An AI self-adaptive adjustment-based STE-KEP-FR flux reconstruction coefficient calculation method

Abstract

The present application relates to the technical field of computational fluid dynamics optimization, and particularly relates to a STE-KEP-FR flux reconstruction coefficient calculation method based on AI adaptive adjustment, comprising the following steps: obtaining a flow field feature, outputting a label and a confidence degree through scene recognition, determining an expert model based on the label, the confidence degree and a trust score, evaluating adaptability, generating a flux reconstruction coefficient if suitable, otherwise entrusting other models to generate by an execution model, loading a solver to solve, forming an experience record to verify, and storing into an experience database after verification. The present application selects a target model by combining flow field scene recognition with a confidence degree and an expert model trust score, introduces adaptability evaluation and a task entrusting mechanism, adaptively generates a flux reconstruction coefficient and loads it into a STE-KEP-FR solver, realizes dynamic perception and scene adaptation for a complex flow field, avoids an unreliable result output by a model beyond the capability boundary, improves numerical precision, convergence and stability, and improves the problem that precision, dissipation and anti-oscillation are difficult to balance.

Description

Technical Field

[0001] This invention relates to the field of computational fluid dynamics optimization technology, specifically to a method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment. Background Technology

[0002] In the field of computational fluid dynamics (CFD) numerical simulation, high-precision numerical methods are crucial for improving the solution accuracy and convergence performance of complex flow problems. Among them, the flux reconstruction (FR) method, as an efficient high-precision numerical scheme, focuses on balancing numerical dissipation, anti-oscillation capability, and computational accuracy by adjusting the flux reconstruction coefficients. However, existing techniques still have significant shortcomings in the adaptive optimization of flux reconstruction coefficients, making it difficult to meet the engineering application requirements of complex flow fields with multi-scale and multi-physics coupling.

[0003] CN115186604A discloses a parallel optimization method for a smoothed solver in computational fluid dynamics based on the Shenwei architecture. This method improves numerical simulation efficiency by converting discontinuous vectors into continuous vector data and distributing them to slave kernels for computation. However, this technical solution primarily focuses on hardware-level parallel computation optimization and does not address adaptive adjustment strategies for flux reconstruction coefficients. Its shortcomings are: first, it lacks dynamic perception of complex flow field characteristics and cannot automatically adjust reconstruction coefficients based on local flow properties; second, it does not consider the differentiated requirements for numerical dissipation and accuracy at different solution stages; and third, it lacks a coefficient optimization mechanism based on flow field characteristics, making it difficult to guarantee a balance between computational accuracy and stability.

[0004] CN120470978A discloses a full-time-domain fluid dynamics simulation optimization method and system, which optimizes the CFD solution process by generating complex frequency domain phase response functions and virtual damping control commands. The shortcomings of this technical solution are: first, although machine learning methods are introduced, no specific design is made for the adaptive adjustment of flux reconstruction coefficients; second, its optimization objective is mainly focused on computational efficiency, without fully considering the impact of flux reconstruction coefficients on numerical accuracy and stability; third, there is a lack of research on coefficient optimization strategies under multi-physics coupling scenarios, making it difficult to adapt to the needs of complex engineering applications.

[0005] The aforementioned existing technologies indicate that the current optimization of flux reconstruction coefficients in computational fluid dynamics still faces the following technical challenges: a lack of dynamic perception and adaptive adjustment capabilities for complex flow field characteristics; existing optimization methods struggle to balance numerical accuracy, convergence, and anti-oscillation capabilities; and a lack of systematic error control and stability assurance mechanisms.

[0006] Therefore, there is an urgent need to provide a method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment to solve the above problems. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment.

[0008] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment, comprising the following steps: S1: Obtain the flow field characteristics of the target flow field; S2: Based on the flow field characteristics, perform scene recognition and output at least one scene label and the confidence level corresponding to each scene label; S3: Based on the scene label, confidence level, and trust scores of multiple expert AI models in the corresponding scene, determine the target expert model; the target expert model performs an adaptability assessment on the current task, and determines whether it is suitable for handling the current task based on the adaptability assessment result; if it is determined to be suitable for handling the current task, the target expert model generates throughput reconstruction coefficients as the execution model; if it is determined to be unsuitable for handling the current task, initiate the task delegation process, with the target expert model acting as a proxy model to delegate the current task to another expert AI model, and the execution model generates throughput reconstruction coefficients. S4: Load the generated flux reconstruction coefficients into the STE-KEP-FR solver to control the flux reconstruction process and perform numerical solutions, and form an experience record for this task based on the solution results; S5: Perform quality verification on the experience records and assign the verified experience records to the experience library dedicated to the execution model.

[0009] Furthermore, the scene recognition adopts a hierarchical labeling system, which includes at least a first-level label and a second-level label; The primary labels include at least one of laminar flow scenario, transitional flow scenario, turbulent flow scenario, shock wave scenario, mixed flow scenario, and unknown scenario; The secondary labels include at least one of the following: stable laminar flow, separated laminar flow, stable turbulent flow, pulsating turbulent flow, near-wall turbulent flow, strong shock wave, weak shock wave, shock wave boundary layer disturbance, laminar-turbulent transition, and turbulent-shock wave coupling; When multiple scene labels are identified, the scene label corresponding to the highest confidence level is used as the primary scene label, and the remaining scene labels are used as auxiliary scene labels to participate in the determination of the target expert model.

[0010] Furthermore, the target expert model is determined based on the comprehensive scoring results; The comprehensive scoring result is based on at least the following two items: main scene label matching, auxiliary scene label matching, scene recognition confidence, trust score of each expert AI model under the corresponding label, historical success rate of each expert AI model, real-time load status of each expert AI model, and response latency of each expert AI model. The expert AI model with the highest overall score was selected as the target expert model.

[0011] Furthermore, the adaptability assessment includes at least one of the following: calculating the similarity between the current task features and historical samples in the expert AI model's experience base, obtaining the expert AI model's trust score for the current scene label, assessing the uncertainty of the current output prediction result, and obtaining the success rate of the expert AI model in recent similar tasks. When the similarity is lower than the first threshold, the scene trust score is lower than the second threshold, the uncertainty of the output prediction result is higher than the third threshold, or the historical success rate is lower than the fourth threshold, the target expert model is determined to be unsuitable for handling the current task.

