lower than the third threshold and risk absorption coefficient higher than the fourth threshold are marked as risk buffer nodes. S660. In the investment decision report, a positive valuation premium is assigned to the investment target corresponding to the risk buffer node.

[0011] Preferably, the interpretable artificial intelligence decision engine described in step S700 specifically includes: S710. Based on the output of the aforementioned risk causal graph, dynamic risk evolution trajectory, and multi-layer network model of industrial chain risk contagion, a multi-dimensional risk assessment matrix is ​​constructed. The risk assessment matrix includes technological risk dimension, supply chain risk dimension, market risk dimension, policy risk dimension, and ESG risk dimension. S720. An attention mechanism is used to assign weights to each risk factor in the multi-dimensional risk assessment matrix to generate a comprehensive risk score. S730. Through the interpretability module, output the risk identification basis in the form of natural language generation. The risk identification basis includes the contribution of each risk factor, the causal path tracing explanation, and the comparative analysis with historical benchmarks. S740: Based on the comprehensive risk score and the user's preset risk preference parameters, an investment decision optimization scheme is generated through a multi-objective optimization algorithm, which includes the ranking of investment targets, investment ratio suggestions, and risk mitigation measures.

[0012] Preferably, the closed-loop decision optimization mechanism for human-machine collaboration described in step S800 specifically includes: S810. After outputting the interpretable investment decision report to the user, receive feedback information from the user on the decision suggestions. The feedback information includes confirmation or correction of risk factors, adjustment of risk tolerance, and preference selection of alternative solutions. S820. Based on the feedback information, the model parameters of the counterfactual reasoning engine, the digital twin, and the multi-layer network model of supply chain risk contagion are incrementally updated using an online learning algorithm. S830. Store user feedback and model update records in the feedback database to build a user decision preference profile; S840. Based on the user decision preference profile, the risk weight allocation and objective function are adaptively adjusted in subsequent decisions to achieve personalized evolution of decision-making strategies.

[0013] Preferably, the environmental and social governance data in step S100 includes carbon emission intensity data, energy consumption data, waste emission data, environmental compliance record data, social responsibility practice data, and corporate governance structure data; the method further includes: integrating the environmental and social governance data with the technology research and development data, supply chain data, market and policy data, and corporate operation data to construct an ESG risk sub-map, and using the counterfactual reasoning engine to identify the causal transmission path between ESG risk factors and financial risk factors, generating an ESG-financial risk linkage early warning signal.

[0014] A data-driven risk identification and decision optimization system for investment in the new materials industry includes: A multi-source heterogeneous data acquisition module is used to collect multi-dimensional data related to investment in the new materials industry in real time; The 3D anchored knowledge graph construction module is used to construct a 3D anchored knowledge graph of "industrial elements-geography-time" based on the collected multidimensional data; The counterfactual reasoning engine module is used to identify the causal path between risk factors based on the three-dimensional anchored knowledge graph through positive causal reasoning, counterfactual reasoning, and forward-looking causal chain prediction, and to generate a risk causal graph and forward-looking risk transmission path prediction results. A digital twin construction module is used to construct a digital twin of the project or enterprise to be evaluated based on the three-dimensional anchored knowledge graph and the risk causal graph. The multi-scenario stress testing module is used to perform multi-scenario stress tests on the digital twin and generate dynamic risk evolution trajectories. The supply chain risk contagion modeling module is used to construct a multi-layer network model of supply chain risk contagion that includes resource layer, processing layer, material layer and application layer. Based on the dynamic risk evolution trajectory, it calculates the risk absorption coefficient of each node and identifies key hub nodes and risk buffer nodes. An interpretable artificial intelligence decision engine module is used to receive the outputs of the risk causal graph, the dynamic risk evolution trajectory, and the multi-layer network model of industrial chain risk contagion, and generate an interpretable investment decision report; The human-computer collaborative interaction module is used to receive user feedback and dynamically adjust decision parameters to form a closed-loop decision optimization mechanism. The modules interact with each other through standardized data interfaces, and a feedback loop for bidirectional data flow is formed between the three-dimensional anchored knowledge graph construction module, the counterfactual reasoning engine module, the digital twin construction module, and the industrial chain risk contagion modeling module.

[0015] The beneficial effects of this invention are as follows: 1. In this invention, by constructing a counterfactual reasoning engine that includes positive causal reasoning, counterfactual reasoning, and causal chain prediction, the transformation from passive response to active prediction is realized. The counterfactual reasoning module sets the risk factors to a non-occurrence state and calculates the intensity of causal effects. The causal chain prediction module generates transmission path predictions before the occurrence of risk events based on the current graph state and historical pattern library, enabling investors to receive early warnings at the nascent stage of risk and significantly improving the foresight of decision-making. 2. In this invention, a three-dimensional anchored knowledge graph of industry elements, geography, and time is constructed. Each industry entity is bound to a geographical anchor and a time anchor. The geographical anchor is connected to the regional policy and resource database to generate policy and regulatory vectors and geopolitical risk indices. The time anchor stores the state time series and sets the relationship time window, so as to realize the accurate tracking of supply chain risks by regional policy changes and solve the technical problem of traditional graphs lacking spatiotemporal coupling capabilities. 3. In this invention, a digital twin containing financial, supply chain, technology evolution and market response models is constructed and multi-scenario stress tests are performed to simulate the risk evolution under extreme scenarios such as supply disruption, policy changes and technology substitution. Each sub-model achieves linkage response through bidirectional data interface, providing a scientific simulation environment that can be repeatedly verified for investment decisions in the new materials industry. 4. In this invention, a risk absorption coefficient is introduced into the multi-layer network model of the industrial chain. The influence of nodes is calculated by embedding learning of graph neural networks and coupling PageRank. The risk destructive power index is calculated by using a cascade failure model. The risk absorption coefficient is calculated by comparing the difference in the total propagation in the two states of including and removing the node. Nodes with low destructive power and high absorption are marked as risk buffer nodes and given a positive valuation premium, thus filling the technical gap in risk absorption node identification and capitalization. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall methodology of a data-driven risk identification and decision optimization method for investment in the new materials industry proposed in this invention. Figure 2 This is a schematic diagram of the three-dimensional anchored knowledge graph structure in this invention; Figure 3 This is a diagram of the data acquisition and fusion system module architecture in this invention; Figure 4 This is a schematic diagram illustrating the working principle of the counterfactual reasoning engine in this invention; Figure 5 This is a diagram illustrating the structure and interaction relationships of the digital twin model in this invention; Figure 6 This is a flowchart of the multi-scenario stress testing and risk evolution trajectory generation process in this invention; Figure 7 This is a diagram of the multi-layered network topology for supply chain risk contagion in this invention. Figure 8 This is a flowchart of the risk absorption coefficient calculation and node identification process in this invention; Figure 9 This is a flowchart illustrating the workflow of the interpretable artificial intelligence decision engine in this invention; Figure 10 This is a structural diagram of the human-machine collaborative closed-loop decision optimization mechanism in this invention; Figure 11 This is a diagram illustrating the overall architecture of a data-driven risk identification and decision optimization system for investment in the new materials industry, as proposed in this invention.

[0017] The diagram shows: 1. Multi-source heterogeneous data acquisition module; 2. 3D anchored knowledge graph construction module; 3. Counterfactual reasoning engine module; 4. Digital twin construction module; 5. Stress testing module; 6. Supply chain risk contagion modeling module; 7. Explainable decision engine module; 8. Human-machine collaborative interaction module. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Reference Figure 1 - Figure 11 A data-driven risk identification and decision optimization method for investment in the new materials industry includes the following steps: In step S100, multi-dimensional data related to investment in the new materials industry are collected in real time through the multi-source heterogeneous data acquisition module 1. The multi-dimensional data includes technology research and development data, supply chain data, market and policy data, enterprise operation data, and environmental and social governance data.

