A land surface process intelligent modeling method and system based on noah-mp, knowledge base and agent

Abstract

The application discloses a land surface process intelligent modeling method and system based on Noah-MP, a knowledge base and an intelligent agent, and specifically is as follows: multi-modal data of Noah-MP is collected, the multi-modal data is abstracted into nine core fields, a knowledge tuple is constructed, and a land surface process parameterization scheme knowledge base is constructed; in a multi-agent system, an intelligent agent reads user natural language and a model running parameter file, extracts geographical environment and model running environment information of a research area, and transmits the information to a retrieval intelligent agent; the retrieval intelligent agent constructs keywords, performs multi-round retrieval in the knowledge base, reasons to obtain a parameterization scheme combination, and then forwards the parameterization scheme combination to an execution intelligent agent; the execution intelligent agent performs physical constraint verification on the parameterization scheme combination, and performs land surface process simulation based on the parameterization scheme combination that passes the verification. The application realizes automatic processing from user natural language to model parameterization scheme configuration and running, and improves the scientificity of parameterization scheme selection.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary application technology of Earth science numerical simulation and artificial intelligence, and in particular relates to an intelligent modeling method and system for land surface processes based on Noah-MP, knowledge base and intelligent agent. Background Technology

[0002] Land surface processes refer to the interactions of energy, moisture, and momentum occurring between the atmosphere, surface, and soil layers. They primarily study energy transfer between the Earth and atmosphere, the redistribution of energy at the Earth's surface, energy transfer within the soil, and the movement of moisture and matter between the Earth and the atmosphere and the subsurface. Land surface process models are mathematical and physical models that quantitatively describe these processes. They are core foundational tools for climate prediction, hydrological simulation, ecological assessment, and disaster early warning, playing an irreplaceable role in global change research and ecological civilization construction.

[0003] As geography continues to develop towards a more holistic, comprehensive, and complex approach, coupling multidisciplinary professional models and constructing customized land surface process simulation schemes have become important means of solving complex Earth system problems. Land surface process models, exemplified by Noah-MP, offer multiple optional parameterization schemes for the same physical process, providing a unified framework for flexibly adapting to the environmental characteristics of different study areas. Researchers need to consider the environmental characteristics of the study area, such as climate, vegetation, soil, and topography, and combine existing observational data, professional knowledge, and engineering experience to select appropriate parameterization schemes and configure corresponding parameters, ultimately completing the model compilation and operation. Despite this, practical modeling work still faces challenges such as a steep learning curve and high operational barriers. On the one hand, land surface process modeling requires users to possess deep geoscientific knowledge, understand key physical processes and the advantages, disadvantages, and adaptability of corresponding parameterization schemes. The selection and determination of the optimal model heavily relies on repeated trial and error and expert knowledge, which is not only time-consuming and laborious with poor repeatability but also highly subjective, leading to significant uncertainties in the scientific validity of the model and the simulation results. On the other hand, the modeling process entirely depends on manual operation of the existing model's source code. This code-level modeling approach has many insurmountable drawbacks: the technical thresholds for parameter file writing, model environment configuration, compilation, and execution are high. Existing modeling processes require users to have experience in Fortran programming, Linux environment configuration, and model compilation, which usually requires several months of systematic training; the entire process involves a huge amount of manual work and is extremely prone to parameter configuration errors or code compilation failures due to human negligence; more importantly, a large amount of the modeler's energy is consumed by tedious technical details, making it difficult to focus on the research of core scientific questions.

[0004] In recent years, large-scale models, with their powerful natural language understanding, planning analysis, and tool usage capabilities, have provided new ideas for solving the aforementioned challenges in modeling land surface processes. However, in highly specialized vertical fields such as Earth sciences, general-purpose large-scale models often face problems such as insufficient domain knowledge reserves and susceptibility to illusions. To improve the reliability and controllability of models in specialized scenarios, the academic community has gradually developed various technical approaches, including cue word engineering, model fine-tuning, efficient parameter fine-tuning, knowledge distillation, and retrieval-Augmented Generation (RAG). Cue word engineering guides model output through carefully designed instruction templates, offering low cost and easy rapid iteration, but suffers from poor stability and struggles to guarantee the strict requirements of physical consistency under long logical chains. Model fine-tuning utilizes domain-specific corpora to retrain open-source large-scale models, effectively improving their performance, but the training process consumes enormous resources and struggles to achieve low-cost, real-time knowledge updates. Efficient parameter fine-tuning is an optimized variant of fine-tuning, updating only a small number of parameters to reduce computational overhead. Its advantage lies in balancing computational resources and model performance, but its generalization ability is sometimes inferior to full fine-tuning. Knowledge distillation transfers knowledge from large models to lightweight models by guiding student models through teacher models, achieving efficient model optimization. However, its accuracy may be limited in complex learning tasks.

[0005] Existing methods for simulating land surface processes face several challenges:

[0006] (a) Complex parameterization scheme configuration and reliance on expert experience: Land surface process models contain multiple physical processes, and different parameterization schemes can be selected for the same physical process. Land surface process models, represented by Noah-MP, have high-dimensional parameter spaces and numerous combinations of physical parameterization schemes. The selection of the optimal parameter combination is highly dependent on human experience and repeated trial and error, which is not only time-consuming and laborious, but also highly subjective. There may also be a lag in the updating of knowledge and experience, resulting in significant uncertainty in the model simulation results and poor repeatability.

[0007] (b) Steep learning curve and high operational threshold: Traditional land surface process modeling requires users to have a strong foundation in numerical simulation and programming, involving complex code environment configuration, parameter file writing, and cumbersome processes such as model compilation and execution. For non-professional users or beginners, the learning cost is extremely high, which seriously limits the promotion and application of the model.

[0008] In summary, due to multiple limitations such as the professional knowledge threshold, the complexity of the model itself, and the model operating environment, land surface process modeling, a knowledge- and technology-intensive process, poses a significant challenge to both beginners and experienced professionals. Existing multi-parameterized model frameworks cannot simultaneously meet the comprehensive requirements of convenient operation procedures, scientifically sound parameterized schemes, and real-time knowledge updates in land surface process modeling. Summary of the Invention

[0009] Purpose of the invention: To address the problems of complex parameterization scheme configuration, steep learning curves, over-reliance on expert experience, large human error, and poor process repeatability in existing land surface process modeling, this invention provides an intelligent land surface process modeling method and system based on Noah-MP, a knowledge base, and intelligent agents.

[0010] Technical solution: This invention is based on Noah-MP, a knowledge base, and an intelligent agent for intelligent modeling of land surface processes, specifically as follows:

[0011] Collect multimodal data covering the Noah-MP land surface process model; extract the core field content from the multimodal data, construct knowledge tuples, perform vectorized storage and indexing of the knowledge tuples, and construct a knowledge base for parameterized schemes of land surface processes. The core fields include: research nature, research objectives, research area, underlying surface type, dominant physical process, improvement content, parameterized scheme combination, applicable conditions, and performance evidence.