[0012] Furthermore, the task delegation process includes: the proxy model sending a delegation request to the execution model, wherein the delegation request includes at least one of the following information: task identifier, current flow field feature summary, scene label, scene confidence, proxy model identifier, and current delegation chain record; After receiving the delegation request, the execution model determines whether it is suitable to handle the current task. If it is suitable to handle the current task, it performs the flux reconstruction coefficient calculation. If it is not suitable to handle the current task, it continues to delegate the current task to other expert AI models or switches to a general backup model.

[0013] Furthermore, it also includes a mechanism to prevent circular delegation: Maintain a delegation chain record for each task, and check whether the target model to be delegated already exists in the current delegation chain record before each delegation; If the detection result indicates that the task exists, it is determined to be a circular delegation, and the current delegation process is interrupted, switching the task to the general backup model.

[0014] Furthermore, the quality verification includes: scoring the experience records for quality, determining duplicates, and determining validity. Determine whether experience records are duplicated by comparing hash values ​​and vector similarity. When the quality score is higher than the first quality threshold, it is judged as a high-quality experience and can be written into the experience library dedicated to the execution model. When the quality score is between the first quality threshold and the second quality threshold, it is determined to be an experience to be observed and written into the observation area; When the quality score is lower than the second quality threshold, it is judged as low-quality experience and is refused to be written into the experience library dedicated to the execution model.

[0015] Furthermore, the attribution rules for the experience records are as follows: If the target expert model generates the flux reconstruction coefficients on its own, the experience records will be written into the target expert model's dedicated experience library. If the proxy model delegates the generation of flux reconstruction coefficients to the execution model, then the experience records will be written into the execution model's dedicated experience library. If there are multiple levels of delegation, the experience records are written into the final execution model experience library of the flux reconstruction coefficients used by the STE-KEP-FR solver in the final output. If the flux reconstruction coefficients are generated by the general backup model, the experience records are written into the general backup model experience library or the public safety net experience library.

[0016] Furthermore, it also includes a trust score update mechanism: Maintain a trust score for each expert AI model under different scenario labels; The trust score is dynamically updated based on the solution effect and empirical quality verification results corresponding to the solution results. When the execution result is better than the preset reference result or reaches the high quality standard, the trust score in the corresponding scenario is increased; When the execution result is worse than the preset reference result, triggers divergence, or produces abnormal oscillations, the trust score in the corresponding scenario is reduced. The updated trust score is used to determine the target expert model for subsequent tasks.

[0017] Furthermore, it also includes at least one of the following fault-tolerance mechanisms: Low confidence backup mechanism: When the maximum confidence of the scene recognition output is lower than the preset threshold, stop selecting the expert AI model and directly call the general backup model to generate the throughput reconstruction coefficients; Timeout retry mechanism: When the target expert model or the execution model fails to return a result within a set time, it will automatically switch to the backup expert AI model with the second highest priority. Abnormal result rollback mechanism: When the flux reconstruction coefficients returned by the execution model trigger abnormal oscillations, non-physical solutions, or divergence warnings during the solution process, the rollback will revert to the previous round of stable flux reconstruction coefficients or switch to conservative backup coefficients.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention identifies the flow field characteristics of the target flow field and determines the target expert model based on scene labels, confidence levels, and trust scores of multiple expert AI models in the corresponding scene. The target expert model performs an adaptability assessment. When it is suitable for handling the current task, it generates flux reconstruction coefficients. When it is not suitable for handling the current task, the execution model completes the generation of flux reconstruction coefficients through a task delegation mechanism. The generated flux reconstruction coefficients are then loaded into the STE-KEP-FR solver for numerical solution. The flux reconstruction coefficients are adaptively calculated according to the specific scene of the complex flow field, thereby improving the dynamic perception and scene adaptability of complex flow field characteristics. At the same time, by introducing an adaptability assessment and task delegation mechanism before coefficient generation, the situation where the model directly outputs unreliable results when it exceeds its capability boundary can be reduced, thus better balancing numerical accuracy, convergence, and stability, and improving the problem of balancing computational accuracy, numerical dissipation, and anti-oscillation capability in the prior art.

[0019] This invention performs quality verification on the experience records generated by the solution, and assigns the verified experience records to the exclusive experience library of the execution model. At the same time, it updates the trust score of each expert AI model in the corresponding scenario based on the solution effect and the experience quality verification results. It also forms a closed-loop control by combining low confidence backup, timeout retry, anti-circular delegation and abnormal result rollback mechanism. Thus, this invention can realize continuous optimization of the flux reconstruction coefficient calculation process, improve the problem of lack of systematic error control and stability guarantee mechanism in the prior art, and improve the continuous operation capability and engineering availability in complex engineering application scenarios. Attached Figure Description

[0020] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a comprehensive flowchart of the task delegation, fault tolerance processing, and experience verification attribution in this invention. Detailed Implementation

[0021] 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.

[0022] Example 1 Please see Figure 1-2 This invention provides a method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment.

[0023] The AI-adaptive adjustment method for calculating STE-KEP-FR flux reconstruction coefficients specifically includes the following steps: S1: Obtain the flow field characteristics of the target flow field; In this embodiment, the flow field characteristics refer to the set of information that characterizes the global state, local state, and current numerical state of the solver of the target flow field.

[0024] In this embodiment, the flow field characteristics include at least three categories: the first category is global physical characteristics, including Reynolds number, Mach number, Prandtl number, inlet velocity, inlet pressure, inlet temperature, boundary condition type, time step, and grid scale level; the second category is local flow characteristics, including velocity gradient, pressure gradient, density gradient, temperature gradient, vorticity, divergence, turbulent kinetic energy, dissipation rate, shock sensor output value, local smoothness index, and element interface difference characteristics; the third category is numerical solution state characteristics, including the current iteration step, current residual level, residual decay rate, local oscillation index, divergence warning index, and the flux reconstruction coefficient of the previous round and its effect feedback.

