A multi-agent driven robotic artificial intelligence chemist system

Abstract

The present invention discloses a multi-agent driven robot artificial intelligence chemist system, which belongs to the field of robot chemical experiment system, including: a task manager agent, which can coordinate other agents to collaboratively complete the corresponding chemical experiment operation according to the chemical experiment target description input by the researcher; a document reader agent, which can mine the document database to accumulate knowledge related to the given chemical experiment operation; an experiment designer agent, which can receive the expected experimental instructions of the corresponding chemical experiment operation of the task manager agent, plan the experimental task, and use the experimental template library to generate the chemical experiment program step by step; a robot operator agent, which can convert the chemical experiment program into codes and commands for operating the robots and workstations of the automated laboratory, so that they can execute and complete the corresponding chemical experiment operation according to the codes and commands. The system can perform complex, multi-step chemical experiments with minimal human intervention, and improve the efficiency of chemical experiments.

Description

Technical Field

[0001] The present invention relates to the technical field of robot chemical experiment systems, and in particular to a multi-agent driven robot artificial intelligence chemist system. Background Art

[0002] Currently, "agents" driven by large language models (LLMs) provide robust capabilities in natural language processing, autonomous task execution, and collaborative problem solving. Their integration with external tools and platforms enables them to perform a wide range of tasks in different fields, including chemistry, where they are revolutionizing the research and discovery process. One of the foreign research examples is ChemCrow, a GPT-4-driven agent designed to simplify the reasoning process of various common chemical tasks, including organic synthesis, drug discovery, and material design. It operates by prompting GPT-4 in turn, guiding it to reason about the current state of the task, consider its relevance to the ultimate goal, and plan the next step accordingly. ChemCrow has proven to be an effective assistant for expert chemists, while also lowering the entry barrier for non-experts by providing a simple interface to obtain accurate chemical knowledge. More generally, an LLM-driven agent is an autonomous system that uses the ability of LLM to interpret, process, and respond to text input in a way that mimics human-like understanding and reasoning. Their integration with external tools and platforms enables them to deliver powerful capabilities in both autonomous task execution and collaborative problem solving, thereby increasing productivity, accelerating innovation, and making advanced research tools more accessible.

[0003] In parallel with the rapid adoption of large language models in chemical and material research, robotics has also made significant progress in laboratory automation, providing different systems to meet the complex needs of autonomous chemical research. These platforms include, but are not limited to, automated platforms that handle high-throughput experiments, table-top fixed robotic arms that perform precise operations, and mobile robots that navigate in laboratory environments to perform and coordinate complex tasks. More complex devices include dual-arm robots that can perform complex, coordinated tasks that mimic the dexterity of human chemists, and multi-robot systems that can seamlessly collaborate to optimize workflow efficiency and output. Naturally, the integration of automated robotic hardware with LLM technology has become a new research direction for autonomous chemical research. For example, Boiko et al. in foreign research example 2 established Co-scientist, a GPT-4-driven artificial intelligence system that is able to semi-autonomously design, plan, and execute experiments through Internet and document search, code execution, and experimental automation.

[0004] Despite these recent advances in autonomous chemistry research driven by large language models, a challenge arises when faced with complex experimental scenarios such as multi-step, multi-laboratory workstation, and / or multi-robot setups: How can large language models be effectively leveraged to build an automated system that can expertly perform complex chemistry tasks? It is impractical to assign overly complex tasks to an LLM in a single request. Similarly, in autonomous chemistry, it is impractical to prompt an LLM to perform a series of experimental tasks at once or in a single conversation, which may involve interpreting a given problem, generating a hypothesis to test, planning an experimental procedure, executing instructions for the robot, adjusting hypotheses, making decisions for subsequent experiments, and so on.

[0005] In view of this, the present invention is proposed. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-agent driven robotic artificial intelligence chemist system that can achieve collaboration and coordination between agents to jointly handle complex chemical experimental operations and achieve certain chemical experimental goals, improve the effectiveness of a single large language model agent, perform complex, multi-step chemical experiments with minimal human intervention, improve chemical experiment efficiency, meet different chemical experiment needs, and thereby solve the above-mentioned technical problems existing in the prior art.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] A multi-agent driven robotic artificial intelligence chemist system, comprising:

[0009] Task manager agent, literature reader agent, experiment designer agent, robot operator agent, literature database, experiment template library and automated laboratory; among them,

[0010] The task manager agent, document reader agent, experiment designer agent, and robot operator agent are all agents based on a large language model;

[0011] The task manager agent is respectively connected to the document reader agent, the experiment designer agent and the robot operator agent in communication, and can interact with the human researcher, and coordinate the document reader agent, the experiment designer agent and the robot operator agent to collaboratively complete the chemical experiment operation corresponding to the target chemical task according to the description information of the target chemical task input by the human researcher;

[0012] The document reader agent is connected to the document database for communication, can receive keywords related to the target chemical task given by the task manager agent, search the document database according to the keywords, mine out knowledge related to the target chemical task, and transmit the knowledge related to the target chemical task to the task manager agent;

[0013] The experiment designer agent is connected to the experiment template library in communication, can receive the knowledge related to the target chemical task excavated by the document reader agent and forwarded by the task manager agent, and generate chemical experiment instructions corresponding to the target chemical task according to the knowledge, plan chemical experiment operations, and generate chemical experiment procedures step by step using the experiment template library, and transmit the chemical experiment procedures to the task manager agent;

[0014] The robot operator intelligent agent is communicatively connected to the automated laboratory, and can receive the chemical experiment program generated by the experiment designer intelligent agent and forwarded by the task manager intelligent agent, and convert the chemical experiment program into code for operating the robot and workstation of the automated laboratory, so that the robot and workstation execute according to the code to complete the chemical experiment operation corresponding to the target chemical task.

[0015] Compared with the prior art, the multi-agent driven robotic artificial intelligence chemist system provided by the present invention has the following beneficial effects:

[0016] By setting up task manager agents, literature reader agents, experiment designer agents and robot operator agents, the task manager agent collaboratively controls the literature reader agents, experiment designer agents and robot operator agents to control the literature database, experiment template library and automated laboratory respectively. Through the formation of this hierarchical multi-agent system based on a large language model, collaboration and coordination between agents are achieved. Each agent has a unique role and related capabilities, and the task manager agent amplifies the effectiveness of a single large language model agent, enabling them to jointly handle complex tasks and achieve certain chemical experiment operation goals, which can well meet the needs of different chemical experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0018] Figure 1A schematic diagram of the structure of a multi-agent driven robotic artificial intelligence chemist system provided in an embodiment of the present invention.

[0019] Figure 2 A schematic diagram of the composition of the literature reader agent of the robotic artificial intelligence chemist system provided in an embodiment of the present invention.

[0020] Figure 3 A schematic diagram of the composition of the experiment designer agent of the robot-human-robot artificial intelligent scientist system provided in an embodiment of the present invention.

[0021] Figure 4 A schematic diagram of the composition of the robot operator agent of the robotic artificial intelligence chemist system provided in an embodiment of the present invention.

[0022] Figure 5 A schematic diagram of the composition of the computing executor agent of the robotic artificial intelligence chemist system provided in an embodiment of the present invention.

