A computing parameter dynamic configuration method and system based on multi-agent log collaborative analysis

CN122867877APending Publication Date: 2026-10-02HANGZHOU DIANZI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610915538.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-24
Publication Date
2026-10-02

AI Technical Summary

Technical Problem

然而,在基于ORCA软件对含镝等具有强相对论效应和开壳层电子结构的复杂稀土配合物开展几何优化和电子结构计算时,仍易出现收敛困难、计算效率低下及人工干预成本高等问题:

Benefits of technology

对于稀土配合物,如镧系单分子磁体,在ORCA软件Auto-TRAH/TRAH-SCF阶段特存在 “持续震荡+误差拖尾”的耦合难题,针对上述问题,本发明构建了一套基于多智能体协同的“监控-决策-自愈”闭环系统,特别是引入了动态校准的震荡判定阈值和分级递进的参数重构策略,替代了传统的人工试错。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122867877A_ABST
    Figure CN122867877A_ABST
Patent Text Reader

Abstract

The application discloses a kind of based on multi-agent log collaborative analysis computing parameter dynamic configuration method and system, belong to quantum chemistry calculation convergence control technical field.For rare earth metal complex because of electronic state highly degenerate, coordination environment is complex and in the process of geometry optimization, it is prone to sustained oscillation, convergence stagnation and high cost of artificial intervention, the problem of serious waste of computing resources, constructs a kind of closed-loop task self-healing scheme.The method comprises the following steps: S110, multi-stage oscillation monitoring and abnormality determination;S120, dynamic parameter regulation and adaptive resubmission;S130, convergence determination and data closed-loop acquisition.The application can improve the convergence efficiency and computing power utilization rate of complex molecular system quantum chemistry calculation, and provide support for efficient calculation of complex systems such as lanthanide monomolecular magnet.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of computer-aided materials discovery and quantum chemical computation technology, specifically to a method and system for dynamic configuration of computational parameters based on multi-agent log collaborative analysis, used to realize anomaly identification, parameter reconstruction and self-healing convergence control in the computation process of complex lanthanide molecular systems. Background Technology

[0002] First-principles calculations have been widely applied to the study of material properties, chemical reaction mechanisms, and molecular magnetism. However, when using ORCA software to perform geometric optimization and electronic structure calculations on complex rare-earth complexes containing dysprosium and other compounds with strong relativistic effects and open-shell electronic structures, problems such as convergence difficulties, low computational efficiency, and high costs of manual intervention still arise. 1. For complex rare earth complexes containing dysprosium and other compounds with strong relativistic effects and open-shell electronic structures, problems such as convergence difficulties, low efficiency, and high cost of manual intervention are likely to occur during the Orca geometry optimization and electronic structure calculation process. 2. Such systems are prone to continuous oscillations of the potential energy surface, error tailing, and abnormal electronic self-consistency convergence during the optimization process. The method of relying on manual log viewing and manual parameter adjustment is lagging and easily leads to a waste of computing power. 3. Existing anomaly handling methods lack a unified judgment, reconstruction, and feedback mechanism, making it difficult to form a closed-loop control process from anomaly identification, task truncation, parameter correction to restart and recovery; 4. Existing technologies typically lack continuous verification of changes in total energy and error norm after a task restart, which can easily lead to repeated restarts, inefficient iterations, or recurring abnormal states.

[0003] Therefore, there is an urgent need for a dynamic configuration scheme that can analyze computation logs in real time, identify continuous oscillations and convergence anomalies, and automatically reconstruct input parameters and resubmit tasks based on anomaly types, in order to improve the convergence efficiency and automation level of computations on complex molecular systems. Summary of the Invention

