Method and apparatus for cross-scale analysis of liquid metals
By employing a multi-agent collaborative scheduling method, code is automatically generated for cross-scale simulation of liquid metals, solving the problem of predicting the morphology of liquid metals and improving simulation efficiency and accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PEKING UNIV
- Filing Date
- 2026-02-25
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies struggle to accurately predict the state of liquid metals, and cross-scale simulations are inefficient, failing to obtain key thermodynamic parameters under extreme conditions.
A multi-agent cooperative scheduling method is adopted. The task type is analyzed by an intent recognition model, and the code is generated by the first and second agents to realize molecular dynamics and finite element simulation, and to automatically complete the microscopic and macroscopic simulation analysis of liquid metal.
It enables rapid and accurate acquisition of thermodynamic parameters and macroscopic properties of liquid metals, improves cross-scale simulation efficiency, and achieves simulation results with an accuracy of over 95%.
Smart Images
Figure CN122135837A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of materials simulation and analysis technology, and more specifically, to a cross-scale analysis method and apparatus for liquid metals. Background Technology
[0002] Room-temperature liquid metals (such as Ga and its eutectic alloys Ga-In, Ga-Sn, and Ga-In-Sn systems) possess both metallic conductivity and fluid deformability. With their advantages of high electrical and thermal conductivity, low toxicity, and tunable surface chemistry, they show great promise for applications in emerging fields such as flexible electronics, soft robotics, and advanced thermal management. However, the morphological evolution of liquid metals is influenced by both atomic-level interfacial interactions and macroscopic fluid dynamics. Accurate prediction of their morphology requires obtaining thermodynamic and interfacial parameters under different environments. Nevertheless, liquid metal alloys exhibit diverse compositions and complex interfacial behaviors, and are subject to the coupling effects of oxidation, temperature gradients, and substrate heterogeneity, resulting in a significant lack of key thermodynamic parameters.
[0003] To address the parameter acquisition problem, relevant technologies are mainly divided into two categories: one is traditional thermodynamic parameter measurement technology, which uses a temperature-controlled experimental platform (such as a checkerboard interface observation system) in an oxygen-free environment such as a nitrogen glove box, and uses contact angle measuring instruments, thermocouples, and other equipment to observe the contact angle and spreading behavior of liquid metal under different temperature gradients and heterogeneous interfaces, thereby deriving thermodynamic parameters. Although this type of technology can obtain reliable data, the experimental process is complex and time-consuming (a single experiment requires several days), and it cannot measure parameters under extreme conditions (such as high temperature and high pressure), thus limiting its applicability. The other type is cross-scale simulation technology, which is based on the MD (Molecular Dynamics)-FEM (Finite Element Method) coupled architecture. By manually writing relevant code, it can achieve performance prediction from the atomic level to the macroscopic scale. However, this technology requires manual parameter matching and code correction, and lacks a unified result verification mechanism, resulting in problems such as low simulation efficiency, poor generalization ability, and large cross-scale data transfer errors.
[0004] There is currently no effective solution to the above problems. Summary of the Invention
[0005] This application provides a method and apparatus for cross-scale analysis of liquid metals, which at least solves the technical problems of difficulty in obtaining thermodynamic parameters of liquid metals and low efficiency of cross-scale simulation.
[0006] According to one aspect of the embodiments of this application, a method for cross-scale analysis of liquid metal is provided, comprising: acquiring a target task instruction for a target liquid metal, and calling an intent recognition model to analyze the target task instruction to obtain a task type and target task data, wherein the target task data includes attribute parameters of the target liquid metal; when the task type is to analyze the macroscopic characteristics of the target liquid metal, calling a first agent to analyze the target task data to determine the type of the missing target thermodynamic parameter in the attribute parameters; calling a second agent to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameter based on the target task data and the type of the target thermodynamic parameter, and calling a molecular dynamics simulation application to execute the first code to obtain the target thermodynamic parameter; calling the first agent to generate second code for performing macroscale simulation of the target liquid metal to determine the macroscopic characteristics of the target liquid metal based on the target task data and the target thermodynamic parameter, and calling a finite element simulation application to execute the second code to obtain the macroscopic characteristics of the target liquid metal.
[0007] Optionally, the training process of the first intelligent agent includes: acquiring a pre-trained initial large language model, wherein the initial large language model includes a first sub-model for analyzing missing parameters and a second sub-model for generating code; acquiring multiple sets of historical task data, and using each set of historical task data as a first training sample, wherein each set of historical task data includes complete attribute parameters of a liquid metal; for each set of historical task data, acquiring structured inference data corresponding to the historical task data, and using the structured inference data as the first sample label of the corresponding first training sample, wherein the structured inference data includes: macroscopic scale analysis of the liquid metal corresponding to the historical task data. The simulation determines the executable code for the macroscopic properties of liquid metal, and the inference chain of the executable code is derived from historical task data. For each set of historical task data, at least one third thermodynamic parameter is randomly selected from the complete attribute parameters in the historical task data. All remaining attribute parameters are used as a second training sample, and the types of all third thermodynamic parameters are used as the second sample labels of the corresponding second training samples. The first sub-model is supervised and fine-tuned based on multiple sets of second training samples and second sample labels. The second sub-model is supervised and fine-tuned based on multiple sets of first training samples and first sample labels. The trained first and second sub-models are integrated into a first agent.
