Intelligent collaborative control method and system for molecular dynamics simulation

Abstract

This invention discloses an intelligent collaborative control method and system for molecular dynamics simulation, belonging to the field of molecular dynamics simulation technology. The method includes: calling a pre-set knowledge graph and a deep learning inference model to decompose received simulation task data, generating an execution logic chain and a target software combination; inputting the structural features of the task data into a reinforcement learning model to output an initial parameter combination, and calling the computational resources to collaboratively run the software combination based on the logic chain; during operation, real-time collection and extraction of three-level indicator data—computational resources, dynamics, and result validity—and comparison with preset three-level monitoring thresholds; when any indicator data exceeds the limit, it is used as state feedback input to the reinforcement learning model to update the running parameters through output parameter fine-tuning actions, or generating resource expansion instructions, performing closed-loop feedback until the task is completed. This achieves automated workflow construction and dynamic closed-loop control, improving computational resource utilization and simulation stability.

Description

Technical Field

[0001] This invention belongs to the field of molecular dynamics simulation technology, specifically relating to an intelligent collaborative control method and system for molecular dynamics simulation. Background Technology

[0002] Molecular dynamics simulation, based on the principles of classical or quantum mechanics, is a core technology that reveals the structure-activity relationship between microstructure and macroscopic properties by calculating the trajectories of atoms and molecules in a molecular system. It plays an irreplaceable role in fields such as new material development (e.g., high-energy-density battery electrode materials, high-strength composite materials), innovative drug design (e.g., analysis of binding modes between target proteins and small molecule drugs), and catalyst performance optimization (e.g., identification of active sites in heterogeneous catalytic reactions).

[0003] With the continuous upgrading of scientific research needs, molecular dynamics simulations are developing towards large-scale, high-precision, and multi-system approaches. However, traditional simulation methods have significant limitations, which seriously restrict the improvement of scientific research efficiency. (1) Traditional manual workflow construction is inefficient: Traditional molecular dynamics simulations require researchers to manually complete multiple steps, including simulation target decomposition, force field selection, software selection (such as GROMACS), input file writing, and cross-software interface adaptation script development. Different software has large differences in input and output formats and inconsistent parameter setting rules, requiring professionals to spend a lot of time to complete the workflow construction. Non-domain experts find it difficult to master the complex operation logic, resulting in a high technical threshold.

[0004] (2) Existing auxiliary tools are limited in function and lack systematic integration: Some molecular simulation auxiliary tools already exist, such as parameter optimization plugins included in some software and trajectory analysis tools developed by third parties. However, these tools are isolated in function and cannot achieve full coverage from task input to result output. At the same time, the data formats between tools are incompatible, making information exchange impossible.

[0005] (3) Parameter selection is highly arbitrary, resulting in a low simulation success rate: Key parameters for molecular dynamics simulations include time step, simulation duration, temperature control method, pressure coupling mode, and force field parameters. Existing parameter recommendations are mostly based on fixed empirical formulas or literature reports, without considering the specificity of the research object, leading to a large number of repetitive experiments and a serious waste of research resources.

[0006] To address this, an intelligent collaborative control method and system for molecular dynamics simulation is proposed. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide an intelligent collaborative control method and system for molecular dynamics simulation, thereby solving the problems in the prior art.

[0008] The objective of this invention can be achieved through the following technical solutions: An intelligent collaborative control method for molecular dynamics simulation includes the following steps: The system receives molecular simulation task data, calls a pre-set knowledge graph to extract the entities and associated features corresponding to the molecular simulation task data, inputs the extracted features into a deep learning inference model, and decomposes the molecular simulation task data through the deep learning inference model to generate an execution logic chain and the corresponding target software combination. The structural features of the molecular simulation task data are extracted and input into the reinforcement learning model. An initial parameter combination is output. Based on the execution logic chain and the initial parameter combination, computing resources are called to co-run the target software combination to generate running status data and trajectory data. During the operation of the target software combination, basic process indicator data is collected in real time, and dynamic indicator data and result validity indicator data are extracted periodically. The collected and extracted indicator data are compared with preset three-level monitoring thresholds, which include: computing resource threshold, dynamic indicator threshold and result validity threshold. When any of the aforementioned indicator data exceeds any of the corresponding three-level monitoring thresholds based on the comparison results, the exceeding indicator data is input as state feedback to the reinforcement learning model. The reinforcement learning model outputs parameter fine-tuning actions to update the running parameters of the current process, or generates a computing resource expansion instruction independent of the reinforcement learning model based on the exceeding indicator data. The above data collection, comparison and feedback are executed in a closed-loop control manner until the preset simulation task completion conditions are met.