[0004] This application provides a method and system for dynamic configuration of computational parameters based on multi-agent log collaborative analysis. Through multi-agent collaboration, it achieves anomaly identification, parameter reconstruction, and closed-loop control for convergence in quantum chemical calculations of complex molecular systems. Auto-TRAH / TRAH-SCF is a convergence algorithm initiated by ORCA software to prevent computational crashes. However, when dealing with complex rare-earth systems, this stage is prone to "continuous oscillations" and "error tailing," leading to computational stagnation and wasted computational resources. The method provided in this invention includes three stages: anomaly monitoring and judgment, parameter adjustment and resubmission, and convergence judgment and data closed-loop acquisition. By real-time analysis of the Auto-TRAH / TRAH-SCF stage logs, convergence parameters are adjusted, improving the convergence efficiency and automation level of complex system calculations. A method for dynamically configuring computational parameters based on multi-agent log collaborative parsing includes the following steps: S110, Multi-level Oscillation Monitoring and Anomaly Detection Algorithm: The log quality inspection agent is used to parse the output log in real time without blocking. After the system enters the self-consistent iterative convergence (Auto-TRAH / TRAH-SCF) stage of electronic structure, the macro-iteration step count, total energy and error norm data are dynamically extracted to construct the feature sequence. The anomaly detection agent uses the aforementioned feature sequence and dynamic calibration threshold to identify energy oscillations, error tailing, or wavefunction convergence anomalies, and generates monitoring logs. The dynamic calibration threshold is adaptively determined based on the middle value of the calibration window to avoid misjudging normal convergence as oscillation. S120, Dynamic Parameter Adjustment and Adaptive Resubmission: After identifying the anomaly, the parameter optimization agent is used to truncate the current process and back up the current configuration file and wave function file; The parameter optimization agent reconstructs the input file based on the anomaly type, loads the backup wave function and adjusts the convergence control parameters, and the execution agent re-initiates the computation task. The input file reconstruction based on anomaly type includes: firstly introducing static damping control to restart the task; if it is determined to be continuous oscillation again, then further introducing energy level shift control and resubmitting the adjusted input file; S130. Convergence Determination and Data Loop Acquisition: The stability recovery determination agent performs stability recovery verification on the reconstruction task. If the conditions are met, the task is determined to be stable. If the conditions are not met, the data closed-loop agent records the abnormal information and triggers a new round of parameter correction, forming a closed-loop correction.

[0005] Preferably, the operation of optimizing feature streaming extraction in S110 includes: S111. The log quality inspection agent scans the computation program output log (orca.log) file in a non-blocking manner. Based on the following fields, it establishes a temporal mapping relationship between the geometric optimization loop and the macro iteration record. S112. Construct a feature sequence for determining continuous oscillations, including the number of macro-iteration steps ( ), total energy at step k ( ), error norm ( ).

[0006] Preferably, the operation of identifying energy oscillations in S110 includes: S113. In the macro-iteration sequence, continuous energy data of a preset length is extracted starting from the current decision step to form the current analysis window; simultaneously, continuous error norm data corresponding to this analysis window is extracted to form the error analysis window; based on the sliding window, the following are calculated respectively: the average change amplitude of total energy within the window, used to characterize the overall energy drift; the fluctuation range of total energy within the window, used to characterize the oscillation intensity of the energy plateau region; the average level of the error norm, used to characterize the error maintenance state; the logarithmic fitting slope of the error norm with the iteration step, used to characterize the error convergence trend; and the number of times the energy change direction reverses, used to characterize the degree of energy oscillation. The sliding window length is preferably set to 20 macro-iteration steps.

[0007] Specifically, the anomaly detection agent truncates a sequence of length [length missing] from the macro-iterative sequence. The sliding window, the macro-iteration sequence is composed of the Trust Region Enhanced Macro-iteration Step Identifier (TRAH Macro) and the Newton Iteration Macro-iteration Step Identifier (NR Macro), denoted as... ; The corresponding error norm window is denoted as: ; Where L=20 is selected, and the average drift within the window is calculated using the following formula. : ; Calculate the energy range within the window using the following formula. ; = ,in, , These represent the maximum and minimum values ​​in the macro-iteration sequence, respectively; Calculate the window mean error norm using the following formula : ; Window error logarithmic fitting slope: Number of window energy reversals: ,in, This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. S114. To avoid misjudging the normal convergence phase of rapid decrease in total energy in the early stage of geometric optimization as continuous oscillation, a continuous oscillation monitoring initial step is introduced. ); only when the macro iteration step reaches the preset starting step ( After that, continuous oscillation recognition is initiated, and the preset starting step ( Not less than 50; S115. After reaching the starting step of continuous oscillation monitoring, multiple judgment windows are selected consecutively to form a calibration window set, and the continuous oscillation judgment threshold is adaptively determined based on the statistical characteristics in the calibration window set.

[0008] The thresholds include: average energy drift threshold; energy fluctuation range threshold; error maintenance threshold; error decline trend threshold; and energy reversal number threshold.

[0009] Specifically: the average energy drift threshold is determined based on the statistical results of the average energy drift in the calibration window; the energy fluctuation range threshold is determined based on the statistical results of the energy fluctuation range in the calibration window; the error maintenance threshold is determined jointly based on the current task convergence error requirements and the error statistics in the calibration window; the error descent trend threshold is determined based on the error convergence trend statistics in the calibration window; and the energy reversal number threshold is determined based on the proportion of the theoretical oscillation upper limit corresponding to the window length.

[0010] The number of calibration windows should preferably be no less than 10.