[0025] In terms of implementation, flow field features can be organized in the form of global feature vectors, local unit feature sets, unit surface feature sets, graph structure features, time series features, or multi-channel feature maps. For structured meshes, local neighborhoods can be encoded as fixed-length tensors; for unstructured meshes, units and adjacency relationships can be encoded as graph structures; for unsteady problems, key features from multiple time steps or multiple iteration steps can be concatenated into a time series. During preprocessing, methods such as minimum-maximum value normalization, mean-variance standardization, outlier truncation, moving average denoising, missing feature imputation, and categorical boundary condition encoding can be used.

[0026] S2: Based on flow field features, perform scene recognition and output at least one scene label and the confidence level corresponding to each scene label; In this embodiment, the scene label refers to the label information obtained after classifying the flow state of the target flow location.

[0027] In this embodiment, scene recognition adopts a hierarchical labeling system, including at least a first-level label and a second-level label. The first-level labels include laminar flow scene, transitional flow scene, turbulent flow scene, shock wave scene, mixed flow scene, and unknown scene. The second-level labels include stable laminar flow, separated laminar flow, stable turbulent flow, pulsating turbulent flow, near-wall turbulent flow, strong shock wave, weak shock wave, shock wave boundary layer interference, laminar-turbulent transition, and turbulent-shock wave coupling.

[0028] In terms of implementation, the scene recognition model can employ one or more of the following: multilayer perceptron, convolutional neural network, graph neural network, Transformer, or temporal modeling network. For grid-type CFD data, a graph neural network or graph convolutional structure is used; for unsteady flow, a temporal modeling network is introduced. The output employs a multi-label output method, enabling the same region to correspond to multiple scene labels simultaneously, each label having a corresponding confidence level. The system uses the scene label corresponding to the highest confidence level as the primary scene label, and the remaining scene labels as auxiliary scene labels to participate in subsequent target expert model determination.

[0029] In this embodiment, the confidence level is obtained by outputting the probability value of the classification head and calibrating it by temperature scaling. The low confidence threshold C1 is determined by the false routing rate corresponding to different confidence intervals in the offline validation dataset. The confidence level at the inflection point where the false routing rate begins to rise significantly is selected as the threshold C1. In a specific embodiment, C1 is 0.35.

[0030] S3: Based on scene labels, confidence levels, and trust scores of multiple expert AI models in the corresponding scene, determine the target expert model; the target expert model performs an adaptability assessment on the current task and determines whether it is suitable for handling the current task based on the adaptability assessment results; if it is determined to be suitable for handling the current task, the target expert model acts as the execution model to generate throughput reconstruction coefficients; if it is determined to be unsuitable for handling the current task, the task delegation process is initiated, and the target expert model acts as the proxy model to delegate the current task to another expert AI model, and the execution model generates throughput reconstruction coefficients. In this embodiment, the trust score refers to the credibility evaluation value maintained for each expert AI model under different scenario labels. The larger the value, the more stable and reliable the historical performance of the expert AI model in the corresponding scenario. The adaptability assessment refers to the technical judgment made on whether an expert AI model is suitable for handling the current task. The proxy model refers to the expert AI model that initiates task delegation but does not directly complete the current task calculation. The execution model refers to the model that finally outputs the flux reconstruction coefficients for use by the STE-KEP-FR solver.

[0031] In this embodiment, the comprehensive score can be determined based on any two, three, or more of the above factors. The comprehensive score is implemented as follows: assuming the comprehensive score of a candidate expert AI model is Score, then: , in, The score is based on the matching of main scene tags. To assist in scene label matching scoring, C represents the scene recognition confidence score, R represents the trust score under the corresponding label, and H represents the historical success rate. This is the normalized real-time load value. This is the normalized response delay value. This is a stability correction term related to the stability of the STE-KEP-FR solver. to As weighting coefficients, in this embodiment, each weighting coefficient satisfies... The weights, normalization coefficients, and penalty coefficients mentioned above are determined through offline validation datasets, feature importance analysis, cross-validation, minimizing the false positive rate, minimizing the abnormality rate, or maximizing the solution success rate. All of the above weights are non-negative and satisfy the normalization constraint in the same formula.

[0032] In this embodiment, the main scene label matching score Assisted scene label matching score and stability correction terms Determine as follows: When the main scene label of a candidate expert AI model is exactly the same as the main scene label of the current task, Set to 1; when they belong to the same parent-level tag but the second-level tags are different. Take 0.6; when there are only successful migration cases of adjacent scenes in historical experience, Take 0.3; when there is no obvious correspondence, Take 0.

[0033] Assisted scene label matching score Determined based on the overlap ratio of auxiliary labels, assuming the current task's auxiliary scene label set is... The set of auxiliary labels covered by a candidate expert AI model's historical high-quality experience is as follows: ,but: , When there are no auxiliary labels Take 0.

[0034] Historical success rate H is statistically analyzed and normalized by time window or task window. In this embodiment, it is assumed that the most recent success rate H is... The number of successful tasks in the same or similar scenarios is ,but , Take 20 to 100.

[0035] Real-time load normalization value The normalized response latency value is obtained by normalizing the current number of waiting tasks, the current computing resource utilization, or a combination of both relative to the system design upper limit. It is obtained by normalizing the current average response delay relative to the preset maximum acceptable delay.

[0036] Stability correction term This is used to reflect the contribution of candidate expert AI models to the stability of the STE-KEP-FR solver in similar scenarios. In this embodiment, it is assumed that the model has made significant contributions in the most recent scenarios. The normalized average residual growth rate value in similar scenario tasks is The normalized value of the average abnormal oscillation index is The normalized result of the average divergence penalty value is ,but: , in, Let be the weight coefficient, and satisfy... .

[0037] Once the target expert model is determined, an adaptation assessment is performed. The adaptation assessment includes at least one of the following: calculating the similarity between the current task features and historical samples in the expert AI model's experience base, obtaining the expert AI model's trust score for the current scene label, assessing the uncertainty of the current output prediction result, and obtaining the success rate of the expert AI model in recent similar tasks.

[0038] In this embodiment, similarity is represented by cosine similarity with feature weights. It is assumed that the current task feature encoding vector is... The feature encoding vector of a certain historical sample is The corresponding feature weights are If k takes values ​​from 1 to m, then the similarity between the two is... for: , in, Let k be the sensitivity weight of the k-th feature to the current scene label. Obtained from feature importance analysis, attention weight statistics, or ablation experiments during offline training, and satisfying the following conditions: .