[0009] Furthermore, after generating the corresponding target software combination, it also includes: Based on the node order of each software in the target software assembly in the execution logic chain, a preset standardized data format protocol is retrieved; Based on the standardized data format protocol, the software adapter corresponding to the target software combination is called to convert the atomic coordinate data and force field parameter data in the molecular simulation task data into the corresponding target software input format, and generate a cross-software interface call script.

[0010] Furthermore, before inputting the structural features into the reinforcement learning model, a step of pre-training the reinforcement learning model using a historical simulation dataset is included. The network control logic of the reinforcement learning model is as follows: The structural features of the molecular simulation task data, the simulation target, the current parameter combination, and the preset historical simulation result data combination are defined as the state space; The direction and magnitude of the adjustment for the current operating parameters are set as the action space; A reward function is constructed with parameters such as quantified simulation success rate, result accuracy, and computational resource consumption. The reward function is then used to guide the reinforcement learning model to iteratively optimize.

[0011] Furthermore, the steps of comparing the collected and extracted indicator data with the preset three-level monitoring thresholds include: For the first level of monitoring, the CPU utilization rate, memory utilization rate, and disk read / write speed data of the computing resources called are collected in real time and compared with the computing resource threshold. For the second-level monitoring dimension, the currently generated trajectory data is extracted, and the average system energy fluctuation and root mean square deviation of atoms are calculated and compared with the dynamic index threshold. For the third-level monitoring dimension, the dynamic feature result data generated in stages is extracted, and the fitting deviation value between the dynamic feature result data and the preset reference experimental data is calculated and compared with the result validity threshold.

[0012] Furthermore, when it is determined from the comparison results that any of the aforementioned indicator data exceeds any of the corresponding three-level monitoring thresholds, the following feedback control action is executed: If the indicator data of the first-level monitoring dimension exceeds the computing resource threshold, a new computing power scheduling expansion instruction for the current process node is sent to the distributed computing platform. If the indicator data of the second-level monitoring dimension exceeds the threshold of the dynamic indicator, the reinforcement learning model is triggered to output the parameter fine-tuning action, generate updated simulation control parameters, update the simulation control parameters corresponding to the current breakpoint state, and resume simulation.

[0013] Furthermore, before invoking the pre-built knowledge graph, the process includes the following steps: Establish an ontology data dictionary covering the research object ontology, software ontology, and force field ontology; Entity words and entity association features are extracted from publicly available literature text databases and historical work records using natural language processing algorithms. After disambiguation and merging of the extracted entity words, they are mapped to the ontology data dictionary to generate the preset knowledge graph.

[0014] Furthermore, after the preset simulation task completion conditions are met, it also includes: Extract the final target software combination and the final parameter combination for completing the iteration corresponding to this execution, add category tags, and store them in the knowledge base; Receive new molecular simulation task data and calculate the feature similarity between the new molecular simulation task data and each historical data in the knowledge base; If the feature similarity is greater than the preset matching threshold, the final target software combination and the final parameter combination corresponding to the historical data with the highest matching degree will be output as the recommended parameter set.

[0015] An intelligent collaborative control system for molecular dynamics simulations, executing the above method, includes: The task decomposition module is used to receive molecular simulation task data, call a preset knowledge graph to extract the entities and associated features corresponding to the molecular simulation task data, input the extracted features into a deep learning inference model, decompose the molecular simulation task data through the deep learning inference model, and generate an execution logic chain and the corresponding target software combination. The parameter initialization and execution module is used to extract the structural features of the molecular simulation task data, input the structural features into the reinforcement learning model, output the initial parameter combination, and, based on the execution logic chain and the initial parameter combination, call computing resources to collaboratively run the target software combination to generate running status data and trajectory data. A multi-level monitoring module is used to collect basic process indicator data in real time during the operation of the target software combination, and periodically extract dynamic indicator data and result validity indicator data. The collected and extracted indicator data are compared with preset three-level monitoring thresholds, including computing resource threshold, dynamic indicator threshold and result validity threshold. The closed-loop intervention module is used to input the exceeding indicator data as state feedback to the reinforcement learning model when it is determined from the comparison results that any of the indicator data exceeds any of the corresponding three-level monitoring thresholds. The reinforcement learning model outputs parameter fine-tuning actions to update the running parameters of the current process, or generates a computing resource expansion instruction independent of the reinforcement learning model based on the exceeding indicator data. The above data collection, comparison and feedback are executed in a closed-loop control manner until the preset simulation task completion conditions are met.