[0011] Specifically, in meeting Then, take the continuous The set of calibration windows consists of several decision windows. And based on the calibration window set, the threshold for determining continuous oscillation is frozen, wherein Not less than 10; the thresholds are defined as follows: Energy average drift threshold: ; Platform energy fluctuation lower limit threshold: ; Error tailing maintenance threshold: ; Minimum slope threshold for effective error reduction: ; Reversal number threshold: ; in, The SCF convergence error threshold is automatically parsed from the current task log; This represents the median statistic over the calibration window set; Preferably, after step S115, the following step is further included: S116. Construct an oscillation judgment index based on the average energy drift degree, energy fluctuation range, error maintenance state, error decline trend, and energy reversal number.

[0012] When the following conditions are met: The overall change in total energy within the window is less than the preset energy drift threshold. The total energy fluctuation range within the window is greater than the preset fluctuation threshold; The average error norm within the sliding window remains above a preset error maintenance threshold. The error reduction trend is weaker than the preset convergence trend threshold; The number of times the direction of total energy change reverses reaches a preset threshold. If the current window is in a state of continuous oscillation, then it is determined that the current window is in a state of continuous oscillation; otherwise, it is determined that the current window is not in a state of continuous oscillation.

[0013] Specifically, define the oscillation determination variable. for: Among them, when When this occurs, it indicates that the current window has entered a plateau phase and is oscillating repeatedly, and the error norm has not continued to be effectively reduced; when When this occurs, it indicates that the current window does not constitute a continuous oscillation window; S117. Define a continuous oscillation counter. When the current judgment window is identified as a continuous oscillation state, the counter increments by one; when the current judgment window is not identified as a continuous oscillation state, the counter is reset to zero and the counting starts again.

[0014] Define a continuous oscillation counter for: ; S118 When the continuous oscillation counter reaches the preset trigger threshold, it is determined that the current task has entered a continuous oscillation state, and the current macro iteration step is determined as the truncation trigger step. ).

[0015] At the same time, according to the trigger threshold ( (Backtrack to determine the starting point of the continuous oscillation.)

[0016] The trigger threshold is preferably set to 5 to 10 consecutive decision windows.

[0017] Specifically, when the continuous oscillation counter satisfies: When the current task enters a state of continuous oscillation, the current step is defined. To truncate the trigger step The corresponding definition of the initial step of continuous oscillation is: Among them, the preferred option is to select .

[0018] Preferably, the operation of truncating the current process and backing up the current configuration file and wavefunction file in S120 includes: S121, in the cutoff trigger step Send a process suspension or termination command to the scheduling system to suspend the current process; S122, Lock-off Trigger Step The most recent successfully written valid structure frame As a restart node; if the output of this frame is incomplete, it backtracks to the previous valid structure frame and extracts the corresponding configuration file and wavefunction file.

[0019] Preferably, the operation of loading the backup wavefunction and adjusting the convergence control parameters in S120 includes: S121, in the truncation trigger step ( ), send a process suspension or termination command to the scheduling system to truncate the current process; S122, Lock-off cutoff trigger step ( The most recently successfully written valid structure frame ( The frame is used as a restart node; if the output of the frame is incomplete, it is backtracked to the previous valid structure frame and the corresponding configuration file and wave function file are extracted.

[0020] The loading of the backup wavefunction and adjustment of the convergence control parameters in S120 specifically includes: S123. Load the historical wavefunction inheritance instruction (MORead) and the wavefunction file read path (%mofile) instruction into the reconstructed input file to inherit the backup wavefunction; S124. Prioritize introducing static damping control in the electronic self-consistent control (%scf) module, and set the control parameters as follows: DampFac=0.85, DampErr=0.05; S125. After completing the damping-assisted restart of S124, S110 is executed again. When the log quality inspection intelligence further identifies the electronic self-consistent iteration completion (SCF CONVERGED AFTER ... CYCLES) flag, it is determined that the current electronic self-consistent cycle has successfully converged. S126. When it is determined that damping control alone is insufficient to eliminate continuous oscillations, energy level shift control is further introduced into the electronic self-consistent control (%scf) module. The control parameters are set as follows: LShift=0.15, ShiftErr=0.10. S127. The damping control and the energy level movement control are written into the electronic self-consistent control (%scf) module in the manner of first executing static damping control and then executing energy level movement control, and automatically exiting after eliminating continuous oscillations. S128. The input file after wave function inheritance and parameter correction is resubmitted to the calculation program for execution, so that the task can continue to enter the subsequent macro-iterative convergence process from the restart node.

