A dynamic multi-fidelity processing method and product for predicting high-temperature alloy creep curves

Abstract

The application provides a dynamic multi-fidelity processing method and product for predicting a high-temperature alloy creep curve, a model library containing a low-fidelity model, a medium-fidelity model and a high-fidelity model is constructed, different precision models in the model library are dynamically called according to a region where the high-temperature alloy is located and a creep strain rate, and thus the goal of 'global high efficiency + local high precision' is achieved, the calculation resource demand is reduced, the calculation efficiency is improved, and meanwhile, the precision of the simulation result is ensured. On this basis, the application also aims at the possible 'breakage' or'mutation' of the creep curve in the model calling and switching process, dynamically solves and optimizes the method, controls the global error, and realizes the continuity of the data output.

Description

Technical Field

[0001] This invention belongs to the field of high-temperature alloy creep performance simulation and prediction technology, specifically involving a dynamic multi-fidelity processing method and product for predicting the creep curve of high-temperature alloys. Background Technology

[0002] High-temperature alloys, as core materials for critical equipment such as aero-engines and gas turbines, directly determine the service life and safety of these equipment due to their creep performance under extreme high-temperature environments. Creep, as the primary failure mechanism of high-temperature alloys during long-term service, exhibits significant time dependence and complex microscopic physical mechanisms. Traditional creep prediction methods based on empirical formulas and semi-empirical models, while possessing certain prediction accuracy under specific conditions, struggle to accurately describe the creep behavior of materials across a wide temperature and stress range, particularly exhibiting limited extrapolation capabilities under extreme conditions. With the development of computational science, numerical simulation has become a crucial tool for studying the creep of high-temperature alloys. Constitutive models based on physical mechanisms, data-driven machine learning methods, and multi-scale coupled modeling techniques offer new avenues for the accurate prediction of the creep performance of high-temperature alloys.

[0003] However, the aforementioned models and methods also have certain limitations in application. Multi-scale coupled modeling techniques, such as the Crystal Plasticity Finite Element Method (CPFEM), incorporate the anisotropy of crystals and the activation of slip systems into the finite element framework. This method describes the plastic deformation of crystals by defining the geometric characteristics and constitutive relations of slip systems, achieving correlation prediction from microscopic mechanisms to macroscopic properties. It can reveal the intrinsic relationship between composition, process, microstructure, and macroscopic properties, but the computational cost is extremely high, requiring a high-performance computing platform. Data-driven models do not rely on creep microscopic mechanisms and directly utilize a large amount of experimental or simulation data. They construct nonlinear mapping relationships between inputs and outputs through machine learning algorithms, featuring fast modeling speed and high computational efficiency, but lack physical meaning and have weak extrapolation capabilities.

[0004] Therefore, the simulation of creep properties of high-temperature alloys still faces the challenges of high computational cost of high-precision models and insufficient accuracy of low-precision models. This results in many challenges for existing numerical simulation methods in terms of accuracy, efficiency and applicability. In particular, how to balance computational efficiency and prediction accuracy under cross-scale and multi-field coupling conditions remains a difficult point. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a dynamic multi-fidelity processing method and product for predicting the creep curve of high-temperature alloys. By dynamically selecting appropriate models and order reduction processing strategies, the accuracy of model calculations can be guaranteed, and the evolution of microstructure and changes in macroscopic properties during alloy creep can be fully described. At the same time, the calculation efficiency can be significantly improved, enabling efficient and accurate prediction of the creep performance of high-temperature alloys.