[0008] Optionally, the training process of the second agent includes: acquiring a pre-trained initial large language model for code generation; acquiring multiple sets of historical task data, wherein each set of historical task data includes complete property parameters of a liquid metal; for each set of historical task data, randomly selecting at least one third thermodynamic parameter from the complete property parameters in the historical task data, using the type of the third thermodynamic parameter and all remaining property parameters as a third training sample; acquiring structured inference data corresponding to the historical task data, and using the structured inference data as the third sample label of the corresponding third training sample, wherein the structured inference data includes... The executable code is used to perform microscale simulations of liquid metal corresponding to historical task data to determine the third thermodynamic parameter. The inference chain of the executable code is obtained from the historical task data. All third training samples and their corresponding third sample labels are divided into a first training subset and a second training subset according to a preset ratio. The number of samples in the first training subset is less than the number of samples in the second training subset. The initial large language model is supervised and fine-tuned based on the first training subset to obtain the first large language model. The first large language model is trained by reinforcement learning using a group relative policy optimization algorithm combined with the second training subset to obtain the second agent.
[0009] Optionally, the first language model is trained using a group relative policy optimization algorithm combined with a second training subset, including: inputting each third training sample from the second training subset into the first language model to generate multiple third codes; constructing a reward function for a target dimension based on the multiple third codes and multiple third sample labels, wherein the target dimension includes at least one of the following: physical and logical consistency, code executability, inference chain rigor, and computational efficiency; and adjusting the model parameters of the first language model based on the reward function combined with a gradient regularization strategy.
[0010] Optionally, when the task type is to determine the target thermodynamic parameters of the target liquid metal, the second agent is directly invoked to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameters based on the target task data and the type of the target thermodynamic parameters, wherein the target thermodynamic parameters are thermodynamic parameters not provided in the attribute parameters; the molecular dynamics simulation application is invoked to execute the first code to obtain the target thermodynamic parameters.
[0011] Optionally, after the first agent analyzes the target task data, if the target task data includes complete property parameters of the target liquid metal, the first agent is directly invoked to generate second code based on the target task data to perform macroscale simulation of the target liquid metal to determine the macroscopic characteristics of the target liquid metal; the finite element simulation application is invoked to execute the second code to obtain the macroscopic characteristics of the target liquid metal.
[0012] Optionally, after obtaining the macroscopic properties of the target liquid metal, the actual macroscopic properties of the target liquid metal are obtained from references or historical experimental data; the similarity between the simulated macroscopic properties and the actual macroscopic properties is calculated; and the macroscopic properties and similarity are output simultaneously.
[0013] Optionally, if the similarity is not less than a preset similarity threshold, the first code and the second code are stored in the target database; the code stored in the target database is used periodically to train the first agent and the second agent, and the model parameters of the first agent and the second agent are updated.
[0014] Optionally, the question-guiding agent is invoked to analyze the task type and generate question-guiding prompts, which are used to guide the target object to ask questions related to the task type.
[0015] According to another aspect of the embodiments of this application, a cross-scale analysis device for liquid metal is also provided, comprising: an acquisition module, configured to acquire a target task instruction for a target liquid metal, and call an intent recognition model to analyze the target task instruction to obtain a task type and target task data, wherein the target task data includes attribute parameters of the target liquid metal; a first analysis module, configured to, when the task type is to analyze the macroscopic characteristics of the target liquid metal, call a first agent to analyze the target task data and determine the type of the missing target thermodynamic parameter in the attribute parameters; a second analysis module, configured to call a second agent to generate a first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameter based on the target task data and the type of the target thermodynamic parameter, and call a molecular dynamics simulation application to execute the first code to obtain the target thermodynamic parameter; and a third analysis module, configured to call a first agent to generate a second code for performing macroscale simulation of the target liquid metal to determine the macroscopic characteristics of the target liquid metal based on the target task data and the target thermodynamic parameter, and call a finite element simulation application to execute the second code to obtain the macroscopic characteristics of the target liquid metal.
[0016] According to another aspect of the embodiments of this application, a computer program product is also provided, the computer program product comprising: a computer program, wherein the computer program, when executed by a processor, implements the above-described method for cross-scale analysis of liquid metals.
[0017] According to another aspect of the embodiments of this application, an electronic device is also provided, the electronic device comprising: a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the above-described method for cross-scale analysis of liquid metals through the computer program.
[0018] In this embodiment, after obtaining the target task instruction for the target liquid metal, the intent recognition model is first invoked to analyze the target task instruction to obtain the task type and target task data, which includes the attribute parameters of the target liquid metal. If the task type is to analyze the macroscopic characteristics of the target liquid metal, a first agent is invoked to analyze the target task data and determine the type of the missing target thermodynamic parameter in the attribute parameters. Then, a second agent is invoked to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameter based on the target task data and the type of the target thermodynamic parameter. A molecular dynamics simulation application is then invoked to execute the first code to obtain the target thermodynamic parameter. Afterward, the first agent is invoked to generate second code for performing macroscale simulation of the target liquid metal to determine its macroscopic characteristics based on the target task data and the target thermodynamic parameter. A finite element simulation application is then invoked to execute the second code to obtain the macroscopic characteristics of the target liquid metal. This solution achieves a fully automated analysis process of "task parsing - code self-generation - simulation self-execution" by coordinating and scheduling multiple models / agents. Through the code generated by the agent, simulation analysis of liquid metals at the micro and macro dimensions can be completed quickly, and its accurate thermodynamic parameters and macroscopic properties can be obtained efficiently. This effectively solves the technical problems of the difficulty in obtaining the thermodynamic parameters of liquid metals and the low efficiency of cross-scale simulation. Attached Figure Description
[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0020] Figure 1 This is a schematic flowchart of an optional cross-scale analysis method for liquid metals according to an embodiment of this application;
[0021] Figure 2 This is a schematic diagram of an optional multi-scale analysis device for liquid metal according to an embodiment of this application;
[0022] Figure 3 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application. Detailed Implementation
[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0024] It should be noted that the terms "first," "second," etc., used in the specification, claims, and drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0025] The information collected in this application embodiment is information and data authorized by the user or fully authorized by all parties. The collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with the relevant laws, regulations and standards of the relevant countries and regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse.