[0039] Specifically, the current task features are first mapped to task vectors through an encoding network. Then, historical successful sample vectors are extracted from the expert AI model's exclusive experience base. The weighted cosine similarity between the current task vector and the nearest neighbor historical sample vector is calculated, and the maximum similarity is used as the model's similarity evaluation value for the current task. The first threshold T1 is determined by offline statistical analysis of the similarity distribution of the model's historical successful tasks, and the 10th percentile of the historical successful task similarity distribution is taken as the first threshold T1.

[0040] In this embodiment, when the scene trust score is lower than the second threshold T2, it is determined that the model has insufficient long-term credibility under the current scene label; the second threshold T2 is obtained by mapping the historical task success rate, anomaly rate and quality verification pass rate. In this embodiment, T2 is 0.62.

[0041] The uncertainty of the output prediction result is jointly represented by prediction dispersion and scene low confidence penalty. It is assumed that the flux reconstruction coefficient prediction results obtained after N forward inferences for the same input feature are as follows: Its mean is: , Specifically, when the execution model outputs the global single flux reconstruction coefficient, It is a scalar; when the model outputs a multi-region piecewise coefficient group, a local coefficient field, or a time-varying coefficient sequence, For vectors; for scalars, the following norm square degenerates into scalar squared differences; for vectors, the following norm square represents the sum of squared differences of the components.

[0042] Predicted dispersion Defined as: , Furthermore, assuming the confidence level of the main scene label is... The total uncertainty U is then defined as: , in, and These are the weighting coefficients, and In this embodiment, Take 0.7, The threshold is set to 0.3 or determined by minimizing the false positive rate on the validation set. The third threshold T3 is determined by the U-distribution of high-quality and low-quality prediction samples in the offline validation set, and the boundary value with the smallest false positive rate is selected as the third threshold T3.

[0043] In this embodiment, when When the vector is used, component weights are introduced based on the importance of different components in the STE-KEP-FR solver. Assuming... The j-th component is Its mean component is The component weights are Then the weighted prediction dispersion Represented as: , Where P is the dimension of the flux reconstruction coefficient vector. Greater than zero, and satisfying .

[0044] In this embodiment, the success rate of similar tasks in the recent period is statistically analyzed using a sliding window method. It is assumed that the recent success rate... Among tasks with the same main scene tag or the same combination of main and auxiliary tags, the number of successful tasks is: The recent success rate for: , in, The value ranges from 20 to 100. The fourth threshold T4 is determined through offline verification. In this embodiment, T4 is set to 0.73.

[0045] When the similarity is lower than the first threshold T1, the scene trust score is lower than the second threshold T2, the uncertainty of the output prediction result is higher than the third threshold T3, or the historical success rate is lower than the fourth threshold T4, the target expert AI model is determined to be unsuitable for handling the current task.

[0046] When the target expert AI model is suitable for handling the current task, it is used as the execution model to generate flux reconstruction coefficients. In this embodiment, the flux reconstruction coefficients adopt any one of four output forms or a combination thereof: the first is a global single coefficient, the second is a multi-region segmented coefficient group, the third is a local coefficient field, and the fourth is a time-varying coefficient sequence.

[0047] When the target expert AI model is not suitable for handling the current task, the task delegation process is initiated. The proxy model delegates the current task to another expert AI model. After confirming that it is suitable for handling the current task, the other expert AI model outputs the throughput reconstruction coefficient as the execution model.

[0048] S4: Load the generated flux reconstruction coefficients into the STE-KEP-FR solver to control the flux reconstruction process and perform numerical solutions, and form an experience record for this task based on the solution results; In this embodiment, the solver interface receives the flux reconstruction coefficient vector output by the execution model and loads the vector into the internal parameters of the STE-KEP-FR solver according to a preset mapping relationship. The flux reconstruction coefficient vector includes at least three components: the first component is used to map the flux reconstruction weight of the unit interface, the second component is used to map the internal reconstruction correction coefficient of the unit, and the third component is used to map the local dissipation adjustment amount. For implementations that require more precise control, a fourth component can be added for anti-oscillation adjustment parameters, a fifth component for smooth reconstruction intensity control parameters, and a sixth component for regional reconstruction strategy switching parameters.

[0049] In this embodiment, the solver reads the current flux reconstruction coefficient vector once in each update cycle. The update cycle is every iteration step, every several iteration steps, every scene switch, or when a stability warning is detected. To make the update triggering conditions clearer, in this embodiment, when the residual growth rate is greater than a preset threshold for two consecutive update cycles, or the local oscillation comprehensive index O exceeds a preset upper limit, or the main scene label changes, or when a preset number of steps is reached, a solver parameter update is triggered. In this specific embodiment, the preset number of steps is 5 steps, 10 steps, or 20 steps.

[0050] The solver call chain is as follows: the solver first reads the conserved variables of the current mesh element and the element interface, then determines the flux reconstruction weight and local dissipation adjustment amount of the element interface based on the current flux reconstruction coefficient, then completes the calculation of the interface numerical flux and the calculation of the internal correction term of the element, and finally updates the residual and state variables of the next iteration step.

[0051] After the solution is completed, an experience record for this task is formed based on the solution results. The experience record includes at least the task identifier, flow field characteristics, scene label and confidence level, surrogate model identifier, execution model identifier, output flux reconstruction coefficient, solution effect evaluation results, convergence index, error index, stability index and timestamp.

[0052] The solution effect is obtained through solution effect evaluation. The solution effect evaluation includes at least one or more of the following: convergence, final residual level, residual decay rate, local numerical oscillation, non-physical solution, divergence, error change relative to the preset reference result, stability change relative to the preset reference result, and computation time. The preset reference result is determined according to the scenario. In steady-state solution scenarios, the result of the fixed flux reconstruction coefficient scheme or the result of the empirical rule switching scheme is used as the preset reference result. In unsteady solution scenarios, the result of the previous round of stable solution is used as the preset reference result. In comparative verification scenarios, the result of the single model prediction scheme is also used as the preset reference result.

[0053] S5: Perform quality verification on the experience records and assign the verified experience records to the experience library dedicated to the execution model.