[0020] The multi-source heterogeneous data acquisition module 1 is equipped with a distributed data crawling engine and a real-time data stream processing framework. This distributed data crawling engine simultaneously connects to global patent databases, academic paper databases, industry research report platforms, raw material futures trading platforms, customs import and export statistics systems, national and local policy release platforms, enterprise credit information disclosure systems, and listed company financial disclosure platforms. It crawls structured, semi-structured, and unstructured data in real time and processes them uniformly. Technology R&D data covers patent application and authorization information, paper publication and citation information, R&D project initiation and progress information, and technology roadmap evolution information in the field of new materials. Supply chain data covers global production capacity and output information, price fluctuation information, inventory level information, logistics and transportation information, and the impact of geopolitical events on the supply pattern of key raw materials. Market and policy data covers the demand scale and growth trend information of downstream application industries, industry competitive landscape information, national and local industrial support policy information, environmental regulations and trade control policy information, and enterprise operation data... The system covers the target company's financial statements, credit ratings, production yield, customer concentration, and management performance information. Environmental and social governance data includes carbon emission intensity, energy consumption intensity, waste discharge and disposal, environmental compliance records, social responsibility practices, and corporate governance structure. The multi-source heterogeneous data acquisition module 1 also includes a data quality assessment submodule and an outlier detection submodule. The data quality assessment submodule assigns a credibility weight to each data source and performs weighted fusion based on this weight in subsequent processing. The outlier detection submodule identifies and marks abnormal data for manual review based on a dual mechanism of statistical outlier detection algorithm and machine learning isolated forest algorithm. All collected raw data undergoes standardization processing such as format unification, field mapping, time alignment, and entity disambiguation, and is then simultaneously output to the 3D anchored knowledge graph construction module 2 and the ESG risk subgraph construction module 600 in both real-time and batch streams. The real-time stream supports dynamic risk warning, while the batch stream supports periodic model retraining.

[0021] In this embodiment, in step S200, a three-dimensional anchored knowledge graph 100 of "industrial elements-geography-time" is constructed based on the collected multidimensional data. Each industrial entity in the knowledge graph 100 is simultaneously bound to a geographical anchor 110 and a time anchor 120. The geographical anchor 110 includes the policy and regulatory vector 111, resource endowment vector 112, and geopolitical risk index 113 of the region where the entity is located. The time anchor 120 includes the state time series 121 and relationship time window 122 of the entity.

[0022] The 3D Anchored Knowledge Graph 100 uses an attribute graph model as its underlying data storage structure and runs on a distributed graph database. In this attribute graph model, each vertex represents an industry entity, each edge represents the relationship between entities, and each vertex and edge is associated with a set of attributes to store multi-dimensional feature information. Industry entities include six core entity types: enterprise entities, material entities, technology entities, supply chain node entities, policy and regulatory entities, and geographical region entities. Enterprise entities are further divided into four subtypes: mineral resource enterprises, raw material processing enterprises, new material manufacturing enterprises, and downstream application enterprises. Material entities are further divided into basic raw material enterprises. The system comprises three subtypes: physical entities, intermediate material entities, and end-material entities. Technical entities are further categorized into four subtypes: preparation technology entities, modification technology entities, characterization technology entities, and application technology entities. Geographic anchors 110 are embedded as a composite attribute structure within the attribute set of each industrial entity. This composite attribute structure includes three dimensions: policy and regulatory vector 111, resource endowment vector 112, and geopolitical risk index 113. The policy and regulatory vector 111 is accessed in real-time through the geographic information system data interface to the target region's industrial support policy database, environmental regulations database, and import / export control policy database, and is vectorized. The resource endowment vector 112 is accessed through the geographic information system data interface. The system accesses a global mineral resource distribution database and a regional energy structure database via an interface, extracting key raw material reserves, grades, mining costs, and energy prices for the region. The geopolitical risk index 113 is calculated by weighting four sub-indicators: political stability, trade barrier levels, infrastructure reliability, and legal environment completeness. The time anchor 120 is embedded as another composite attribute structure within the attribute set of each industry entity. This composite attribute structure includes two dimensions: state time series 121 and relationship timeliness window 122. The state time series 121 is based on a time-series database. The system stores the status changes of each entity on the historical timeline and supports backtracking queries and trend analysis at any time granularity. The relationship validity window 122 assigns a start timestamp and an end timestamp to each relationship to define the effective interval of the relationship in the time dimension. When the current time exceeds the effective interval, the relationship automatically becomes invalid or is marked as pending review. The knowledge graph 100 also has an entity-relationship-attribute triplet index structure and a spatiotemporal joint query interface. This spatiotemporal joint query interface supports simultaneous specification of industry element conditions, geographical region conditions, and time interval conditions for composite retrieval, thereby quickly locating the set of industry entities and their relationship network that meet specific spatiotemporal conditions.

[0023] In this embodiment, in step S300, a counterfactual reasoning engine 200 is constructed based on the three-dimensional anchored knowledge graph 100. The counterfactual reasoning engine 200 includes a forward causal reasoning module 210, a counterfactual reasoning module 220, and a causal chain prediction module 230. The forward causal reasoning module 210 identifies the causal structure between risk factors from multidimensional data through a causal discovery algorithm and generates a risk causal graph 211. The counterfactual reasoning module 220 sets the value of a specific risk factor to a "not occurred" state through an intervention operation 221 and drives the structural causal model to recalculate the counterfactual results 222. Based on the counterfactual results 222, the causal effect strength 223 of each risk factor is calculated. The causal chain prediction module 230 predicts the risk transmission path before the risk event actually occurs, based on the current knowledge graph state 231 and the historical causal pattern library 232, through pattern matching and extrapolation 233.

[0024] The counterfactual reasoning engine 200 is deployed and runs on a high-performance computing cluster using a microservice architecture. The forward causal reasoning module 210, the counterfactual reasoning module 220, and the causal chain prediction module 230 each run as independent microservices and communicate and exchange data asynchronously via message queues. The forward causal reasoning module 210 employs a dual causal discovery strategy combining a constraint-based PC algorithm and a score-based GES algorithm. First, it performs skeleton graph learning on variables in multidimensional data through conditional independence testing. Then, it uses a greedy search to determine the direction of the skeleton graph to generate a complete directed acyclic graph structure. In the acyclic graph structure, each node represents a risk factor variable, and each directed edge represents a direct causal relationship between variables. The forward causal reasoning module 210 also includes a bootstrap sampling unit and a stability selection unit. The bootstrap sampling unit evaluates the confidence of causal relationships by resampling the original data multiple times and repeating the causal discovery process. The stability selection unit only retains causal relationship edges that stably appear in the resampling results above a preset proportion. Each causal relationship edge in the final output risk causal graph 211 is accompanied by a confidence score and a causal direction certainty score. The counterfactual reasoning module 220 receives the risk causal graph. 211 is used as input and inference calculations are performed based on the third level of the Pearl causal hierarchy model, namely the counterfactual level. The execution flow of the counterfactual inference module 220 includes three sub-stages: the first sub-stage is the intervention operation 221, which isolates the target risk factor variable from the set of perturbed variables in the structural causal model based on the Do-calculus rule and forces its value to be set to the "not occurred" state or the baseline state value; the second sub-stage is to perform Monte Carlo simulation based on the intervened structural causal model to generate a large-scale counterfactual sample space. This Monte Carlo simulation is performed by extracting the counterfactual sample space from the structural causal model. The exogenous variable distribution is randomly sampled multiple times and propagated forward along the causal directed edge to calculate and generate the endogenous variable value under each sample, thereby constructing the probability distribution of the counterfactual result 222. The third sub-stage is to compare the difference between the probability distribution of the counterfactual result 222 and the probability distribution of the original result and calculate the causal effect strength 223 of each risk factor. The causal effect strength 223 is output in two forms: average causal effect and quantile causal effect. The average causal effect reflects the mean difference of the result caused by the occurrence or non-occurrence of the risk factor, and the quantile causal effect reflects the differentiated impact of the risk factor at different risk quantiles.The causal chain prediction module 230 runs independently after the counterfactual reasoning module 220. This module 230 has a built-in historical causal pattern library 232, which is formed offline by annotating and vectorizing causal chains of risk event cases in the new materials industry over the past five years. The causal chain prediction module 230 first extracts the current knowledge graph state 231 from the current three-dimensional anchored knowledge graph 100 and encodes this state as a graph embedding vector. Then, it uses a similarity retrieval algorithm to match the K most similar historical causal patterns in the historical causal pattern library 232 to this graph embedding vector. Finally, it uses a pattern extrapolation mechanism to extrapolate the subsequent actual causal chains from the K matched historical patterns to the current state using a similarity-weighted fusion method, thereby generating a forward-looking risk transmission path prediction result. This prediction result is output in the form of a causal chain sequence, with each predicted causal relationship edge accompanied by a prediction confidence score and an expected occurrence time window.