[0026] To better understand the embodiments of this application, the following is a translation and explanation of some nouns or terms that appear in the description of the embodiments of this application:
[0027] MD: Molecular dynamics is a computer simulation method based on classical mechanics (Newton's equations of motion) used to study the motion behavior of atoms and molecules over a certain period of time. It obtains information such as the thermodynamic properties, structural dynamics, diffusion coefficient, and reaction paths of the system by numerically solving the force and motion equations of all particles in the system and tracking the evolution of particle positions and velocities over time.
[0028] FEM: Finite Element Analysis is a numerical computation technique used to solve partial differential equations (such as heat conduction, structural mechanics, fluid dynamics, electromagnetic fields, etc.). Its core idea is to discretize a continuous physical domain into a finite number of "elements", approximate the unknown field variables in each element with a simple function (such as a polynomial), and then establish a system of algebraic equations through variational principles or weighted residual methods to finally solve the response of the entire system.
[0029] LAMMPS (Large-scale Atomic / Molecular Massively Parallel Simulator) is an open-source, highly parallel molecular dynamics simulation software designed for large-scale atomic / molecular systems. It supports various force fields, boundary conditions, and integral algorithms, and can simulate solid, liquid, gas, polymer, and biomolecular systems.
[0030] COMSOL Multi-physics is a multiphysics simulation platform based on finite element analysis, which allows users to couple multiple physical fields (such as thermal-structural, fluid-electric, piezoelectric, chemical reaction, etc.) in the same environment to perform highly integrated numerical simulations.
[0031] SFT (Supervised Fine-Tuning) is an important stage in the training process of large language models, typically occurring after pre-training and before reinforcement learning based on human feedback. Its goal is to enable the model to learn to follow human instructions and generate responses that conform to the expected format and content.
[0032] GRPO (Group Relative Policy Optimization) is a novel reinforcement learning algorithm used to solve the alignment problem in large language models. It optimizes the policy through intra-group relative feedback, avoiding the noise problem of scoring individual samples in traditional reinforcement learning based on human feedback.
[0033] CoT (Chain-of-Thought): A prompting technology that improves the performance of complex reasoning tasks by guiding large language models to "think step by step" (i.e., outputting intermediate reasoning steps) before answering questions.
[0034] Example 1
[0035] According to an embodiment of this application, a method for cross-scale analysis of liquid metals is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.
[0036] Figure 1 This is a schematic flowchart of a cross-scale analysis method for liquid metals according to an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps S102-S106:
[0037] Step S102: Obtain the target task instruction for the target liquid metal, and call the intent recognition model to analyze the target task instruction to obtain the task type and target task data, wherein the target task data includes the attribute parameters of the target liquid metal.
[0038] Step S104: When the task type is to analyze the macroscopic properties of the target liquid metal, the first intelligent agent is invoked to analyze the target task data and determine the type of the missing target thermodynamic parameter in the attribute parameters.
[0039] Step S106: The second agent is invoked to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameters based on the target task data and the type of target thermodynamic parameters. The molecular dynamics simulation application is then invoked to execute the first code to obtain the target thermodynamic parameters.
[0040] Step S108: The first intelligent agent is invoked to generate second code based on the target task data and target thermodynamic parameters to perform macroscopic simulation of the target liquid metal to determine the macroscopic properties of the target liquid metal, and the finite element simulation application is invoked to execute the second code to obtain the macroscopic properties of the target liquid metal.
[0041] This method achieves a fully automated analysis process of "task parsing - code self-generation - simulation self-execution" by coordinating the scheduling of multiple models / agents. The agent generates executable code based on physical constraints and structured reasoning, which can quickly complete the simulation analysis of liquid metals in micro / macro dimensions without human intervention, efficiently obtain its thermodynamic parameters and macroscopic properties, and the simulation results have been verified by experiments to have an accuracy of over 95%, which can significantly improve the accuracy and efficiency of analysis for liquid metals.
[0042] The following section describes each step of the cross-scale analysis method for liquid metals, using a specific implementation process as an example.
[0043] As an optional implementation, the system can receive target task instructions for the target liquid metal input by the user in the human-computer interaction interface, and then call the intent recognition model to analyze the target task instructions to obtain the task type and target task data.
[0044] The target task instruction can be in the form of natural language text or a structured form. The intent recognition model can be any conventional model that has natural language processing and intent recognition capabilities; no further restrictions are imposed here.
[0045] The task type can be to analyze the macroscopic properties of the target liquid metal (such as droplet morphology, contact angle, etc.). In this case, the target task data obtained by analysis usually includes the property parameters of the target liquid metal (such as surface energy, thermal conductivity, dynamic viscosity, melting point, etc.). These property parameters may be complete or some thermodynamic parameters may be missing. The task type can also be to determine the unknown thermodynamic parameters of the target liquid metal. In this case, the target task data obtained by analysis usually includes the known thermodynamic parameters of the target liquid metal.
[0046] Optionally, when the task type is to analyze the macroscopic properties of the target liquid metal, the system can first call the first intelligent agent to analyze the target task data, determine whether the attribute parameters are complete, and whether the macroscopic dimension simulation of the target liquid metal can be completed directly based on the current attribute parameters.