[0016] A computer storage medium storing a readable program, which, when executed, instructs a computing device to perform an intelligent collaborative control method for molecular dynamics simulation as described above.

[0017] An electronic device includes: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform an operation corresponding to the intelligent cooperative control method for molecular dynamics simulation described above.

[0018] The beneficial effects of this invention are: 1. This invention performs semantic parsing of molecular simulation task data based on a pre-built knowledge graph and automatically decomposes the task using a deep learning inference model to generate an execution logic chain. Simultaneously, this invention retrieves standardized data format protocols and software adapters to automatically convert data such as atomic coordinates into the target software's input format. This eliminates the tedious work of researchers manually writing input files and cross-software interface conversion scripts, enabling data flow and seamless collaboration between multiple heterogeneous simulation software programs, effectively shortening the time required for initial workflow setup.

[0019] 2. This invention introduces a reinforcement learning closed-loop control architecture. The system defines the structural features of the molecular system and the simulation target as the state space, and during the integral operation, it dynamically outputs parameter fine-tuning actions (such as adjusting the time step) based on the dynamic indicators fed back by the system. This scheme enables the system to dynamically optimize in the parameter space based on the specificity of different research objects, reducing computational divergence caused by improper initial parameter settings and significantly reducing the waste of resources caused by ineffective computation.

[0020] 3. During the operation of the target software combination, this invention collects three levels of monitoring data in real time, covering basic computing resources, dynamic indicators, and result validity, and compares them with preset thresholds. When the process indicators are found to be out of bounds, the system can automatically trigger parameter fine-tuning instructions for the reinforcement learning model or send computing power expansion instructions to the distributed computing platform based on the out-of-bounds data. Through this multi-dimensional closed-loop feedback intervention mechanism, the system can promptly save the simulation process from collapse, avoid program interruption caused by drastic fluctuations in local energy or underlying memory overflow, and ensure the continuity of computing tasks.

[0021] 4. After the simulation task is completed, the system of this invention extracts the verified final target software combination and the parameter combination that has completed the iteration and stores them in the knowledge base. When a new molecular simulation task is received, the system calculates feature similarity, matches and recommends the closest historical solution. This solution breaks down data silos in the traditional scientific research model, enabling subsequent research on similar molecular systems to directly inherit existing high-quality computational experience, thereby shortening the theoretical calculation and solution exploration cycle in the early stages of new material and new drug development. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the intelligent collaborative control system structure for molecular dynamics simulation of the present invention; Figure 2 This is the workflow execution logic diagram of the system of this invention; In the diagram, 1: Multi-source heterogeneous data acquisition and preprocessing system, 1-1: Public data acquisition unit, 1-2: Private data access unit, 1-3: Data preprocessing unit; 2: Knowledge graph workflow construction system, 2-1: Knowledge graph calling unit, 2-2: Task decomposition and reasoning unit, 2-3: Software adaptation script generation unit; 3: Parameter intelligent recommendation system, 3-1: Research object feature extraction unit, 3-2: Reinforcement learning reasoning unit; 4: Simulation execution and dynamic monitoring system, 4-1: Computational resource scheduling unit, 4-2: Process monitoring unit, 4-3: Anomaly adjustment unit; 5: Result analysis and reuse system, 5-1: Simulation result parsing unit, 5-2: Knowledge base storage unit, 5-3: Similar scheme matching unit; 6: User interface; 7: Distributed computing platform; 8: External data source. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 This embodiment provides an intelligent collaborative control system for molecular dynamics simulation. The system adopts a hierarchical architecture design, which is divided into a data layer, an algorithm layer, a functional layer and an application layer from bottom to top. Each layer works collaboratively to solve the problems of low efficiency in building traditional molecular dynamics simulation workflows, poor parameter matching, difficulty in reusing results and lag in process monitoring. It is applicable to molecular-scale simulation research in multiple fields such as materials science, biomedicine, energy and chemical engineering, and semiconductor research and development.