[0021] Preferably, the stability recovery verification operation in S130 includes: S131. The stability recovery determination agent performs continuity verification on the task restarted after S120 during the subsequent macro-iteration convergence phase. S132. The stability recovery determination agent continuously monitors the stability of the task after restarting. If in subsequent continuous... Within a given decision window, if the following conditions are met simultaneously: no further continuous oscillation is identified; the logarithmic slope of the error is less than a preset decreasing threshold (when the logarithmic slope of the error is less than the preset decreasing threshold, it indicates that the error is in a continuous decreasing state); and the error level is lower than a preset maintenance threshold, then the current task is determined to have escaped the continuous oscillation state and resumed stable convergence.

[0022] Specifically, if in subsequent consecutive Within the decision window corresponding to each macro iteration step, all conditions are met: If so, it is determined that the current task has broken free from continuous oscillation and resumed stable convergence; among which, Preferred selection ; S133. If the stability recovery condition described in S132 is not met, the data closed-loop agent records the abnormal characteristics, parameter correction actions and restart results of the current round, and returns to execute S120 to truncate the current process, and backs up the current configuration file and wave function file. Based on the recorded abnormal characteristics, the parameter optimization agent makes a new round of adjustments to the damping control parameters and / or energy level movement control parameters and resubmits the calculation task.

[0023] Preferably, in S118 In S132 .

[0024] This invention also provides a dynamic configuration system for computational parameters based on multi-agent log collaborative parsing. The system is based on a method for dynamic configuration of computational parameters based on multi-agent log collaborative parsing, comprising: The multi-level oscillation monitoring and anomaly judgment unit is used to use the log quality inspection agent to parse the output log in real time without blocking. After the system enters the Auto-TRAH / TRAH-SCF stage, it dynamically extracts the macro iteration steps, total energy and error norm data to construct feature sequences. The anomaly detection agent uses the aforementioned feature sequence and dynamic calibration threshold to identify energy oscillations, error tailing, or wavefunction convergence anomalies, and generates monitoring logs. The dynamic parameter control and adaptive resubmission unit is used to identify anomalies, use the parameter optimization agent to truncate the current process, and back up the current configuration file and wave function file; The parameter optimization agent reconstructs the input file based on the anomaly type, loads the backup wave function and adjusts the convergence control parameters, and the execution agent re-initiates the computation task. The input file reconstruction based on anomaly type includes: firstly introducing static damping control to restart the task; if it is determined to be continuous oscillation again, then further introducing energy level shift control and resubmitting the adjusted input file; The convergence determination and data closed-loop acquisition unit is used for the stability recovery determination agent to perform stability recovery verification on the reconstruction task. When the conditions are met, the task is determined to be stable. When the conditions are not met, the data closed-loop agent records the abnormal information and triggers a new round of parameter correction to form a closed-loop correction. Each unit interacts with the others via a message bus and shares a distributed historical optimization knowledge base. This invention offers at least the following advantages: For rare earth complexes, such as lanthanide single-molecule magnets, there is a unique coupling problem of "continuous oscillation + error tailing" in the Auto-TRAH / TRAH-SCF stage of ORCA software. To address this problem, this invention constructs a closed-loop system of "monitoring-decision-self-healing" based on multi-agent collaboration. In particular, it introduces a dynamically calibrated oscillation judgment threshold and a graded and progressive parameter reconstruction strategy to replace the traditional manual trial and error.

[0025] Specifically, traditional methods use a fixed threshold to determine convergence, which can easily misjudge the normal convergence phase of rapid decrease in total energy in the early stages of geometric optimization as "continuous oscillation," leading to premature truncation and wasted computational time. This invention introduces a dynamic calibration threshold mechanism, for example, by setting a monitoring starting step k. s Based on the median of statistics from the previous few windows, a dynamic calibration threshold is calculated, enabling adaptive judgment. This distinguishes between "effective decline" and "pathological oscillation," avoiding invalid restarts due to misjudgment and ensuring the continuity of the calculation process.

[0026] Traditional solutions for dealing with oscillations typically employ a "one-size-fits-all" approach of restarting or modifying the system, lacking specificity. In rare earth systems, adjusting a single parameter is often insufficient to eliminate oscillations, leading to repeated failures. This application provides a tiered, progressive self-healing strategy, achieving precise intervention. It avoids the waste of computational resources caused by over-adjustment while ensuring that persistent oscillations are effectively eliminated, significantly improving the convergence success rate of rare earth complexes.

[0027] After an abnormal truncation, traditional solutions typically restart from the initial structure, losing completed iteration information; or the inherited files may be mismatched, leading to restart failure. This invention, however, can backtrack to the most recent valid structural frame before the truncation, enabling breakpoint continuation of computation, preserving electronic structure information, avoiding wasted computational power starting from scratch, and significantly shortening the convergence path after restart.