[0006] In a first aspect, the present invention provides a dynamic multi-fidelity processing method for predicting the creep curve of high-temperature alloys, the technical solution of which includes the following steps: 1. A multi-fidelity model library is constructed for the creep properties of high-temperature alloys. The multi-fidelity model library includes high-fidelity models, medium-fidelity models, and low-fidelity models. The high-fidelity models adopt a multi-scale modeling strategy to predict the creep deformation behavior and microstructure evolution characteristics of high-temperature alloys with different crystal orientations under different temperatures and loads. The medium-fidelity models adopt surrogate models or machine learning models to describe the coupled change process between elasticity and creep of high-temperature alloys under a wide range of stress, temperature, and time scales, and can achieve the ability to predict complex, nonlinear, and multi-factor coupled creep behavior. The low-fidelity models adopt empirical formulas to describe the deformation behavior of materials in the steady-state creep stage, predict the creep strain rate and creep strain field distribution of high-temperature alloys under different working conditions, and can quickly estimate the deformation of high-temperature alloys under long-term loads.

[0007] 2. First, a low-fidelity model is used to distinguish whether the region where the high-temperature alloy is located is a critical region based on the operating parameters of the high-temperature alloy. These operating parameters include the temperature conditions and stress level of the high-temperature alloy, and the strain gradient of the high-temperature alloy. When the creep strain rate is >3%, the high-temperature alloy is in a critical region. A medium-fidelity model is used to calculate the creep strain rate of the high-temperature alloy at the initial moment; when the strain gradient of the high-temperature alloy... When the creep strain rate is ≤3%, the high-temperature alloy is in a non-critical region. A low-fidelity model is used to calculate the creep strain rate of the high-temperature alloy at the initial moment.

[0008] 3. Based on the creep strain rate of the high-temperature alloy, select a suitable model from the multi-fidelity model library for iterative simulation of the creep strain rate; the model selection criterion is: when the creep strain rate of the high-temperature alloy... When the high-temperature alloy is in the accelerated creep stage or critical region, a high-fidelity model is used to recalculate; when Creep strain rate of high-temperature alloys At that time, the high-temperature alloy was in the steady-state creep stage, and the calculation was re-performed using a medium-fidelity model; when the creep strain rate of the high-temperature alloy... At that time, the high-temperature alloy was in the elastic stage or the initial creep stage, and a low-fidelity model was used for calculation.

[0009] 4. As time iterates, based on the creep strain rate at each moment and according to the model selection criteria in step 3, the model in the multi-fidelity model library is dynamically called in real time to obtain the creep strain change curve of the high-temperature alloy over time.

[0010] Furthermore, high-fidelity models can employ full-order coupled crystal plastic finite element (CPFEM) models, medium-fidelity models can employ surrogate models (such as the Chaboche model based on unified creep-plastic constitutive model) or machine learning models (such as the Gaussian process regression GP model), and low-fidelity models can employ Norton's power-law creep equation.

[0011] Furthermore, in step 3, a dynamic weighted solution and optimization method is adopted. First, based on the working conditions and regional characteristics of the high-temperature alloy, and the initial weights assigned to each high-fidelity model, the predicted values ​​of each model are calculated. Then, the predicted values ​​of each model are compared with the simulation results of the high-fidelity model or the actual results of the experimental data to calculate the prediction error of each high-fidelity model. Next, based on the Bayesian adaptive weighting algorithm, combined with the error likelihood probability and the initial weights, the posterior weights of each model are calculated to achieve dynamic adjustment of the weights of each high-fidelity model. The predicted values ​​of each model are multiplied by the posterior weights to obtain the optimal creep strain value after fusion. If the posterior weight of a certain model is continuously lower than the threshold and simultaneously meets the creep strain rate condition in model selection, model weight adjustment is triggered. The model weight adjustment method is as follows: 1. If the weights W of the low-fidelity model LF <0.1, and Creep strain rate of high-temperature alloys At this time, the calculation of the low-fidelity model is turned off, and the weight of the medium-fidelity model is increased; 2. If the weights W of the medium-fidelity model MF <0.1, and the creep strain rate of the high-temperature alloy At that time, increase the weight of the high-fidelity model.