[0012] A multi-agent system is constructed, including a parsing agent, a retrieval agent, and an execution agent. The parsing agent reads the natural language input by the user and the model operation parameter information from the driver file of the Noah-MP land surface process model, and then identifies the geographical environment and model operation environment characteristics of the study area. The natural language is used to describe the geographical environment of the study area. The parsing results of the parsing agent are transmitted to the retrieval agent. The retrieval agent constructs keywords based on the received information, searches in the land surface process parameterization scheme knowledge base, and forwards the retrieved parameterization scheme combination to the execution agent. The execution agent performs physical constraint verification on the received parameterization scheme combination. The parameterization scheme combination is a combination of parameterization schemes corresponding to all physical processes in the Noah-MP land surface process model. The parameterization scheme combination that has passed the physical constraint verification by the execution agent is written back to the Noah-MP land surface process model to execute the simulation task.

[0013] Furthermore, the parameterization scheme combination directly maps the parameters corresponding to the physical processes in the Noah-MP land surface process model in a structured manner;

[0014] The study describes the types of knowledge tuples;

[0015] The underlying surface type describes environmental characteristics and provides underlying surface constraints for parametric scheme combinations;

[0016] The dominant physical process extracts the physical mechanism, providing logical constraints for the physical reasoning process of multi-agent systems;

[0017] The applicable conditions clearly define the applicable time, region, and underlying surface environment boundaries of the parameterized scheme;

[0018] Performance evidence provides quantitative verification, making the recommendation results of multi-agent systems interpretable and traceable;

[0019] Improve the annotation of multimodal data to determine whether it involves improvements to the parameterization scheme and the parameters to be improved, so as to help large models distinguish between conventional applications and optimization schemes.

[0020] Furthermore, information extraction and standardized annotation are performed on the collected multimodal data. For PDF documents, the data is segmented at the paragraph level with a segment size of 1024 characters. For TXT documents, the data is segmented at the paragraph level with a segment size of 4096 characters. The segmentation process involves splitting the document into text fragments suitable for the embedded model, with each fragment retaining its semantic context.

[0021] An embedding model is used to vectorize text fragments, converting each text fragment into a dense vector;

[0022] Dense vectors are stored in a vector database, and a cosine similarity index is constructed to form a knowledge base for land surface process parameterization schemes that supports semantic retrieval.

[0023] Furthermore, the parsing agent, retrieval agent, and execution agent communicate through the asynchronous message passing mechanism provided by the AgentScope framework, sharing a unified context session window. Each agent adopts the ReAct mode, and each agent is configured with a large language model as the inference engine.

[0024] Furthermore, the retrieval agent is registered with a knowledge base retrieval tool, and system prompts are set for the retrieval agent, specifically including the following specifications:

[0025] Core principle: All search results must be obtained from the land surface process parameterization scheme knowledge base based on search technology. It is prohibited to generate search results directly without searching the land surface process parameterization scheme knowledge base.

[0026] Search criteria: Searches should only include keywords related to the geographical environment and model operating parameters of the study area;

[0027] Decision-making criteria: The parameterized scheme should meet the preset physical constraints;

[0028] Output format requirements: Output the final result in JSON object format. The final result includes the parameterization scheme combination of the Noah-MP land surface process model and the reason field for each parameterization scheme. The reason field is used to explain the selection basis.

[0029] The retrieval agent performs multiple rounds of retrieval, and then combines the information parsed by the parsing agent to perform fusion reasoning, outputting a combination of parameterized schemes; the execution agent performs physical constraint verification on the combination of parameterized schemes, specifically: determining whether the related parameterized schemes meet the physical constraints; if the physical constraint verification passes, the simulation task is executed based on the parameterized scheme combination that has passed the physical constraint verification, otherwise a restart command is sent to the parsing agent to restart the entire process.

[0030] Furthermore, the parameterized schemes that perform physical constraint verification by the intelligent agent will be combined and written back into the Noah-MP land surface process model, specifically as follows:

[0031] Step A: Extract the JSON object from the text returned by the executing agent, use regular expression matching to locate the string boundaries of the JSON object, parse the parameterization scheme and reason field in the JSON object, and verify the parameterization scheme. If the verification fails, extract it again; otherwise, go to step B.

[0032] Step B: Read the current driver file, use line-by-line scanning to locate the marker at the start position of the parameterization scheme, identify the original parameterization scheme, replace the original parameterization scheme with the parameterization scheme in the searched parameterization scheme combination according to the original format, complete the replacement of all parameterization schemes, and update the driver file;

[0033] Step C: Locate the directory containing the Noah-MP land surface process model, find the Noah-MP land surface process model executable program, pass the updated driver file as a command-line parameter to the executable program, capture the standard output and error output in real time during the Noah-MP land surface process model runtime, monitor the simulation progress, and troubleshoot any runtime anomalies.

[0034] Furthermore, if the parameterization scheme is of integer type and the combination of parameterization schemes is complete and without omissions, then the verification passes.

[0035] A land surface process intelligent modeling system based on Noah-MP, knowledge base, and intelligent agents, comprising:

[0036] A multi-source data and knowledge input layer is used to collect multimodal data covering the Noah-MP land surface process model;

[0037] The knowledge base construction layer is used to extract the core fields and corresponding values ​​from multimodal data, construct knowledge tuples, perform vectorized storage and index construction on the knowledge tuples, and construct a knowledge base for the parameterized scheme of land surface processes.

[0038] The agent decision-making and execution layer includes a parsing agent, a retrieval agent, and an execution agent. The parsing agent reads the natural language input from the user and the model operation parameter information from the Noah-MP land surface process model's driver file, and then identifies the geographical environment and model operation environment characteristics of the study area. The parsing agent transmits the parsing results to the retrieval agent, which constructs keywords based on the received information, searches in the land surface process parameterization scheme knowledge base, and forwards the retrieved parameterization scheme combination to the execution agent. The execution agent performs physical constraint verification on the received parameterization scheme combination.

[0039] The model execution, output, and evaluation layer writes back the parameterized scheme combination, which performs physical constraint verification of the agent, to the Noah-MP land surface process model. The Noah-MP land surface process model executes the simulation task and outputs the results; and evaluates the land surface process system.

[0040] Furthermore, the land surface process system is evaluated, including verifying the accuracy of the decision results, the reliability of the land surface process system, and its usability.