[0028] This application constructs a multi-agent collaborative architecture to achieve anomaly identification, parameter reconstruction, and closed-loop control in the quantum chemical calculation process of complex molecular systems. Compared with the traditional method that relies on manual log viewing and parameter tuning, this invention can perform real-time analysis of the output logs of the Auto-TRAH / TRAH-SCF stage, extract feature data such as total energy and error norm, and promptly identify anomalies such as continuous oscillations, error tailing, and electronic self-consistent convergence. After identifying anomalies, the system can truncate the current task, back up the valid configuration file and wavefunction file, reconstruct the input file based on the anomaly type, and resubmit the calculation task. After restarting, the system further continuously verifies the stability of the task, and continues to trigger closed-loop correction if stability is not restored. Through the above technical solution, this invention improves the convergence efficiency, automation level, and computing power utilization of complex molecular system calculations, and has good application value. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the specific embodiments of the present invention or the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Some specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings in an exemplary and non-limiting manner. The same reference numerals in the drawings indicate the same or similar parts or components. These drawings are not necessarily drawn to scale. In the drawings: Figure 1 A flowchart illustrating a method for dynamically configuring computational parameters based on collaborative parsing of multi-agent logs, provided in an embodiment of this application; Figure 2 A schematic diagram of the framework of a multi-agent computing parameter dynamic configuration system provided in this application embodiment; Figure 3 This is a diagram illustrating the energy oscillation pattern during the calculation process; Figure 4 A magnified view of the energy oscillations and error oscillations from steps 20 to 73; Figure 5 To optimize the changes in molecular structure after success. Detailed Implementation

[0030] Please see Figure 1 This invention provides a method for dynamically configuring computational parameters based on multi-agent log collaborative parsing, used to achieve closed-loop control in quantum chemical calculations of complex molecular systems. This invention also provides a system for dynamically configuring computational parameters based on multi-agent log collaborative parsing, the system framework of which is as follows: Figure 2 As shown, the various units work together to achieve dynamic intervention and stability restoration of the computing task. The following section, in conjunction with a specific calculation task, uses the bis(pentaisopropylcyclopentadienyl)dysprosium(III) cation [Dy Ⅲ (Cp iPr5 ) 2 ] + Taking an example, this paper illustrates the complete application process of the dynamic configuration method for computational parameters based on multi-agent log collaborative parsing in practical materials simulation calculations. Here, Cp represents cyclopentadienyl; iPr 5 Represents pentaisopropyl substitution; Dy Ⅲ This represents dysprosium, a trivalent lanthanide (rare earth) metal. Dysprosium (Dy) is a heavy rare earth element with unfilled 4f electrons and extremely strong spin-orbit coupling. In ORCA calculations, the SCF convergence of this type of system is difficult, and it is prone to "persistent oscillations" and "error tailing" problems. The pentaisopropyl group is a very large group, resulting in a highly flexible molecular structure and a complex potential energy surface, further increasing the difficulty of geometry optimization.

[0031] S100 User Input Phase The user uploaded [Dy] through the interface. Ⅲ (Cp iPr5 ) 2 ] + The system uses .log files and orca's .inp format input files, records relevant information, and prepares to enter the intelligent analysis process.

[0032] S110, Multi-level Oscillation Monitoring and Anomaly Detection Algorithm The log quality inspection agent scans orca.log in a non-blocking manner to locate the fields related to GEOMETRY OPTIMIZATIONCYCLE and Auto-TRAH / TRAH-SCF. Constructing a feature sequence for determining continuous oscillations: ; The anomaly detection agent truncates a length of [length missing] from the TRAH iteration sequence. Sliding window: ; And the corresponding error norm window: ; Select L=20 and calculate the average drift within the window: ; Calculate the energy range within the calculation window: = ; Calculate the window mean error norm: ; Window error logarithmic fitting slope: ; Number of window energy reversals: ; To avoid misjudging the normal convergence phase, where the total energy decreases rapidly in the early stages of geometry optimization, as continuous oscillation, a continuous oscillation monitoring initial step is introduced. Only if the decision step satisfies At that time, continuous oscillation identification is initiated, among which... The preferred value is 50; In satisfying Then, take the continuous =10 decision windows constitute the calibration window set The continuous oscillation determination threshold is frozen based on the calibration window set; the thresholds are defined as follows: Energy average drift threshold: ; Platform energy fluctuation lower limit threshold: ; Error tailing maintenance threshold: ; Minimum slope threshold for effective error reduction: ; Reversal number threshold: ; Define oscillation determination variables for: Substituting the data from this embodiment window by window, we find that all seven consecutive windows from k = 67 to k = 73 satisfy the above criteria. This invention uses... For 50 and consecutive =10 decision windows constitute a calibration window set, which is used to calculate the specific threshold, so as to determine the truncation at step 73.