[0012] Furthermore, in step 3, the initial weights assigned to each fidelity model are as follows: 1. When the high-temperature alloy is in the elastic stage or the initial creep stage, and the high-temperature alloy is in a non-critical region, the initial weight W of the low-fidelity model is... LF（0） =0.7, the initial weights W of the medium-fidelity model MF（0） =0.3; 2. When the high-temperature alloy is in the steady-state creep stage or critical region, the initial weight W of the low-fidelity model LF（0） =0.2, the initial weights W of the medium-fidelity model MF（0） =0.7, the initial weights W of the high-fidelity model HF（0） =0.1; 3. When the high-temperature alloy is in the accelerated creep stage and critical region, the initial weight W of the medium-fidelity model MF（0） =0.2, the initial weights W of the high-fidelity model HF（0） =0.8.

[0013] Furthermore, in the process of dynamic weighted summation, error analysis can be used to optimize the calculation of posterior weights, and active learning can be used to optimize the model parameters of low-fidelity and medium-fidelity models.

[0014] The method of optimizing the calculation of posterior weights using error analysis involves the following steps: A. Real-time monitoring of the local normalized error and global error of the fused creep strain. The local normalization error ,in Representing the units in the model High-temperature alloys in Local normalization error at time t, Representing the units in the model High-temperature alloys in At time i, the creep strain is predicted using the i-model; This refers to units retrieved from real experimental data. High-temperature alloys in The actual creep strain value at the given time; if no actual test value is available, a high-fidelity model prediction unit is used. High-temperature alloys in Creep strain at a given moment; This represents the maximum absolute error of prediction allowed by the predefined model i; The global error includes the global average error. and global maximum error Global average error Global maximum error , where N represents the total number of high-temperature alloy units in the model; B. Based on the error calculation results and the magnitude of the creep strain rate, formulate different model adjustment strategies. When local normalization error If the creep strain rate meets the selection criteria of the current model, the current model weights are maintained, the model is used again, and the simulation proceeds to the next time step. When local normalization error If the creep strain rate does not meet the selection criteria of the current model, the model accuracy is insufficient. The current model is closed, and a higher-fidelity model is called for prediction. When the global average error When the overall structural accuracy is not up to standard, the weights of each model are adjusted based on the Bayesian algorithm, and the creep strain after fusion is recalculated. When the global maximum error When this occurs, it indicates that there is a prediction blind zone in the current region or working condition. A high-fidelity model is then used to predict the creep strain in that region or working condition. The creep strain data from the prediction blind zone is then fed into a Bayesian algorithm to recalibrate the error standard deviation of the likelihood probability. Optimize the sensitivity of weight updates.

[0015] Furthermore, in the process of optimizing the calculation of posterior weights using error analysis, to avoid frequent model switching or repeated weight adjustments caused by sudden abnormal fluctuations in error, which could lead to oscillations in prediction results and decreased computational efficiency, this invention also employs a smoothing method to process the model's prediction error. The specific processing method is as follows: Using smoothed error The error calculation results are judged, among which... Indicates the current At this moment, for the units in the model Smoothing error after smoothing high-temperature alloys. Indicates the previous moment At this moment, for the units in the model Smoothing error after smoothing high-temperature alloys. Indicates the current At this moment, for the units in the model The original error of high-temperature alloys It is the smoothing coefficient, and .

[0016] Furthermore, when the high-temperature alloy is in the creep steady-state stage, When high-temperature alloys are in the accelerated creep stage, .