[0025] In this embodiment, in step S400, a digital twin 400 of the new material project or enterprise to be evaluated is constructed based on the three-dimensional anchored knowledge graph 100 and the risk causal graph 211. The digital twin 400 includes a financial model 410, a supply chain model 420, a technology evolution model 430, and a market response model 440. Data interaction between the sub-models is achieved through a bidirectional data interface 450.

[0026] The digital twin 400 is constructed and deployed on a cloud simulation platform using an agent-based modeling framework and system dynamics methods. The construction process of the digital twin 400 first extracts all attribute information and related entity information of the target project or enterprise from the 3D anchored knowledge graph 100, and extracts all risk factors and causal path information involved in the target project or enterprise from the risk causal graph 211. Then, the extracted information is injected into the initialization parameter sets of four sub-models: financial model 410, supply chain model 420, technology evolution model 430, and market response model 440. The financial model 410 uses historical financial data of the enterprise, industry benchmark data, and macroeconomic data... Based on economic data, a complete financial simulation framework is constructed, including profit and loss statement forecasting, balance sheet forecasting, and cash flow statement forecasting units. This framework uses Monte Carlo stochastic simulation to predict the probability distribution of multiple future financial indicators for enterprises and supports outputting forecast results at three time granularities: monthly, quarterly, and annual. Financial Model 410 also includes a built-in financial ratio analysis unit and a financial vulnerability assessment unit. The financial ratio analysis unit automatically calculates key financial ratios such as the current ratio, quick ratio, debt-to-equity ratio, interest coverage ratio, return on equity, and total asset turnover. The financial vulnerability assessment unit is based on Altman... The Z-score model and the Ohlson O-score model are used to assess the probability of corporate bankruptcy. The supply chain model 420 is based on raw material supply data, logistics data, inventory data and geopolitical risk index 113 to build a supply chain simulation framework including supply network topology unit, inventory dynamic management unit and logistics route planning unit. The supply network topology unit builds a multi-level supplier network structure based on the upstream and downstream relationships of the supply chain in knowledge graph 100 and supports the simulation of the cascading impact of the interruption event of any supplier node on the entire network. The inventory dynamic management unit simulates the optimal inventory strategy of enterprises under different supply conditions and its corresponding capital occupation cost based on economic order quantity model and safety stock model. The logistics route planning unit simulates the timeliness and cost distribution of different logistics routes based on global shipping data, port congestion data and land transportation data. The supply chain model 420 also supports the alternative supply scheme simulation function, that is, when the main supply route is interrupted, it automatically evaluates the feasibility and cost increment of alternative suppliers and alternative logistics routes.The technology evolution model 430 constructs a technology simulation framework based on patent data, R&D investment data, technology life cycle data, and technology substitution data. This framework includes a patent valuation unit, a technology roadmap prediction unit, and a technology obsolescence risk warning unit. The patent valuation unit quantifies the technological and market value of the patents held by the target company using patent citation network analysis and patent claim width analysis. The technology roadmap prediction unit simulates the maturity evolution and performance improvement trajectory of the target technology roadmap at different future time points based on technology life cycle curves and technology substitution curves. The technology obsolescence risk warning unit monitors the patent growth rate and paper growth rate of emerging technologies to determine the existence of technology substitution threats and issues a warning signal when the substitution probability exceeds a preset threshold. Market response model 4... Based on downstream market demand data, competitive landscape data, and price elasticity data, a market simulation framework is constructed, comprising a demand forecasting unit, a market share simulation unit, and a price game unit. The demand forecasting unit uses time series analysis and causal regression methods to predict future demand in various downstream application areas and provides forecast intervals. The market share simulation unit uses a Markov chain model to simulate the dynamic changes in the target company's market share in its competition with rivals. The price game unit uses the Bertrand competition model and a differentiated product pricing model to simulate the company's optimal pricing strategy under different market structures and its impact on profits. The sub-models achieve real-time data synchronization and event-triggered communication through a standardized bidirectional data interface 450, which is based on RESTful APIs. The API and message queue mechanisms enable both synchronous calls and asynchronous notifications. When the state of any sub-model changes significantly, this change is broadcast to all other sub-models via bidirectional data interface 450, triggering corresponding response calculations. For example, when the supply chain model 420 detects a sharp increase in raw material prices, it simultaneously transmits the price change information to the financial model 410 via bidirectional data interface 450, triggering cash flow revaluation, and to the market response model 440, triggering product pricing strategy adjustments. This achieves collaborative evolution and overall linkage among the sub-models of the digital twin 400.

[0027] In this embodiment, step S500 involves performing multi-scenario stress tests on the digital twin 400 to generate a dynamic risk evolution trajectory.

[0028] The multi-scenario stress testing module 5 establishes a connection with the digital twin 400 through a standardized API interface and performs large-scale scenario simulation calculations in batch processing. The multi-scenario stress testing module 5 first calls the scenario generator to generate a multi-dimensional stress test scenario library based on a historical extreme event database and an expert knowledge base. This stress test scenario library includes at least the following scenarios: supply disruption scenario (setting a sudden drop in the supply of key raw materials by 10% to 50%), demand plunge scenario (setting a sudden drop in demand in major downstream application areas by 15% to 40%), and policy change scenario (setting a sudden introduction of restrictive industrial policies in the target region). The scenarios include sudden cancellation of supportive subsidies, technology substitution scenarios (where competitive alternative materials suddenly achieve performance breakthroughs or significant cost reductions), and combined impact scenarios (where two or more of the above scenarios occur simultaneously). Each stress test scenario is configured with impact intensity parameters, impact start time parameters, and impact duration parameters to cover different levels of extreme situations. The multi-scenario stress test module 5 inputs all the above stress test scenarios one by one into the digital twin 400 and drives the financial model 410, supply chain model 420, technology evolution model 430, and market response model 440 to perform collaborative simulations under each scenario. Each scenario simulation is repeated a preset number of times to eliminate randomness and ensure statistical stability. During the simulation, each sub-model progresses step by step according to a preset time step, recording the state values ​​of key risk indicators at each step. After the simulation is completed, the multi-scenario stress test module 5 automatically summarizes the key risk indicator value sequences for each time step under each scenario and generates a dynamic risk evolution trajectory. This dynamic risk evolution trajectory includes the risk probability change curve (i.e., the dynamic change of the probability of occurrence of each risk indicator on the time axis), the risk loss distribution curve (i.e., the probability distribution of the loss amount of each risk indicator at each time point), and the threshold surge of key risk indicators. Breaking the time sequence refers to the set of time points at which each risk indicator first breaks through the preset warning threshold and crisis threshold; the multi-scenario stress test module 5 also calculates the risk value under each scenario, i.e., the value at risk and the conditional risk value, i.e., the expected tail loss, based on the above simulation results, and presents the risk intensity of different scenarios, different time points and different risk indicators on a two-dimensional or three-dimensional heat map through heat map visualization. This dynamic risk heat map uses color depth to represent the magnitude of risk intensity, time on the horizontal axis to represent the evolution process, and scenarios on the vertical axis to represent different stress conditions, thereby intuitively displaying the overall risk profile of the target project or enterprise in the multi-dimensional stress space.