[0033] Taking a trigger window of k = 73 as an example, the statistics are as follows: Clearly, all five conditions above are met simultaneously: ; Therefore, ξ(73) = 1 Define a continuous oscillation counter: ; In this embodiment, the consecutive windows k = 67, 68, 69, 70, 71, 72, 73 all satisfy ξ(k) = 1, therefore: This embodiment adopts the preferred value. = 7. When the continuous oscillation counter satisfies ≥ When the current task enters a state of continuous oscillation, the current step is defined as the truncation trigger step. Since this embodiment satisfies the condition for the first time... The window with k = 7 is k = 73, therefore: = 73 The corresponding starting step for continuous oscillation is: = 67; the truncation trigger step is step 73. S120, Dynamic Parameter Adjustment and Adaptive Resubmission After the system identifies the anomaly described in S118, it sends a process suspension or termination instruction to the scheduling system to truncate the current inefficient process. Lock cutoff trigger step The most recent successfully written valid structure frame As a restart node; if the output of this frame is incomplete, backtrack to the previous valid structure frame and extract the corresponding configuration file and wavefunction file; Load MORead with the %mofile dy.gbw command in the reconstructed input file to inherit the backup wavefunction; First, introduce static damping control in the %scf module, setting: DampFac=0.85, DampErr=0.05; After completing a damped-assisted restart, S111 to S118 are executed again, and the SCFCONVERGED AFTER 28 CYCLES flag is further identified in the log, indicating that the current electronic self-consistent cyclic convergence is successful, the current task has broken away from continuous oscillation and resumed stable convergence, and a stable molecular structure is obtained. To further illustrate the practical application effects of the present invention, the following description is provided in conjunction with the appendix. Figure 3-5 Each figure will be explained in detail.

[0034] Figure 3 This diagram illustrates the energy and error oscillations from steps 1 to 73. It shows the overall trend of the total energy and error norm of the optimization task as a function of the number of iterations from step 1 to step 73. The left diagram reflects a rapid decrease in total energy at the beginning of the computation, followed by a gradual entry into a low-amplitude fluctuation range; the right diagram reflects a significant decrease in the error norm initially, followed by some oscillating fluctuations. This diagram is used to illustrate that the present invention can identify the transition from the rapid convergence phase to the plateau oscillation phase from the entire log, providing basic data support for subsequent continuous oscillation determination. Figure 4 This is a magnified view of the energy oscillations and error oscillations from steps 20 to 73. Compared to... Figure 3 The results of the entire process of change, Figure 4 Further focusing on the iterative behavior within the later platform range, it more clearly demonstrates the repeated fluctuations of total energy within a narrow range and the change characteristics of the error norm not being effectively reduced. Through this magnified view, it can be intuitively seen that the system no longer exhibits normal monotonic convergence at this stage, but rather a state of continuous oscillation and error tailing, thus serving as an important basis for triggering anomaly identification, task truncation, and parameter reconstruction; Figure 5 This figure shows a comparison of the molecular structure before and after optimization. It illustrates the structural changes in the system configuration after adaptive parameter adjustment and resubmission. The figure shows that the optimized molecular structure remains stable in its overall framework, while local atomic coordinates and configurational details are further adjusted, indicating that the computational task can obtain reasonable optimization results after achieving stable convergence. This figure structurally verifies the effectiveness and feasibility of the proposed method in the geometric optimization of complex molecular systems.

Claims

1. A method for dynamically configuring computational parameters based on multi-agent log collaborative parsing, characterized in that, Includes the following steps: S110, Multi-level Oscillation Monitoring and Anomaly Detection Algorithm: The log quality inspection agent is used to parse the output logs in real time without blocking. After the system enters the self-consistent iterative convergence stage of electronic structure, the macro-iteration step count, total energy and error norm data are dynamically extracted to construct the feature sequence. The anomaly detection agent uses the aforementioned feature sequence and dynamic calibration threshold to identify energy oscillations, error tailing, or wavefunction convergence anomalies, and generates monitoring logs. The dynamic calibration threshold is adaptively determined based on the middle value of the calibration window to avoid misjudging normal convergence as oscillation. S120, Dynamic Parameter Adjustment and Adaptive Resubmission: After identifying the anomaly, the parameter optimization agent is used to truncate the current process and back up the current configuration file and wave function file; The parameter optimization agent reconstructs the input file based on the anomaly type, loads the backup wave function and adjusts the convergence control parameters, and the execution agent re-initiates the computation task. The input file for reconstruction based on anomaly type includes: prioritizing the introduction of static damping control restart tasks; If the oscillation is determined to be continuous again, energy level shift control will be introduced, and the adjusted input file will be resubmitted. S130, Convergence Determination and Data Loop Acquisition: The stability recovery determination agent performs stability recovery verification on the reconstruction task. If the conditions are met, the task is determined to be stable. If the conditions are not met, the data closed-loop agent records the abnormal information and triggers a new round of parameter correction, forming a closed-loop correction.