[0017] Furthermore, creep strain data from the prediction blind zone can be used to train low-fidelity and medium-fidelity models. An active learning strategy can be adopted to improve the prediction accuracy of low-fidelity and medium-fidelity models. The specific optimization process is as follows: For low-fidelity model parameter optimization: a Gaussian regression surrogate model is used to construct the working conditions. , The mapping relationship between creep rate error and the simulation results of the high-fidelity model is selected in each round, and the Norton power law coefficient A, stress exponent n, and activation energy Q in the low-fidelity model parameters are adjusted by Bayesian optimization algorithm. When the relative change of the high-fidelity model parameters is <2% or when 50 high-fidelity model calls are reached, the parameter optimization of the low-fidelity model is stopped. For parameter optimization of the medium-fidelity model: Real-time recording of dislocation density in key regions of the high-fidelity model. A dynamic dataset is constructed based on local strain rates; an incremental training mode is adopted based on a Long Short-Term Memory (LSTM) neural network, with the input being... The model outputs creep strain rate; when the prediction confidence is <90%, high-fidelity calculation is triggered and the training set is expanded; after every 5 iterations, the model is evaluated using an independent test set. If the relative change in RMSE between the predicted value and the high-fidelity simulation value or experimental test value of the sample is ≥2% compared to the previous iteration, the parameters are rolled back to prevent overfitting. N is the number of samples in the independent test set. For the first High-fidelity simulation values ​​or experimental test values ​​for a single sample These are the predicted values ​​obtained by LSTM based on a mid-fidelity model.

[0018] Secondly, the present invention also provides a dynamic multifidelity device for predicting the creep curve of high-temperature alloys, comprising: The test data storage and retrieval module is used to store and retrieve test and simulation data of the creep performance of high-temperature alloys under various operating conditions; A multi-fidelity model library is used to store and run low-fidelity, medium-fidelity, and high-fidelity models, and output simulation results such as creep strain rate, creep strain, and dislocation density based on the simulation results of each model. The creep strain output module dynamically calls models from the multi-fidelity model library in real time based on the creep strain rate at each moment to obtain the creep strain curve of the high-temperature alloy over time.

[0019] Furthermore, the dynamic multi-fidelity device for predicting the creep curve of high-temperature alloys of the present invention also includes a dynamic weighted solution and optimization module, which can dynamically call the prediction results of each fidelity model according to the model calling and switching criteria, and compare the prediction results of each fidelity model with the simulation results of the high-fidelity model or the actual results of experimental data. Based on the comparison results, according to its own stored dynamic weighted solution and optimization strategy, the simulation results calculated by each model are weighted and fused to obtain the optimal creep strain value after fusion, and fed back to the creep strain output module.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention addresses the dilemma that a single model cannot simultaneously achieve both computational efficiency and simulation accuracy. By constructing a multi-fidelity model library and implementing a defined model calling strategy, it enables dynamic calling of low-fidelity, medium-fidelity, and high-fidelity models, achieving the goal of "global high efficiency + local high accuracy." This reduces computational resource requirements, improves computational efficiency, and ensures the accuracy of simulation results.

[0021] 2. Employing a dynamic weighted solution and optimization method not only controls global errors and improves computational efficiency, but also ensures the continuity of data output when switching between different models through weighted fusion, enabling collaborative work among various fidelity models. Especially in the transition regions between different stages of creep, the fusion of different models eliminates the "breaks" or "abrupt changes" in the creep curve during creep model switching.

[0022] 3. To address the potential for sudden increases in computational prediction errors across different time steps, and to mitigate model prediction blind spots in certain regions or operating conditions, error analysis methods are employed to optimize the weights of each model during the dynamic weighted summation process, enhancing the adaptability of the weighting strategy. Prediction errors are smoothed to avoid frequent model switching, ensuring the stability and rationality of model call decisions. Furthermore, leveraging data collected from model prediction blind spots, active learning is used to optimize the model parameters of both low-fidelity and medium-fidelity models. Continuous use of the model library further improves the prediction accuracy of both models. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0024] Figure 1 This is a schematic diagram illustrating the operation of the dynamic multifidelity model built in this invention for solving creep curves.

[0025] Figure 2 This is a flowchart illustrating the creep curve calculation based on a dynamic multifidelity model library, as described in this invention.

[0026] Figure 3 This is a creep-time curve of the CMSX-4 alloy at different temperatures and under different stress conditions at 950°C, according to an embodiment of the present invention. Detailed Implementation

[0027] The embodiments of the technical solution of this application will be described in detail below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of this application, and are therefore merely examples and should not be used to limit the scope of protection of this application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and the foregoing description of the accompanying drawings are intended to cover non-exclusive inclusion.