[0054] In this embodiment, quality verification includes quality scoring, deduplication judgment, and validity judgment. Let the convergence score be... The error score is Stability score: The calculated cost score is The physical rationality score is Then the quality score Q of the experience record is: , in, For the weighting coefficients, satisfying .

[0055] In this embodiment, Take values ​​of 0.25, 0.20, 0.25, 0.15, and 0.15 respectively.

[0056] In this embodiment, Defined as: , in, For the final residual, This serves as the upper limit of the reference residual.

[0057] Defined as: , in, This is the current error. This is the upper limit of the reference error.

[0058] Defined as: , in, The current solution time, This is the upper limit for reference time consumption.

[0059] Assuming the comprehensive index of abnormal oscillations is 0, the upper limit of the abnormal oscillation reference is: Then the stability score Defined as: , When no divergence warning is triggered and the comprehensive local oscillation index is significantly lower than the upper limit of the reference range. Approaching 1; when the abnormal oscillation composite index approaches or exceeds the upper limit of the reference range, If the value is close to 0 and a divergence warning is triggered during the solution process, then directly set the value to 0. Take 0, It can be determined by the upper quartile of the comprehensive index of abnormal oscillations in historical high-quality tasks, or by the mean plus a standard deviation of the verified centered stable solution sample.

[0060] The Physical Reasonableness Score (Sphys) reflects whether the solution meets the basic physical constraints. It is assumed that the number of monitored physical constraint violations is [number missing]. The total number of monitoring times was ,but: , Among them, physical constraint violations include, but are not limited to, density being less than the preset lower limit, pressure being less than the preset lower limit, temperature exceeding the preset physical range, or other cases where conserved variables cross the bounds. When there are no physical constraint violations, take 1; when the number of violations increases, gradually decrease; when the violation ratio reaches or exceeds 1, take 0, is the total number of constraint checks performed on all monitored physical quantities during this solution process, including the sum of the number of density checks, pressure checks, temperature checks, and other key conserved variable checks.

[0061] The first quality threshold Q1 and the second quality threshold Q2 are determined by offline statistical analysis of the historical experience score distribution. An example is to calculate the score mean and standard deviation , and take - as the first quality threshold Q1; calculate the score mean and standard deviation , and take as the second quality threshold Q2. When Q > Q1, it is determined as a high-quality experience and is allowed to be written into the exclusive experience library of the execution model; when Q2 ≤ Q ≤ Q1, it is determined as an experience to be observed and is written into the observation area; when Q < Q2, it is determined as a low-quality experience and is rejected from being written into the exclusive experience library of the execution model.

[0062] The duplicate judgment is jointly implemented by hash comparison and vector similarity comparison. Hash comparison is used to identify completely duplicate records, and vector similarity comparison is used to identify highly similar but not completely identical records. In addition to the above methods, other technical means capable of identifying duplicate experiences or highly similar experiences are also used for duplicate judgment.

[0063] The validity judgment is used to identify experience records that, although having a high score, have engineering risks. If obvious divergence warnings, non-physical solutions, or abnormal oscillations were triggered during the solution process, then even if such an experience performs well in some individual indicators, it can be marked as invalid or low-quality. In addition to the above methods, other technical means capable of identifying low-quality experiences, risk experiences, or non-reusable experiences are also used for validity judgment.

[0064] The experience records written into the observation area are regularly reviewed according to preset rules; the observation area is used to temporarily store experience records with quality scores in the middle range that require subsequent re-verification or delayed storage. For each experience record to be observed, a review is performed after accumulating M new task results of the same type of scenario, where M ranges from 5 to 20. The same type of scenario refers to a scenario where the main scenario label is the same and the overlapping ratio of the auxiliary scenario labels is higher than the preset threshold; Without auxiliary scene labels, similar scenes must at least have consistent main scene labels and local feature similarity higher than a preset similarity threshold. If the flux reconstruction coefficient structure corresponding to the experience record is verified multiple times in subsequent similar tasks, and its average quality score in similar scenes is higher than the first quality threshold Q1, or is confirmed by manual review to have a stable positive effect, then it is transferred to the execution model's dedicated experience library. If, within a preset observation period, its average quality score in similar scenes is consistently lower than the second quality threshold Q2, or is repeatedly associated with abnormal oscillations, non-physical solutions, or divergence warnings, then it is deleted from the observation area. If its performance is consistently in the middle range, it remains in the observation area and awaits subsequent review.

[0065] Once the experience record passes quality verification, its attribution rules are as follows: If the target expert model generates the flux reconstruction coefficients on its own, the experience record is written into the target expert model's dedicated experience library; if the proxy model delegates the generation of flux reconstruction coefficients to the execution model, the experience record is written into the execution model's dedicated experience library, not the proxy model's experience library; if there are multiple levels of delegation, the experience record is written into the experience library of the final execution model that outputs the flux reconstruction coefficients used by the STE-KEP-FR solver; the final execution model refers to the model that ultimately outputs the flux reconstruction coefficients for the STE-KEP-FR solver in the multi-level delegation chain, excluding proxy models that only pass on tasks without outputting the final coefficients. If the flux reconstruction coefficients are generated by the general backup model, the experience record is written into the general backup model's experience library or the public backup experience library.

[0066] The task delegation process includes the proxy model sending a delegation request to the execution model. The delegation request contains at least one of the following information: task identifier, current flow field feature summary, scene label, scene confidence, proxy model identifier, and current delegation chain record. After receiving a delegation request, the execution model determines whether it is suitable to handle the current task. If it is suitable, it performs flux reconstruction coefficient calculation. If it is not suitable, it continues to delegate the current task to other expert AI models or switches to a general backup model.

[0067] In this embodiment, the delegation request sent by the proxy model to the execution model includes at least the task identifier, current flow field feature summary, scene label, scene confidence, proxy model identifier, and current delegation chain record. If necessary, it may also include a list of models that have been tried, current computation context information, and adaptability evaluation results.

[0068] After receiving a delegation request, the execution model determines whether it is suitable for handling the current task. If it is suitable for handling the current task, it performs the flux reconstruction coefficient calculation; if it is not suitable for handling the current task, it continues to delegate to other expert AI models or switches to a general backup model.