2. The method for dynamically configuring computational parameters based on multi-agent log collaborative parsing according to claim 1, characterized in that, The operation of constructing the feature sequence in S110 includes: S111. The log quality inspection agent scans the log files output by the computation program in a non-blocking manner. Based on the following geometric structure optimization loop identifier, Newton iteration macrostep identifier, and trust region enhancement macro iteration step identifier, a temporal mapping relationship is established between the geometric optimization loop and the macro iteration record. S112. Construct a feature sequence for determining continuous oscillations, including the number of macro-iteration steps. Total energy at step k Error norm .

3. The method for dynamically configuring computational parameters based on multi-agent log collaborative parsing according to claim 2, characterized in that, In S110, the operation of identifying energy oscillations includes: S113. In the macro-iteration sequence, continuous energy data of a preset length is extracted starting from the current decision step to form the current analysis window; at the same time, continuous error norm data corresponding to the analysis window is extracted to form the error analysis window. The analysis window is a sliding window. Based on the sliding window, the following calculations are performed: The average change in total energy within the window is used to characterize the overall degree of energy drift. The range of total energy fluctuation within the window is used to characterize the oscillation intensity of the energy plateau region; The average level of the error norm is used to characterize the error maintenance state; The slope of the logarithmic fit of the error norm as a function of the iteration step is used to characterize the error convergence trend. The number of times the direction of energy change reverses is used to characterize the degree of energy oscillation. The preferred length of the sliding window is 20 macro iteration steps. S114. To avoid misjudging the normal convergence phase of rapid decrease in total energy in the early stage of geometric optimization as continuous oscillation, a continuous oscillation monitoring start step is introduced; the continuous oscillation identification process is started only when the macro iteration step reaches the preset start step, and the preset start step is not less than 50. S115. After reaching the starting step of continuous oscillation monitoring, multiple judgment windows are selected consecutively to form a calibration window set, and the continuous oscillation judgment threshold is adaptively determined based on the statistical characteristics in the calibration window set. The thresholds include: average energy drift threshold, energy fluctuation range threshold, error maintenance threshold, error decline trend threshold, and energy reversal number threshold; in: The energy average drift threshold is determined based on the statistical results of the energy average drift degree within the calibration window; The energy fluctuation range threshold is determined based on the statistical results of the energy fluctuation range in the calibration window; The error maintenance threshold is determined jointly based on the current task convergence error requirements and the error statistics results in the calibration window; The error descent trend threshold is determined based on the statistical results of the error convergence trend in the calibration window; The threshold for the number of energy reversals is determined based on the proportion of the theoretical upper limit of oscillation corresponding to the window length; The number of calibration windows shall not be less than 10.

4. The method for dynamically configuring computational parameters based on multi-agent log collaborative parsing according to claim 3, characterized in that, Following S115, the following steps are also included: S116. Construct an oscillation determination index based on the average energy drift degree, energy fluctuation range, error maintenance state, error decrease trend, and energy reversal number; When the following conditions are met: If the overall change in total energy within the window is less than a preset energy drift threshold, the fluctuation range of total energy within the window is greater than a preset fluctuation threshold, the average value of the error norm within the sliding window remains above a preset error maintenance threshold, the error decreasing trend is weaker than a preset convergence trend threshold, and the number of times the direction of total energy change reverses reaches a preset number threshold, then the current window is determined to be in a continuous oscillation state; otherwise, the current window is determined not to be in a continuous oscillation state. S117. Define a continuous oscillation counter. When the current judgment window is identified as a continuous oscillation state, the counter is incremented by one. When the current judgment window is not identified as a continuous oscillation state, the counter is reset to zero and the counting starts again; S118. When the continuous oscillation counter reaches the preset trigger threshold, it is determined that the current task has entered the continuous oscillation state, and the current macro iteration step is determined as the truncation trigger step. The starting position of the continuous oscillation is determined by backtracking based on the trigger threshold; the trigger threshold is set to 5 to 10 consecutive judgment windows.