[0047] If the target thermodynamic parameters required for macroscopic dimension simulation are missing from the attribute parameters, the first agent outputs the type of the target thermodynamic parameter. The system can then call the second agent to generate first code based on the target task data and the type of the target thermodynamic parameter. This first code is used to perform microscopic-scale simulation of the target liquid metal to determine the target thermodynamic parameter. The molecular dynamics simulation application can be LAMMPS, and the first code is the corresponding LAMMPS code.
[0048] Subsequently, the system can invoke the first intelligent agent to generate second code based on the target task data and target thermodynamic parameters. This second code is used to perform macroscopic simulations of the target liquid metal to determine its macroscopic properties. Then, the system calls the finite element simulation application to execute the second code, obtaining the macroscopic properties of the target liquid metal. The finite element simulation application can use COMSOL Multi-physics in conjunction with MATLAB, and the second code is the corresponding COMSOL-MATLAB joint code.
[0049] If the attribute parameters include the complete attribute parameters of the target liquid metal, the system can directly call the first agent to generate second code based on the target task data to perform macroscopic simulation of the target liquid metal to determine the macroscopic characteristics of the target liquid metal, and then call the finite element simulation application to execute the second code to obtain the macroscopic characteristics of the target liquid metal.
[0050] Optionally, when the task type is to determine unknown target thermodynamic parameters of a target liquid metal, the system can directly call the second agent to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameters based on the target task data and the type of target thermodynamic parameters. Then, the system can call the molecular dynamics simulation application to execute the first code and obtain the target thermodynamic parameters.
[0051] Understandably, in the above-mentioned automated process, to ensure the accuracy of the target thermodynamic parameters or macroscopic characteristics obtained from the simulation, the first and second intelligent agents need to be able to accurately generate the relevant code required. Therefore, the training process of the first and second intelligent agents has been specifically optimized in this embodiment.
[0052] As an optional implementation, the first agent can be trained through the following steps S1-S7:
[0053] S1, Obtain the pre-trained initial large language model, which includes a first sub-model for analyzing missing parameters and a second sub-model for generating code.
[0054] S2, acquire multiple sets of historical task data, and use each set of historical task data as a first training sample, wherein each set of historical task data includes complete property parameters of a liquid metal.
[0055] Understandably, in order to improve the versatility of the first intelligent agent, historical task data corresponding to various liquid metals can be obtained here for training.
[0056] S3. For each set of historical task data, obtain the structured inference data corresponding to the historical task data, and use the structured inference data as the first sample label of the corresponding first training sample. The structured inference data includes: executable code used to perform macroscale simulation of liquid metal corresponding to historical task data to determine the macroscopic characteristics of liquid metal, and inference chain of executable code obtained by inference from historical task data.
[0057] Unlike the conventional approach of directly using historically generated executable code as sample labels, in order to ensure the reliability of the agent during the reasoning process, this application proposes a "reasoning-answer" structured data, composed of the complete reasoning chain obtained from historical task data and the finally generated executable code, as sample labels based on the thinking chain approach. This data simultaneously guides the model's reasoning process and the final generated content.
[0058] S4, based on multiple sets of first training samples and first sample labels, performs supervised fine-tuning of the second sub-model.
[0059] The code generated by the second sub-model is primarily a COMSOL-MATLAB joint code with a relatively simple structure. Therefore, simple supervised fine-tuning of the second sub-model using the first training samples and their labels is sufficient to ensure its output accuracy. Of course, users can continue to use other methods, such as reinforcement learning, to further optimize and train the second sub-model as needed; no specific limitations are made here.
[0060] S5. For each set of historical task data, randomly select at least one third thermodynamic parameter from the complete attribute parameters in the historical task data, take all the remaining attribute parameters as a second training sample, and take the type of all the third thermodynamic parameters as the second sample label of the corresponding second training sample.
[0061] S6, based on multiple sets of second training samples and second sample labels, performs supervised fine-tuning of the first sub-model.
[0062] Through the above two-step training process, the first sub-model can accurately identify whether the property parameters of the input liquid metal are complete, and accurately determine the type of missing thermodynamic parameters if they are incomplete.
[0063] S7 integrates the trained first and second sub-models into the first agent.
[0064] As an optional implementation, the second agent can be trained through the following steps S1-S6:
[0065] S1, Obtain the pre-trained initial large language model for code generation.
[0066] S2, acquire multiple sets of historical task data, where each set of historical task data includes complete property parameters of a liquid metal.
[0067] S3, for each set of historical task data, randomly select at least one third thermodynamic parameter from the complete attribute parameters in the historical task data, take the type of the third thermodynamic parameter and all the remaining attribute parameters as a third training sample, obtain the structured inference data corresponding to the historical task data, and use the structured inference data as the third sample label of the corresponding third training sample. The structured inference data includes: executable code used to perform microscale simulation of liquid metal corresponding to the historical task data to determine the third thermodynamic parameter, and inference chain of the executable code obtained by inferring from the historical task data.
[0068] Similar to training the first intelligent agent, this method also uses the thought chain approach. The complete reasoning chain from historical task data to obtain executable code, along with the final generated executable code, forms a "reasoning-answer" structured data as sample labels to simultaneously guide the model's reasoning process and the final generated content.
[0069] S4. Divide all third training samples and their corresponding third sample labels into a first training subset and a second training subset according to a preset ratio, wherein the number of samples in the first training subset is less than the number of samples in the second training subset.