[0026] like Figure 1As shown, this system specifically comprises: a multi-source heterogeneous data acquisition and preprocessing system 1, a knowledge graph workflow construction system 2, a parameter intelligent recommendation system 3, a simulation execution and dynamic monitoring system 4, a result analysis and reuse system 5, a user interface 6, a distributed computing platform 7, and external data sources 8. The specific structure and functional logic of each subsystem and its internal units are as follows: Multi-source heterogeneous data acquisition and preprocessing system 1: This system serves as the data foundation of the overall architecture, responsible for breaking down data silos. Specifically, it comprises: a public data acquisition unit 1-1, a private data access unit 1-2, and a data preprocessing unit 1-3. The public data acquisition unit 1-1 is responsible for crawling data from external data sources 8, including international molecular simulation databases, simulation literature published in authoritative journals, and official software parameter manuals. The private data access unit 1-2 provides API interfaces and file upload functions, supporting users to access local simulation data and workflow templates, and employs data anonymization technology to ensure the security of users' private data. The data preprocessing unit 1-3 performs cleaning (removing duplicate, invalid, and abnormal data), standardization (converting atomic coordinates and trajectory data of different formats to a unified format), and structured annotation (tagging templates and parameter schemes) on the aforementioned data, thereby constructing a unified molecular simulation knowledge base and providing standardized input for upper-level spectra and models.

[0027] Knowledge Graph Workflow Construction System 2: This system is the core component that breaks through the bottleneck of traditional manual workflow construction. It comprises a knowledge graph invocation unit 2-1, a task decomposition and reasoning unit 2-2, and a software adaptation script generation unit 2-3. The knowledge graph invocation unit 2-1 maintains and invokes a knowledge graph in the field of molecular dynamics simulation based on ontology methods. The task decomposition and reasoning unit 2-2 internally deploys a deep learning reasoning model. When a user inputs requirements through the user interface 6, this unit can automatically parse and decompose the simulation task, generating an execution logic chain containing the node order. The software adaptation script generation unit 2-3 matches the most suitable combination of simulation software based on the generated execution logic chain and acts as a standardized interface adaptation layer, automatically generating software invocation scripts and cross-software interface adaptation scripts to achieve seamless data connection between different mainstream simulation software.

[0028] Parameter-based intelligent recommendation system 3: This system aims to overcome the shortcomings of manual trial-and-error parameter tuning. It comprises a research object feature extraction unit 3-1 and a reinforcement learning inference unit 3-2. The research object feature extraction unit 3-1 is responsible for quantifying and analyzing the intrinsic properties of the molecular system to be simulated. The reinforcement learning inference unit 3-2, based on a deep reinforcement learning framework, uses a pre-trained model to dynamically output and recommend optimal parameter combinations (such as force field type, time step, and temperature and pressure control parameters) according to the specificity of the research object and real-time data of the simulation process, replacing the traditional static recommendation based on fixed empirical formulas.

[0029] Simulation Execution and Dynamic Monitoring System 4: This system is responsible for the efficient operation and crash prevention of the simulation task, and includes a computing resource scheduling unit 4-1, a process monitoring unit 4-2, and an anomaly adjustment unit 4-3. The computing resource scheduling unit 4-1 interfaces with and manages the underlying distributed computing platform 7, supporting multi-threaded and multi-node parallel computing to fully utilize idle resources. Throughout the simulation execution, the process monitoring unit 4-2 synchronously collects indicators such as energy fluctuations and atomic trajectory stability from the distributed computing platform 7 in real time. The anomaly adjustment unit 4-3 continuously compares the collected data with the set multi-level early warning thresholds; once parameter anomalies or resource insufficiency are encountered, it automatically triggers parameter fine-tuning or resource expansion commands to achieve closed-loop management.

[0030] Results Analysis and Reuse System 5: This system is responsible for the accumulation of scientific research assets, comprising a simulation result analysis unit 5-1, a knowledge base storage unit 5-2, and a similar scheme matching unit 5-3. Upon completion of a task, the simulation result analysis unit 5-1 performs multi-dimensional analysis of structural, energy, and dynamic characteristics and generates a visual report, which is then sent back to the interactive interface. The knowledge base storage unit 5-2 extracts the optimal workflow and extreme parameters of the successful task and stores them in the database. When new requirements arise, the similar scheme matching unit 5-3 retrieves historical schemes using a high-dimensional similarity matching algorithm, enabling efficient reuse of experience.