[0041] Verification of the accuracy of decision results includes accuracy verification and matching rate verification:

[0042] Accuracy Verification: For a specific region, several land surface process simulation scenarios were selected, covering all spatial scales, all underlying surface types, and all climate types. For each simulation scenario, an exhaustive method was used to traverse all parameterization scheme combinations. For each parameterization scheme combination, the Noah-MP land surface process model was run under the same driving file, initial conditions, and boundary conditions. The simulation results of the Noah-MP land surface process model were compared with actual observation data, and the Kling-Gupta efficiency coefficient (KGE) was calculated. For each simulation scenario, the parameterization scheme combinations were sorted from highest to lowest KGE value, and the top k parameterization scheme combinations were selected to form the baseline scheme set, denoted as [missing information]. ;

[0043] Using the land surface process system, N parameterized scheme combinations are obtained for the same region and the same scenario. The set of these N parameterized scheme combinations is denoted as... Calculate the hit rate H;

[0044] ;

[0045] ;

[0046] in, For the i-th parameterization scheme combination retrieved by the land surface process system, For indicator functions;

[0047] If H is greater than the preset threshold, it indicates that the accuracy of the land surface process system decision results meets the requirements.

[0048] Matching rate verification: X parameterization schemes corresponding to physical processes are manually selected as key parameterization schemes.

[0049] exist Extract the first A set of key parameterization schemes constitutes a numerical set. , , among which, if The first in The values ​​of the key parameterization schemes are in Medium; Matching rate Add 1, If the match rate exceeds the preset threshold, it indicates that the matching rate meets the requirements.

[0050] The reliability verification specifically involves using correlation coefficient, root mean square error, and average deviation for reliability verification.

[0051] The usability verification specifically involves: recruiting several users with different backgrounds and randomly dividing them into a control group and an experimental group; assigning all users the same land surface process simulation task, which is for users to complete the entire process from parameter configuration to running the Noah-MP land surface process model;

[0052] The control group completed the task using the traditional method: manually editing the driver files and environmental parameters of the Noah-MP land surface process model, including modifying the parameterization scheme, setting output variables, compiling the model, and executing the executable program.

[0053] The experimental group used the land surface process system. Users described the environmental characteristics of the study area, and the land surface process system automatically completed the subsequent operations.

[0054] Record and calculate three indicators: task completion time, number of operation steps, and subjective satisfaction score; calculate the mean score and standard deviation of the three indicators for the control group and the experimental group respectively.

[0055] A computer device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the aforementioned intelligent modeling method for land surface processes based on Noah-MP, a knowledge base, and an intelligent agent.

[0056] Beneficial effects:

[0057] (1) This invention innovatively introduces retrieval-enhanced generation technology into the field of parameterization scheme configuration for land surface process models, constructs a knowledge base for land surface process parameterization schemes, and transforms unstructured knowledge scattered in literature, technical reports, and expert experience into structured knowledge that can be queried and reasoned about by machines. The system can automatically match the optimal parameterization scheme combination according to the environmental characteristics of the study area, effectively solving the configuration problem brought about by high-dimensional parameter spaces such as the Noah-MP land surface process model, significantly reducing human trial and error and subjective bias, and improving the scientific nature of parameter selection for physical schemes.

[0058] (2) This invention constructs a modular multi-agent system based on the AgentScope framework, realizing full-process automation. It transforms complex code operations and professional configuration processes into intuitive natural language interaction, greatly reducing the learning cost and usage threshold of land surface process models, enabling non-professional users to quickly complete high-quality land surface process simulations, while freeing professionals from tedious technical details and allowing them to focus on the research of core scientific issues.

[0059] (3) This invention adopts a lightweight technical solution that combines RAG (Retrieval-Augmented Generation) with a large language model. It does not require full fine-tuning of the large model, consumes low computational resources, and facilitates knowledge updates. Simply adding new literature, technical reports, or parameter configuration experience to the knowledge base can achieve continuous upgrades of the system's capabilities. It can quickly adapt to version iterations of the land surface process model and updates of domain knowledge, and has good timeliness and scalability.

[0060] (4) The system architecture of the present invention has high versatility and portability. The knowledge base supports multimodal data input such as PDF and plain text. The intelligent agent system adopts a modular design and can be quickly adapted to other mainstream land surface process models by replacing the knowledge base and model call interface. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the overall framework provided for an embodiment of the present invention.

[0062] Figure 2 This is a graph showing the relationship between the results obtained using the method of the present invention and the actual observation results. Detailed Implementation

[0063] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0064] like Figure 1As shown, this embodiment provides a land surface process system based on the combination of a knowledge base and an intelligent agent. The system adopts a four-layer architecture design, from top to bottom: a multi-source data and knowledge input layer, a RAG parameterization scheme knowledge base construction layer, an AgentScope intelligent agent decision-making and execution layer, and a model operation, output, and evaluation layer. The intelligent agent system of this invention is built on the AgentScope framework and adopts the ReAct (reasoning-action) mode. Through the coordinated work of three functional links—planning reasoning, knowledge retrieval, and parameter configuration and execution—it achieves a closed-loop process from natural language commands to automatic model operation.

[0065] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0066] Step 1: Construction of the knowledge base for land surface process parameterization schemes.

[0067] Step 1-1: Collect multimodal datasets. Systematically collect multimodal literature (including application, evaluation, and improvement studies) covering the Noah-MP land surface process model, including journal articles, dissertations, official technical manuals, version update notes, and open-source community documents, to form an initial dataset. The dataset includes academic literature and operation manuals in PDF and plain text formats, stored in a pre-defined literature directory.

[0068] Steps 1-2: Extract the content of the corresponding core fields from the multimodal data and construct field-content knowledge tuples. The knowledge tuples are extracted from text fragments of unstructured literature, as shown in Table 1. A structured knowledge tuple template containing the following core fields is designed: research nature, research objectives, research area, underlying surface type, dominant physical process, improvement content, parameterization scheme combination, applicable conditions, and performance evidence, totaling 9 core fields.

[0069] Table 1. Core Fields and Their Meanings

[0070]

[0071] The design of the above nine core fields fully considers the knowledge consumption and reasoning needs of large models, forming a complete knowledge expression and decision support system in terms of functionality.

[0072] The study of property fields helps agents quickly determine the type of knowledge tuples retrieved, providing a basis for decision-making in large models.

[0073] The four fields—research objective, research area, underlying surface type, and dominant physical process—together constitute a multi-dimensional environmental feature vector, providing accurate case matching basis for large-scale models. When a user inputs a natural language description of the research area, the retrieval agent can quickly recall the most relevant historical parameterization scheme combinations from the knowledge base by calculating the semantic similarity of these fields, significantly improving the accuracy and efficiency of the retrieval. Specifically, the research objective field summarizes the core research purpose of the literature, providing task-oriented matching basis; the research area field records geographical location and spatial scale information, supporting semantic retrieval based on geographical features; the underlying surface type field describes environmental characteristics such as vegetation cover and soil texture, providing underlying surface constraints for parameter configuration. The dominant physical process field extracts core physical mechanisms that play a decisive role in the research objective, such as freeze-thaw processes and evapotranspiration processes, providing physical logic constraints for the reasoning process of large-scale models, ensuring that the output parameter combinations meet the physical consistency requirements of a specific environment, and effectively avoiding the physically unreasonable configurations that may occur with purely data-driven methods.

[0074] Improving the annotation of content fields to determine whether the literature involves improvements to the parameterization scheme and its improved parameters helps large models distinguish between conventional applications and scheme optimizations.