[0070] Considering that the code generated by the second agent is mainly LAMMPS code with a relatively complex structure, simple supervised fine-tuning may not guarantee the reliability of the final training results. Therefore, this application proposes a training method that combines cold start training (i.e. supervised fine-tuning) with reinforcement learning. To this end, the acquired training samples are first divided into two parts: the first training subset is used for the cold start training stage, which requires less data, such as 15%; the second training subset is used for the reinforcement learning training stage, which requires more data, such as 75%.
[0071] S5, based on the first training subset, performs supervised fine-tuning of the initial large language model to obtain the first large language model.
[0072] S6. The first large language model is trained using a group relative policy optimization algorithm combined with the second training subset to obtain the second agent.
[0073] Optionally, when using the group relative policy optimization algorithm to train the first language model for reinforcement learning, each third training sample in the second training subset can be input into the first language model to generate multiple third codes; then, a reward function of the target dimension can be constructed based on the multiple third codes and the multiple third sample labels, wherein the target dimension includes at least one of the following: physical logic consistency, code executability, inference chain rigor, and computational efficiency; the model parameters of the first language model are adjusted based on the reward function combined with the gradient regularization strategy.
[0074] Taking a target dimension that includes all four dimensions mentioned above as an example, we can first construct four sub-reward functions from the dimensions of physical logic consistency, code executability, inference chain rigor, and computational efficiency, and then determine the weight coefficients of each dimension, for example, 0.3, 0.3, 0.2, and 0.2 respectively. Then, we can obtain the target reward function by weighted summation of the four sub-reward functions.
[0075] As an optional implementation, to verify the accuracy of the final simulation results, after simulating the macroscopic properties of the target liquid metal, the system can also obtain the actual macroscopic properties of the target liquid metal from references or historical experimental data. Then, it calculates the similarity between the simulated macroscopic properties and the actual macroscopic properties. This step, following the task instruction parsing, code self-generation, and simulation self-execution described above, achieves self-verification of the simulation results, forming a closed loop in the entire automated process. Finally, the system can simultaneously output the simulated macroscopic properties and the calculated similarity to demonstrate the reliability of the simulation results to the user.
[0076] It should be noted that the above similarity calculation and the approach to outputting the results are also applicable to scenarios where the task type is to determine the unknown thermodynamic parameters of the target liquid metal, which will not be elaborated on here.
[0077] Optionally, if the calculated similarity is not less than a preset similarity threshold, the system can also store the generated first code and second code in the target database; then periodically use the code stored in the target database to train the first agent and the second agent, update the model parameters of the first agent and the second agent, thereby optimizing the performance of the first agent and the second agent.
[0078] Optionally, the system can also invoke a question-guiding agent to analyze the task type and generate question-guiding prompts, which are used to guide the target object to ask questions related to the task type.
[0079] Understandably, in order to improve the user experience when using the system to perform cross-scale analysis of liquid metals, a question-guided agent can be pre-trained. After completing the relevant simulation analysis corresponding to the user's task instructions, the question-guided agent can perform correlation analysis on the task type of the task instructions, thereby guiding the user to ask more in-depth questions to understand the relevant properties of liquid metals.
[0080] The following example illustrates the complete workflow of the cross-scale analysis method for liquid metals proposed in this application, including steps S1-S7, using a specific application scenario:
[0081] S1, User inputs natural language request: "Predict Ga under a temperature gradient of 30-60℃..." 75 In 25 Morphological changes of droplets at silicon / gallium nitride heterostructure interfaces;
[0082] S2, the intent recognition model determines that it is a COMSOL simulation task and transmits the requirement to the FEM Agent (first agent).
[0083] S3, the FEM Agent detected some thermodynamic parameters (such as Ga). 75 In 25 The loss of surface energy at different temperatures triggers the MD Agent (second intelligent agent) to execute the simulation;
[0084] S4, MD Agent automatically extracts relevant parameters, generates LAMMPS code, calls the LAMMPS engine to calculate missing thermodynamic parameters, and outputs the results to FEM Agent;
[0085] S5, after the FEM Agent confirms that all parameters are complete, generates the experimental plan and COMSOL-MATLAB joint code, and calls the COMSOL Multi-physics software to perform the simulation and automatically draw the three views of the droplet profile;
[0086] S6 compares the simulation results with the experimental data of the checkerboard interface observation system and outputs a similarity assessment (≥95%).
[0087] S7 provides users with simulation results, three-view diagrams, and verification reports, and guides them to ask more in-depth questions (such as "compare the differences in interface behavior of gallium-based alloys with different compositions under the same conditions").
[0088] In this embodiment, after obtaining the target task instruction for the target liquid metal, the intent recognition model is first invoked to analyze the target task instruction to obtain the task type and target task data, which includes the attribute parameters of the target liquid metal. If the task type is to analyze the macroscopic characteristics of the target liquid metal, a first agent is invoked to analyze the target task data and determine the type of the missing target thermodynamic parameter in the attribute parameters. Then, a second agent is invoked to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameter based on the target task data and the type of the target thermodynamic parameter. A molecular dynamics simulation application is then invoked to execute the first code to obtain the target thermodynamic parameter. Afterward, the first agent is invoked to generate second code for performing macroscale simulation of the target liquid metal to determine its macroscopic characteristics based on the target task data and the target thermodynamic parameter. A finite element simulation application is then invoked to execute the second code to obtain the macroscopic characteristics of the target liquid metal. This solution achieves a fully automated analysis process of "task parsing - code self-generation - simulation self-execution" by coordinating and scheduling multiple models / agents. Through the code generated by the agent, simulation analysis of liquid metals at the micro and macro dimensions can be completed quickly, and its accurate thermodynamic parameters and macroscopic properties can be obtained efficiently. This effectively solves the technical problems of the difficulty in obtaining the thermodynamic parameters of liquid metals and the low efficiency of cross-scale simulation.