[0023] Figure 6 Schematic diagram of experimental results of the robotic artificial intelligence chemist system processing Task 1 provided in Example 2 of the present invention.

[0024] Figure 7 Schematic diagram of the experimental results of the robotic artificial intelligence chemist system processing Task 2 provided in Example 3 of the present invention.

[0025] Figure 8 This is an experimental flow chart of the robotic artificial intelligence chemist system processing task 3 provided in Example 4 of the present invention.

[0026] Fig. 9 Schematic diagram of the experimental results of the robotic artificial intelligence chemist system processing Task 3 provided in Example 4 of the present invention. DETAILED DESCRIPTION

[0027] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in combination with the specific content of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments, which does not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0028] First, the terms that may be used in this article are explained as follows:

[0029] The term “and / or” means that either or both of them can be realized at the same time. For example, X and / or Y means both “X” or “Y” and “X and Y”.

[0030] The terms "include", "comprises", "contains", "has" or other descriptions with similar semantics should be interpreted as non-exclusive inclusion. For example, including certain technical feature elements (such as raw materials, components, ingredients, carriers, dosage forms, materials, dimensions, parts, components, mechanisms, devices, steps, procedures, methods, reaction conditions, processing conditions, parameters, algorithms, signals, data, products or products, etc.) should be interpreted as including not only certain technical feature elements explicitly listed, but also other technical feature elements known in the art that are not explicitly listed.

[0031] The term "consisting of..." means excluding any technical feature elements not explicitly listed. If this term is used in a claim, it will make the claim closed, so that it does not contain technical feature elements other than the technical feature elements explicitly listed, except for the conventional impurities related to them. If this term only appears in a clause of a claim, it only limits the elements explicitly listed in the clause, and the elements recorded in other clauses are not excluded from the overall claim.

[0032] Unless otherwise specified or limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this article can be understood according to specific circumstances.

[0033] The orientation or position relationship indicated by terms such as "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc. are based on the orientation or position relationship shown in the drawings and are only for the convenience and simplification of description, and do not explicitly or implicitly indicate that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation of this document.

[0034] The scheme provided by the present invention is described in detail below. The contents not described in detail in the embodiments of the present invention belong to the prior art known to professionals in the field. If the specific conditions are not specified in the embodiments of the present invention, they are carried out according to the conventional conditions in the field or the conditions recommended by the manufacturer. If the manufacturers are not specified in the reagents or instruments used in the embodiments of the present invention, they are all conventional products that can be purchased commercially.

[0035] like Figure 1 As shown, an embodiment of the present invention provides a multi-agent driven robotic artificial intelligence chemist system, comprising:

[0036] Task manager agent, literature reader agent, experiment designer agent, robot operator agent, literature database, experiment template library and automated laboratory; among them,

[0037] The task manager agent, the document reader agent, the experiment designer agent, and the robot operator agent are all agents based on a large language model;

[0038] The task manager agent is respectively connected to the document reader agent, the experiment designer agent and the robot operator agent in communication, and can interact with the human researcher, and coordinate the document reader agent, the experiment designer agent and the robot operator agent to collaboratively complete the chemical experiment operation corresponding to the target chemical task according to the description information of the target chemical task input by the human researcher;

[0039] The document reader agent is connected to the document database for communication, can receive keywords related to the target chemical task given by the task manager agent, search the document database according to the keywords, mine out knowledge related to the target chemical task, and transmit the knowledge related to the target chemical task to the task manager agent;

[0040] The experiment designer agent is connected to the experiment template library in communication, can receive the knowledge related to the target chemical task excavated by the document reader agent and forwarded by the task manager agent, and generate chemical experiment instructions corresponding to the target chemical task according to the knowledge, plan chemical experiment operations, and generate chemical experiment procedures step by step using the experiment template library, and transmit the chemical experiment procedures to the task manager agent;

[0041] The robot operator intelligent agent is communicatively connected to the automated laboratory, and can receive the chemical experiment program generated by the experiment designer intelligent agent and forwarded by the task manager intelligent agent, and convert the chemical experiment program into code for operating the robot and workstation of the automated laboratory, so that the robot and workstation execute according to the code to complete the chemical experiment operation corresponding to the target chemical task.

[0042] Preferably, in the above system, the large language model used by the task manager agent, the document reader agent, the experiment designer agent, and the robot operator agent is any one of an open source large language model and a commercial large language model, preferably an open source large language model Llama-3.1-70B;

[0043] The target chemical tasks include: preparation and measurement chemical experiment tasks, exploration and screening chemical experiment tasks. Among them, the preparation and measurement chemical experiment tasks refer to the synthesis of the required compounds and / or the characterization of the given compounds; the exploration and screening chemical experiment tasks refer to the exploration and / or screening of the experimental parameters of the required functional materials.

[0044] Preferably, in the above system, the literature database is a database containing titles and abstracts of a plurality of scientific publications in the disciplines of chemistry and materials science;

[0045] like Figure 2 As shown, the document reader agent is provided with a document retrieval tool and a document mining tool respectively. The document reader agent can receive keywords related to the target chemical task provided by the task manager agent, and retrieve and collect publication titles and abstracts related to the target chemical task from the document database using the received keywords through the document retrieval tool; and transmit the retrieved and collected publication titles and abstracts to the document mining tool, which uses an unsupervised syntactic distance analysis method to mine the data in the documents corresponding to the publication titles and abstracts, extract information on chemical substances, physicochemical properties, chemical and / or material functions, and experimental conditions from the data, and generate statistical data from the extracted information as prior knowledge for designing chemical experimental operations, and transmit the statistical data as knowledge related to the target chemical task to the task manager agent.

[0046] Preferably, in the above system, the experiment template library includes chemical experiment program templates of historical experiments in the automated laboratory and various configuration protocols of various workstations in the automated laboratory, and the various configuration protocols are stored in XML format, and the experimental steps and parameters of the chemical experiment program templates are described in natural language;

[0047] like Figure 3 As shown, the experiment designer agent includes: a protocol writer and a protocol reviewer; wherein,

[0048] The protocol writer is connected to the task manager agent in communication, is an agent based on a large language model, and is provided with a protocol search tool and a site query tool. It can call the protocol search tool to search for the best matching chemical experiment program template in the experimental template library according to the description information of the target chemical task received from the task manager agent. If the best matching chemical experiment program template is found, the best matching chemical experiment program template is referenced, and the experimental steps and parameters of the chemical experiment program template are adjusted to formulate an experimental program for the chemical experiment operation corresponding to the target chemical task; if there is no existing chemical experiment program template, the site query tool is called to obtain a site list of workstations that can be used in the automated laboratory, and the experimental workflow is planned according to the site list, and a chemical experiment program for the chemical experiment operation corresponding to the target chemical task is generated based on the large language model function;

[0049] The protocol reviewer, which is in communication connection with the protocol writer, is an intelligent agent based on a large language model, which can review, criticize and improve the chemical experiment program derived by the protocol writer according to predefined program checking expert rules, and output the final chemical experiment program in a format unified with the chemical experiment program template.