[0075] The parameterization scheme combination field directly maps the parameterization scheme options in the physical processes of the Noah-MP model in a structured manner (this embodiment has 12 physical processes, including: opt_dveg (vegetation model), opt_crs (canopy stomatal resistance scheme), opt_btr (stomatal resistance soil moisture correction scheme), opt_run (runoff and groundwater scheme), opt_sfc (surface drag coefficient scheme), opt_frz (supercooled water scheme), opt_inf (permafrost permeability coefficient scheme), opt_rad (radiative transfer scheme), opt_alb (snow surface albedo scheme), opt_snf (rainfall and snowfall classification scheme), opt_tbot (soil temperature boundary scheme), and opt_stc (snow cover and soil temperature time integration scheme). Each physical process has 2-4 optional parameterization schemes. This eliminates the semantic ambiguity that may be caused by natural language descriptions, enabling large models to accurately generate directly executable parameter configuration files and achieving a seamless connection from knowledge reasoning to model operation.

[0076] The applicable conditions and performance evidence fields together form a closed loop for decision verification. The former clarifies the applicable time, region, and underlying surface environment boundaries of the parameterized scheme, preventing the large model from recommending incorrect schemes in mismatched scenarios. The latter provides quantitative verification evidence such as correlation coefficients and root mean square errors, making the recommendation results of the large model interpretable and traceable, and enhancing users' trust in the intelligent decision-making results.

[0077] The above field structure transforms unstructured document knowledge into structured knowledge that machines can query and reason about, laying the foundation for subsequent intelligent parameter decision-making.

[0078] Steps 1-3: Knowledge Tuple Construction. Based on the knowledge tuple template defined in Step 1-2, information extraction and standardized annotation are performed on the dataset collected in Step 1-1. For PDF documents, a document reader is used to segment the data at the paragraph level, with a segment size of 1024 characters; for TXT text documents, a text reader is used to segment the data at the paragraph level, with a segment size of 4096 characters. This segmentation process breaks long documents into text fragments suitable for the embedding model, with each fragment retaining its complete semantic context, forming a standardized set of knowledge tuples.

[0079] Steps 1-4: Vectorized Storage and Index Construction. An open-source text embedding model is used to vectorize the knowledge tuple text fields generated in Steps 1-3, converting each text fragment into a 1024-dimensional dense vector. Formally, let the set of text fragments be... N is the total number of text fragments, obtained through an embedding model. The process of performing vectorization mapping can be represented as:

[0080] ;

[0081] in, For the generated dense vector, No. A text fragment (knowledge tuple). It is an open-source text embedding model (encoder). Let d be a real vector space. For vector dimensions (in this embodiment) ).

[0082] Specifically, this embodiment uses the text-embedding-v4 model provided by the DashScope platform as the text embedding model, with an output vector dimension of 1024. The vectorized knowledge tuples are persistently stored in the Qdrant vector database to construct a vector index based on cosine similarity. The Qdrant vector database is deployed as an independent service, providing vector storage and retrieval interfaces via the HTTP protocol, and supporting classification management by specifying set names. After completing the above steps, a knowledge base for land surface process parameterization schemes supporting semantic retrieval is formed.

[0083] Step 2: Build an intelligent agent system based on the AgentScope framework.

[0084] The multi-agent system comprises three cooperating sub-agents, specifically defined as follows:

[0085] The Planner is configured with first system prompts to receive natural language commands from the user, extract key environmental information of the study area such as geographical location, climate type, and underlying surface features, and convert it into standardized environmental feature descriptions, which are then output to the Retriever.

[0086] Retriever: Registered with a knowledge base retrieval tool interface, used to receive and parse the environmental features output by the agent, dynamically construct retrieval query statements, call the constructed vectorized knowledge base to perform semantic similarity matching, and forward the Top-K parameterized scheme combination returned by the retrieval to the Executor agent;

[0087] The Executor receives the parameterized scheme combination output by the retrieval agent, performs fusion inference based on the large language model, verifies the physical consistency of the parameterized scheme combination, and finally outputs a JSON format configuration result containing 12 parameterized scheme options.

[0088] The three sub-agents communicate with each other through the asynchronous message passing mechanism provided by the AgentScope framework, share a unified context session window, and work together to complete the inference action loop.

[0089] Step 2-1: Initialize the agent's runtime environment. Initialize the runtime environment of each sub-agent based on the open-source AgentScope framework, configuring the asynchronous message passing mechanism and tool invocation interface. Each sub-agent adopts a reasoning-action mode, which enables each sub-agent to autonomously plan its reasoning steps after receiving instructions, alternately executing a loop of thinking, using tools, and invoking observation until it completes its assigned sub-task. The multi-agent system achieves functional expansion through a tool registration mechanism, encapsulating functions such as knowledge base retrieval, parameter file operations, and system command invocation into independent tool functions, which are registered to the tool sets of the corresponding sub-agents for autonomous invocation.

[0090] Step 2-2: Configure the large language model. Configure a large language model as the inference engine for each agent. This embodiment supports two large language model interfaces: first, using the Qwen series models provided by the DashScope platform, called through the DashScope native interface; second, using a third-party large language model compatible with the OpenAI interface protocol, called through configuring a custom API endpoint address. The large language model is responsible for receiving system prompts and user messages, performing core inference tasks such as intent recognition, keyword extraction, knowledge fusion reasoning, and parameter decision-making, and outputting a structured response in JSON format containing the values ​​of 12 parameterized scheme options and the decision basis.

[0091] Steps 2-3: Register the knowledge base retrieval tool. The retrieval function of the parameterized scheme knowledge base constructed in Steps 1-4 is encapsulated as a tool function and registered to the agent toolset. The workflow of the retrieval tool is as follows: it receives the retrieval query string from the agent, calls the text embedding model to convert it into a 1024-dimensional vector representation, performs cosine similarity retrieval through the Qdrant vector database, sets the similarity threshold to 0.5, and returns the top K knowledge tuples with the highest similarity scores as the retrieval results, where the default value of K is set to 5 to 8.

[0092] Specifically, let the vector of the search query string be... The candidate knowledge tuple vector in the land surface process parameterization scheme knowledge base is: The formula for calculating the cosine similarity between the two is:

[0093] ;

[0094] in, Represents the query vector With knowledge tuple vector The cosine similarity between them ranges from 1 to 10. A larger value indicates greater similarity. The query vector is formed by embedding the retrieval query string (based on the search keywords set by the user in natural language input) into the model. The generated 1024-dimensional dense vector; Representing vectors and The dot product (inner product) is calculated using the following formula: ; and Representing vectors respectively and The L2 norm is calculated using the following formulas: and ; The dimension index of the vector, from 1 to 1024 (vector dimension in this embodiment) ); and They are query vectors and knowledge tuple vector In the Numerical components in each dimension.