5. The method for dynamically configuring computational parameters based on multi-agent log collaborative parsing according to claim 4, characterized in that, S120 involves truncating the current process and backing up the current configuration file and wavefunction file, specifically including: S121. In the truncation trigger step, send a process suspension or termination instruction to the scheduling system to truncate the current process; S122. Lock the most recently successfully written valid structure frame before the truncation trigger step as the restart node; if the output of this frame is incomplete, backtrack to the previous valid structure frame and extract the corresponding configuration file and wavefunction file.

6. The method for dynamically configuring computational parameters based on multi-agent log collaborative parsing according to claim 5, characterized in that, The loading of the backup wavefunction and adjustment of the convergence control parameters in S120 specifically includes: S123. Load the historical wavefunction inheritance instruction and wavefunction file read path instruction into the reconstruction input file to inherit the backup wavefunction; S124. Prioritize the introduction of static damping control in the electronic self-consistent control module, and set the control parameters as follows: DampFac=0.85, DampErr=0.05; S125. After completing the damping-assisted restart of S124, S110 is executed again. When the log quality inspection intelligence further identifies the electronic self-consistent iteration completion mark, it is determined that the current electronic self-consistent loop has successfully converged. S126. When it is determined that damping control alone is insufficient to eliminate continuous oscillations, energy level shift control is further introduced into the electronic self-consistent control module. The control parameters are set as follows: LShift=0.15, ShiftErr=0.

10. S127. The damping control and the energy level movement control are written into the electronic self-consistent control module in the manner of first executing static damping control and then executing energy level movement control, and automatically exiting after eliminating continuous oscillations. S128. The input file after wave function inheritance and parameter correction is resubmitted to the calculation program for execution, so that the task can continue to enter the subsequent macro-iterative convergence process from the restart node.

7. The method for dynamically configuring computational parameters based on multi-agent log collaborative parsing according to claim 1, characterized in that, The stability recovery verification in S130 specifically includes: S131. The stability recovery determination agent performs continuity verification on the task restarted after S120 during the subsequent macro-iteration convergence phase. S132. The stability recovery determination agent continuously monitors the stability of the task after restarting; if the stability is restored in subsequent continuous... Within each decision window, the following conditions must be met simultaneously: If a persistent oscillation is not identified again, the logarithmic slope of the error fit is less than a preset descent threshold, and the error level is lower than a preset maintenance threshold, then the current task is determined to have escaped the persistent oscillation and resumed stable convergence; among which, Set to 5-20; S133. If the stability recovery condition described in S132 is not met, the data closed-loop agent records the abnormal characteristics, parameter correction actions and restart results of the current round, and returns to execute S120 to truncate the current process, and backs up the current configuration file and wave function file. Based on the recorded abnormal characteristics, the parameter optimization agent makes a new round of adjustments to the damping control parameters and / or energy level movement control parameters and resubmits the calculation task.

8. The method for dynamically configuring computational parameters based on multi-agent log collaborative parsing according to claim 4, characterized in that, The trigger threshold in S118 is 7.

9. The method for dynamically configuring computational parameters based on multi-agent log collaborative parsing according to claim 7, characterized in that, In S132 .

10. A system for dynamically configuring computational parameters based on multi-agent log collaborative parsing, the system being used to execute the method for dynamically configuring computational parameters as described in any one of claims 1 to 6, characterized in that, include: The multi-level oscillation monitoring and anomaly judgment unit is used to use the log quality inspection agent to parse the output log in real time without blocking. After the system enters the self-consistent iterative convergence stage of electronic structure, it dynamically extracts the macro-iteration step number, total energy and error norm data to construct a feature sequence. The anomaly detection agent uses the aforementioned feature sequence and dynamic calibration threshold to identify energy oscillations, error tailing, or wavefunction convergence anomalies, and generates monitoring logs. The dynamic parameter control and adaptive resubmission unit is used to identify anomalies, use the parameter optimization agent to truncate the current process, and back up the current configuration file and wave function file; The parameter optimization agent reconstructs the input file based on the anomaly type, loads the backup wave function and adjusts the convergence control parameters, and the execution agent re-initiates the computation task. The input file reconstruction based on anomaly type includes: firstly introducing static damping control to restart the task; if it is determined to be continuous oscillation again, then further introducing energy level shift control and resubmitting the adjusted input file; The convergence determination and data closed-loop acquisition unit is used for the stability recovery determination agent to perform stability recovery verification on the reconstruction task. When the conditions are met, the task is determined to be stable. When the conditions are not met, the data closed-loop agent records the abnormal information and triggers a new round of parameter correction to form a closed-loop correction. Each unit interacts with other units via a message bus and shares a distributed historical optimization knowledge base.