[0050] Preferably, in the above system, the program checking expert rules used by the protocol reviewer are as follows:

[0051] (1) The first step in checking the chemical experiment procedure is to "retrieve the rack from the sample rack workstation";

[0052] (2) Check that the last step in the chemical experiment procedure is "Put the rack into the sample rack workstation";

[0053] (3) After adding the sample, ensure that all experimental parameters do not exceed the maximum capacity of the sample bottle during the process;

[0054] (4) The experimental parameters of the solid dispensing workstation and the liquid dispensing workstation are in units of mass;

[0055] (5) The workstation in front of the gas chromatography workstation is the packaging station;

[0056] (6) Before using the electrocatalytic workstation, make sure there is liquid in the test bottle;

[0057] (7) Before using the UV-Vis spectroscopy workstation, make sure there is liquid in the test bottle.

[0058] Preferably, in the above system, the automated laboratory comprises: a plurality of robots, a plurality of automated workstations and a robot application interface; wherein the plurality of robots comprise: a mobile robot and a desktop robot;

[0059] The plurality of automated workstations include: a sample rack workstation, a liquid dispensing workstation, a solid dispensing workstation, a magnetic stirring workstation, a centrifugal workstation, a calcination workstation, a liquid aspiration workstation, a drying workstation, an ultrasonic workstation, a packaging workstation, a photocatalytic workstation, a gas chromatography workstation, a sample preparation workstation, a Raman spectroscopy workstation, an infrared spectroscopy workstation, a UV-visible spectroscopy workstation, a fluorescence spectroscopy workstation, an X-ray diffraction workstation, a laser-induced breakdown spectroscopy workstation, and an electrocatalytic workstation;

[0060] The robot application interface includes: a visual positioning application interface, a motion planning application interface, an experimental operation application interface and an instrument communication application interface;

[0061] like Figure 4 As shown, the robot operator agent includes: a code writer, a code reviewer and a code proofreader; wherein,

[0062] The code writer is respectively connected to the robot application interface and the code reviewer of the automated laboratory, is an intelligent agent based on a large language model, and is provided with a robot application interface query tool. It can obtain the usable robot application interface of each robot in the automated laboratory through the robot application interface query tool according to the chemical experiment program generated by the received experiment designer intelligent agent, and write a robot operation code corresponding to the chemical experiment program by using the function of the large language model according to the received chemical experiment program and the usable robot application interface queried;

[0063] The code reviewer, which is in communication connection with the code proofreader, is an intelligent agent based on a large language model and can check, correct, criticize and improve the code written by the code writer according to the code checking expert rules;

[0064] The code proofreader is respectively communicated with the robot application interface and the code reviewer of the automation laboratory. It is an intelligent agent based on a large language model. It can use the robot application interface query tool to query and obtain the usable robot application interface of each robot in the automation laboratory, proofread the program code received from the code reviewer, and transmit the proofread code to the automation laboratory, and the robot in the automation laboratory will execute the proofread code.

[0065] Preferably, in the above system, the code checking expert rules are as follows:

[0066] (1) Visual positioning must be performed before operating at a new workstation;

[0067] (2) After calling the 'locate()' function, the previous operation posture obtained by the 'locate()' function will be invalid;

[0068] (3) If the workstations operated in two adjacent steps are the same, there is no need to repeat movement and visual positioning;

[0069] (4) Only the solid dispensing workstation, liquid dispensing workstation, centrifuge workstation, encapsulation workstation, photocatalytic workstation and gas chromatography workstation need to place test bottles; other workstations should directly place sample racks as required;

[0070] (5) Each sample rack can hold 10 test bottles, and each synthesis experiment corresponds to a specific test bottle; please do not change the bottle number to avoid errors; be careful when calling functions that require specifying the bottle number, because the numbering starts from 0;

[0071] (6) Rack numbering also starts at 0, and each synthesis experiment corresponds to a test vial in a specific rack; ensure consistency in rack numbering throughout the program to avoid errors;

[0072] (7) The experimental parameters of the solid dispensing workstation and the liquid dispensing workstation must be expressed in mass units;

[0073] (8) The time parameters of all workstations, including magnetic stirring workstations and centrifuge workstations, must be in seconds;

[0074] (9) Calling the 'wait()' function takes significantly longer than calling other functions; however, depending on the restrictions, you still need to call the 'wait()' function when necessary;

[0075] (10) The liquid dispensing workstation has a capacity of 10 channels. Adjacent liquid sampling operations do not require repeated picking up and placing of test bottles. Only instrument communication is required to change the liquid type and continue sampling.

[0076] (11) You need to strictly follow the given chemical experiment operation sequence and minimize the running time.

[0077] Preferably, the above system further comprises: a computational executor agent and a theoretical model library; wherein,

[0078] The computational executor agent is respectively communicatively connected with the task manager agent and the theoretical model library, and is an agent based on a large language model. It can search for a suitable pre-trained model in the theoretical model library according to the relevant information of the target chemical task transmitted by the task manager agent, and fuse the pre-trained model with the experimental data obtained by the robot operator agent through model fusion to obtain a fusion model, and optimize the obtained fusion model through Bayesian optimization to obtain optimized experimental data, and transmit the obtained optimized experimental data to the task manager agent for subsequent operations of the robot operator agent.

[0079] Preferably, in the above system, the theoretical model library stores pre-trained machine learning models of historical publications, machine learning models developed for robotic artificial intelligence chemists, and open source machine learning models;

[0080] like Figure 5 As shown, the computing executor agent includes: a computing task manager agent, a model search tool, a model fusion tool, a Bayesian optimizer and a deep learning computing platform; wherein,

[0081] The computing task manager agent is respectively communicated with the task manager agent, the model search tool, the deep learning computing platform, the model fusion tool and the Bayesian optimizer. It is an agent based on a large language model and can receive input information sent by the task manager agent. The input information includes: keywords of the target chemical task, detailed description information of the target chemical task and experimental result information of the robot operator agent. The model search tool is called to search the theoretical model library with keywords to obtain a suitable pre-trained machine learning model. The model fusion tool is called to fuse the pre-trained machine learning model with the experimental data obtained by the robot operator agent to obtain a fusion model. The Bayesian optimizer is called to optimize the fusion model and generate a machine learning code. The deep learning computing platform receives and executes the generated machine learning code to obtain the prediction results of the chemical experiment operation parameters corresponding to the target chemical task as the optimized experimental data, and the optimized experimental data is transmitted to the task manager agent for subsequent operations of the robot operator agent.

[0082] Preferably, in the above system, the target chemical task is a chemical experiment task of discovery and optimization, and the chemical experiment task of discovery and optimization refers to discovering functional materials by combining literature mining, computational modeling and closed-loop optimization.