[0095] Let the complete set of the knowledge base be... Then, after passing the threshold The final set of search results after filtering and Top-K sorting satisfy:

[0096] ;

[0097] in; This indicates selecting the top results based on similarity from highest to lowest. The operation of each result; the threshold can be dynamically set according to the knowledge base density or similarity distribution, and is set to 0.5 in this embodiment. The final result... This refers to the set of search results after threshold filtering and Top-K sorting.

[0098] The retrieval agent autonomously initiates multiple rounds of retrieval based on task complexity, using related words from the previous round as the current retrieval keywords. These related words include synonyms, near-synonyms, or higher-level concepts. The retrieval keywords are organized in the form of core terms separated by Chinese commas, until the values ​​of all parameterization schemes meet preset requirements (these requirements are the basis for selecting parameterization schemes). The retrieval results are returned to the agent in the form of a text list as a reference for subsequent fusion reasoning.

[0099] Steps 2-4: Configure Retriever agent system prompts. Set system prompts for the agent, defining its role, task objectives, and behavioral constraints, specifically including the following specifications:

[0100] (a) Core principle: All answers must be based on information obtained from the knowledge base of land surface process parameterization schemes using retrieval enhancement generation technology. It is prohibited to generate content directly without retrieving the knowledge base of land surface process parameterization schemes.

[0101] (b) Search guidelines: Search queries should only contain keywords related to the research background and simulation environment, separated by commas, and all search queries should be in Chinese;

[0102] (c) Decision criteria: All information must be considered to ensure that the parametric schemes that are related to each other in the parametric scheme combination meet the physical constraints. The physical constraints in this embodiment include the requirement that freeze-thaw related schemes must be used in permafrost areas, that dynamic vegetation schemes must be matched with the corresponding canopy stomatal resistance schemes, and that runoff schemes must be adapted to the hydrological characteristics of the study area.

[0103] (d) Output format requirements: Output the final result in JSON object format. The JSON object must contain 12 parameterized scheme option fields (opt_dveg, opt_crs, opt_btr, opt_run, opt_sfc, opt_frz, opt_inf, opt_rad, opt_alb, opt_snf, opt_tbot, opt_stc). All parameterized scheme options are integers and include a string field named reason to explain the selection criteria.

[0104] Step 3: User interaction and intelligent parameter decision-making, the steps of which are as follows:

[0105] Step 3-1: User Input. The user inputs natural language commands through the command-line interactive interface deployed on the server to describe the background environmental characteristics of the study area. The background environmental characteristics include, but are not limited to: the geographical location and latitude and longitude of the study area, the type of underlying surface (in this embodiment, the type of underlying surface includes farmland, grassland, forest, permafrost, etc.), climate type, vegetation cover, soil type, research objectives, and simulation spatiotemporal scale, etc.

[0106] Step 3-2: Constructing the agent's input message. The system reads the environmental parameter information of the current Noah-MP land surface process model, including the start and end dates, latitude and longitude, soil layer thickness parameters, initial values ​​of soil temperature and moisture content, land use dataset type, vegetation type index, and soil type index from the METADATA_NAMELIST section, as well as the current values ​​of 12 parameterization scheme options and the descriptions of the available schemes for each option in the parameter configuration section. The user background environment description and the driver file information are concatenated and assembled into a structured input message, which contains three parts: a user background environment description, a summary of the driver file content, and a JSON format output requirement description.

[0107] Step 3-3: Agent Reasoning. After receiving the input message constructed in step 3-2, the agent enters the reasoning stage and executes the following process:

[0108] (a) Thinking phase: The parsing agent analyzes the user's background environment description, identifies key environmental features of the study area, including geographical location, climate zone, underlying surface type, dominant physical processes, etc., and extracts atmospheric forcing elements such as wind speed and temperature from the driving file as auxiliary decision-making information. The parsing results are then sent to the retrieval agent.

[0109] (b) Retrieval phase: The retrieval agent receives and analyzes the environmental features output by the retrieval agent, constructs retrieval keywords based on them, calls the knowledge base retrieval tool registered in step 2-3, and retrieves historical cases and literature evidence that best match the current research area environment from the parameterization scheme knowledge base. The retrieval results include parameterization scheme combinations that have been verified to be effective in existing studies and their applicable conditions.

[0110] (c) Inference Phase: The retrieval agent integrates the results of multiple rounds of retrieval with the environmental parameters in the driver file, analyzes the environmental features output by the agent, performs fusion inference based on the reasoning capabilities of the large language model, outputs the recommended parameterized scheme combination, and forwards it to the execution agent.

[0111] (d) Output Phase: The executing agent performs physical constraint verification on the parameterized schemes, judging the physical rationality of each parameterized scheme option to ensure that the parameter combination meets the physical constraints, including the requirement to enable freeze-thaw related schemes in permafrost areas, the need for dynamic vegetation schemes to match the corresponding canopy stomatal resistance schemes, and the need for runoff schemes to adapt to the hydrogeological characteristics of the study area. If the verification passes: The executing agent generates a JSON object containing the final recommended values ​​for the 12 parameterized scheme options, and generates detailed explanations of the selection criteria in the reason field. If the verification fails: A restart command is sent to the parsing agent to restart the entire process.

[0112] The entire process of intelligent agent fusion reasoning and parameter decision-making can be abstracted into a multivariable mapping function. :

[0113]

[0114] in, This indicates the background environment described by the user. This indicates the information extracted from the driver file. This represents the prior knowledge of historical cases obtained through multiple rounds of retrieval. Output results. This represents a vector of optimal values ​​for 12 physical parameterization schemes, each consisting of integers. This is the corresponding logical reasoning text description.

[0115] Step 4: Automatic write-back of parameter files and model execution. The steps are as follows:

[0116] Step 4-1: Parse and execute the agent's output. The system extracts a JSON object from the text response returned by the agent, uses regular expression matching to locate the boundaries of the JSON string, and parses and extracts the integer values ​​of the 12 parameterized scheme options and the reason field. A validity check is performed on each parameterized scheme option to verify the completeness of the fields (all 12 required fields exist) and the correctness of the data types (all are integers).

[0117] Step 4-2: Driver File Parameter Write-back. The system reads the contents of the current driver file and locates the starting position of the parameterization scheme segment by scanning line by line. In this embodiment, the starting position is marked as "!MPOptions". For each line in the parameterization scheme segment, a regular expression is used for pattern matching. The matching pattern of the regular expression is: indentation + parameter name + equal sign + integer value + line end comment. After identifying the opt_* parameter lines (including opt_dveg, opt_crs, opt_btr, opt_run, opt_sfc, opt_frz, opt_inf, opt_rad, opt_alb, opt_snf, opt_tbot, opt_stc), the original values ​​are replaced with the new values ​​parsed in Step 4-1, while maintaining the original indentation format, parameter names, and line end comments. The scanning of the parameter configuration segment ends when a blank line or the Fortran name list terminator " / " is encountered. After replacing all physical scheme selection parameters, write the updated complete text back to the driver file, and record the change details of each physical scheme selection parameter (including the physical scheme selection parameter name, original value and new value).