[0089] Example 2
[0090] According to an embodiment of this application, a multi-scale analysis apparatus for liquid metals is also provided for implementing the multi-scale analysis method for liquid metals in Embodiment 1, such as... Figure 2 As shown, the multi-scale analysis device for liquid metal includes at least: an acquisition module 21, a first analysis module 22, a second analysis module 23, and a third analysis module 24, wherein:
[0091] The acquisition module 21 is used to acquire the target task instruction for the target liquid metal, and call the intent recognition model to analyze the target task instruction to obtain the task type and target task data, wherein the target task data includes the attribute parameters of the target liquid metal.
[0092] The first analysis module 22 is used to call the first intelligent agent to analyze the target task data and determine the type of missing target thermodynamic parameters in the attribute parameters when the task type is to analyze the macroscopic properties of the target liquid metal.
[0093] The second analysis module 23 is used to call the second intelligent agent to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameters based on the target task data and the type of target thermodynamic parameters, and to call the molecular dynamics simulation application to execute the first code to obtain the target thermodynamic parameters.
[0094] The third analysis module 24 is used to call the first intelligent agent to generate second code based on the target task data and target thermodynamic parameters to perform macroscopic simulation of the target liquid metal to determine the macroscopic properties of the target liquid metal, and to call the finite element simulation application to execute the second code to obtain the macroscopic properties of the target liquid metal.
[0095] The following section describes the functions of each module of the multiscale analysis device for liquid metals, using the specific implementation process as an example.
[0096] Optionally, the cross-scale analysis device for liquid metal in this application embodiment further includes a first training module, used to train a first agent in the following manner: acquiring a pre-trained initial large language model, wherein the initial large language model includes a first sub-model for analyzing missing parameters and a second sub-model for generating code; acquiring multiple sets of historical task data, and using each set of historical task data as a first training sample, wherein each set of historical task data includes complete attribute parameters of a liquid metal; for each set of historical task data, acquiring structured inference data corresponding to the historical task data, and using the structured inference data as the first sample label of the corresponding first training sample, wherein the structured inference data includes: data for analyzing historical task data; and for each set of historical task data, acquiring structured inference data corresponding to the historical task data, and using the structured inference data as the first sample label of the corresponding first training sample, wherein the structured inference data includes: data for analyzing historical task data; and for analyzing missing parameters in the historical task data. The executable code for determining the macroscopic properties of liquid metal is obtained by performing macroscopic-scale simulations of the liquid metal corresponding to the task data. The inference chain of the executable code is derived from historical task data. For each set of historical task data, at least one third thermodynamic parameter is randomly selected from the complete attribute parameters in the historical task data. All remaining attribute parameters are used as a second training sample, and the types of all third thermodynamic parameters are used as the second sample labels of the corresponding second training samples. The first sub-model is supervised and fine-tuned based on multiple sets of second training samples and second sample labels. The second sub-model is supervised and fine-tuned based on multiple sets of first training samples and first sample labels. The trained first and second sub-models are integrated into a first intelligent agent.
[0097] Optionally, the cross-scale analysis device for liquid metals in this application embodiment further includes a second training module, used to train a second agent in the following manner: acquiring a pre-trained initial large language model for generating code; acquiring multiple sets of historical task data, wherein each set of historical task data includes complete property parameters of a liquid metal; for each set of historical task data, randomly dividing at least one third thermodynamic parameter from the complete property parameters in the historical task data, taking the type of the third thermodynamic parameter and all remaining property parameters as a third training sample, acquiring structured inference data corresponding to the historical task data, and taking the structured inference data as the third sample of the corresponding third training sample. This tag, in which the structured inference data includes: executable code used to perform microscale simulation of liquid metal corresponding to historical task data to determine the third thermodynamic parameter, and an inference chain for inferring the executable code from historical task data; dividing all third training samples and their corresponding third sample labels into a first training subset and a second training subset according to a preset ratio, wherein the number of samples in the first training subset is less than the number of samples in the second training subset; supervised fine-tuning of the initial large language model based on the first training subset to obtain the first large language model; and reinforcement learning training of the first large language model using a group relative policy optimization algorithm combined with the second training subset to obtain the second agent.
[0098] In the second training module, when using the group relative policy optimization algorithm to perform reinforcement learning training on the first large language model, each third training sample in the second training subset can be input into the first large language model to generate multiple third codes. Based on the multiple third codes and multiple third sample labels, a reward function for the target dimension is constructed, wherein the target dimension includes at least one of the following: physical and logical consistency, code executability, inference chain rigor, and computational efficiency. The model parameters of the first large language model are adjusted based on the reward function combined with the gradient regularization strategy.
[0099] Optionally, when the acquisition module determines that the task type is to determine the target thermodynamic parameters of the target liquid metal, the second analysis module directly calls the second agent to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameters based on the target task data and the type of the target thermodynamic parameters, wherein the target thermodynamic parameters are thermodynamic parameters not provided in the attribute parameters; the molecular dynamics simulation application is called to execute the first code to obtain the target thermodynamic parameters.
[0100] Optionally, after the first analysis module calls the first intelligent agent to analyze the target task data, if it determines that the target task data includes the complete property parameters of the target liquid metal, it directly calls the first intelligent agent to generate second code based on the target task data to perform macroscale simulation of the target liquid metal to determine the macroscopic characteristics of the target liquid metal; and calls the finite element simulation application to execute the second code to obtain the macroscopic characteristics of the target liquid metal.