[0083] In summary, the multi-agent driven robotic artificial intelligence chemist system of the embodiment of the present invention constitutes a hierarchical multi-agent system based on a large language model by setting up a task manager agent, a document reader agent, an experiment designer agent, a robot operator agent and a computational executor agent. These agents - document readers, experiment designers, robot operators and computational executors - receive text instructions, perform specific operations (for example, using external tools, writing codes, etc.), and generate outputs in a suitable format, which can be directly used as input instructions for other agents or interpreted by human researchers. The task manager agent collaboratively controls the document reader agent, The experimental designer agent, the robot operator agent and the computational executor agent respectively control and / or utilize the literature database, the experimental template library, the automated laboratory and the theoretical model library to realize the collaboration and coordination among multiple agents. Each agent has a unique role and related capabilities. In this way, the task manager agent coordinates other agents with specific roles, cooperates with the integrated external software tools and the robot hardware to autonomously plan, manage and execute complex chemical experimental tasks. The task manager agent amplifies the effectiveness of a single large language model agent, enabling them to jointly handle complex tasks and achieve determined chemical experimental operation goals, which can well meet different chemical experimental needs. The system of the present invention can meet the performance requirements and achieve good results in handling the following three types of experimental tasks with increasing complexity: (1) performing the synthesis of the desired compound and / or the characterization of the given compound, which can be called the chemical experimental task of manufacturing and measurement; (2) exploring and / or screening the experimental parameters of the desired functional material, which can be called the chemical experimental task of exploration and screening; (3) combining literature mining, computational modeling and closed-loop optimization to discover functional materials, which can be called the chemical experimental task of discovery and optimization.

[0084] In order to more clearly demonstrate the technical solution and technical effects provided by the present invention, the solution provided by the embodiment of the present invention is described in detail with specific embodiments below.

[0085] Example 1

[0086] This embodiment provides a multi-agent driven robotic artificial intelligence chemist system, which is based on the open source Llama-3.1-70B as the basis of the large language model. The overall architecture of the system is as follows: Figure 1As shown, the task manager agent serves as the central coordinating entity, which interacts with human researchers and communicates with several role-specific agents under management - literature reader agents, experiment designer agents, robot operator agents, and computational executor agents. The task manager agent has its own preset system prompts, in which its roles, tasks, specific functions and interface descriptions, as well as message formats for workflow and communication are defined for all entities it can use (including lower-level agents and their tools). By leveraging the power of the large language model, the task manager agent interacts with human researchers in natural language. It interprets requirement descriptions, plans and decomposes tasks, and instructs and manages role-specific agents to perform specified operations using the corresponding basic resources, facilitating automated experimental workflows described by requirements. Each role-specific agent has its own large language model instance and built-in tools that operate on one of the four basic resources of the system, see Figure 1 . The literature reader agent receives keywords related to the target chemical task from the task manager agent, searches the literature database, and mines knowledge related to the target chemical task from the literature data. The experiment designer agent receives instructions about the intended experiment from the task manager agent, plans the experimental task, and generates the chemical experiment program step by step by utilizing the experimental template library. The robot operator agent converts the experimental program of the experiment designer agent into codes and commands for operating the robots and workstations of the automated laboratory, facilitating and completing the execution of robotic chemical experiments. The computational executor agent searches for suitable pre-trained machine learning models in the theoretical model library, enhances the model by integrating the experimental results of the robot operator agent, and performs Bayesian optimization on this basis to obtain the parameter optimization results corresponding to the target chemical task.

[0087] Furthermore, the composition and function of the document reader agent in the above system are as follows:

[0088] The literature database used by the Literature Reader agent is a local literature database constructed, which contains the titles and abstracts of approximately 1.2 million scientific publications in the disciplines of chemistry and materials science, downloaded from Crossref. These publications cover a wide variety of research topics, themes, and fields. The Literature Reader agent mines the literature database and analyzes the resulting data to extract knowledge.

[0089] The composition of the document reader agent is as follows Figure 2As shown in the figure, there are two built-in tools: (1) a literature search tool, which can be called LiteratureSearch, which can search the literature database based on the input keywords and return relevant literature data; (2) a literature mining tool, which can be called LiteratureMine, which uses unsupervised syntactic distance analysis to mine literature data and can extract information such as chemical substances, physicochemical properties, chemical / material functions, and experimental conditions. The workflow of the literature reader agent is as follows: it receives keywords from the task manager agent and then uses the literature search tool to collect relevant publication titles and abstracts. These data are then passed to the literature mining tool, which processes the data and generates statistics. These statistics can be used as prior knowledge for designing chemical experiments.

[0090] It can be known that the above-mentioned document retrieval tools and document mining tools can adopt the required software tools that can realize the corresponding functions. For example, the document retrieval tool can be used and packaged according to the functional interface provided by the document database, and the document mining tool can be designed and implemented according to the relevant technologies in the field of natural language processing. Open source document mining tools can also be used.

[0091] Furthermore, the composition and function of the experiment designer agent in the above system are as follows:

[0092] The experimental template library used by the experimental designer agent is a template library accumulated based on historical experiments conducted in the automated laboratory. The experimental template library includes chemical experimental program templates of previous experiments and protocols of various configurations from 20 experimental workstations in the automated laboratory. These protocols are stored in XML format and describe the chemical experimental steps in natural language. The workstations include sample rack workstation, liquid dispensing workstation, solid dispensing workstation, magnetic stirring workstation, centrifugation workstation, calcination workstation, liquid aspiration workstation, drying workstation, ultrasonic workstation, encapsulation workstation, photocatalysis workstation, gas chromatography workstation, sample preparation workstation, Raman spectroscopy workstation, infrared spectroscopy workstation, ultraviolet-visible spectroscopy workstation, fluorescence spectroscopy workstation, X-ray diffraction workstation, laser induced breakdown spectroscopy workstation and electrocatalysis workstation. These protocols and workstations together define the experimental capabilities of the automated laboratory in chemical and material synthesis, characterization and performance testing.

[0093] The architecture of the experiment designer agent is as follows Figure 3As shown in Figure 1, it is a multi-agent system that integrates two agents based on a large language model: a protocol writer based on a large language model (also known as Protocol writer) and a protocol critic based on a large language model (also known as Protocol critic). The protocol writer functions through a pre-prompted large language model using two built-in tools: (1) a protocol search tool, also known as ProtocolSearch, which searches the experimental template library based on keywords and returns the best matching chemical experimental program template; (2) a station query tool, also known as StationQuery, which allows the protocol writer to query detailed information about the workstations available in the automated laboratory, such as specific functions and related parameters. The protocol critic is another instance of the large language model with its own pre-set prompts. These prompts outline its specific task, which is to check and correct chemical experimental programs using pre-defined program checking expert rules based on the operating constraints of different workstations.

[0094] The workflow of the experiment designer agent is as follows: First, the protocol writer receives the description of the chemical experiment corresponding to the target chemical task from the task manager agent. Then, it calls the protocol search tool to find the best matching chemical experiment program template in the experiment template library. If a matching chemical experiment program template is found, the protocol writer will reference the relevant chemical experiment program template and adjust the steps and parameters of the chemical experiment program template to formulate a chemical experiment program for the chemical experiment corresponding to the target chemical task. If there is no existing chemical experiment program template, the protocol writer will call the site query tool to obtain the site list of available workstations in the automated laboratory, plan the experimental workflow, and generate a corresponding chemical experiment program relying only on the functions of the large language model. Next, the protocol reviewer conducts a reflection process to check the chemical experiment program generated by the protocol writer. It reviews, criticizes and improves the chemical experiment program according to predefined program checking expert rules. Finally, the protocol reviewer formats and outputs the checked chemical experiment program in the same format as the chemical experiment program template to ensure that the generated chemical experiment program is consistent with the expected experiment and the capabilities of the automated laboratory.