[0029] In this embodiment, step S600 constructs a multi-layer network model 300 for the contagion of supply chain risks. The multi-layer network model 300 includes a resource layer 310, a processing layer 320, a material layer 330, and an application layer 340, with the layers connected by directed weighted edges 350. Based on the dynamic risk evolution trajectory, the multi-layer network model 300 simulates the propagation process of risks between the layers of the supply chain, calculates the risk absorption coefficient 361 of each node 360, and identifies key hub nodes 390 and risk buffer nodes 391.

[0030] The multi-layered network model 300 for supply chain risk contagion is constructed based on enterprise entity nodes and supply chain relationships in the three-dimensional anchored knowledge graph 100. This multi-layered network model 300 adopts a heterogeneous network architecture with inter-layer coupling. The resource layer 310 includes global mineral resource enterprise nodes and edges representing competition and cooperation relationships among enterprises within this layer. The processing layer 320 includes processing enterprise nodes that convert mineral resources into industrial-grade raw materials and edges representing capacity sharing and order allocation relationships among enterprises within this layer. The materials layer 330 includes manufacturing enterprise nodes that further process industrial-grade raw materials into new materials and edges representing technology licensing and joint R&D relationships among enterprises within this layer. The application layer 340 includes downstream manufacturing processes that apply new materials to end products. The system constructs enterprise nodes and establishes customer sharing and market segmentation relationships among enterprises within each layer. Directed weighted edges 350 between adjacent layers are built based on real transaction data and supply and demand contract data. The weight of each directed weighted edge 350 is calculated comprehensively based on the proportion of transaction amount and the degree of supply dependence. No direct connections are established between non-adjacent layers to ensure that risk propagation must be transmitted layer by layer to simulate the hierarchical transmission characteristics in a real industrial chain. After the multi-layer network model 300 is constructed, the industrial chain risk contagion modeling module 6 loads the time series of risk states of each node in the dynamic risk evolution trajectory as the initial risk injection amount. Then, a graph neural network is used to perform structural embedding learning on each node 360 ​​in the multi-layer network model 300 to obtain the topological features of each node. The system uses vectors and attribute feature vectors, and then iteratively calculates the influence score of each node both within the same layer and across layers using a coupled PageRank algorithm. This coupled PageRank algorithm considers the degree to which a node is referenced by other nodes within its own layer and the degree to which it is referenced by nodes in adjacent layers in each iteration, updating the node's global influence score accordingly. Based on this, the supply chain risk contagion modeling module 6 simulates single-point failure scenario 370 (setting only one node to a completely failed state each time and observing the propagation range and depth of this failure in the multi-layer network) and multi-point cascading scenario 380 (simultaneously setting multiple nodes to a failed state in different combinations and observing the synergistic amplification effect between multiple failures). In each scenario simulation, the system records the total downstream risk propagation 381 caused by each node as a failure source and uses this total as an indicator of the node's risk destructive power. The supply chain risk contagion modeling module 6 also calculates the risk absorption coefficient 361 of the node by comparing the difference in the total downstream risk propagation in the two scenarios: the network state containing the specific node and the network state after removing the node. Specifically, the risk absorption coefficient 361 is equal to the total upstream risk propagating downstream when the node is included minus the total upstream risk propagating downstream after the node is removed, divided by the total propagation when the node is included. The value of the risk absorption coefficient 361 ranges from zero to one. The larger the value, the stronger the node's buffering and absorption capacity for upstream risks.The supply chain risk contagion modeling module 6 finally executes a dual-threshold classification decision: Nodes with a risk destructive power index higher than the first preset threshold and a risk absorption coefficient 361 higher than the second preset threshold are identified and marked as key hub nodes 390. These nodes play a dual role as risk amplifiers and risk buffers in the supply chain, thus being both key tools for risk prevention and control and key targets for investment due diligence. Nodes with a risk destructive power index lower than the third preset threshold and a risk absorption coefficient 361 higher than the fourth preset threshold are identified and marked as risk buffer nodes 391. These nodes play a natural risk barrier role in the supply chain, and their corresponding investment targets should receive a positive valuation premium in investment decisions. The supply chain risk contagion modeling module 6 outputs the classification labels of all nodes, along with the risk destructive power index and risk absorption coefficient 361 of each node, to the interpretable artificial intelligence decision engine module 7. Simultaneously, the list and topological location information of key hub nodes 390 and risk buffer nodes 391 are presented visually on a multi-layer network topology diagram for user reference.

[0031] In this embodiment, in step S700, the output results of the risk causal graph 211, the dynamic risk evolution trajectory, and the multi-layer network model of industrial chain risk contagion are input into the interpretable artificial intelligence decision engine 500 to generate an interpretable investment decision report containing risk identification basis, transmission path description, and investment advice.

[0032] The explainable AI decision engine 500 adopts a hierarchical processing architecture and is deployed on a server equipped with a graphics processing unit accelerator card. This engine first performs unified data alignment and normalization on all causal relationship edges and their confidence scores extracted from the risk causal graph 211, time-series data of risk indicators under various scenarios extracted from the dynamic risk evolution trajectory, and risk destructive power indicators and risk absorption coefficients 361 of each node output by the multi-layer network model of industrial chain risk contagion 300. The normalization process employs both minimum-maximum standardization and quantile transformation methods in parallel to ensure that data from different sources and with different dimensions can be uniformly incorporated into the evaluation framework. All processed data flows into the construction unit of the multi-dimensional risk assessment matrix 510. This construction unit classifies and groups all risk factors according to five dimensions: technology risk, supply chain risk, market risk, policy risk, and ESG risk. Each dimension contains a set of sub-risk factors specific to that dimension. Each sub-risk factor is accompanied by a set of root cause nodes and a causal path chain traced back from the risk causal graph 211, thus ensuring that every risk assessment result can be traced back to the original data evidence. After the multi-dimensional risk assessment matrix 510 is constructed, it can explain the AI ​​decision engine 500 calling the attention mechanism calculation module, which is based on multi-head self-attention. The multi-head self-attention network architecture assigns weights to each risk factor in the multi-dimensional risk assessment matrix 510. This network captures the interaction between risk factors by calculating the pairwise correlations of each risk factor with all other risk factors. Then, it converts the correlation scores into attention weight scores using a trainable parameter matrix. Finally, it weights and sums the results of each attention head to obtain the comprehensive attention weight for each risk factor and generates the overall risk score. The interpretability module 520 starts running after the risk score is generated. This module includes a template-based natural language generation engine and a feature importance analysis-based interpretation unit. The interpretation unit of the importance analysis uses the SHAP value analysis method to calculate the marginal contribution value of each risk factor to the final comprehensive risk score and sorts the contribution values ​​from largest to smallest in order to filter out the top-ranked key contribution factors. The template-based natural language generation engine automatically fills the name of the key contribution factor, the contribution value, the description of the upstream and downstream causal relationship of the factor in the risk causal diagram 211, the description of the historical change trend of the factor in the dynamic risk evolution trajectory, and the list of related nodes involved in the factor in the multi-layer network model 300 of the industrial chain risk contagion into the corresponding positions of the template according to the preset text template, thereby generating readable text containing risk identification basis and causal path tracing explanation;The interpretable AI decision engine 500 ultimately calls a multi-objective optimization solver. This solver uses minimizing the overall risk score and maximizing the expected return on investment as two parallel optimization objectives. It uses user-preset risk preference parameters (via the human-computer interaction module 8) as constraints and the investment proportions of each asset in the portfolio as decision variables. It employs the decomposition-based multi-objective evolutionary algorithm MOEA / D to search for the optimal portfolio solution set on the Pareto optimal frontier. Based on user-preset preferences, it selects the optimal compromise solution from the Pareto frontier. This optimal compromise solution includes a ranking list of investment assets, suggested investment proportions for each asset, expected risk-adjusted returns, and risk mitigation measures proposed for each identified high-risk factor. All of the above information is formatted and output as a complete interpretable investment decision report.