[0101] Optionally, the cross-scale analysis device for liquid metal in this application embodiment further includes a verification module, which is used to obtain the actual macroscopic characteristics of the target liquid metal from references or historical experimental data after obtaining the macroscopic characteristics of the target liquid metal; calculate the similarity between the simulated macroscopic characteristics and the actual macroscopic characteristics; and output the macroscopic characteristics and similarity at the same time.
[0102] Optionally, if the similarity is not less than a preset similarity threshold, the verification module can store the first code and the second code in the target database; the first training module and the second training module can periodically use the code stored in the target database to train the first agent and the second agent, and update the model parameters of the first agent and the second agent.
[0103] Optionally, the cross-scale analysis device for liquid metal in this application embodiment further includes a guidance module, which is used to call a question guidance agent to analyze the task type and generate question guidance prompts, wherein the question guidance prompts are used to guide the target object to raise questions related to the task type.
[0104] It should be noted that each module in the liquid metal cross-scale analysis device in this application corresponds one-to-one with each implementation step of the liquid metal cross-scale analysis method in Example 1. Since Example 1 has been described in detail, some details not shown in this embodiment can be referred to Example 1, and will not be elaborated further here.
[0105] Example 3
[0106] According to an embodiment of this application, a computer program product is also provided, which includes a computer program, wherein when the computer program is executed by a processor, it implements the cross-scale analysis method for liquid metal in Embodiment 1.
[0107] According to an embodiment of this application, a non-volatile storage medium is also provided, which includes a stored computer program, wherein the device containing the non-volatile storage medium executes the cross-scale analysis method for liquid metal in Embodiment 1 by running the computer program.
[0108] According to an embodiment of this application, a processor is also provided for running a computer program, wherein the computer program executes the cross-scale analysis method for liquid metal in Embodiment 1.
[0109] According to an embodiment of this application, an electronic device is also provided, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the cross-scale analysis method for liquid metals in Embodiment 1 via the computer program.
[0110] Specifically, the computer program executes the following steps during runtime: It acquires a target task instruction for the target liquid metal and calls an intent recognition model to analyze the instruction, obtaining the task type and target task data, where the target task data includes the attribute parameters of the target liquid metal; when the task type is to analyze the macroscopic characteristics of the target liquid metal, it calls a first agent to analyze the target task data and determine the type of the missing target thermodynamic parameter in the attribute parameters; it calls a second agent to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameter based on the target task data and the type of the target thermodynamic parameter, and calls a molecular dynamics simulation application to execute the first code to obtain the target thermodynamic parameter; it calls the first agent to generate second code for performing macroscale simulation of the target liquid metal to determine the macroscopic characteristics of the target liquid metal based on the target task data and the target thermodynamic parameter, and calls a finite element simulation application to execute the second code to obtain the macroscopic characteristics of the target liquid metal.
[0111] As an alternative implementation, the above-mentioned electronic device may exist in the form of a mobile terminal, a computer terminal, or a similar computing device. Figure 3 A hardware block diagram of an electronic device for implementing a cross-scale analysis method for liquid metals is shown. Figure 3 As shown, the electronic device 30 may include one or more (shown as 302a, 302b, ..., 302n in the figure) processors 302 (processors 302 may include, but are not limited to, processing devices such as microprocessors or programmable logic devices), a memory 304 for storing data, and a transmission device 306 for communication functions. In addition, it may also include: a display, an input / output interface, a universal serial bus port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 3 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, electronic device 30 may also include... Figure 3 The more or fewer components shown, or having the same Figure 3 The different configurations shown.
[0112] It should be noted that the aforementioned one or more processors 302 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element of the electronic device 30. As involved in the embodiments of this application, the data processing circuit serves as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).
[0113] The memory 304 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the cross-scale analysis method of liquid metal in this embodiment of the application. The processor 302 executes various functional applications and data processing by running the software programs and modules stored in the memory 304, thereby realizing the aforementioned cross-scale analysis method of liquid metal. The memory 304 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 304 may further include memory remotely located relative to the processor 302, and these remote memories can be connected to the electronic device 30 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.
[0114] The transmission device 306 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the electronic device 30. In one example, the transmission device 306 includes a network adapter that can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 306 may be a radio frequency module used for wireless communication with the Internet.
[0115] The display can be, for example, a touchscreen LCD display that allows the user to interact with the user interface of the electronic device 30.
[0116] The sequence numbers of the above embodiments are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0117] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0118] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.
[0119] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0120] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0121] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory, random access memory, portable hard drive, magnetic disk, or optical disk.
[0122] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. A cross-scale analysis method for liquid metals, characterized in that, include: Obtain the target task instruction for the target liquid metal, and call the intent recognition model to analyze the target task instruction to obtain the task type and target task data, wherein the target task data includes the attribute parameters of the target liquid metal; When the task type is to analyze the macroscopic properties of the target liquid metal, the first intelligent agent is invoked to analyze the target task data and determine the type of the missing target thermodynamic parameter in the attribute parameters; The second intelligent agent is invoked to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameters based on the target task data and the type of the target thermodynamic parameters. The molecular dynamics simulation application is then invoked to execute the first code to obtain the target thermodynamic parameters. The first intelligent agent is invoked to generate second code based on the target task data and the target thermodynamic parameters for performing macroscopic simulation of the target liquid metal to determine the macroscopic properties of the target liquid metal, and the finite element simulation application is invoked to execute the second code to obtain the macroscopic properties of the target liquid metal.