[0095] The predefined program checking expert rules used by the above protocol critic are as follows:

[0096] (1) The first step in checking the chemical experiment procedure is to "retrieve the rack from the sample rack workstation";

[0097] (2) Check that the last step in the chemical experiment procedure is "Put the rack into the sample rack workstation";

[0098] (3) After adding the sample, ensure that all experimental parameters do not exceed the maximum capacity of the sample bottle during the process;

[0099] (4) The experimental parameters of the solid dispensing workstation and the liquid dispensing workstation are in units of mass;

[0100] (5) The workstation in front of the gas chromatography workstation is the packaging station;

[0101] (6) Before using the electrocatalytic workstation, make sure there is liquid in the test bottle;

[0102] (7) Before using the UV-Vis spectroscopy workstation, make sure there is liquid in the test bottle.

[0103] Furthermore, the composition and function of the robot operator agent in the above system are as follows:

[0104] The automated laboratory controlled by the robot operator agent includes two robots and 20 automated experimental workstations with agent-machine interaction interfaces and a back-end control system for all hardware. Of the two existing robots, one is a mobile robot and the other is a desktop robot, which can be a desktop manipulator. The mobile robot can move freely in the laboratory and mainly handles tasks involving multiple instruments. The desktop robot is placed on a platform in the center of the laboratory and moves on tracks to handle the characterization and performance test instruments on the platform, including Raman spectroscopy, X-ray diffraction, infrared spectroscopy, UV-visible spectroscopy, fluorescence spectroscopy, and electrochemical testing. The back-end control system schedules tasks, manages data, and communicates with each robot and workstation through the HTTP protocol. In order to improve scalability and flexibility, the operation methods of the robots and workstations are abstracted and encapsulated, and standard operation commands and common interface specifications are established. These commands contain robot application interfaces, including: visual positioning application interface, motion planning application interface, experimental operation application interface, and instrument communication reference interface. These application interfaces allow the hardware to be utilized through the code of the robot operator agent, thereby ensuring the accurate and efficient execution of experimental operations.

[0105] The architecture of the robot operator agent is as follows Figure 4As shown in Figure 1, it is a multi-agent system consisting of three agents based on a large language model - code writer, code reviewer and code proofreader, which is integrated with the automation laboratory. The code writer is a pre-prompted large language model instance equipped with a built-in tool: the robot application interface query tool, which can be called RobotAPIQuery, which is used to obtain high-level application interface functions available to the robot. The code reviewer is another pre-prompted large language model instance, responsible for checking, correcting, criticizing and improving the robot code for robot execution according to predefined program checking expert rules. The code proofreader uses the robot application interface query tool to perform a final code check on the grammar, spelling and formatting errors of the robot code according to the robot application interface specifications and code standards.

[0106] The workflow of the robot operator agent is as follows: First, the code writer receives the chemical experiment program generated by the experiment designer agent. Then, it calls the robot application interface query tool to obtain the list of available application interfaces of the robot and uses the power of the large language model to write a robot-executed code. Next, the code reviewer will review and improve the code according to the predefined code expert rules. The improved code is then passed to the code proofreader, which calls the robot application interface query tool again to retrieve the list of available robot application interfaces and proofreads the code for further enhancement. Finally, the automation laboratory receives the final code for executing the automated chemical experiment for robot execution to complete the corresponding experimental operation.

[0107] The code review expert rules used by the above code reviewer are as follows:

[0108] (1) Visual positioning must be performed before operating at a new workstation;

[0109] (2) After calling the 'locate()' function, the previous operation posture obtained by the 'locate()' function will be invalid;

[0110] (3) If the workstations operated in two adjacent steps are the same, there is no need to repeat movement and visual positioning;

[0111] (4) Only the solid dispensing workstation, liquid dispensing workstation, centrifuge workstation, encapsulation workstation, photocatalytic workstation and gas chromatography workstation need to place test bottles; other workstations should directly place sample racks as required;

[0112] (5) Each sample rack can hold 10 test bottles, and each synthesis experiment corresponds to a specific test bottle; please do not change the bottle number to avoid errors; be careful when calling functions that require specifying the bottle number, because the numbering starts from 0;

[0113] (6) Rack numbering also starts at 0, and each synthesis experiment corresponds to a test vial in a specific rack; ensure consistency in rack numbering throughout the program to avoid errors;

[0114] (7) The experimental parameters of the solid dispensing workstation and the liquid dispensing workstation must be expressed in mass units;

[0115] (8) The time parameters of all workstations, including magnetic stirring workstations and centrifuge workstations, must be in seconds;

[0116] (9) Calling the 'wait()' function takes significantly longer than calling other functions; however, depending on the restrictions, you still need to call the 'wait()' function when necessary;

[0117] (10) The liquid dispensing workstation has a capacity of 10 channels. Adjacent liquid sampling operations do not require repeated picking up and placing of test bottles. Only instrument communication is required to change the liquid type and continue sampling.

[0118] (11) You need to strictly follow the given chemical experiment operation sequence and minimize the running time.

[0119] The robot application interface of the above-mentioned automated laboratory is as follows:

[0120] (1) The function of the visual positioning application interface is to allow the robot to use computer vision processing related technologies to obtain the position information of the experimental site to be operated, including the 'locate()' function, which receives the name of the experimental site as a parameter and returns the robot's operating posture;

[0121] (2) The function of the motion planning application interface is to allow the robot to move between experimental sites using mobile navigation related technologies, including the 'move()' function, which receives the experimental site name as a parameter and controls the robot to move to the corresponding position;

[0122] (3) The function of the experimental operation application interface is to allow the robot to perform chemical experimental operations, including the 'take_bottle()', 'put_bottle()', 'take_rack()', and 'put_rack()' functions. These functions receive the experimental station name, sample rack number, and sample bottle number as parameters, and control the robot to perform experimental operations such as taking sample bottles, placing sample bottles, taking sample racks, and placing sample racks respectively;

[0123] (4) The function of the instrument communication application interface is to allow the robot to communicate with the instruments at the experimental site, including the 'command()' and 'wait()' functions. The 'command()' function receives the experimental site name and chemical experiment parameters as parameters to control the chemical instruments at the experimental site to perform experiments. The 'wait()' function receives the experimental site name as a parameter to wait for the instrument to complete the experimental operation.

[0124] The above application interface is implemented through abstract and encapsulated robot high-level API functions, using PYTHON as the programming language.

[0125] Furthermore, the composition and function of the computational executor agent in the above system are as follows:

[0126] Most of the on-demand experimental tasks that human researchers request of the robotic AI chemist system of the present invention follow the pattern of preparing measurement tests (e.g., measuring the infrared spectrum of a molecule) or involve exploring a prescribed design space (e.g., conducting a full factorial experiment on certain variables). For such tasks, it is sufficient for the task manager agent to call only the experimental designer agent and the robotic operator agent. However, research tasks that require data-driven discovery or closed-loop iterative optimization (e.g., discovering high-performance catalysts in a vast search space) require additional computing power. To meet this need, a library of theoretical models is compiled, most of which are pre-trained models from previous publications or models developed for the robotic AI chemist system. Each pre-trained model provides auxiliary functions for processing input data, ensures consistency in data pre-processing (such as normalization and handling missing values), and can be used for direct prediction. These models come with annotations detailing their technical specifications and origins, so that human researchers or agents based on large language models can query them. Further, the theoretical model library can also include open source models that can be extended and trained with experimental data collected by the robotic AI chemist according to the requirements of specific research tasks.