[0033] In this embodiment, in step S800, the decision parameters of the explainable artificial intelligence decision engine 500 are dynamically adjusted based on user interaction feedback to form a closed-loop decision optimization mechanism for human-machine collaboration.

[0034] The human-machine collaborative interaction module 8 is presented on the user terminal device in the form of a visual dashboard. This dashboard simultaneously displays the core conclusions of the investment decision report, an interactive and scalable view of the risk causal graph 211, a curve of the dynamic risk evolution trajectory, and the topology diagram of the multi-layer network model 300 for risk contagion in the industrial chain. Users can view detailed information of any risk factor, the complete chain of any causal path, and the risk status of any node in the industrial chain through interactive operations such as clicking, dragging, and zooming. While displaying the decision report, the human-machine collaborative interaction module 8 provides a feedback input interface to the user. This interface supports user confirmation or correction of each risk factor; users can click to confirm the existence of a risk factor identified by the system or modify the system's severity rating for that risk factor. It also supports user adjustment of risk tolerance; users can adjust the upper limit of tolerance for each of the five dimensions—technology risk, supply chain risk, market risk, policy risk, and ESG risk—using a slider. Finally, it supports user preference selection of alternative solutions; users can choose a preferred solution from the top Pareto optimal solutions generated by multi-objective optimization and explain their reasons for preference. The human-machine collaborative interaction module 8 will... All collected user feedback information is structured and encoded before being transmitted to the online learning engine and the feedback database. The online learning engine employs two parallel mechanisms: an online gradient descent algorithm and a Bayesian incremental update algorithm. The online gradient descent algorithm incrementally fine-tunes the causal effect strength parameters in the counterfactual reasoning engine 200 and the attention weight parameters in the interpretable artificial intelligence decision engine 500. The Bayesian incremental update algorithm performs posterior updates on the prior distribution parameters of each sub-model of the digital twin 400 and the graph neural network weight parameters of the multi-layer network model 300 for supply chain risk contagion. The update results of the two algorithms are cross-validated and then fused into the final parameter update values. The feedback database synchronously receives user feedback information and model parameter update records and archives them using the user ID as the primary key. The feedback database also runs a user behavior clustering analysis module and a preference evolution tracking module. The user behavior clustering analysis module clusters users' historical feedback patterns based on the K-means algorithm to identify users' decision-making style types. The preference evolution tracking module tracks the temporal change trends of users' tolerance parameters for each risk dimension based on time series analysis to determine the drift direction of user preferences.Based on the aforementioned user decision-making preference profile, the human-computer collaborative interaction module 8 automatically executes an adaptive adjustment strategy in subsequent decisions. This adaptive adjustment strategy includes adjusting the initial weight allocation of each risk dimension in the multi-dimensional risk assessment matrix 510 according to the risk tolerance distribution in the user preference profile; adjusting the preference-oriented parameters in the multi-objective optimization algorithm, i.e., the preference weight vector in the objective function, according to the historical scheme selection tendency in the user preference profile; and adjusting the learning rate parameter of the online learning algorithm according to the feedback correction frequency in the user preference profile to avoid over-adaptation to noisy feedback. Through the above adaptive adjustment strategy, the system can achieve a certain degree of personalized evolution after each decision cycle, thus forming a complete closed-loop decision optimization mechanism from decision report output to user feedback reception, model parameter update, and strategy adaptive adjustment.

[0035] In this embodiment, the environmental and social governance data in step S100 includes carbon emission intensity data, energy consumption data, waste emission data, environmental compliance record data, social responsibility practice data, and corporate governance structure data. The method also includes integrating environmental and social governance data with technology research and development data, supply chain data, market and policy data, and corporate operation data to construct an ESG risk sub-map 600, and using a counterfactual reasoning engine 200 to identify the causal transmission path between ESG risk factors and financial risk factors, generating an ESG-financial risk linkage early warning signal.

[0036] The ESG Risk Subgraph 600 is independent of the 3D Anchored Knowledge Graph 100 but shares the same underlying graph database storage engine and query interface. This ESG Risk Subgraph 600 uses enterprise entities as core nodes and constructs a unique set of ESG attributes around each enterprise entity. Carbon emission intensity data includes three sub-items: direct carbon emissions (Scope 1 emissions), indirect carbon emissions (Scope 2 emissions), and supply chain carbon emissions (Scope 3 emissions), stored as a time series on an annual basis. Energy consumption data includes four sub-items: electricity consumption, heat consumption, fossil fuel consumption, and the proportion of renewable energy use, stored as a time series on a monthly basis. Waste emissions... The data includes four sub-items: hazardous waste emissions, harmless waste emissions, waste recycling rate, and waste treatment compliance rate. Environmental compliance record data includes four sub-items: environmental administrative penalty records, environmental litigation records, environmental permit holdings, and environmental management system certification status. Social responsibility practice data includes four sub-items: employee safety incident rate, community relations evaluation, product safety liability incidents, and public welfare investment amount. Corporate governance structure data includes four sub-items: board independence indicators, executive compensation reasonableness indicators, audit transparency indicators, and shareholder rights protection indicators. The ESG risk sub-map 600 also constructs the internal relationships between ESG risk factors. Correlationships, such as the positive correlation between high carbon emission intensity and high energy consumption, the causal relationship between environmental administrative penalty records and corporate governance structure deficiencies, and the temporal correlation between the lack of social responsibility practices and the increase in employee safety incident rates; after the ESG risk sub-map 600 is constructed, the ESG risk factor data is synchronously input into the counterfactual reasoning engine 200 through a standardized data interface. The positive causal reasoning module 210 in the counterfactual reasoning engine 200 combines the ESG risk factor data with financial risk factor data, such as corporate profit margin fluctuations, cash flow tension, and changes in debt-to-equity ratio, to perform joint causal discovery, thereby identifying the relationship between ESG risk factors and financial risks. Cross-domain causal transmission paths between risk factors, such as the causal path of tightening carbon emission policies leading to increased corporate environmental expenditures and thus compressing profit margins, the causal path of corporate governance deficiencies leading to decreased financial reporting quality and thus increasing financing costs, and the causal path of supply chain environmental violations leading to the loss of core customers and thus decreasing revenue; the counterfactual reasoning module 220 in the counterfactual reasoning engine 200 further calculates the causal effect strength 223 of each ESG risk factor in the above cross-domain causal transmission paths, that is, quantifies the specific difference value of whether or not the occurrence of the ESG risk factor affects the financial performance indicators, and the causal effect strength 223 is output in the form of monetary amount or percentage;The ESG risk sub-map 600 includes an early warning rule engine. This engine contains multiple sets of combined early warning rules based on the values ​​of ESG risk factors and their causal effect strength 223. When the current measured value of an ESG risk factor exceeds a preset threshold and its causal effect strength 223 exceeds a preset effect threshold, the early warning rule engine automatically triggers an ESG-financial risk linkage early warning signal. This signal is then pushed as a high-priority message to the explainable artificial intelligence decision engine 500 and the human-machine collaborative interaction module 8. The linkage early warning signal includes the name of the ESG risk factor that triggered the warning, the deviation of the factor's current measured value from its historical benchmark value, the expected financial impact amount derived from causal inference, and a list of recommended mitigation measures.