[0118] Step 4-3: Invoke the Noah-MP land surface process model to perform the simulation. The system detects the pre-compiled Noah-MP executable program in the directory where the Noah-MP land surface process model is located. This executable program is compiled from Fortran source code and relies on the NetCDF library for runtime support. The system starts the executable program as a subprocess, passing the updated driver file path from Step 4-2 as a command-line parameter. Simultaneously, it captures the standard output and standard error streams for real-time monitoring of simulation progress and anomaly diagnosis. After the simulation is completed, the executable program outputs a simulation result file in NetCDF format, containing time-series data of key variables in land surface physical processes such as soil temperature, soil moisture content, sensible heat flux, latent heat flux, and runoff.

[0119] The operating mechanism of the Noah-MP land surface process model at this time can be equivalent to the following equation:

[0120] ;

[0121] in It is a collection of various land surface physical process state variables output by the model (Noah-MP land surface process model). This represents the global dynamic mapping operator for the Noah-MP land surface process model. Data representing external driving forces (such as meteorological elements) that change over time. These are the initial state parameters (such as initial values ​​of soil temperature and water content) for the Noah-MP land surface process model. Configure vectors for the 12 parameterized schemes injected after agent-optimized inference.

[0122] The system persistently records information such as user input, agent inference content, parameter configuration results, and simulation running status during the interaction process to a log file for process traceability and scientific repeatability verification.

[0123] Step 5: System Validation and Indicator Evaluation

[0124] To verify the scientific nature of the parameter configuration, the reliability of the simulation, and the ease of use of the system of this invention, three types of verification indicators—accuracy, reliability, and ease of use—are set up to quantitatively evaluate the parameter decision results and the model's performance.

[0125] Step 5-1: Accuracy Verification

[0126] This validation aims to evaluate whether the combination of parameterization schemes recommended by the agent system belongs to the set of optimal parameterization schemes obtained through exhaustive search. The Kling-Gupta Efficiency (KGE) index is used as the screening criterion, with the top 5% of scheme combinations by KGE value serving as the benchmark.

[0127] (a) Typical scenario selection: Several representative land surface process simulation scenarios were selected as test cases. The simulation scenarios cover different spatial scales (point scale, regional scale), underlying surface types (grassland, farmland, forest, permafrost), and climatic conditions (arid, humid, alpine) to ensure the universality of the evaluation results. Each simulation scenario is equipped with complete driving data and corresponding observation data (soil temperature and humidity, surface flux, etc. observed at the stations).

[0128] (b) Obtaining the Baseline Scheme Set – Exhaustive Screening Based on KGE: For each simulation scenario, an exhaustive method is used to traverse all possible combinations of parameterized schemes. The Noah-MP land surface process model contains 12 parameterized scheme options, each with 2 to 5 possible values, resulting in an exhaustive space of tens of thousands to hundreds of thousands of combinations. For each candidate scheme combination, the Noah-MP land surface process model is run under the same driving data, initial conditions, and boundary conditions. The simulation results are compared with the observed data time-step by time, and the KGE index is calculated as an evaluation index. The formula for calculating the KGE index is:

[0129] in, The Pearson correlation coefficient is the ratio between simulated and observed values, which measures the consistency of their trends. and These are the standard deviations of the simulated values ​​and the observed values, respectively. Their ratio reflects the deviation of the simulated variability from the observed variability. and These are the means of the simulated value and the observed value, respectively; their ratio reflects systematic bias. The KGE index ranges from [value missing]. The closer the value is to 1, the better the simulation effect.

[0130] For each simulation scenario, all candidate solutions are sorted from highest to lowest according to their KGE values, and the top 5% of solutions are selected to form the baseline solution set, denoted as . Each combination of parameterization schemes in this set represents a high-performance configuration under given observation constraints.

[0131] (c) Intelligent agent scheme generation: The environmental feature description of the test scenario is input into the system of this invention for processing. Each independent simulation generates N recommended combinations of parameterized schemes, denoted as the scheme to be tested, which are automatically generated by the agent. .

[0132] (d) Metric Calculation: Define Hit Rate The percentage of system-generated parameter schemes that fall into the high-performance parameter set is calculated using the following formula:

[0133] ;

[0134] in,

[0135] ;

[0136] In the formula, This is an indicator function. It takes the value 1 when the parameterized scheme generated by the system falls into the top 5% high-performance parameter set obtained by traversing the parameter space in advance; otherwise, it takes the value 0. The solution to be tested The i-th scheme to be tested is the i-th parameterized scheme combination.

[0137] Hit rate This is used to measure the ability of the parameter scheme generated by this system to enter the high-performance parameter region. The higher the value, the more accurately the system can provide high-performance parameterization schemes and obtain near-optimal Noah-MP parameter combinations without exhaustive calculations across the entire space, thus verifying the effectiveness and stability of the system in the parameter optimization of the Noah-MP land surface process model.

[0138] (e) Key parameterized scheme matching rate (optional auxiliary indicator): To further analyze the similarity between the agent scheme and the set of benchmark schemes in the configuration of core physical processes, the key parameterized scheme matching rate can be calculated. Based on the physical mechanisms of land surface process simulation, a set of key parameters that significantly affect the simulation results are pre-selected from 12 parameterization schemes (in this embodiment, the dynamic vegetation scheme is selected). Canopy stomatal resistance scheme Runoff and Groundwater Scheme Permafrost permeability scheme and radiative transfer schemes ).exist Extract the first A set of key parameterization schemes constitutes a numerical set. , X is the total number of key parameters, where if The first in The values ​​of the key parameterization schemes are in Medium; Matching rate Add 1, If the match rate exceeds the preset threshold, it indicates that the matching rate meets the requirements.

[0139] Step 5-2: Reliability Verification

[0140] This verification aims to evaluate the consistency between the simulation results of the Noah-MP land surface process model driven by the parameterization scheme configured by the intelligent agent system and real observation data, thereby verifying the scientific reliability of the simulation results output by the system of this invention.

[0141] (a) Model running: using the test scheme generated by the agent in step 5-1. Run the Noah-MP model under given driving data, initial conditions, and boundary conditions to obtain simulation results.

[0142] (b) Variable Extraction: Key diagnostic variables are extracted from the NetCDF format output file of the model, including: sensible surface heat flux, latent surface heat flux, soil temperature at different depths (e.g., 5cm, 20cm, 50cm), soil moisture content at different depths, surface runoff, and snowmelt equivalent. The corresponding spatiotemporal value sequences of these variables are extracted and denoted as follows: ,in This serves as a time step index. Simultaneously, it retrieves the actual observation data sequence for the corresponding time period. .

[0143] (c) Indicator Calculation: The following three statistical indicators are used to quantitatively evaluate the consistency between simulated and observed values:

[0144] (1) Correlation coefficient ( ): Used to measure the synchronicity of the changing trends between simulated and observed values, the calculation formula is:

[0145] in, is the mean of the sequence. The closer the value is to 1, the stronger the correlation between the simulation and the observation.