[0031] At the overall development and deployment level of the system, the data layer uses distributed databases and graph databases to store heterogeneous data and knowledge graphs respectively; the algorithm layer uses Python combined with TensorFlow and PyTorch frameworks to develop deep learning and reinforcement learning models; and the functional layer adopts a microservice architecture to decouple modules. The system has flexible deployment modes, supporting both cloud deployment for small and medium-sized research teams (providing SaaS services, accessible via a web interface without local installation) and local server deployment for large research institutions to meet the highest level of data privacy protection requirements.

[0032] Example 2 This embodiment, based on the system provided in Embodiment 1, offers a corresponding intelligent collaborative control method for molecular dynamics simulation. This method fully integrates artificial intelligence, big data, and domain knowledge, achieving a highly automated collaborative closed loop from input to output.

[0033] like Figure 2 As shown, the method is executed according to the following logical steps: S1, Task Reception and Decomposition / Reconstruction The system first receives molecular simulation task data, calls a pre-set knowledge graph to extract the entities and associated features corresponding to the molecular simulation task data, and inputs the extracted features into a deep learning inference model. Specifically, the deep learning inference model adopts a sequence-to-sequence (Seq2Seq) network architecture based on an attention mechanism, such as the Transformer model. The system transforms the features such as molecular entities and force field types extracted from natural language processing into a high-dimensional input vector sequence through an embedding layer. The decoder of the model outputs a directed acyclic graph (DAG) structure sequence, where each node corresponds to a specific simulation subtask (e.g., energy minimization, NVT balancing, NPT balancing, product kinetics operation), and the edges of the graph represent the data flow direction, thereby generating a precise execution logic chain. The molecular simulation task data is decomposed through the deep learning inference model to generate the execution logic chain and the corresponding target software combination.

[0034] Before invoking the knowledge graph in this step, the system pre-constructs the underlying knowledge graph. Specifically, the system establishes an ontology data dictionary covering the research object ontology (such as small molecules, proteins, and polymers), software ontology, and force field ontology. Then, using named entity recognition and relation extraction techniques from Natural Language Processing (NLP) algorithms, it extracts entity words and entity association features from publicly available literature text databases (external data source 8) and historical work files. Next, it performs rigorous disambiguation and merging on the extracted entity words to eliminate heterogeneous ambiguities, and finally maps them to the ontology data dictionary to generate a pre-built knowledge graph. The knowledge graph is dynamically updated, continuously evolving with new literature and local data.

[0035] After generating the target software suite using deep learning, to address data incompatibility issues between different software programs, the system invokes a pre-defined standardized data format protocol based on the node order of each software program within the target software suite's execution logic chain. Based on this protocol, the system calls the software-specific adapter corresponding to the target software suite to convert the underlying data (such as atomic coordinate data and force field parameter data) in the molecular simulation task data into a proprietary input format recognizable by each target software. The pre-defined standardized data format protocol is specifically a set of intermediate metadata schemas based on JSON or XML. For example, for the spatial coordinates of atoms, the standardized protocol uniformly represents them as nanometer-level floating-point arrays of [x, y, z]. When the target software is GROMACS, the adapter automatically extracts the floating-point array according to pre-defined mapping rules and performs forced alignment and reconstruction according to the fixed column width format of the .gro file. When the target software is LAMMPS, its adapter is invoked to convert it into a .data format containing atom type IDs and coordinates, thereby achieving lossless parsing and reconstruction of the underlying heterogeneous data. Cross-software interface call scripts are also automatically generated. This mechanism completely eliminates the tedious process of manually writing conversion code, enabling seamless integration between different mainstream software.

[0036] S2, Intelligent Parameter Initialization and Resource Co-operation Execution The system extracts the structural features of the molecular simulation task data and inputs the structural features into the reinforcement learning model, outputting an initial parameter combination. Subsequently, based on the execution logic chain generated in the first step and the initial parameter combination, the system calls distributed computing resources to collaboratively run the target software combination to initiate dynamic integral calculations and generate running status data and trajectory data.