[0037] A data-driven risk identification and decision optimization system for investment in the new materials industry includes: a multi-source heterogeneous data acquisition module 1 for real-time acquisition of multi-dimensional data related to investment in the new materials industry; a three-dimensional anchored knowledge graph construction module 2 for constructing a three-dimensional anchored knowledge graph 100 of "industry elements-geography-time" based on the acquired multi-dimensional data; a counterfactual reasoning engine module 3 for identifying causal paths between risk factors based on the three-dimensional anchored knowledge graph 100 through forward causal reasoning, counterfactual reasoning, and forward-looking causal chain prediction, generating a risk causal graph 211 and forward-looking risk transmission path prediction results; a digital twin construction module 4 for constructing a digital twin 400 of the project or enterprise to be evaluated based on the three-dimensional anchored knowledge graph 100 and the risk causal graph 211; and a multi-scenario stress testing module 5 for performing multi-scenario stress tests on the digital twin 400 to generate a dynamic risk evolution trajectory; and a supply chain risk transmission module. The modeling module 6 is used to construct a multi-layered network model 300 for the contagion of risks in the industrial chain, which includes a resource layer 310, a processing layer 320, a material layer 330, and an application layer 340. Based on the dynamic risk evolution trajectory, it calculates the risk absorption coefficient 361 of each node 360 ​​and identifies key hub nodes 390 and risk buffer nodes 391. The interpretable artificial intelligence decision engine module 7 is used to receive the output of the risk causal graph 211, the dynamic risk evolution trajectory, and the multi-layered network model 300 for the contagion of risks in the industrial chain, and generate an interpretable investment decision report. The human-machine collaborative interaction module 8 is used to receive user feedback and dynamically adjust decision parameters to form a closed-loop decision optimization mechanism. Among them, the modules interact with each other through standardized data interfaces, and a feedback loop of two-way data flow is formed between the three-dimensional anchored knowledge graph construction module 2, the counterfactual reasoning engine module 3, the digital twin construction module 4, and the industrial chain risk contagion modeling module 6.

[0038] The system is deployed using a microservice architecture. Each of the eight core modules runs as an independent containerized service on a cloud server cluster. Dynamic addressing and load balancing between modules are achieved through service registration and discovery mechanisms. Standardized data interfaces between modules are defined based on the OpenAPI specification and uniformly use JSON format for data serialization and deserialization. Interface calls support both synchronous request-response and asynchronous event subscription-publishing modes to accommodate different types of inter-module interaction needs. The multi-source heterogeneous data acquisition module 1 serves as the data entry point for the entire system, continuously feeding data into the 3D anchored knowledge graph construction module 2. The ESG risk sub-graph 600 pushes cleaned and standardized multi-dimensional data. Upon receiving new data, the 3D anchored knowledge graph construction module 2 triggers an incremental graph update process. This dynamically merges new entities, relationships, and attributes into the existing graph without disrupting its structure, and simultaneously updates the time-series records of geographic anchor 110 and time anchor 120. After the graph update is complete, the 3D anchored knowledge graph construction module 2 broadcasts the graph change notification to the counterfactual reasoning engine module 3 and the digital twin construction module 4 via an event publishing mechanism. Upon receiving the graph change notification, the counterfactual reasoning engine module 3 automatically triggers an incremental causal reasoning process. Causal discovery and counterfactual reasoning are re-executed only for the entities and relationships that have changed to update the risk causal graph 211. Upon receiving a graph change notification, the digital twin construction module 4 automatically triggers an incremental update process for the digital twin 400, i.e., recalibrating the relevant parameters of the financial model 410, supply chain model 420, technology evolution model 430, and market response model 440 only for the risk factors that have changed. A bidirectional data channel is established between the digital twin construction module 4 and the multi-scenario stress testing module 5. The current state of the digital twin 400 is synchronized to the multi-scenario stress testing module 5 in real time through this bidirectional data channel as the initial data for stress testing. Under the given conditions, the dynamic risk evolution trajectory generated by the multi-scenario stress test module 5 is simultaneously transmitted back to the digital twin construction module 4 through the bidirectional data channel to drive the state updates of each sub-model of the digital twin 400 and to the supply chain risk contagion modeling module 6 as the input driving force for risk propagation simulation. After completing the risk propagation simulation, the supply chain risk contagion modeling module 6 simultaneously pushes the risk destructive power indicators and risk absorption coefficients 361 of each node output by it to the three-dimensional anchored knowledge graph construction module 2 to enhance the risk attribute information of enterprise entities in the knowledge graph and to the interpretable artificial intelligence decision engine module 7 as the risk assessment basis for investment decisions.The decision report generated by the explainable AI decision engine module 7 is output to the human-computer collaborative interaction module 8 via a standardized data interface for user review and feedback. The user feedback information collected by the human-computer collaborative interaction module 8 is transmitted back to the counterfactual reasoning engine module 3 via the standardized data interface to trigger online updates of causal effect strength parameters; to the digital twin construction module 4 to trigger Bayesian corrections of the prior distributions of each sub-model; to the industrial chain risk contagion modeling module 6 to trigger incremental retraining of graph neural network weights; and to the 3D anchored knowledge graph construction module 2 to trigger adjustments to entity attribute confidence. This two-way data flow feedback loop ensures information sharing, parameter coordination, and overall consistency among the system modules, enabling the entire system to achieve a complete closed loop from data acquisition to knowledge discovery, decision output, and strategy optimization during continuous operation.

[0039] The working principle of this invention lies in constructing a complete intelligent link from multi-source heterogeneous data acquisition to human-machine collaborative decision optimization. This link is based on data fusion, causal reasoning, digital twins as a simulation sandbox, multi-layered networks as a risk propagation analysis framework, and explainable artificial intelligence as the decision output port. Ultimately, it achieves continuous evolution of decision-making strategies through human-machine interaction. First, the multi-source heterogeneous data acquisition module captures massive amounts of data in real time from dimensions such as technology research and development, supply chain, market policies, enterprise operations, and environmental and social governance. After cleaning, standardization, and alignment, this data is simultaneously input into the three-dimensional anchored knowledge graph construction module and ES. The G-risk sub-graph construction module; secondly, the three-dimensional anchored knowledge graph construction module binds each industry entity to both geographical and temporal anchors, forming a "industry element-geography-time" three-dimensional anchored structure. This structure enables each industry entity to possess traceable spatial policy attributes, resource endowment attributes, geopolitical risk attributes, as well as traceable state time series attributes and relationship timeliness attributes, thus providing a spatiotemporally coupled data foundation for subsequent causal reasoning and risk propagation analysis; thirdly, the counterfactual reasoning engine performs bidirectional causal reasoning based on this three-dimensional anchored knowledge graph, and its forward causal reasoning module learns the causal structure from the data and generates... The system constructs a risk causal graph. Its counterfactual reasoning module intervenes to set specific risk factors to a "not yet occurred" state and recalculates the counterfactual results to quantify the causal effect strength of each risk factor. Its causal chain prediction module performs pattern matching and extrapolation based on the current knowledge graph state and historical causal pattern library, thereby generating a forward-looking risk transmission path prediction before the risk event actually occurs. Then, the digital twin construction module constructs a digital twin of the project or enterprise to be evaluated based on the three-dimensional anchored knowledge graph and the risk causal graph. This digital twin includes four sub-models: a financial model, a supply chain model, a technology evolution model, and a market response model. The sub-models interact and respond in unison through standardized bidirectional data interfaces, enabling any change in the state of a sub-model to trigger a chain reaction in other sub-models. This allows for a highly realistic simulation of the complex dynamics of the real industrial environment in the virtual space. Next, the multi-scenario stress testing module injects various extreme scenarios, such as supply disruptions, sudden drops in demand, policy changes, technological substitution, and compound shocks, into the digital twin. This drives the sub-models to work together to generate the dynamic risk evolution trajectory of the project or enterprise under each scenario. This trajectory includes multi-dimensional information such as risk probability change curves, risk loss distribution curves, and the timing of key risk indicator threshold breaches.Simultaneously, the supply chain risk contagion modeling module constructs a four-layer heterogeneous network model comprising resource, processing, material, and application layers. This multi-layer network model represents upstream and downstream supply and demand relationships through directed weighted edges. Based on a dynamic risk evolution trajectory-driven cascading failure model, it simulates the propagation process of risk between layers, calculates the risk destructive power index and risk absorption coefficient of each node, and then marks nodes with both high destructive power and high absorption capacity as key hub nodes, and nodes with low destructive power but high absorption capacity as risk buffer nodes. In investment decisions, risk buffer nodes are given a positive valuation premium. Subsequently, the interpretable AI decision engine aggregates all outputs from the risk causal graph, dynamic risk evolution trajectory, and the supply chain risk contagion multi-layer network model into a multi-dimensional risk assessment matrix. This matrix covers five dimensions: technological risk, supply chain risk, market risk, policy risk, and ESG risk. The decision engine uses an attention mechanism to process each risk factor. The system assigns weights to generate a comprehensive risk score. Then, an interpretability module outputs the risk identification criteria, causal path tracing explanations, and comparative analysis with historical benchmarks in natural language. Finally, based on user-preset risk preference parameters, a multi-objective optimization algorithm generates an interpretable investment decision report containing investment target rankings, investment ratio suggestions, and risk mitigation measures. The human-computer collaborative interaction module outputs the decision report to the user, receiving feedback such as confirmation or correction of risk factors, adjustment of risk tolerance, and preference for alternative solutions. An online learning algorithm incrementally updates the model parameters of the counterfactual reasoning engine, digital twin, and supply chain risk contagion modeling module. User feedback and model updates are recorded and stored in a feedback database to construct a user decision preference profile. Based on this profile, subsequent decisions adaptively adjust risk weight allocation and optimize the objective function, thus achieving a complete closed loop from data collection to strategy evolution.