[0127] The architecture of the computational agent is as follows Figure 5As shown in the figure, it integrates two agents based on large language models, namely the computational task manager agent and the Bayesian optimizer, two built-in tools, and a deep learning computing platform; the two built-in tools are: (1) a model search tool, which can be called ModelSearch, which can use keywords to perform searches in the theoretical model library and return a matching model; (2) a model fusion tool, which can be called ModelFuse, which extends a pre-trained neural network model by adding additional hidden layers to create a fusion model. The Bayesian optimizer is an agent based on a large language model, which uses the task description provided by human researchers to the task manager agent to write Bayesian optimization code; the completion of its task depends entirely on the capabilities of the large language model. The deep learning computing platform provides a complete operating environment for the computational executor agent, such as the TensorFlow environment, which is necessary to run machine learning code and obtain results.

[0128] The workflow of the computational executor is as follows: First, the computational task manager agent receives input information, including experimental task keywords, detailed task descriptions, and robot experimental results. Then, it calls the model search tool to obtain a suitable pre-trained model, calls the model fusion tool to fuse the pre-trained model with the experimental data obtained by the robot operator agent to obtain a fusion model, calls the Bayesian optimizer to optimize the fusion model and generate machine learning code, and finally, by executing the generated code on the deep learning computing platform, the optimized prediction results are obtained.

[0129] The multi-agent driven robotic artificial intelligence chemist system of the embodiment of the present invention can perform various chemical experimental operations as needed, and its main limitation is the availability and experimental capacity of the automated workstations in the laboratory. It generates appropriate workflows based on understanding the task descriptions provided by human researchers, and selects and combines the necessary agents and workstations according to the needs of specific workflows. To demonstrate this, the performance of the robotic artificial intelligence chemist system of the present invention is demonstrated below through different embodiments, by instructing the robotic artificial intelligence chemist system of the present invention to perform several experimental tasks of different complexity and complexity.

[0130] Example 2

[0131] This embodiment is to process chemical experiment tasks of the experimental category of preparation and measurement by the multi-agent driven robotic artificial intelligence chemist system of embodiment 1. This chemical experiment task of preparation and measurement is the most direct type of experimental task that can be completed by the multi-agent driven robotic artificial intelligence chemist system. After receiving the task description, the task manager agent will interpret and understand the instructions, pass the necessary information to the experiment designer agent, and the experiment designer agent will find and use the appropriate experimental template to generate the appropriate experimental program. The robot operator agent will then convert the experimental program into the required robot executable code to perform robot synthesis and testing in the automated laboratory, and obtain the experimental results returned to the task manager agent.

[0132] Task 1, addressed in this example, belongs to this category of on-demand research tasks. Task 1 involves experiments to characterize three azobenzene molecules using Fourier transform infrared (FT-IR) spectroscopy. Azobenzenes are a class of organic molecules characterized by the presence of an N-=-N double bond connecting two phenyl rings. They are known for their photo-switchable properties, which enable them to change their molecular structure and chemical properties upon irradiation. The Task Manager Agent was instructed to measure the infrared spectra of three azobenzene molecules using the following prompt input: "You have been assigned a scientific research task to measure the infrared spectra of azobenzene molecules. You need to perform one measurement of each of the following solid powders: azobenzene, 1-(benzoxazolyl)naphthalene-2-amine, and 4,4'-dihydroxymethylbenzoxazolylbenzene." After receiving this task description, the Robotic AI Chemist System successfully planned, coordinated, and collected high-quality infrared spectra of the molecules. The results showed the expected characteristic peaks of stretching vibrations N=N at 1600-1500 cm⁻¹, and C=C and CH of the benzene ring at 1500-1400 cm⁻¹ and 3100-3000 cm⁻¹, respectively. Figure 6 shown.

[0133] Example 3

[0134] This embodiment is a multi-agent driven robotic artificial intelligence chemist system in Example 1 that processes chemical experiment tasks with the experimental category of exploration and screening. Compared with manufacturing and measurement tasks, chemical experiment tasks of exploration and screening require more advanced large language model capabilities. For example, the task manager agent must be able to interpret task instructions that require in-depth knowledge of specific fields, such as understanding and executing instructions to "conduct a full-factor experiment." This involves recognizing the concept of full-factor experiments and systematically changing multiple factors to evaluate the importance of their impact. Similarly, the experimental designer agent must also show flexibility and creativity when adapting to existing chemical experiment program templates. This may include assigning different values ​​to variables according to experimental requirements, and customizing the experimental program by expanding or combining these variables to accommodate more complex experimental designs.

[0135] Task 2 dealt with in this example belongs to this type of on-demand research task. Task 2 involves a robotic AI chemist conducting a complete full-factorial experiment to synthesize graphite carbon nitride (g-C3N4) and test their hydrogen evolution reaction (HER) performance. For demonstration purposes, the multi-agent-driven robotic AI chemist system of the present invention was asked to consider three heating temperatures (500°C, 550°C, and 600°C) and three heating times (3 hours, 4 hours, and 5 hours). By measuring the HER performance of the resulting g-C3N4 material, the different sample qualities produced under different synthesis conditions are illustrated, highlighting the importance of screening these synthesis conditions for optimizing the HER performance of g-C3N4. Figure 7 As shown, the results indicate that g-C3N4 degrades or over-sinters at higher temperatures, which is particularly evident at 600 °C. In addition, it is clear that increasing the heating time has a negative impact on the HER performance at all three temperatures.

[0136] Example 4

[0137] This embodiment is to process chemical experiment tasks of the experimental category of discovery and optimization through the multi-agent driven robotic artificial intelligence chemist system of Example 1. The chemical experiment tasks of discovery and optimization are the most advanced research tasks that can be completed by the multi-agent driven robotic artificial intelligence chemist system at present, and require the use of all four role-specific agents and their corresponding basic resources. A typical workflow includes (1) the literature reader agent mines experimental knowledge from the literature database, (2) the experiment designer agent customizes the experimental program from the experiment template library, (3) the robot operator agent generates code to control the automatic laboratory to conduct machine experiments, and (4) the computational executor agent uses the theoretical model library to analyze data and generate optimization goals. Task 3 processed in this embodiment belongs to this type of on-demand research task.