[0069] It also includes a mechanism to prevent circular delegation: Maintain a delegation chain record for each task, and check whether the target model to be delegated already exists in the current delegation chain record before each delegation. If the result is that it exists, it is determined to be a circular delegation, and the current delegation process is interrupted, and the task is switched to the general backup model.

[0070] In this embodiment, each task maintains a delegation chain record. Before each delegation, the system checks whether the target model to be delegated already exists in the current delegation chain. If it does, it is determined to be a circular delegation, and the current delegation process is interrupted, and the task is switched to the general backup model.

[0071] The delegation chain records are stored using a dual structure of ordered list and hash set. The ordered list is used to preserve the delegation order, and the hash set is used to quickly determine whether there is a duplicate. In addition to duplicate detection, a maximum delegation level is also set. When the delegation level exceeds the preset limit, even if no duplicate is detected, it will directly switch to the general backup model.

[0072] Quality verification includes: scoring the quality of experience records, determining duplicates, and assessing validity. Determine whether experience records are duplicated by comparing hash values ​​and vector similarity. When the quality score is higher than the first quality threshold, it is judged as a high-quality experience and can be written into the experience library dedicated to the execution model. When the quality score is between the first quality threshold and the second quality threshold, it is determined to be an experience to be observed and written into the observation area; When the quality score is lower than the second quality threshold, it is judged as low-quality experience and is refused to be written into the experience library dedicated to the execution model.

[0073] In this embodiment, the experience records in the observation area do not participate in immediate positive training or immediate retrieval enhancement. Instead, after accumulating a sufficient number, the decision on whether to transfer them to the formal dedicated experience database is made through subsequent performance verification on similar tasks, manual sampling verification, or stability statistics within a time window.

[0074] The rules for attributing experience records are as follows: If the target expert model generates the flux reconstruction coefficients on its own, the experience records will be written into the target expert model's dedicated experience library. If the proxy model delegates the generation of flux reconstruction coefficients to the execution model, then the experience records will be written into the execution model's dedicated experience library. If there are multiple levels of delegation, the experience records are written into the final execution model experience library of the flux reconstruction coefficients used by the STE-KEP-FR solver in the final output. If the flux reconstruction coefficients are generated by the general backup model, the experience records are written into the general backup model experience library or the public safety net experience library.

[0075] In this embodiment, before writing the experience record, the experience attribution control reads the task link information to determine whether the task has been delegated, whether multi-level delegation has occurred, and which model will ultimately output the flux reconstruction coefficients for the STE-KEP-FR solver to actually execute. Then, the writing operation is performed according to the attribution rules.

[0076] It also includes a trust score update mechanism: Maintain a trust score for each expert AI model under different scenario labels; The trust score is dynamically updated based on the solution effect and empirical quality verification results corresponding to the solution results. When the execution result is better than the preset reference result or reaches the high quality standard, the trust score in the corresponding scenario is increased; When the execution result is worse than the preset reference result, triggers divergence, or produces abnormal oscillations, the trust score in the corresponding scenario is reduced. The updated trust score is used to determine the target expert model for subsequent tasks.

[0077] In this embodiment, the trust score is maintained in two dimensions: model and scene label. Its initial value is set through offline verification results, or it can be given a unified initial value and gradually converged during online operation.

[0078] Trust score updates use an exponential moving average method, assuming the current trust score is... The evaluation value for this task is Then the updated trust score for: , in, As the attenuation factor, Furthermore, Defined as: , in, This represents the performance improvement relative to the preset reference result. This is the normalized value for the quality score. To solve for the stability gain, For abnormal penalty amount, For the weighting coefficients, satisfying .

[0079] It also includes at least one of the following fault tolerance mechanisms: Low confidence backup mechanism: When the maximum confidence of the scene recognition output is lower than the preset threshold, stop selecting the expert AI model and directly call the general backup model to generate the throughput reconstruction coefficients; Timeout retry mechanism: When the target expert model or the execution model fails to return a result within a set time, it will automatically switch to the backup expert AI model with the second highest priority. Abnormal result rollback mechanism: When the flux reconstruction coefficients returned by the execution model trigger abnormal oscillations, non-physical solutions, or divergence warnings during the solution process, the rollback will revert to the previous round of stable flux reconstruction coefficients or switch to conservative backup coefficients.

[0080] In this embodiment, the low confidence backup mechanism is used to handle the situation where scene recognition is unreliable. When the maximum confidence of the scene recognition output is lower than the preset threshold C1, the system stops selecting the expert AI model and directly calls the general backup model to generate the throughput reconstruction coefficients.

[0081] In this embodiment, the timeout retry mechanism is used to handle the situation where the model response fails. When the target expert model or the execution model does not return a result within a set time, the system automatically switches to the backup expert AI model with the second highest priority. If the number of retries exceeds the threshold, it switches to the general backup model.

[0082] Suppose that a certain physical quantity within a local monitoring window takes the value of K consecutive iterations. The main indicator of local oscillation for: , in, This represents the mean of the physical quantity over the K iteration steps. To prevent the use of tiny positive numbers with a denominator of zero, we further assume that the residuals of K consecutive iterations are respectively Then the residual volatility index for: , The final abnormal oscillation composite index O is defined as follows: , in, For the weighting coefficients, satisfying When O exceeds the preset oscillation threshold, abnormal oscillation is determined to have occurred. Abnormal oscillation refers to the unexpected high-frequency oscillation or continuous amplification of amplitude of local variables during the solution process. Non-physical solution refers to the solution variable exceeding the physical allowable range, such as density, pressure or temperature showing obviously unreasonable values. Divergence warning refers to the situation where the residual continues to grow, the oscillation index exceeds the preset threshold, or the monitoring index indicates that the solution may be unstable. In addition, the residual growth rate is greater than the preset threshold for several consecutive update cycles, density, pressure or temperature exceeds the physical allowable range, and the divergence warning index exceeds the preset upper limit as trigger conditions. Once triggered, the current flux reconstruction coefficient is immediately stopped and the system reverts to the previous round of stable flux reconstruction coefficient, or switches to the conservative backup coefficient output by the general backup model.