[0028] In this document, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" may be used interchangeably.

[0029] Figure 1 This is a schematic diagram illustrating the operation of the dynamic multi-fidelity model built in this invention for solving creep curves. Figure 2 This is a flowchart illustrating the creep curve calculation based on a dynamic multi-fidelity model library, as described in this invention. (Combined with...) Figure 1 and 2 Through the explanation of the embodiments, it can be understood that the operation process of the dynamic multi-fidelity processing method for predicting the creep curve of high-temperature alloys in this invention and the control strategy for calling each model are as follows.

[0030] Taking aero-engine blades made of high-temperature alloys as an example, based on their location, they can be roughly divided into the blade tip, the middle of the blade body, and the blade root tenon area. According to past experience, the blade tip is a non-load-bearing area, where the high-temperature alloy creep is relatively slow. The blade root tenon area, however, is a stress concentration area, where creep occurs earlier and is the most prone to performance failure. Using a single creep life prediction model presents a dilemma: insufficient accuracy or low computational efficiency. Therefore, this invention provides a dynamic multi-fidelity processing method for predicting the creep curve of high-temperature alloys. By using models with different levels of precision, calculations can be performed on different regions of the high-temperature alloy engine blade. The method provided by this invention includes the following steps: 1. A multi-fidelity model library is constructed for the creep properties of high-temperature alloys. The multi-fidelity model library includes high-fidelity models, medium-fidelity models, and low-fidelity models. The high-fidelity models adopt a multi-scale modeling strategy to predict the creep deformation behavior and microstructure evolution characteristics of high-temperature alloys with different crystal orientations under different temperatures and load conditions. The medium-fidelity models adopt surrogate models or machine learning models to describe the coupled change process between elasticity and creep of high-temperature alloys under a wide range of stress, temperature, and time scales, and can achieve the ability to predict complex, nonlinear, and multi-factor coupled creep behavior. The low-fidelity models adopt empirical formulas to describe the deformation behavior of materials in the steady-state creep stage, predict the creep strain rate and creep strain field distribution of high-temperature alloys under different working conditions, and can quickly estimate the deformation of high-temperature alloys under long-term loads.

[0031] 2. First, a low-fidelity model is used to distinguish whether the area where the high-temperature alloy is located is a critical area based on the operating parameters. These operating parameters include the temperature conditions and stress level of the high-temperature alloy, and the strain gradient of the high-temperature alloy. When the creep strain rate is >3%, the high-temperature alloy is in a critical region. A medium-fidelity model is used to calculate the creep strain rate of the high-temperature alloy at the initial moment; when the strain gradient of the high-temperature alloy... When the creep strain rate is ≤3%, the high-temperature alloy is in a non-critical region. A low-fidelity model is used to calculate the creep strain rate of the high-temperature alloy at the initial moment.

[0032] 3. Based on the creep strain rate of the high-temperature alloy, select a suitable model from the multi-fidelity model library for iterative simulation of the creep strain rate; the model selection criterion is: when the creep strain rate of the high-temperature alloy... When the high-temperature alloy is in the accelerated creep stage or critical region, a high-fidelity model is used to recalculate; when Creep strain rate of high-temperature alloys At that time, the high-temperature alloy was in the steady-state creep stage, and the calculation was re-performed using a medium-fidelity model; when the creep strain rate of the high-temperature alloy... At that time, the high-temperature alloy was in the elastic stage or the initial creep stage, and a low-fidelity model was used for calculation.

[0033] 4. As time iterates, based on the creep strain rate at each moment and according to the model selection criteria in step 3, the model in the multi-fidelity model library is dynamically called in real time to obtain the creep strain change curve of the high-temperature alloy over time.