[0146] (2) Root Mean Square Error (RMSE): Used to quantify the average deviation between simulated and observed values. It is particularly sensitive to larger errors. The calculation formula is:

[0147] A smaller RMSE value indicates a smaller simulation error. RMSE has the same physical units as the original variable.

[0148] (3) Average bias: This measures whether the simulation results are systematically too high or too low. The calculation formula is:

[0149]

[0150] The closer the absolute value of Bias is to 0, the better the unbiasedness of the simulation. A positive bias indicates that the simulated value overestimates the observed value, while a negative bias indicates that it underestimates the observed value.

[0151] (d) Multi-scenario summary statistics: Calculate the above indicators for each of the multiple typical scenarios selected in step 5-1, and summarize the diagnostic variables for each scenario in tabular form. The mean and standard deviation of RMSE and Bias are used to comprehensively evaluate the reliability performance of the system of the present invention under different environmental conditions.

[0152] Step 5-3: Usability Verification

[0153] This verification aims to quantitatively evaluate the actual effectiveness of the system of the present invention in reducing the learning cost and operational threshold of land surface process models through actual user operation tests.

[0154] (a) User Recruitment: Recruit several users with different backgrounds to participate in the test. The users are divided into three groups: the first group consists of domain experts, the second group consists of people with a certain level of domain knowledge, and the third group consists of beginners. Each group should recruit at least 5 users.

[0155] (b) Task Design: Each user is assigned the same standardized land surface process simulation task. The task requires the user to complete the entire process from parameter configuration to model execution, and finally obtain the simulation results of soil temperature and humidity and surface fluxes for a specific region. The task description provides the geographical location, climate type, underlying surface characteristics and driving data path of the study area.

[0156] (c) Operational Procedure: Users were randomly assigned to a control group and an experimental group. The control group was required to complete the task using the traditional method: manually editing the driver file and parameter configuration file of the Noah-MP land surface process model, including modifying the values ​​of 12 parameterization scheme options, setting output variables, compiling the model, and executing the executable program. The experimental group was required to use the system of this invention to complete the task through the natural language interactive interface described in step 3-1: users only needed to describe the environmental characteristics of the study area in Chinese (which could be: permafrost region of the Qinghai-Tibet Plateau, grassland cover, average annual temperature -2℃, simulation period 2015-2016), and the system would automatically complete the subsequent parameter configuration and model operation.

[0157] (d) Indicator Collection and Calculation: Record and calculate the following three indicators:

[0158] (1) Task Completion Time: The total time, in minutes, from when the user begins executing the task (reading the task description) to successfully obtaining the required simulation results (NetCDF file generated and complete). This is recorded via system logs or manual timing. For the experimental group, this also includes the time spent by the user inputting a natural language description and confirming the system's recommended parameters.

[0159] (2) Number of operation steps: The number of key operation steps that the user needs to perform from the start to the completion of the task. In the traditional method, each file opening, parameter editing, saving, compilation command input, run command input, error checking, etc., is counted as one operation step; in the system of this invention, each round of natural language interaction (inputting a description statement, viewing the parameter recommendations returned by the system, and confirming execution) is counted as one operation step.

[0160] (3) Subjective satisfaction rating: After the task was completed, an anonymous questionnaire was distributed to each user using a 5-point Likert scale (1-very dissatisfied, 2-satisfied, 3-neutral, 4-satisfied, 5-very satisfied) to collect their subjective feelings about the system in terms of ease of use, learning curve, task completion efficiency, and result reliability. The mean scores and standard deviations of the control group and the experimental group were calculated separately.

[0161] Figure 2 This is a comparison curve of the simulated temperature results and observed values ​​under different schemes in this embodiment. The graph shows a time series comparison of daily temperature changes throughout the year, with the horizontal axis representing day sequence and the vertical axis representing temperature. It includes four curves: blue represents the actual observed values, serving as baseline verification data; red represents the simulation results using this system with RAG retrieval and containing specific information; green represents the simulation results using this system with RAG knowledge retrieval but without a single prompt containing a description of specific environmental information (a certain thickness of snow accumulation in the area); and purple represents the simulation results using this system without RAG knowledge retrieval and without prompts containing a description of specific environmental information (a certain thickness of snow accumulation in the area). Figure 2 It can be seen from the results that the simulation results of this system are in good agreement with the actual observations.

[0162] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A method for intelligent modeling of land surface processes based on Noah-MP, a knowledge base, and intelligent agents, characterized in that, Specifically: Collect multimodal data covering the Noah-MP land surface process model; extract the content of core fields from the multimodal data, construct knowledge tuples, perform vectorized storage and indexing of the knowledge tuples, and construct a knowledge base for parameterized schemes of land surface processes. The core fields include: research nature, research objectives, research area, underlying surface type, dominant physical process, improvement content, parameterized scheme combination, applicable conditions, and performance evidence. The parameterization scheme combination directly maps the parameters corresponding to the physical processes in the Noah-MP land surface process model in a structured manner; The study describes the types of knowledge tuples; The underlying surface type describes environmental characteristics and provides underlying surface constraints for parametric scheme combinations; The dominant physical process extracts the physical mechanism, providing logical constraints for the physical reasoning process of multi-agent systems; The applicable conditions clearly define the applicable time, region, and underlying surface environment boundaries of the parameterized scheme; Performance evidence provides quantitative verification, making the recommendation results of multi-agent systems interpretable and traceable; Improve the annotation of multimodal data to determine whether it involves improvements to the parameterization scheme and the improved parameters, so as to help multi-agent systems distinguish between conventional applications and optimized schemes; A multi-agent system is constructed, including a parsing agent, a retrieval agent, and an execution agent. The parsing agent reads the natural language input by the user and the model operation parameter information from the driver file of the Noah-MP land surface process model, and then identifies the geographical environment and model operation environment characteristics of the study area. The natural language is used to describe the geographical environment of the study area. The parsing results of the parsing agent are transmitted to the retrieval agent. The retrieval agent constructs keywords based on the received information, searches in the land surface process parameterization scheme knowledge base, and forwards the retrieved parameterization scheme combination to the execution agent. The execution agent performs physical constraint verification on the received parameterization scheme combination. The parameterization scheme combination is a combination of parameterization schemes corresponding to all physical processes in the Noah-MP land surface process model. The parameterization scheme combination that has passed the physical constraint verification by the execution agent is written back to the Noah-MP land surface process model to execute the simulation task. The parameterized schemes, which are verified by performing physical constraints on the intelligent agent, are written back into the Noah-MP land surface process model, specifically as follows: Step A: Extract the JSON object from the text returned by the executing agent, use regular expression matching to locate the string boundaries of the JSON object, parse the parameterization scheme and reason field in the JSON object, and verify the parameterization scheme. If the verification fails, extract it again; otherwise, go to step B. Step B: Read the current driver file, use line-by-line scanning to locate the marker at the start position of the parameterization scheme, identify the original parameterization scheme, replace the original parameterization scheme with the parameterization scheme in the searched parameterization scheme combination according to the original format, complete the replacement of all parameterization schemes, and update the driver file; Step C: Locate the directory containing the Noah-MP land surface process model, find the Noah-MP land surface process model executable program, pass the updated driver file as a command-line parameter to the executable program, capture the standard output and error output in real time during the Noah-MP land surface process model runtime, monitor the simulation progress, and troubleshoot any runtime anomalies.

