Environment chemical question and answer system and method based on large model fine tuning

Abstract

The application discloses an environment chemical question and answer system and method based on large model fine tuning, belongs to the technical field of cross of environment science, atmospheric chemistry and large language model, and comprises the following modules: an environment data retrieval and collection module, which is used for acquiring environment chemical multi-source data; an environment chemical mechanism knowledge base construction module, which is used for analyzing data and extracting chemical mechanism information, and constructing a structured knowledge base; a large model fine tuning module, which is used for efficiently fine tuning parameters of a pre-trained large language model based on the knowledge base; a chemical reasoning generation module, which is used for receiving user questions, calling the fine-tuned model to generate a chemical reaction chain and mechanism explanation through a prompt word package; an upper computer provides an interactive interface; and a lower computer executes a calculation task.The application can generate natural language mechanism explanations in real time without relying on experimental simulation cabins or chemical transmission models, significantly reduces analysis cost, realizes fast and interactive environment chemical mechanism analysis, and is suitable for atmospheric pollution cause diagnosis and regulatory decision support.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of environmental science, atmospheric chemistry, and large language models, and more specifically to an environmental chemistry question-and-answer system and method based on large model fine-tuning. Background Technology

[0002] With the continuous aggravation of environmental pollution, regional compound pollution has become a core challenge in atmospheric environmental management. In recent years, the concentration levels of typical pollutants such as ozone, volatile organic compounds (VOCs), and nitrogen oxides have exhibited characteristics such as multi-factor coupling, significant diurnal variation, and frequent regional transport, drawing significant attention to the pollutant formation mechanisms. The formation and transformation of these pollutants in the atmosphere involve complex free radical chain reaction processes, including HOx-NOx-RO2 chains, oxidation chains of olefins and aromatic hydrocarbons, multiphase NOx loss chains, and their feedback effects. These reactions are characterized by strong nonlinearity, strong time-varying nature, and high species coupling, and photochemical reaction chains often exhibit mutually promoting or inhibiting effects. Therefore, using pollutant observation data for pollution cause diagnosis and control strategy formulation has always been a major technical challenge for environmental scientists and management departments.

[0003] Most existing technical solutions analyze observational data through simulation or experiments. For example, an integrated atmospheric environment simulation chamber with temperature and humidity control, light regulation, gas circulation, and pollutant generation and sampling systems is constructed for studying the mechanisms of secondary atmospheric pollution. Another example is the use of WRF-CHEM to diagnose and analyze the impact of large point pollution sources on the regional atmospheric environment through parameterized comparison, emission inventory construction, and point source overlay simulation.

[0004] However, simulation chamber experiments require the construction of large, costly, sealed reaction spaces, along with controllable light sources, temperature and humidity control systems, and multiple sets of online monitoring instruments. They typically occupy a large area, have high construction and maintenance costs, and long experimental cycles, often taking several hours to days per experiment. Furthermore, simulation chambers require the manual setting of experimental scenarios such as initial pollutant concentrations, lighting conditions, and meteorological conditions, a process that is extremely time-consuming. Therefore, while simulation chamber experiments can reflect some photochemical reaction chains, they are difficult to directly use for rapid, scenario-based mechanistic analysis of real observation data. Moreover, the data output from simulation chamber experiments are mostly concentration change curves or reaction rate inferences, while CTM model outputs high-dimensional concentration fields, contribution matrices, or sensitivity indicators. These results all require secondary interpretation by atmospheric chemistry experts to infer reaction chains. Traditional methods lack the ability to directly generate structured chemical reaction chains from observational data and cannot explain pollution causes in natural language, thus failing to meet the needs of environmental management departments for convenient, intuitive, and real-time mechanistic interpretation.

[0005] Existing chemical transport models require high-precision emission inventories, boundary conditions, meteorological field simulations, and numerical solutions to hundreds to thousands of chemical reaction equations. Model operation places extremely high demands on computational resources, typically requiring multi-core servers or even clusters for extended periods of computation. Furthermore, the models involve uncertainties such as emission inventory quality, meteorological biases, and multiphase chemical parameters, necessitating repeated parameter tuning and post-processing by skilled technicians, making the overall process cumbersome and complex. Therefore, this approach struggles to achieve rapid response to observational data and real-time reaction chain analysis. Summary of the Invention