[0034] Among them, the high-fidelity model can adopt the full-order coupled model of crystal plastic finite element (CPFEM), the medium-fidelity model can adopt the surrogate model (such as the Chaboche model based on unified creep-plastic constitutive model) or the machine learning model (such as the Gaussian process regression GPR model), and the low-fidelity model adopts the Norton power-law creep equation.

[0035] In step 2, the high-temperature alloy stress in each region of the engine blade is initially calculated using a low-fidelity model. This helps the system initially identify critical and non-critical regions. Then, depending on whether a region is in a critical region, the system decides whether to call a low-fidelity or medium-fidelity model. Since critical regions are in stress concentration areas, they are more prone to accelerated high-temperature creep. Therefore, a medium-fidelity model is used to obtain a more accurate creep strain rate, which helps to better determine whether to call a high-fidelity model, which has higher accuracy but lower computational efficiency.

[0036] In step 3, when different models are called based on the creep strain rate of the high-temperature alloy, the differences in accuracy between models may lead to significant deviations in the calculation results before and after model switching in some areas, or discontinuous output results in certain transition regions. Therefore, a dynamic weighted solution and optimization method is used in step 3.

[0037] A dynamic weighted solution and optimization method is adopted. First, based on the working conditions and regional characteristics of the high-temperature alloy, the predicted values ​​of each model are calculated according to the initial weights assigned to each high-fidelity model. Then, the predicted values ​​of each model are compared with the simulation results of the high-fidelity model or the actual results of the experimental data to calculate the prediction error of each high-fidelity model. Next, based on the Bayesian adaptive weighting algorithm, combined with the error likelihood probability and the initial weights, the posterior weights of each model are calculated to realize the dynamic adjustment of the weights of each high-fidelity model. The predicted values ​​of each model are multiplied by the posterior weights to obtain the optimal creep strain value after fusion. If the posterior weight of a certain model is continuously lower than the threshold and the creep strain rate condition in the model selection is met, the model weight adjustment is triggered. The model weight adjustment method is as follows: 1. If the weights W of the low-fidelity model LF <0.1, and Creep strain rate of high-temperature alloys At this time, the calculation of the low-fidelity model is turned off, and the weight of the medium-fidelity model is increased; 2. If the weights W of the medium-fidelity model MF <0.1, and the creep strain rate of the high-temperature alloy At that time, increase the weight of the high-fidelity model.

[0038] Specifically, in step 3, the initial weights assigned to each fidelity model can be as follows: 1. When the high-temperature alloy is in the elastic stage or the initial creep stage, and the high-temperature alloy is in a non-critical region, the initial weight W of the low-fidelity model is... LF（0） =0.7, the initial weights W of the medium-fidelity model MF（0） =0.3; 2. When the high-temperature alloy is in the steady-state creep stage or critical region, the initial weight W of the low-fidelity model LF（0） =0.2, the initial weights W of the medium-fidelity model MF（0） =0.7, the initial weights W of the high-fidelity model HF（0） =0.1; 3. When the high-temperature alloy is in the accelerated creep stage and critical region, the initial weight W of the medium-fidelity model MF（0） =0.2, the initial weights W of the high-fidelity model HF（0） =0.8.

[0039] By allocating the initial weights as described above, we can take into account the different stages of the high-temperature alloy. When the high-temperature alloy is in the elastic stage or the initial creep stage, we can give full play to the high computational efficiency of the low-fidelity model. When the high-temperature alloy is in the steady-state creep stage or the critical region, we can adopt the medium-fidelity model as the main model, taking into account both the computational accuracy and efficiency of the model. When the high-temperature alloy is in the accelerated creep stage or the critical region, we can adopt the high-fidelity model as the main model, which can more accurately describe the creep change process of the alloy in this region and ensure the prediction accuracy of creep damage. To address the challenge that a single model cannot meet the needs of the entire process: high-fidelity models are accurate but slow, low-fidelity models are fast but inaccurate, and medium-fidelity models fall in between, dynamic calling of a multi-fidelity model library achieves the goal of "global efficiency + local high accuracy." By adopting a dynamic weighted solution and optimization method, global errors can be controlled and computational efficiency can be improved. Weighted fusion can also ensure the continuity of data output when switching between different models, especially in the transition region of creep at different stages. By fusing different models, the "break" or "abrupt change" of creep curves when switching creep models can be eliminated.