[0037] In this step, before inputting structural features into the model, the system has completed offline pre-training of the reinforcement learning model using historical simulation datasets, and conducts online learning in conjunction with real-time data during operation. This reinforcement learning model abandons traditional static configuration, and its network control logic is rigorously defined as follows: the structural features of the molecular simulation task data, the simulation target, the current running parameter combination, and the preset historical simulation result data are combined to define the "state space". The state space refers to the system's rigorous numerical definition of abstract data when constructing the reinforcement learning model: the structural features of the state space are quantized into the SMILES expression matrix of molecules or the three-dimensional coordinate tensor of atoms, and the simulation target is quantized into a one-hot encoded vector of a specific dimension; The physical adjustment direction and magnitude for the current operating parameters are limited to the "action space." Simultaneously, a "reward function" is constructed, using quantified "simulation success rate," "result accuracy," and "computational resource consumption" as comprehensive parameters. This reward function continuously guides the model to iteratively seek the optimal solution within the vast parameter space. In the construction of the reinforcement learning model, the reward function... R The design is a multi-objective weighted summation, as detailed below: in, U This represents the resource consumption rate, defined as the ratio of the currently used resources to the user's preset maximum budget, with a value ranging from 0 to 1. P This represents the simulation success rate, with a value ranging from 0 to 1, and is determined by the ratio of the number of convergences of the current parameter combination in historical similar systems to the total number of attempts. ε The result accuracy deviation rate is defined as the relative deviation between the current simulation output key properties and the standard reference value. α and β Positive weighting coefficients, default α =1.0, β =0.5.

[0038] S3, multi-level dynamic monitoring throughout the entire process During the operation of the target software combination, the system collects basic process indicator data in real time and periodically extracts dynamic indicator data and result validity indicator data. The collected and extracted indicator data are compared with preset three-level monitoring thresholds, which are strictly divided into computing resource threshold, dynamic indicator threshold and result validity threshold.

[0039] To address this comparison step, the system has constructed a robust three-dimensional monitoring network: For the first level of monitoring, the system collects underlying resource data such as CPU utilization, memory usage, and disk I / O read / write speed of the called distributed computing nodes in real time, and compares them with preset computing resource thresholds in real time to ensure that the computing power base is sufficient.

[0040] For the second level of monitoring, the system extracts the currently generated trajectory data from the process in real time, and calculates core dynamic characteristics such as the mean energy fluctuation of the physical system and the root mean square deviation of atoms (RMSD) at high frequency, and compares them with the safety dynamic index threshold.

[0041] For the third-level monitoring dimension, the system extracts the phased dynamic characteristic result data at specific checkpoints in the simulation, calculates the fitting deviation value between the calculation result and the preset reference experimental data (or theoretical expected structure-property relationship), and compares it with the result validity threshold to prevent the system from falling into invalid false convergence calculation.

[0042] S4, Closed-Loop Feedback and Automated Intervention for Anomalies When the system determines, based on the comparison results, that any of the aforementioned indicator data exceeds any of the corresponding three-level monitoring thresholds, the system immediately inputs the exceeding indicator data as status feedback into the reinforcement learning model. The reinforcement learning model then outputs parameter fine-tuning actions to hot-update the current process's running parameters, or generates a computational resource expansion instruction independent of the reinforcement learning model based on the exceeding indicator data. The system executes the above data acquisition, comparison, and feedback in a closed-loop control manner until the preset simulation task completion conditions are met.

[0043] The specific intervention mechanisms strictly correspond to the levels of exceeding limits: if the indicator data of the first-level monitoring dimension exceeds the threshold (such as a memory overflow alarm), the system will send an instruction to the distributed computing platform to increase the computing power of the current process node; if the indicator data of the second-level or third-level monitoring dimension exceeds the threshold (such as an energy divergence or structural collapse warning), the system immediately triggers the reinforcement learning model, re-infers the current state space, and outputs a parameter fine-tuning action. In actual intervention execution, the system has built-in explicit hard trigger thresholds and action mapping rules. For example, for the dynamic indicators of the second-level monitoring dimension, when the system calculates that the slope of the atom root mean square deviation (RMSD) mutation rate within 1000 consecutive steps is greater than the preset safety threshold, or when the second derivative of the system's total energy shows a sharp increase, the system determines that the conformation is about to collapse. At this point, the reinforcement learning model is triggered and outputs a priority control instruction. According to this instruction, the system updates the integration time step parameter dt of the current process according to a preset ratio, and restarts the integrator by overwriting the current breakpoint file with the updated smaller step size. This achieves precise anti-crash physical intervention (such as shortening the time step or adjusting the temperature and pressure control weights), generates updated simulation control parameters, and then directly updates the simulation control parameters corresponding to the current breakpoint state and seamlessly resumes the simulation. This effectively avoids the waste of computing power and time caused by simulation crashes, and improves resource utilization by more than 50%.