[0138] Task 3 involves having the Robotic AI Chemist system discover metal-organic high entropy catalysts (MO-HECs) for the oxygen evolution reaction (OER). The process starts with a prompt to the Task Manager agent: "You have been assigned a scientific research task to discover high-performance metal-organic high entropy catalysts (MO-HECs) composed of five metal elements for the oxygen evolution reaction (OER). Please use the Literature Reader agent to discover popular metals for OER in the literature and select the top five suggestions. The automated laboratory is equipped with the necessary precursor solutions and other required chemicals. Please randomly select 100 different metal ratios with the ratio of each metal ranging from 5% to 35% and perform the synthesis and overpotential measurements. After obtaining the experimental data from the 100 automated experiments , use a computational executor agent to optimize the experiments. Using the compositions and measured overpotentials of these 100 experiments, along with a pre-trained model, build a fusion model to predict the overpotential from the ratios. This fusion model is then combined with Bayesian optimization to find the combination that gives the lowest predicted overpotential. The constraints of Bayesian optimization are that the metal ratios must be between 5% and 35% and the sum must reach 100%. After the computational executor agent provides the results, your final experiment is to synthesize and test this optimal combination. ", the workflow of the multi-agent driven robotic artificial intelligence chemist system is as follows Figure 8 As shown in the figure, specifically: the task manager agent first calls the literature reader agent to read a large number of literature on oxygen evolution reaction, and the five metal elements recommended are cobalt (Co), nickel (Ni), iron (Fe), manganese (Mn) and copper (Cu); then the task manager agent passes the experimental content and parameters to the experimental designer agent to design a suitable chemical experimental program; then the task manager agent passes the experimental program to the robot operator agent, the robot operator agent writes code according to the program, and conducts robot chemical experiments to obtain experimental data; the task manager agent passes the experimental content and experimental data to the computational executor agent, and the computational executor agent establishes a fusion model to give the optimized parameter results; finally, the task manager agent again gives the parameter information to the experimental designer agent to generate the experimental program, and then the robot operator agent conducts a robot experiment according to the experimental program to obtain a metal organic high entropy catalyst with superior performance. In the experiment, the multi-agent-driven robotic artificial intelligence chemist system first synthesized 100 metal organic high entropy catalysts according to the needs of human researchers and measured their overpotentials. All of these overpotentials are within 10mA / cm 2 After optimization using the computational executor agent, the catalyst with the best metal composition showed an overpotential of 266.1 mV, a significant improvement over 100 catalysts with randomly selected metal compositions. Fig. 9The optimal metal-organic high-entropy catalyst was confirmed to maintain a high degree of stability, with performance decreasing by less than 2% after 500 hours of continuous operation.

[0139] In summary, the multi-agent driven robotic artificial intelligence chemist system of the present invention utilizes the capabilities of a large language model, such as the open source Llama-3.1-70B. This hierarchical multi-agent system integrates various professional agents based on a large language model, namely, a literature reader agent, an experimental designer agent, a robot operator agent and a literature reader agent, an experimental designer agent, and a robot operator agent, which can perform complex chemical research tasks alone or in collaboration. It includes a task manager agent, which interacts with human researchers and communicates and connects to manage agents with four specific roles - a literature reader agent, an experimental designer agent, a computational executor agent, and a robot operator agent. This architecture enhances the effectiveness of individual agents through collaboration and coordination, enabling them to jointly handle complex tasks and achieve determined goals. The system has four basic resources: a comprehensive literature database, an extensive experimental template library, a general theoretical model library, and a state-of-the-art automated laboratory. These resources enable the robotic artificial intelligence chemist system to perform a wide range of chemical experimental operations, from literature mining and experimental process generation to computational modeling and robotic execution of targeted chemical experiments. It can well complete three types of experimental tasks: preparation and measurement chemical experimental tasks, exploration and screening chemical experimental tasks, and discovery and optimization chemical experimental tasks. These experimental tasks highlight the adaptability and precision of intelligent agents in generating experimental procedures, executing robotic experiments, and using computational models for data-driven discovery and optimization. It is worth noting that the robotic artificial intelligence chemist system successfully explored the huge chemical space of five possible ingredients to discover high-performance metal organic high entropy catalysts (MO-HECs) for electrocatalytic oxygen evolution reaction. Through literature mining, computational modeling, and Bayesian optimization, the system discovered and experimentally verified an optimal MO-HEC composition, achieving a significant improvement in overpotential performance. The system makes it possible to perform complex, multi-step experiments with minimal human intervention, thus opening up new avenues for scientific inquiry.

[0140] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0141] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims. The information disclosed in the background technology section of this article is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as an admission or in any form that the information constitutes prior art known to those skilled in the art.

Claims

1. A multi-agent driven robotic artificial intelligence chemist system, characterized in that: include: Task manager agent, literature reader agent, experiment designer agent, robot operator agent, literature database, experiment template library and automated laboratory; among them, The task manager agent, the document reader agent, the experiment designer agent, and the robot operator agent are all agents based on a large language model; The task manager agent is respectively connected to the document reader agent, the experiment designer agent and the robot operator agent in communication, and can interact with the human researcher, and coordinate the document reader agent, the experiment designer agent and the robot operator agent to collaboratively complete the chemical experiment operation corresponding to the target chemical task according to the description information of the target chemical task input by the human researcher; The document reader agent is connected to the document database for communication, can receive keywords related to the target chemical task given by the task manager agent, search the document database according to the keywords, mine out knowledge related to the target chemical task, and transmit the knowledge related to the target chemical task to the task manager agent; The experiment designer agent is connected to the experiment template library in communication, can receive the knowledge related to the target chemical task excavated by the document reader agent and forwarded by the task manager agent, and generate chemical experiment instructions corresponding to the target chemical task according to the knowledge, plan chemical experiment operations, and generate chemical experiment procedures step by step using the experiment template library, and transmit the chemical experiment procedures to the task manager agent; The robot operator intelligent agent is communicatively connected to the automated laboratory, and can receive the chemical experiment program generated by the experiment designer intelligent agent and forwarded by the task manager intelligent agent, and convert the chemical experiment program into code for operating the robot and workstation of the automated laboratory, so that the robot and workstation execute according to the code to complete the chemical experiment operation corresponding to the target chemical task.

2. The multi-agent driven robotic artificial intelligence chemist system according to claim 1, characterized in that: The large language model used by the task manager agent, the document reader agent, the experiment designer agent, and the robot operator agent is any one of an open source large language model and a commercial large language model; The target chemical tasks include: chemical experiment tasks of preparation and measurement, and chemical experiment tasks of exploration and screening.

3. The multi-agent driven robotic artificial intelligence chemist system according to claim 1 or 2, characterized in that: The literature database is a database containing titles and abstracts of a plurality of scientific publications in the disciplines of chemistry and materials science; The document reader agent is provided with a document search tool and a document mining tool respectively, and the document reader agent can receive keywords related to the target chemical task provided by the task manager agent, and retrieve and collect publication titles and abstracts related to the target chemical task from the document database using the received keywords through the document search tool; The retrieved and collected publication titles and abstracts are transmitted to a literature mining tool, which uses an unsupervised syntactic distance analysis method to mine data in the literature corresponding to the publication titles and abstracts, extract information on chemical substances, physicochemical properties, chemical and / or material functions, and experimental conditions from the data, and generate statistical data from the extracted information as prior knowledge for designing chemical experimental operations, and transmit the statistical data as knowledge related to the target chemical task to the task manager intelligent agent.

4. The multi-agent driven robotic artificial intelligence chemist system according to claim 1 or 2, characterized in that: The experiment template library contains chemical experiment program templates of historical experiments in the automated laboratory and various configuration protocols of various workstations in the automated laboratory. The various configuration protocols are stored in XML format and describe the experimental steps and parameters of the chemical experiment program templates in natural language. The experimental designer agent includes: a protocol writer and a protocol reviewer; wherein, The protocol writer is connected to the task manager agent in communication, is an agent based on a large language model, and is provided with a protocol search tool and a site query tool. It can call the protocol search tool to search for the best matching chemical experiment program template in the experimental template library according to the description information of the target chemical task received from the task manager agent. If the best matching chemical experiment program template is found, the best matching chemical experiment program template is referenced, and the experimental steps and parameters of the chemical experiment program template are adjusted to formulate an experimental program for the chemical experiment operation corresponding to the target chemical task; if there is no existing chemical experiment program template, the site query tool is called to obtain a site list of workstations that can be used in the automated laboratory, and the experimental workflow is planned according to the site list, and a chemical experiment program for the chemical experiment operation corresponding to the target chemical task is generated based on the large language model function; The protocol reviewer, which is in communication connection with the protocol writer, is an intelligent agent based on a large language model, which can review, criticize and improve the chemical experiment program derived by the protocol writer according to predefined program checking expert rules, and output the final chemical experiment program in a format unified with the chemical experiment program template.