[0006] In view of this, the present invention provides an environmental chemistry question-and-answer system and method based on large model fine-tuning, which aims to solve the problem of difficulty in performing chemical mechanism-level reaction chain analysis on observation data in existing environmental chemistry research. In particular, it overcomes the shortcomings of traditional simulation chamber experiments, such as high cost, long cycle, and limited conditions, as well as the cumbersome simulation process, huge amount of computation, and inability to provide intuitive structured mechanism explanations of chemical transport models.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: On the one hand, this invention provides an environmental chemistry question-answering system based on large model fine-tuning, comprising: An environmental data retrieval and acquisition module is used to acquire multi-source data in the field of environmental chemistry and convert the multi-source data into a preset text format. The environmental chemical mechanism knowledge base construction module is used to analyze the acquired multi-source data, extract chemical mechanism information, and construct a structured environmental chemical mechanism knowledge base. The large model fine-tuning module is used to efficiently fine-tune the parameters of the pre-trained large language model based on the environmental chemical mechanism knowledge base, so as to obtain the optimized large language model. The chemical reasoning generation module receives natural language questions input by the user, calls an optimized large language model through prompt word packaging technology, and generates chemical reaction chains and mechanism explanations. The host computer is used to provide a user interface, receive natural language questions input by the user, and return the chemical reaction chain and mechanism explanation to the user in the form of natural language text. The lower-level machine communicates with the upper-level machine and is used to call the environmental data retrieval and acquisition module, the environmental chemical mechanism knowledge base construction module, the large model fine-tuning module, and the chemical reasoning generation module to perform calculation tasks.

[0008] Preferably, the acquisition of multi-source data in the field of environmental chemistry includes: Data is obtained from at least one of the following sources: public environmental monitoring databases, air quality monitoring platforms, and online academic literature platforms, using at least one of web crawlers and application programming interfaces (APIs).

[0009] Preferably, the environmental chemical mechanism knowledge base construction module extracts information on photochemical reaction chains, free radical cycle pathways, and precursor action mechanisms from multi-source data using natural language processing technology, and constructs them into structured knowledge pairs for mechanism explanation questions and answers.

[0010] Preferably, the large model fine-tuning module uses a multi-GPU parallel approach to fine-tune the parameters of the pre-trained model, and the multi-GPU parallel approach includes data parallelism, pipeline parallelism, or tensor parallelism.

[0011] Preferably, the parameter fine-tuning method includes one or more combinations of LoRA, AdaLoRA, QLoRA, Prefix-Tuning, and Prompt-Tuning.

[0012] Preferably, the generation of the chemical reaction mechanism explanation corresponding to the user query content includes: the optimized large language model generating a multi-step logical explanation in a chain reasoning manner, wherein the multi-step logical explanation includes at least one piece of information in the pollutant generation path, key intermediates, transformation relationships, and free radical cycle process.

[0013] On the other hand, the present invention provides a method for applying the above-mentioned environmental chemistry question-answering system based on large model fine-tuning, comprising the following steps: Automatically acquire multi-source data and convert the multi-source data into a preset text format; Chemical reaction mechanism information is extracted from the multi-source data to construct an environmental chemical mechanism knowledge base; Using the aforementioned environmental chemical mechanism knowledge base as training corpus, a parameter fine-tuning method is employed to train the pre-trained large language model, enabling the pre-trained large language model to possess the ability to infer chemical reaction chains and generate mechanism explanations. Receive user query content, construct inference template based on user query content, input the inference template into a fine-tuned large language model, and generate an explanation of the chemical reaction mechanism corresponding to the user query content; The explanation of the chemical reaction mechanism is returned to the user.

[0014] As can be seen from the above technical solutions, compared with existing technologies, this invention discloses a method that does not rely on experimental simulation chambers and chemical transport models. It directly generates chemical reaction chains from observational data through a knowledge base and fine-tuned large language model, eliminating the need for physical experimental equipment and cumbersome simulations, significantly reducing costs and analysis time, and meeting the needs for rapid diagnosis of pollution events. Traditional methods output numerical curves or concentration fields, requiring secondary interpretation by experts. This invention generates structured natural language interpretations, with intuitive, consistent, and reproducible results. Furthermore, the environmental chemical mechanism knowledge base of this invention can continuously and automatically expand its knowledge from the latest literature, covering actual observational data from different regions, time periods, and pollutant combinations, making it suitable for multi-scenario pollution event diagnosis. Through web front-end and GPU parallel inference, a mechanism explanation can be generated within seconds after the user inputs a question, achieving true real-time analysis and enabling rapid response to sudden pollution events. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the structure provided by the present invention.