[0044] S5, Simulation Result Analysis and Knowledge Retention and Reuse After the preset simulation task completion conditions are met, the system performs a post-mortem analysis of the run. The system extracts the final target software combination corresponding to this execution and the final optimal parameter combination after countless closed-loop iterations of the reinforcement learning model. The system automatically performs multi-dimensional analysis on the results (including structural analysis, energy analysis, etc.), adds classification tags, and stores them in the knowledge base to form domain accumulation.

[0045] When the system receives new molecular simulation task data in the future, it will calculate the feature similarity between the new task data and historical data in the knowledge base. If the feature similarity is greater than a preset matching threshold, the system will directly extract and output the final target software combination and final parameter combination corresponding to the historical data with the highest matching degree (most similar) as a recommended parameter set. This achieves efficient reuse of high-quality simulation experience across time and space, significantly reducing the cycle of new drug screening or new material development for subsequent similar projects. Feature Similarity Sim Specifically as follows: in, x i Indicates new task data in i Numerical values ​​in each feature dimension; y i This indicates that a certain historical record in the knowledge base is in the [number]th [section]. i The corresponding numerical values ​​on each feature dimension; w i For the first i The weight coefficients of each feature, which characterize the degree of influence of the feature on the simulation results, are automatically generated by the system based on the mutual information statistics between the feature and the simulation results in historical data; σ i For the first i The standard deviation of each feature in the entire knowledge base is used for dimensionless processing to eliminate the difference in unit order of magnitude between different features. n The total number of feature dimensions; exp The natural exponential function maps the weighted squared distance to the interval (0,1).

[0046] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0047] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. An intelligent collaborative control method for molecular dynamics simulation, characterized in that, Includes the following steps: The system receives molecular simulation task data, calls a pre-set knowledge graph to extract the entities and associated features corresponding to the molecular simulation task data, inputs the extracted features into a deep learning inference model, and decomposes the molecular simulation task data through the deep learning inference model to generate an execution logic chain and the corresponding target software combination. The structural features of the molecular simulation task data are extracted and input into the reinforcement learning model. An initial parameter combination is output. Based on the execution logic chain and the initial parameter combination, computing resources are called to co-run the target software combination to generate running status data and trajectory data. During the operation of the target software combination, basic process indicator data is collected in real time, and dynamic indicator data and result validity indicator data are extracted periodically. The collected and extracted indicator data are compared with preset three-level monitoring thresholds, which include: computing resource threshold, dynamic indicator threshold and result validity threshold. When any of the aforementioned indicator data exceeds any of the corresponding three-level monitoring thresholds based on the comparison results, the exceeding indicator data is input as state feedback to the reinforcement learning model. The reinforcement learning model outputs parameter fine-tuning actions to update the running parameters of the current process, or generates a computing resource expansion instruction independent of the reinforcement learning model based on the exceeding indicator data. The above data collection, comparison and feedback are executed in a closed-loop control manner until the preset simulation task completion conditions are met.

2. The intelligent collaborative control method for molecular dynamics simulation according to claim 1, characterized in that, After generating the corresponding target software combination, it also includes: Based on the node order of each software in the target software assembly in the execution logic chain, a preset standardized data format protocol is retrieved; Based on the standardized data format protocol, the software adapter corresponding to the target software combination is called to convert the atomic coordinate data and force field parameter data in the molecular simulation task data into the corresponding target software input format, and generate a cross-software interface call script.

3. The intelligent collaborative control method for molecular dynamics simulation according to claim 1, characterized in that, Before inputting the structural features into the reinforcement learning model, the method further includes a step of pre-training the reinforcement learning model using a historical simulation dataset. The network control logic of the reinforcement learning model is as follows: The structural features of the molecular simulation task data, the simulation objective, the current parameter combination, and the preset historical simulation result data combination are defined as the state space; The direction and magnitude of the adjustment for the current operating parameters are set as the action space; A reward function is constructed with parameters such as quantified simulation success rate, result accuracy, and computational resource consumption. The reward function is then used to guide the reinforcement learning model to iteratively optimize.