[0083] Example 2 The system receives a request to calculate the flux reconstruction coefficient from a numerical simulation of flow around an airfoil. It acquires flow field features such as inlet velocity, Reynolds number, Mach number, local velocity gradient, pressure gradient, turbulent kinetic energy, residual level, and feedback from the previous flux reconstruction coefficient. These features are then standardized and graphically encoded. The scene recognition model infers from the encoded features, outputting the main scene label as "turbulent" and the auxiliary scene label as "near-wall turbulence," with a maximum confidence score of 0.90. Model routing combines the matching results of the main scene label and the auxiliary scene label, the confidence score, the trust score of the expert AI model in the turbulent scene, historical success rate, and response latency to comprehensively score multiple expert AI models. Expert model B, with the highest score, is identified as the target expert. The model is then evaluated. Expert model B performs an adaptability assessment based on the similarity between the current task characteristics and historical samples in its experience base, the trust score in turbulent scenarios, the uncertainty of the output prediction results, and the recent success rate. The results show that it is suitable for handling the current task. Therefore, it is directly used as the execution model to generate flux reconstruction coefficients. After being loaded into the STE-KEP-FR solver, these flux reconstruction coefficients are applied to the flux reconstruction weights and local dissipation adjustment of the cell interface. After solving, the residuals continue to decrease, and there are no obvious abnormal oscillations at the local monitoring points. Based on this, the system forms an experience record for this task and writes it into the expert model B's exclusive experience base when the quality score is higher than Q1. At the same time, the trust score of expert model B under the turbulence and near-wall turbulence labels is improved.

[0084] Example 3 The system receives a request to calculate flux reconstruction coefficients for a shock-turbulence coupled scenario. After scenario identification and model routing, the task is initially assigned to expert model A. After adaptability evaluation, expert model A finds that the similarity between the current task and its laminar flow historical samples is lower than the first threshold T1, and its trust score in the turbulence scenario is lower than the second threshold T2. Therefore, it determines that it is not suitable to handle the current task. As a proxy model, expert model A sends a delegation request to expert model C, which includes a task identifier, flow field feature summary, scenario label, scenario confidence, proxy model identifier, and current delegation chain record. After receiving the delegation request, expert model C confirms that it is suitable to handle the current task. It then generates flux reconstruction coefficients that enhance local dissipation in the shock region and reduce unnecessary dissipation in the smooth region. After the solution is completed, the system forms an experience record and completes quality verification. Since the flux reconstruction coefficients for this task are generated by expert model C and actually executed by the STE-KEP-FR solver, the experience record is written into the dedicated experience library of expert model C, not the dedicated experience library of expert model A.

[0085] Example 4 When the system processes a complex transitional flow field task, the proxy model first delegates the task to expert model B. After receiving the delegation request, expert model B finds that its success rate under this label has recently decreased and its output uncertainty is high. Therefore, it continues to delegate the current task to expert model C. If expert model C attempts to delegate the same task back to expert model A, and expert model A already exists in the current delegation chain, the anti-circular delegation mechanism detects circular delegation, interrupts the current delegation process, and switches to the general backup model. The general backup model outputs conservative flux reconstruction coefficients, and the solver uses these coefficients to continue running, thereby avoiding the system from getting stuck due to circular delegation.

[0086] Example 5 The system continuously accumulates experience records during long-term operation. Whenever a new experience record is formed, the experience quality verification first excludes exact duplicate records through hash comparison, and then judges whether it is too similar to existing high-quality experiences through vector similarity comparison. If the record is not duplicated, a quality score is formed based on convergence, error, stability, computational cost, and physical rationality. Records above Q1 are directly written into the execution model's dedicated experience library, records in the middle range are written into the observation area, and records below Q2 are rejected from being included in the library. At the same time, the trust score management dynamically updates the model's trust score under the corresponding label based on the current solution effect and quality verification results. If a model generates high-quality experiences multiple times in a row, its trust score gradually increases and it will obtain higher priority in subsequent routing. If a model triggers abnormal oscillations or divergence warnings continuously, its trust score gradually decreases and its probability of being selected in the future decreases.

[0087] Example 6 When the system processes a critical mixed scenario of laminar and turbulent flow, the scenario identification outputs multiple labels, but the maximum confidence level is only 0.32, which is lower than the low confidence backup threshold C1. At this point, the system stops selecting the expert AI model and directly calls the general backup model to output conservative flux reconstruction coefficients. Subsequently, during the solution process, if the solution result corresponding to the conservative coefficients is stable, an experience record is formed and written into the general backup model experience library or the public safety experience library. If the residual abnormally increases, the local oscillation amplitude amplifies, or a divergence warning occurs when using flux reconstruction coefficients output by other models, the system immediately reverts to the previous stable coefficient or calls the general backup model again to ensure that the solution can continue.

[0088] Example 7 Before deployment, the system first constructs an offline verification sample set, which includes high-quality solving tasks, low-quality solving tasks, and failed tasks. For each expert AI model, the similarity distribution of historical successful and failed tasks is statistically analyzed to determine the first threshold T1. The relationship between trust scores and task success rates under each scenario label is statistically analyzed to determine the second threshold T2. The uncertainty distribution of high-quality and low-quality prediction samples is statistically analyzed to determine the third threshold T3. The most recent... The relationship between the success rate of the second-order similar task and the subsequent actual success rate is used to determine the fourth threshold T4; the error routing rate corresponding to the recognition confidence interval of different scenarios is statistically analyzed to determine the low confidence backup threshold C1; the quality score distribution is statistically analyzed to determine Q1 and Q2; and the distribution of the abnormal oscillation comprehensive index is statistically analyzed to determine the oscillation threshold.