5. The multi-agent driven robotic artificial intelligence chemist system according to claim 4, characterized in that: The program checking expert rules used by the protocol critic are as follows: (1) The first step in checking the chemical experiment procedure is "retrieve the rack from the sample rack workstation"; (2) Check that the last step in the chemical experiment procedure is "Put the rack into the sample rack workstation"; (3) After adding the sample, ensure that all experimental parameters do not exceed the maximum capacity of the sample bottle during the process; (4) The experimental parameters of the solid dispensing workstation and the liquid dispensing workstation are in units of mass; (5) The workstation in front of the gas chromatography workstation is the packaging station; (6) Before using the electrocatalytic workstation, make sure there is liquid in the test bottle; (7) Before using the UV-Vis spectroscopy workstation, make sure there is liquid in the test bottle.

6. The multi-agent driven robotic artificial intelligence chemist system according to claim 1 or 2, characterized in that: The automated laboratory comprises: a plurality of robots, a plurality of automated workstations and a robot application interface; wherein the plurality of robots comprise: a mobile robot and a desktop robot; The plurality of automated workstations include: a sample rack workstation, a liquid dispensing workstation, a solid dispensing workstation, a magnetic stirring workstation, a centrifugal workstation, a calcination workstation, a liquid aspiration workstation, a drying workstation, an ultrasonic workstation, a packaging workstation, a photocatalytic workstation, a gas chromatography workstation, a sample preparation workstation, a Raman spectroscopy workstation, an infrared spectroscopy workstation, a UV-visible spectroscopy workstation, a fluorescence spectroscopy workstation, an X-ray diffraction workstation, a laser-induced breakdown spectroscopy workstation, and an electrocatalytic workstation; The robot application interface includes: a visual positioning application interface, a motion planning application interface, an experimental operation application interface and an instrument communication application interface; The robot operator agent includes: a code writer, a code reviewer and a code checker; wherein, The code writer is respectively connected to the robot application interface and the code reviewer of the automated laboratory, is an intelligent agent based on a large language model, and is provided with a robot application interface query tool. It can obtain the usable robot application interface of each robot in the automated laboratory through the robot application interface query tool according to the chemical experiment program generated by the received experiment designer intelligent agent, and write a robot operation code corresponding to the chemical experiment program by using the function of the large language model according to the received chemical experiment program and the usable robot application interface queried; The code reviewer, which is in communication connection with the code proofreader, is an intelligent agent based on a large language model, and can check, correct, criticize and improve the code written by the code writer according to the code checking expert rules; The code proofreader is respectively communicated with the robot application interface and the code reviewer of the automation laboratory. It is an intelligent agent based on a large language model. It can use the robot application interface query tool to query and obtain the usable robot application interface of each robot in the automation laboratory, proofread the program code received from the code reviewer, and transmit the proofread code to the automation laboratory, and the robot in the automation laboratory will execute the proofread code.

7. The multi-agent driven robotic artificial intelligence chemist system according to claim 6, characterized in that: The code checking expert rules are as follows: (1) Visual positioning must be performed before operating at a new workstation; (2) After calling the 'locate()' function, the previous operation posture obtained by the 'locate()' function will be invalid; (3) If the workstations operated in two adjacent steps are the same, there is no need to repeat movement and visual positioning; (4) Only the solid dispensing workstation, liquid dispensing workstation, centrifuge workstation, encapsulation workstation, photocatalytic workstation and gas chromatography workstation need to place test bottles; other workstations should directly place sample racks as required; (5) Each sample rack can hold 10 test vials, and each synthesis experiment corresponds to a specific test vial; Please do not change the bottle number to avoid mistakes; Be careful when calling functions that require a bottle number, because numbering starts at 0; (6) The sample rack number also starts from 0, and each synthesis experiment corresponds to a test bottle in a specific sample rack; Ensure consistency in rack numbering throughout the program to avoid errors; (7) The experimental parameters of the solid dispensing workstation and the liquid dispensing workstation must be expressed in mass units; (8) The time parameters of all workstations, including magnetic stirring workstations and centrifuge workstations, must be in seconds; (9) Calling the 'wait()' function takes significantly longer than calling other functions; however, depending on the restrictions, you still need to call the 'wait()' function when necessary; (10) The liquid dispensing workstation has a capacity of 10 channels, and adjacent liquid sampling operations do not require repeated picking and placing of test bottles; Only instrument communication is required to change liquid type and continue sampling; (11) You need to strictly follow the given chemical experiment operation sequence and minimize the running time.

8. The multi-agent driven robotic artificial intelligence chemist system according to claim 1 or 2, characterized in that: Also includes: Computational executor agents and theoretical model libraries; among them, The computational executor agent is respectively communicatively connected with the task manager agent and the theoretical model library, and is an agent based on a large language model. It can search for a suitable pre-trained model in the theoretical model library according to the relevant information of the target chemical task transmitted by the task manager agent, and use the searched pre-trained model according to a corresponding method, and fuse the pre-trained model with the experimental data obtained by the robot operator agent through model fusion to obtain a fusion model, and optimize the obtained fusion model through Bayesian optimization to obtain optimized experimental data, and transmit the obtained optimized experimental data to the task manager agent for subsequent operations of the robot operator agent.

9. The multi-agent driven robotic artificial intelligence chemist system according to claim 8, characterized in that: The theoretical model library stores pre-trained machine learning models from historical publications, machine learning models developed for robotic artificial intelligence chemists, and open source machine learning models; The computational executor agent comprises: Computational task manager agent, model search tool, model fusion tool, Bayesian optimizer and deep learning computing platform; among them, The computing task manager agent is respectively communicated with the task manager agent, the model search tool, the deep learning computing platform, the model fusion tool and the Bayesian optimizer. It is an agent based on a large language model and can receive input information sent by the task manager agent. The input information includes: keywords of the target chemical task, detailed description information of the target chemical task and experimental result information of the robot operator agent. The model search tool is called to search the theoretical model library with keywords to obtain a suitable pre-trained machine learning model. The model fusion tool is called to fuse the pre-trained machine learning model with the experimental data obtained by the robot operator agent to obtain a fusion model. The Bayesian optimizer is called to optimize the fusion model and generate a machine learning code. The deep learning computing platform receives and executes the generated machine learning code to obtain the prediction results of the chemical experiment operation parameters corresponding to the target chemical task as the optimized experimental data, and the optimized experimental data is transmitted to the task manager agent for subsequent operations of the robot operator agent.

10. The multi-agent driven robotic artificial intelligence chemist system according to claim 8, characterized in that: The target chemical task is a chemical experimental task of discovery and optimization.

Images