[0040] The above provides a detailed description of a data-driven risk identification and decision optimization method for investment in the new materials industry, as provided by this invention. Specific embodiments have been used to illustrate the principles and implementation methods of this invention. These embodiments are merely illustrative and are intended to aid in understanding the method and its core concepts. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the scope of protection of the claims of this invention.

Claims

1. A data-driven risk identification and decision optimization method for investment in the new materials industry, characterized in that, Includes the following steps: S100. Real-time collection of multi-dimensional data related to investment in the new materials industry through the multi-source heterogeneous data acquisition module (1), including technology research and development data, supply chain data, market and policy data, enterprise operation data and environmental and social governance data; S200. Based on the collected multidimensional data, construct a three-dimensional anchored knowledge graph (100) of "industrial elements-geography-time". Each industrial entity in the knowledge graph (100) is simultaneously bound to a geographical anchor (110) and a time anchor (120). The geographical anchor (110) includes the policy and regulatory vector (111), resource endowment vector (112), and geopolitical risk index (113) of the region where the entity is located. The time anchor (120) includes the state time series (121) and relationship time window (122) of the entity. S300. Based on the three-dimensional anchored knowledge graph (100), a counterfactual reasoning engine (200) is constructed. The counterfactual reasoning engine (200) includes a positive causal reasoning module (210), a counterfactual reasoning module (220), and a causal chain prediction module (230). The positive causal reasoning module (210) identifies the causal structure between risk factors from the multidimensional data through a causal discovery algorithm and generates a risk causal graph (211). The counterfactual reasoning module (220) sets the value of a specific risk factor to a "not occurred" state through an intervention operation (221) and drives the structural causal model to recalculate the counterfactual results (222). Based on the counterfactual results (222), it calculates the causal effect strength of each risk factor (223). The causal chain prediction module (230) predicts the risk transmission path before the risk event actually occurs, based on the current knowledge graph state (231) and the historical causal pattern library (232) through pattern matching and extrapolation (233). S400. Based on the three-dimensional anchored knowledge graph (100) and the risk causal graph (211), a digital twin (400) of the new material project or enterprise to be evaluated is constructed. The digital twin (400) includes a financial model (410), a supply chain model (420), a technology evolution model (430), and a market response model (440). Data interaction between the sub-models is achieved through a two-way data interface (450). S500, Perform multi-scenario stress tests on the digital twin (400) to generate a dynamic risk evolution trajectory; S600. Construct a multi-layer network model for risk contagion in the industrial chain (300). The multi-layer network model (300) includes a resource layer (310), a processing layer (320), a material layer (330), and an application layer (340). The layers are connected by directed weighted edges (350). Based on the dynamic risk evolution trajectory, the multi-layer network model (300) is used to simulate the risk propagation process between the layers of the industrial chain, calculate the risk absorption coefficient (361) of each node (360), and identify key hub nodes (390) and risk buffer nodes (391). S700, Input the output results of the risk causal graph (211), the dynamic risk evolution trajectory and the multi-layer network model of industrial chain risk contagion (300) into the interpretable artificial intelligence decision engine (500) to generate an interpretable investment decision report containing risk identification basis, transmission path description and investment advice; S800: Based on user interaction feedback, the decision parameters of the explainable artificial intelligence decision engine (500) are dynamically adjusted to form a closed-loop decision optimization mechanism for human-machine collaboration.

2. The data-driven risk identification and decision optimization method for investment in the new materials industry according to claim 1, characterized in that, The construction of the "industrial elements-geography-time" three-dimensional anchored knowledge graph (100) described in step S200 specifically includes: S210. Perform entity recognition and relationship extraction on the collected multidimensional data to identify enterprise entities, material entities, technology entities, supply chain node entities, policy and regulatory entities, and geographical area entities in the new materials industry; S220. Construct geographic anchor points for each entity (110), and access the regional policy database, mineral resource distribution database and geopolitical risk assessment database through the geographic information system data interface to generate policy and regulation vector (111), resource endowment vector (112) and geopolitical risk index (113) of the region where each entity is located. S230, construct time anchors for each entity (120), store the state time series of each entity based on the time series database (121), and set relationship validity windows (122) for each entity relationship. The relationship validity window (122) is used to define the effective range of the relationship in the time dimension. S240. The geographic anchor (110) and the time anchor (120) are embedded as independent dimensions into the entity-relationship-attribute triple structure of the knowledge graph (100) to form a three-dimensional anchoring graph.

3. The data-driven risk identification and decision optimization method for investment in the new materials industry according to claim 1, wherein the counterfactual reasoning engine (200) in step S300 is constructed in the following manner: S310. A structural causal model based on a directed acyclic graph is adopted as the mathematical framework of the positive causal reasoning module (210). The causal structure between variables is learned from the multidimensional data to generate a risk causal graph (211). S320, The counterfactual reasoning module (220) is based on the Pearl causal hierarchy model and performs the intervention operation (221) through Do-calculus to intervene the value of the target risk factor to the "not occurred" state; S330. A counterfactual sample space is generated using Monte Carlo simulation, and the structural causal model is driven to recalculate the counterfactual results on the counterfactual sample space (222). S340. Compare the counterfactual results (222) with the original results, and calculate the causal effect strength (223) of each risk factor, wherein the causal effect strength (223) characterizes the difference in results caused by the occurrence or non-occurrence of the risk factor; S350, the causal chain prediction module (230) performs pattern matching (233) based on the current knowledge graph state (231) and the historical causal pattern library (232), identifies historical causal patterns similar to the current state, extrapolates the causal chains that appear later in the historical patterns to the current state, and generates a forward-looking risk transmission path prediction result.