[0089] In a more specific embodiment, before deployment, the system uses an offline verification sample set to calibrate key thresholds. Taking the similarity threshold T1 as an example, the weighted similarity distribution of historical successful tasks of the target expert AI model is statistically analyzed, and the 10th percentile value is taken as T1. Taking the scene trust score threshold T2 as an example, it is set to 0.62; taking the recent success rate threshold T4 as an example, it is set to 0.73; taking the low confidence backup threshold C1 as an example, it is set to 0.35. For the quality score thresholds Q1 and Q2, they are determined based on the mean and standard deviation of the scores of the high-quality experience set and the failure experience set, respectively. For the experience records in the observation area, the system performs a review after accumulating 10 results of similar tasks. If the average quality score of similar tasks after review is higher than Q1, it is transferred to the formal dedicated experience library; if it is lower than Q2, it is deleted; otherwise, it is retained. Through the above values ​​and rules, those skilled in the art can complete the system deployment and operation.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment, characterized in that, Includes the following steps: S1: Obtain the flow field characteristics of the target flow field; S2: Based on the flow field characteristics, perform scene recognition and output at least one scene label and the confidence level corresponding to each scene label; S3: Based on the scene label, confidence level, and trust scores of multiple expert AI models in the corresponding scene, determine the target expert model; the target expert model performs an adaptability assessment on the current task, and determines whether it is suitable for handling the current task based on the adaptability assessment result; if it is determined to be suitable for handling the current task, the target expert model generates throughput reconstruction coefficients as the execution model; if it is determined to be unsuitable for handling the current task, initiate the task delegation process, with the target expert model acting as a proxy model to delegate the current task to another expert AI model, and the execution model generates throughput reconstruction coefficients. S4: Load the generated flux reconstruction coefficients into the STE-KEP-FR solver to control the flux reconstruction process and perform numerical solutions, and form an experience record for this task based on the solution results; S5: Perform quality verification on the experience records and assign the verified experience records to the experience library dedicated to the execution model.

2. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 1, characterized in that, The scene recognition adopts a hierarchical labeling system, which includes at least a first-level label and a second-level label. The primary labels include at least one of laminar flow scenario, transitional flow scenario, turbulent flow scenario, shock wave scenario, mixed flow scenario, and unknown scenario; The secondary labels include at least one of the following: stable laminar flow, separated laminar flow, stable turbulent flow, pulsating turbulent flow, near-wall turbulent flow, strong shock wave, weak shock wave, shock wave boundary layer disturbance, laminar-turbulent transition, and turbulent-shock wave coupling; When multiple scene labels are identified, the scene label corresponding to the highest confidence level is used as the primary scene label, and the remaining scene labels are used as auxiliary scene labels to participate in the determination of the target expert model.

3. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 1, characterized in that, The target expert model is determined based on the comprehensive scoring results; The comprehensive scoring result is based on at least the following two items: main scene label matching, auxiliary scene label matching, scene recognition confidence, trust score of each expert AI model under the corresponding label, historical success rate of each expert AI model, real-time load status of each expert AI model, and response latency of each expert AI model. The expert AI model with the highest overall score was selected as the target expert model.

4. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 1, characterized in that, The adaptability assessment includes at least one of the following: calculating the similarity between the current task features and historical samples in the expert AI model's experience base, obtaining the expert AI model's trust score for the current scene label, assessing the uncertainty of the current output prediction result, and obtaining the success rate of the expert AI model in recent similar tasks. When the similarity is lower than the first threshold, the scene trust score is lower than the second threshold, the uncertainty of the output prediction result is higher than the third threshold, or the historical success rate is lower than the fourth threshold, the target expert model is determined to be unsuitable for handling the current task.

5. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 1, characterized in that, The task delegation process includes: the proxy model sending a delegation request to the execution model, wherein the delegation request contains at least one of the following information: task identifier, current flow field feature summary, scene label, scene confidence, proxy model identifier, and current delegation chain record; After receiving the delegation request, the execution model determines whether it is suitable to handle the current task. If it is suitable to handle the current task, it performs the flux reconstruction coefficient calculation. If it is not suitable to handle the current task, it continues to delegate the current task to other expert AI models or switches to a general backup model.

6. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 5, characterized in that, It also includes a mechanism to prevent circular delegation: Maintain a delegation chain record for each task, and check whether the target model to be delegated already exists in the current delegation chain record before each delegation; If the detection result indicates that the task exists, it is determined to be a circular delegation, and the current delegation process is interrupted, switching the task to the general backup model.

7. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 1, characterized in that, The quality verification includes: scoring the quality of experience records, determining duplicates, and determining validity. Determine whether experience records are duplicated by comparing hash values ​​and vector similarity. When the quality score is higher than the first quality threshold, it is judged as a high-quality experience and can be written into the experience library dedicated to the execution model. When the quality score is between the first quality threshold and the second quality threshold, it is determined to be an experience to be observed and written into the observation area; When the quality score is lower than the second quality threshold, it is judged as low-quality experience and is refused to be written into the experience library dedicated to the execution model.

8. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 1, characterized in that, The rules for attributing the experience records are as follows: If the target expert model generates the flux reconstruction coefficients on its own, the experience records will be written into the target expert model's dedicated experience library. If the proxy model delegates the generation of flux reconstruction coefficients to the execution model, then the experience records will be written into the execution model's dedicated experience library. If there are multiple levels of delegation, the experience records are written into the final execution model experience library of the flux reconstruction coefficients used by the STE-KEP-FR solver in the final output. If the flux reconstruction coefficients are generated by the general backup model, the experience records are written into the general backup model experience library or the public safety net experience library.

9. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 1, characterized in that, It also includes a trust score update mechanism: Maintain a trust score for each expert AI model under different scenario labels; The trust score is dynamically updated based on the solution effect and empirical quality verification results corresponding to the solution results. When the execution result is better than the preset reference result or reaches the high quality standard, the trust score in the corresponding scenario is increased; When the execution result is worse than the preset reference result, triggers divergence, or produces abnormal oscillations, the trust score in the corresponding scenario is reduced. The updated trust score is used to determine the target expert model for subsequent tasks.

10. The method for calculating STE-KEP-FR flux reconstruction coefficients based on AI adaptive adjustment according to claim 1, characterized in that, It also includes at least one of the following fault tolerance mechanisms: Low confidence backup mechanism: When the maximum confidence of the scene recognition output is lower than the preset threshold, stop selecting the expert AI model and directly call the general backup model to generate the throughput reconstruction coefficients; Timeout retry mechanism: When the target expert model or the execution model fails to return a result within a set time, it will automatically switch to the backup expert AI model with the second highest priority. Abnormal result rollback mechanism: When the flux reconstruction coefficients returned by the execution model trigger abnormal oscillations, non-physical solutions, or divergence warnings during the solution process, the rollback will revert to the previous round of stable flux reconstruction coefficients or switch to conservative backup coefficients.

Images