2. The method according to claim 1, characterized in that, The training process of the first agent includes: Obtain a pre-trained initial large language model, wherein the initial large language model includes a first sub-model for analyzing missing parameters and a second sub-model for generating code; Acquire multiple sets of historical task data and use each set of historical task data as a first training sample. Each set of historical task data includes complete property parameters of a liquid metal. For each set of historical task data, obtain the structured inference data corresponding to the historical task data, and use the structured inference data as the first sample label of the corresponding first training sample. The structured inference data includes: executable code for performing macroscale simulation of liquid metal corresponding to the historical task data to determine the macroscopic characteristics of the liquid metal, and an inference chain for inferring the executable code from the historical task data. For each set of historical task data, at least one third thermodynamic parameter is randomly selected from the complete attribute parameters in the historical task data, and all remaining attribute parameters are used as a second training sample, and the types of all third thermodynamic parameters are used as the second sample labels of the corresponding second training sample. The first sub-model is supervised and fine-tuned based on multiple sets of the second training samples and the second sample labels. The second sub-model is supervised and fine-tuned based on multiple sets of the first training samples and the first sample labels; The first and second sub-models, after training, are integrated into the first intelligent agent.
3. The method according to claim 1, characterized in that, The training process for the second agent includes: Obtain a pre-trained initial large language model for code generation; Acquire multiple sets of historical task data, each set of historical task data including complete property parameters of a liquid metal; For each set of historical task data, at least one third thermodynamic parameter is randomly selected from the complete attribute parameters in the historical task data. The type of the third thermodynamic parameter and all remaining attribute parameters are used as a third training sample. The structured inference data corresponding to the historical task data is obtained, and the structured inference data is used as the third sample label of the corresponding third training sample. The structured inference data includes: executable code for performing microscale simulation of liquid metal corresponding to the historical task data to determine the third thermodynamic parameter, and an inference chain for inferring the executable code from the historical task data. All third training samples and their corresponding third sample labels are divided into a first training subset and a second training subset according to a preset ratio, wherein the number of samples in the first training subset is less than the number of samples in the second training subset. The initial large language model is subjected to supervised fine-tuning based on the first training subset to obtain the first large language model. The first large language model is trained using a group relative policy optimization algorithm combined with the second training subset to obtain the second agent.
4. The method according to claim 3, characterized in that, The first large language model is trained using a group-relative policy optimization algorithm combined with the second training subset, including: Each third training sample in the second training subset is input into the first large language model to generate multiple third codes; A reward function for a target dimension is constructed based on the plurality of third codes and the plurality of third sample labels, wherein the target dimension includes at least one of the following: physical and logical consistency, code executability, inference chain rigor, and computational efficiency; The model parameters of the first large language model are adjusted based on the reward function combined with the gradient regularization strategy.
5. The method according to claim 1, characterized in that, The method further includes: When the task type is to determine the target thermodynamic parameters of the target liquid metal, the second agent is directly invoked to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameters based on the target task data and the type of the target thermodynamic parameters, wherein the target thermodynamic parameters are thermodynamic parameters not provided in the attribute parameters; The first code is executed by calling a molecular dynamics simulation application to obtain the target thermodynamic parameters.
6. The method according to claim 1, characterized in that, After invoking the first intelligent agent to analyze the target task data, the method further includes: If the target task data includes complete property parameters of the target liquid metal, the first agent is directly invoked to generate second code based on the target task data for performing macroscopic simulation of the target liquid metal to determine the macroscopic properties of the target liquid metal; The second code is executed by calling a finite element simulation application to obtain the macroscopic properties of the target liquid metal.
7. The method according to claim 1, characterized in that, After obtaining the macroscopic properties of the target liquid metal, the method further includes: Obtain the actual macroscopic properties of the target liquid metal from references or historical experimental data; The similarity between the macroscopic characteristics obtained from the simulation and the actual macroscopic characteristics is calculated. Simultaneously output the macroscopic characteristics and the similarity.
8. The method according to claim 7, characterized in that, The method further includes: If the similarity is not less than a preset similarity threshold, the first code and the second code are stored in the target database; The first agent and the second agent are periodically trained using the code stored in the target database, and the model parameters of the first agent and the second agent are updated.
9. The method according to claim 1, characterized in that, The method further includes: The question guidance agent is invoked to analyze the task type and generate question guidance prompts, wherein the question guidance prompts are used to guide the target object to ask a question related to the task type.
10. A multi-scale analysis device for liquid metals, characterized in that, include: The acquisition module is used to acquire the target task instruction for the target liquid metal, and call the intent recognition model to analyze the target task instruction to obtain the task type and target task data, wherein the target task data includes the attribute parameters of the target liquid metal; The first analysis module is used to, when the task type is to analyze the macroscopic properties of the target liquid metal, call the first intelligent agent to analyze the target task data and determine the type of the missing target thermodynamic parameter in the attribute parameters; The second analysis module is used to call the second intelligent agent to generate first code for performing microscale simulation of the target liquid metal to determine the target thermodynamic parameters based on the target task data and the type of the target thermodynamic parameters, and to call the molecular dynamics simulation application to execute the first code to obtain the target thermodynamic parameters; The third analysis module is used to call the first agent to generate second code based on the target task data and the target thermodynamic parameters for performing macroscopic simulation of the target liquid metal to determine the macroscopic properties of the target liquid metal, and to call the finite element simulation application to execute the second code to obtain the macroscopic properties of the target liquid metal.
11. A computer program product, characterized in that, include: A computer program, wherein when executed by a processor, the computer program implements the cross-scale analysis method for liquid metals according to any one of claims 1 to 9.
12. An electronic device, characterized in that, include: A memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute, via the computer program, the method for cross-scale analysis of liquid metals according to any one of claims 1 to 9.