2. The intelligent land surface process modeling method based on Noah-MP, knowledge base, and intelligent agent as described in claim 1, characterized in that, Information extraction and standardized annotation are performed on the collected multimodal data. For PDF documents, the data is divided into blocks at the paragraph level, with a block size of 1024 characters. For TXT documents, the data is divided into blocks at the paragraph level, with a block size of 4096 characters. The block processing involves splitting the document into text segments suitable for embedded model processing, with each segment retaining its semantic context. An embedding model is used to vectorize text fragments, converting each text fragment into a dense vector; Dense vectors are stored in a vector database, and a cosine similarity index is constructed to form a knowledge base for land surface process parameterization schemes that supports semantic retrieval.

3. The intelligent modeling method for land surface processes based on Noah-MP, a knowledge base, and an intelligent agent as described in claim 1, characterized in that, The parsing agent, retrieval agent, and execution agent communicate through the asynchronous message passing mechanism provided by the AgentScope framework, sharing a unified context session window. Each agent adopts the ReAct pattern, and each agent is configured with a large language model as the inference engine.

4. The intelligent modeling method for land surface processes based on Noah-MP, knowledge base, and intelligent agent as described in claim 3, is characterized in that... The retrieval agent is registered with a knowledge base retrieval tool, and system prompts are set for the retrieval agent, specifically including the following specifications: Core principle: All search results must be obtained from the land surface process parameterization scheme knowledge base based on search technology. It is prohibited to generate search results directly without searching the land surface process parameterization scheme knowledge base. Search criteria: Searches should only include keywords related to the geographical environment and model operating parameters of the study area; Decision-making criteria: The parameterized scheme should meet the preset physical constraints; Output format requirements: Output the final result in JSON object format. The final result includes the parameterization scheme combination of the Noah-MP land surface process model and the reason field for each parameterization scheme. The reason field is used to explain the selection basis. The retrieval agent performs multiple rounds of retrieval, and then combines the information parsed by the parsing agent to perform fusion reasoning, outputting a combination of parameterized schemes; the execution agent performs physical constraint verification on the combination of parameterized schemes, specifically: determining whether the related parameterized schemes meet the physical constraints; if the physical constraint verification passes, the simulation task is executed based on the parameterized scheme combination that has passed the physical constraint verification, otherwise a restart command is sent to the parsing agent to restart the entire process.

5. The intelligent land surface process modeling method based on Noah-MP, knowledge base, and intelligent agent as described in claim 1, characterized in that, If the parameterization scheme is of integer type and the combination of parameterization schemes is complete and without omissions, then the verification passes.

6. A system applied to the intelligent land surface process modeling method based on Noah-MP, knowledge base, and intelligent agent as described in claim 1, characterized in that, include: A multi-source data and knowledge input layer is used to collect multimodal data covering the Noah-MP land surface process model; The knowledge base construction layer is used to extract the core fields and corresponding values ​​from multimodal data, construct knowledge tuples, perform vectorized storage and index construction on the knowledge tuples, and construct a knowledge base for the parameterized scheme of land surface processes. The agent decision-making and execution layer includes a parsing agent, a retrieval agent, and an execution agent. The parsing agent reads the natural language input by the user and the model operation parameter information in the driver file of the Noah-MP land surface process model, and then identifies the geographical environment and model operation environment characteristics of the study area. The parsing results of the parsing agent are transmitted to the retrieval agent. The retrieval agent constructs keywords based on the received information, searches in the land surface process parameterization scheme knowledge base, and forwards the retrieved parameterization scheme combination to the execution agent. The execution agent performs physical constraint verification on the received parameterization scheme combination. The model execution, output, and evaluation layer is used to write back the parameterized scheme combination, which has been verified by performing physical constraints on the agent, to the Noah-MP land surface process model. The Noah-MP land surface process model performs simulation tasks and outputs results; and evaluates the land surface process system.

7. The system according to claim 6, characterized in that, The evaluation of the land surface process system includes verifying the accuracy of decision results, the reliability of the land surface process system, and its usability. Verification of the accuracy of decision results includes accuracy verification and matching rate verification: Accuracy Verification: For a specific region, several land surface process simulation scenarios were selected, covering all spatial scales, all underlying surface types, and all climate types. For each simulation scenario, an exhaustive method was used to traverse all parameterization scheme combinations. For each parameterization scheme combination, the Noah-MP land surface process model was run under the same driving file, initial conditions, and boundary conditions. The simulation results of the Noah-MP land surface process model were compared with actual observation data, and the Kling-Gupta efficiency coefficient (KGE) was calculated. For each simulation scenario, the parameterization scheme combinations were sorted from highest to lowest KGE value, and the top k parameterization scheme combinations were selected to form the baseline scheme set, denoted as [missing information]. ; Using the land surface process system, N parameterized scheme combinations are obtained for the same region and the same scenario. The set of these N parameterized scheme combinations is denoted as... Calculate the hit rate H; ； ； in, For the i-th parameterization scheme combination retrieved by the land surface process system, For indicator functions; If H is greater than the preset threshold, it indicates that the accuracy of the land surface process system decision results meets the requirements. Matching rate verification: X parameterization schemes corresponding to physical processes are manually selected as key parameterization schemes. Extract the first A set of key parameterization schemes constitutes a numerical set. , , among which, if The first in The values ​​of the key parameterization schemes are in Medium; Matching rate Add 1, If the match rate exceeds the preset threshold, it indicates that the matching rate meets the requirements. The reliability verification specifically involves using correlation coefficient, root mean square error, and average deviation for reliability verification. The usability verification specifically involves: recruiting several users with different backgrounds and randomly dividing them into a control group and an experimental group; assigning all users the same land surface process simulation task, which is for users to complete the entire process from parameter configuration to running the Noah-MP land surface process model; The control group completed the task using the traditional method: manually editing the driver files and environmental parameters of the Noah-MP land surface process model, including modifying the parameterization scheme, setting output variables, compiling the model, and executing the executable program. The experimental group used the land surface process system. Users described the environmental characteristics of the study area, and the land surface process system automatically completed the subsequent operations. Record and calculate three indicators: task completion time, number of operation steps, and subjective satisfaction score; calculate the mean score and standard deviation of the three indicators for the control group and the experimental group respectively.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements a land surface process intelligent modeling method based on Noah-MP, a knowledge base, and an intelligent agent as described in any one of claims 1 to 5.

Images