4. The data-driven risk identification and decision optimization method for investment in the new materials industry according to claim 1, wherein the sub-models of the digital twin (400) in step S400 are specifically as follows: The financial model (410) is constructed based on the company’s historical financial data, industry benchmark data and macroeconomic data, and supports cash flow forecasting, asset and liability simulation and profitability assessment. The supply chain model (420) is constructed based on raw material supply data, logistics data, inventory data and geopolitical risk data (113), and supports supply disruption simulation and alternative supply scheme deduction; The technology evolution model (430) is constructed based on patent data, R&D investment data, technology life cycle data and technology substitution data, and supports technology route evolution prediction and technology obsolescence risk assessment. The market response model (440) is constructed based on downstream market demand data, competitive landscape data and price elasticity data, and supports market demand fluctuation simulation and competitive situation deduction. Data interaction between the sub-models is achieved through a standardized bidirectional data interface (450), so that a change in the state of any sub-model can trigger a linkage response in other sub-models.

5. The data-driven risk identification and decision optimization method for investment in the new materials industry according to claim 1, wherein the multi-scenario stress test in step S500 specifically includes: S510. Based on historical extreme event data and expert experience, generate a stress test scenario library that includes supply disruption scenarios, demand plunge scenarios, policy change scenarios, technology substitution scenarios, and compound shock scenarios. S520. Input each stress scenario into the digital twin (400), drive each sub-model to run the simulation, and generate the risk evolution trajectory of the project or enterprise under each scenario; S530, The risk evolution trajectory includes a risk probability change curve, a risk loss distribution curve, and a threshold breakthrough time sequence of key risk indicators in the time dimension. S540. Based on the simulation results, calculate the value at risk and value at risk for each scenario, and generate a dynamic risk heat map.

6. The data-driven risk identification and decision optimization method for investment in the new materials industry according to claim 1, wherein step S600, which involves constructing a multi-layered network model for risk contagion in the industrial chain (300) and identifying key hub nodes (390) and risk buffer nodes (391), specifically includes: S610. Based on the three-dimensional anchored knowledge graph (100), construct a multi-layer network topology structure for the industrial chain. The resource layer (310) of the multi-layer network model (300) includes mineral resource enterprise nodes, the processing layer (320) includes raw material processing enterprise nodes, the material layer (330) includes new material manufacturing enterprise nodes, and the application layer (340) includes downstream application enterprise nodes. S620. A graph neural network is used to perform embedding learning on each node (360) in the multi-layer network model (300), and the influence scores of each node (360) within and between layers are calculated by coupling the PageRank algorithm. S630. Simulate single-point failure scenario (370) and multi-point cascading scenario (380) through cascading failure model, calculate the total downstream risk propagation (381) caused by the failure of each node (360), and use it as the risk damage index of the node. S640. By comparing the difference in the total amount of risk propagation under the two network states of "including the node" and "removing the node", calculate the risk absorption coefficient (361) of each node (360). The risk absorption coefficient (361) represents the buffering capacity of the node against upstream risks. S650, mark nodes with risk damage index higher than the first threshold and risk absorption coefficient (361) higher than the second threshold as critical hub nodes (390), and mark nodes with risk damage index lower than the third threshold and risk absorption coefficient (361) higher than the fourth threshold as risk buffer nodes (391). S660. In the investment decision report, a positive valuation premium is given to the investment target corresponding to the risk buffer node (391).

7. The data-driven risk identification and decision optimization method for investment in the new materials industry according to claim 1, wherein the interpretable artificial intelligence decision engine (500) in step S700 specifically includes: S710. Based on the output of the risk causal graph (211), dynamic risk evolution trajectory and multi-layer network model of industrial chain risk contagion (300), a multi-dimensional risk assessment matrix (510) is constructed. The risk assessment matrix (510) includes the dimensions of technology risk, supply chain risk, market risk, policy risk and ESG risk. S720. An attention mechanism is used to assign weights to each risk factor in the multi-dimensional risk assessment matrix (510) to generate a comprehensive risk score. S730. Through the interpretability module (520), output the risk identification basis in the form of natural language generation. The risk identification basis includes the contribution of each risk factor, the causal path tracing explanation and the comparative analysis with historical benchmarks. S740: Based on the comprehensive risk score and the user's preset risk preference parameters, an investment decision optimization scheme is generated through a multi-objective optimization algorithm, which includes the ranking of investment targets, investment ratio suggestions, and risk mitigation measures.

8. The data-driven risk identification and decision optimization method for investment in the new materials industry according to claim 1, wherein the human-machine collaborative closed-loop decision optimization mechanism in step S800 specifically includes: S810. After outputting the interpretable investment decision report to the user, receive feedback information from the user on the decision suggestions. The feedback information includes confirmation or correction of risk factors, adjustment of risk tolerance, and preference selection of alternative solutions. S820. Based on the feedback information, the model parameters of the counterfactual reasoning engine (200), the digital twin (400), and the multi-layer network model of supply chain risk contagion (300) are incrementally updated using an online learning algorithm. S830. Store user feedback and model update records in the feedback database to build a user decision preference profile; S840. Based on the user decision preference profile, the risk weight allocation and objective function are adaptively adjusted in subsequent decisions to achieve personalized evolution of decision-making strategies.

9. The data-driven risk identification and decision optimization method for investment in the new materials industry according to claim 1, wherein the environmental and social governance data in step S100 includes carbon emission intensity data, energy consumption data, waste emission data, environmental compliance record data, social responsibility practice data, and corporate governance structure data; the method further includes: The environmental and social governance data is integrated with the technology research and development data, supply chain data, market and policy data and enterprise operation data to construct an ESG risk sub-map (600). The counterfactual reasoning engine (200) is used to identify the causal transmission path between ESG risk factors and financial risk factors and generate an ESG-financial risk linkage early warning signal.

10. A data-driven risk identification and decision optimization system for investment in the new materials industry, characterized in that, include: A multi-source heterogeneous data acquisition module (1) is used to collect multi-dimensional data related to investment in the new materials industry in real time. The three-dimensional anchored knowledge graph construction module (2) is used to construct a three-dimensional anchored knowledge graph (100) based on the collected multi-dimensional data. The counterfactual reasoning engine module (3) is used to identify the causal path between risk factors based on the three-dimensional anchored knowledge graph (100) through positive causal reasoning, counterfactual reasoning and forward-looking causal chain prediction, and generate a risk causal graph (211) and forward-looking risk transmission path prediction results. The digital twin construction module (4) is used to construct a digital twin (400) of the project or enterprise to be evaluated based on the three-dimensional anchored knowledge graph (100) and the risk causal graph (211). The multi-scenario stress testing module (5) is used to perform multi-scenario stress testing on the digital twin (400) and generate a dynamic risk evolution trajectory. The supply chain risk contagion modeling module (6) is used to construct a multi-layer network model (300) of supply chain risk contagion, which includes a resource layer (310), a processing layer (320), a material layer (330), and an application layer (340). Based on the dynamic risk evolution trajectory, the risk absorption coefficient (361) of each node (360) is calculated, and key hub nodes (390) and risk buffer nodes (391) are identified. An interpretable artificial intelligence decision engine module (7) is used to receive the outputs of the risk causal graph (211), the dynamic risk evolution trajectory and the multi-layer network model of industrial chain risk contagion (300), and generate an interpretable investment decision report; The human-computer collaborative interaction module (8) is used to receive user feedback and dynamically adjust decision parameters to form a closed-loop decision optimization mechanism; Among them, the modules interact with each other through standardized data interfaces, and a feedback loop of two-way data flow is formed between the three-dimensional anchored knowledge graph construction module (2), the counterfactual reasoning engine module (3), the digital twin construction module (4), and the industrial chain risk contagion modeling module (6).

Images