[0017] Figure 2 The flowchart illustrates an environmental chemistry question-and-answer method based on large model fine-tuning provided by this invention. Detailed Implementation

[0018] The technical solutions of 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention discloses an environmental chemistry question-answering system based on large model fine-tuning, such as... Figure 1 As shown, it includes: The environmental data retrieval and acquisition module is used to acquire multi-source data in the field of environmental chemistry and convert the multi-source data into a preset text format. Specifically, it acquires data from at least one source, including public environmental monitoring databases, air quality monitoring platforms, and online academic literature platforms, through at least one method using web crawlers and application programming interfaces. The acquired data includes environmental monitoring data and related academic paper texts. The academic paper data can be converted into structured or semi-structured text formats using methods such as HTML parsing, PDF parsing, and text extraction. This data is then uniformly received and cached by the lower-level machine as the input data source for subsequently constructing the knowledge base of the mechanism.

[0020] The environmental chemistry mechanism knowledge base construction module is used to analyze acquired multi-source data, extract chemical mechanism information, and construct a structured environmental chemistry mechanism knowledge base. Specifically, it performs text parsing on acquired literature data, employing natural language processing technology to extract mechanistic information such as photochemical reaction chains, free radical cycle pathways, and precursor action mechanisms from the literature; and constructs this information into structured "mechanism explanation question-and-answer pairs" of knowledge, which are then written into the environmental chemistry mechanism knowledge base. This step can further improve the accuracy of the mechanistic knowledge through manual verification, expert rules, and mechanistic model calibration.

[0021] The large-scale model fine-tuning module is used to efficiently fine-tune the parameters of a pre-trained large language model based on the aforementioned environmental chemical mechanism knowledge base, obtaining an optimized large language model. Specifically, the large-scale model fine-tuning module employs a multi-GPU parallel approach to efficiently fine-tune the parameters of the pre-trained model. This module reads the constructed chemical mechanism knowledge base as training corpus and calls the GPU cluster connected to the lower-level machine to perform model fine-tuning. Fine-tuning methods may include, but are not limited to, LoRA, AdaLoRA, QLoRA, Prefix-Tuning, Prompt-Tuning, or combinations thereof to achieve efficient parameter updates. Training performance is significantly improved through strategies such as multi-GPU parallel training, pipelined parallelism, or tensor parallelism, enabling the pre-trained large language model to possess the ability to infer environmental chemical reaction chains, understand chain mechanisms, and explain multi-step causality. The fine-tuned model is stored on the lower-level machine for subsequent inference.

[0022] The chemical reasoning generation module receives user-input natural language questions and uses prompt word packaging technology to call an optimized large language model to generate chemical reaction chains and mechanistic explanations. Specifically, when a user submits a query, the host computer structures the user-input natural language question and optional observation data, constructs a prompt input template, and sends it to the slave computer via a communication interface. The slave computer calls the finely tuned large language model and deploys it on the GPU for parallel reasoning, generating mechanistic explanations including pollutant reaction chains, key intermediates, and free radical cycle pathways. This module supports chained reasoning and multi-step logical explanations, thereby enabling mechanistic-level analysis of complex pollution processes.

[0023] The host computer includes a user interface for receiving user input and returning analysis results; the slave computer is used to call core functions such as the environmental data retrieval module, the large model fine-tuning module, and the chemical mechanism inference module, and to execute data processing tasks. Specifically, the user interface outputs mechanism explanation results, realizing natural language question answering. The user interface is built through a front-end web application to achieve natural language interaction between the user and the system. Users can directly input queries into the interface based on observational data, self-described pollution event information, or questions about specific pollutants. After receiving the user input, the host computer communicates with the slave computer, encapsulates the user's question and relevant observational data into a structured request, and calls the chemical mechanism inference module for reasoning analysis. The chemical mechanism inference module generates a complete chemical mechanism-level reaction chain analysis based on the encapsulated input, including the pollutant formation pathway, intermediate transformation relationships, free radical cycle processes, and other mechanism explanations. Finally, the host computer returns the inferred mechanism explanation results to the user interface in natural language text form for the user to view and further analyze.