4. The intelligent collaborative control method for molecular dynamics simulation according to claim 1, characterized in that, The steps of comparing the collected and extracted indicator data with the preset three-level monitoring thresholds include: For the first level of monitoring, the CPU utilization rate, memory utilization rate, and disk read / write speed data of the computing resources called are collected in real time and compared with the computing resource threshold. For the second-level monitoring dimension, the currently generated trajectory data is extracted, and the average system energy fluctuation and root mean square deviation of atoms are calculated and compared with the dynamic index threshold. For the third-level monitoring dimension, the dynamic feature result data generated in stages is extracted, and the fitting deviation value between the dynamic feature result data and the preset reference experimental data is calculated and compared with the result validity threshold.

5. The intelligent collaborative control method for molecular dynamics simulation according to claim 4, characterized in that, When it is determined from the comparison results that any of the aforementioned indicator data exceeds any of the corresponding three-level monitoring thresholds, the following feedback control action is executed: If the indicator data of the first-level monitoring dimension exceeds the computing resource threshold, a new computing power scheduling expansion instruction for the current process node is sent to the distributed computing platform. If the indicator data of the second-level monitoring dimension exceeds the threshold of the dynamic indicator, the reinforcement learning model is triggered to output the parameter fine-tuning action, generate updated simulation control parameters, update the simulation control parameters corresponding to the current breakpoint state, and resume simulation.

6. The intelligent collaborative control method for molecular dynamics simulation according to claim 1, characterized in that, Before calling the pre-built knowledge graph, there are also steps involved in constructing the knowledge graph: Establish an ontology data dictionary covering the research object ontology, software ontology, and force field ontology; Entity words and entity association features are extracted from publicly available literature text databases and historical work records using natural language processing algorithms. After disambiguation and merging of the extracted entity words, they are mapped to the ontology data dictionary to generate the preset knowledge graph.

7. The intelligent collaborative control method for molecular dynamics simulation according to claim 1, characterized in that, After the preset simulation task completion conditions are met, the following are also included: Extract the final target software combination and the final parameter combination for completing the iteration corresponding to this execution, add category tags, and store them in the knowledge base; Receive new molecular simulation task data and calculate the feature similarity between the new molecular simulation task data and each historical data in the knowledge base; If the feature similarity is greater than the preset matching threshold, the final target software combination and the final parameter combination corresponding to the historical data with the highest matching degree will be output as the recommended parameter set.

8. An intelligent collaborative control system for molecular dynamics simulation, executing the method according to any one of claims 1-7, characterized in that, include: The task decomposition module is used to receive molecular simulation task data, call a preset knowledge graph to extract the entities and associated features corresponding to the molecular simulation task data, input the extracted features into a deep learning inference model, decompose the molecular simulation task data through the deep learning inference model, and generate an execution logic chain and the corresponding target software combination. The parameter initialization and execution module is used to extract the structural features of the molecular simulation task data, input the structural features into the reinforcement learning model, output the initial parameter combination, and, based on the execution logic chain and the initial parameter combination, call computing resources to collaboratively run the target software combination to generate running status data and trajectory data. A multi-level monitoring module is used to collect basic process indicator data in real time during the operation of the target software combination, and periodically extract dynamic indicator data and result validity indicator data. The collected and extracted indicator data are compared with preset three-level monitoring thresholds, including computing resource threshold, dynamic indicator threshold and result validity threshold. The closed-loop intervention module is used to input the exceeding indicator data as state feedback to the reinforcement learning model when it is determined from the comparison results that any of the indicator data exceeds any of the corresponding three-level monitoring thresholds. The reinforcement learning model outputs parameter fine-tuning actions to update the running parameters of the current process, or generates a computing resource expansion instruction independent of the reinforcement learning model based on the exceeding indicator data. The above data collection, comparison and feedback are executed in a closed-loop control manner until the preset simulation task completion conditions are met.

9. A computer storage medium storing a readable program, characterized in that, When the program is running, it can instruct the computing device to execute an intelligent collaborative control method for molecular dynamics simulation as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform an operation corresponding to the intelligent cooperative control method for molecular dynamics simulation as described in any one of claims 1-7.

Images