[0024] On the other hand, the present invention provides a method for applying to the above-mentioned environmental chemistry question-answering system based on large model fine-tuning, such as... Figure 2 As shown, it includes the following steps: Automatically acquire multi-source data and convert the multi-source data into a preset text format; Chemical reaction mechanism information is extracted from the multi-source data to construct an environmental chemical mechanism knowledge base; Using the aforementioned environmental chemical mechanism knowledge base as training corpus, a parameter fine-tuning method is employed to train the pre-trained large language model, enabling the pre-trained large language model to possess the ability to infer chemical reaction chains and generate mechanism explanations. Receive user query content, construct inference template based on user query content, input the inference template into a fine-tuned large language model, and generate an explanation of the chemical reaction mechanism corresponding to the user query content; The explanation of the chemical reaction mechanism is returned to the user.

[0025] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0026] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An environmental chemistry question-answering system based on large model fine-tuning, characterized in that, include: An environmental data retrieval and acquisition module is used to acquire multi-source data in the field of environmental chemistry and convert the multi-source data into a preset text format. The environmental chemical mechanism knowledge base construction module is used to analyze the acquired multi-source data, extract chemical mechanism information, and construct a structured environmental chemical mechanism knowledge base. The large model fine-tuning module is used to efficiently fine-tune the parameters of the pre-trained large language model based on the environmental chemical mechanism knowledge base, so as to obtain the optimized large language model. The chemical reasoning generation module receives natural language questions input by the user, calls an optimized large language model through prompt word packaging technology, and generates chemical reaction chains and mechanism explanations. The host computer is used to provide a user interface, receive natural language questions input by the user, and return the chemical reaction chain and mechanism explanation to the user in the form of natural language text. The lower-level machine communicates with the upper-level machine and is used to call the environmental data retrieval and acquisition module, the environmental chemical mechanism knowledge base construction module, the large model fine-tuning module, and the chemical reasoning generation module to perform calculation tasks.

2. The environmental chemistry question-answering system based on large model fine-tuning according to claim 1, characterized in that, The acquisition of multi-source data in the field of environmental chemistry includes: Data is obtained from at least one of the following sources: public environmental monitoring databases, air quality monitoring platforms, and online academic literature platforms, using at least one of web crawlers and application programming interfaces (APIs).

3. The environmental chemistry question-answering system based on large model fine-tuning according to claim 1, characterized in that, The environmental chemical mechanism knowledge base construction module extracts information on photochemical reaction chains, free radical cycle pathways, and precursor action mechanisms from multi-source data using natural language processing technology, and constructs them into structured knowledge in the form of mechanism explanation question-and-answer pairs.

4. The environmental chemistry question-answering system based on large model fine-tuning according to claim 1, characterized in that, The large model fine-tuning module uses a multi-GPU parallel approach to fine-tune the parameters of the pre-trained model. The multi-GPU parallel approach includes data parallelism, pipeline parallelism, or tensor parallelism.

5. An environmental chemistry question-and-answer system based on large model fine-tuning according to claim 4, characterized in that, The parameter fine-tuning methods include one or more combinations of LoRA, AdaLoRA, QLoRA, Prefix-Tuning, and Prompt-Tuning.

6. The environmental chemistry question-answering system based on large model fine-tuning according to claim 1, characterized in that, The generation of the chemical reaction mechanism explanation corresponding to the user query content includes: the optimized large language model generates a multi-step logical explanation in a chain reasoning manner, and the multi-step logical explanation includes at least one piece of information in the pollutant generation path, key intermediates, transformation relationship and free radical cycle process.

7. A method for applying an environmental chemistry question-answering system based on large model fine-tuning as described in any one of claims 1 to 6, characterized in that, Includes the following steps: Automatically acquire multi-source data and convert the multi-source data into a preset text format; Chemical reaction mechanism information is extracted from the multi-source data to construct an environmental chemical mechanism knowledge base; Using the aforementioned environmental chemical mechanism knowledge base as training corpus, a parameter fine-tuning method is employed to train the pre-trained large language model, enabling the pre-trained large language model to possess the ability to infer chemical reaction chains and generate mechanism explanations. Receive user query content, construct inference template based on user query content, input the inference template into a fine-tuned large language model, and generate an explanation of the chemical reaction mechanism corresponding to the user query content; The explanation of the chemical reaction mechanism is returned to the user.

Images