[0040] The specific dynamic weighted solution and optimization process is illustrated in the following example: In the Bayesian dynamic weighted algorithm, the likelihood probability , Representation Model Model weights, It represents the absolute error of the model prediction, that is, the absolute value of the difference between the model's predicted value and the simulation result of the high-fidelity model or the actual result of the experimental data; Indicates the normalization error. Representation Model The absolute error of the prediction Indicates a pre-defined model Maximum permissible absolute error in prediction; Let be the standard deviation of the error, representing the model. The standard deviation of historical prediction errors is obtained by statistical analysis of a large amount of training or simulation data.

[0041] If the maximum absolute error of the low-fidelity model is pre-defined... The maximum absolute error of the medium-fidelity model prediction The maximum absolute error of prediction in high-fidelity models When the high-fidelity model At that time, the medium-fidelity model Under the same error standard, the medium-fidelity model has a higher likelihood probability. That is, the model with a smaller error has a higher likelihood probability, and its weight will be automatically increased in the subsequent posterior weight. Conversely, the model with a larger error will have its posterior weight actively reduced.

[0042] Posterior weights The sum of the posterior weights of all models For example, when the high-temperature alloy is in the elastic stage or the initial creep stage, and the high-temperature alloy is in a non-critical region, the initial weights of the low-fidelity model... Initial weights of the medium-fidelity model Initial weights of the high-fidelity model The likelihood probability of the low-fidelity model calculated in the simulation Likelihood probability of medium-fidelity model Likelihood probability of high-fidelity model (At this point, the high-fidelity model is not being computed.) After posterior weight adjustment, the posterior weights of the low-fidelity model are... Posterior weights of the medium-fidelity model Posterior weights of high-fidelity models .

[0043] The optimal creep strain after fusion is obtained by multiplying the predicted values ​​of each model by their posterior weights and dynamically weighting the sum. ,in Representation Model The calculated creep strain.

[0044] It can be seen that the dynamic adjustment of the posterior weights can optimize and balance the prediction errors of each model, and obtain the fused creep strain through dynamic weighted summation, eliminating the bias of a single model and solving the possible "abrupt" or "fracture" situation of creep strain when switching between models, thus ensuring the smoothness of the subsequent creep curve output.

[0045] Furthermore, during the dynamic weighted summation process, error analysis can be used to further optimize the calculation of posterior weights, and active learning can be employed to optimize the model parameters of the low-fidelity and medium-fidelity models.

[0046] The method of optimizing the calculation of posterior weights using error analysis is as follows: A. Real-time monitoring of the local normalized error and global error of the fused creep strain, wherein the local normalized error... ,in Representing the units in the model High-temperature alloys in Local normalization error at time t, Representing the units in the model High-temperature alloys in At time i, the creep strain is predicted using the i-model; This refers to units retrieved from real experimental data. High-temperature alloys in The actual creep strain value at the given time; if no actual test value is available, a high-fidelity model prediction unit is used. High-temperature alloys in Creep strain at a given moment; This represents the maximum absolute error of prediction allowed by the predefined model i; The global error includes the global average error. and global maximum error Global average error Global maximum error , where N represents the total number of high-temperature alloy units in the model.

[0047] B. Based on the error calculation results and the magnitude of the creep strain rate, formulate different model adjustment strategies. When local normalization error If the creep strain rate meets the selection criteria of the current model, the current model weights are maintained, the model is used again, and the simulation proceeds to the next time step. When local normalization error If the creep strain rate does not meet the selection criteria of the current model, the model accuracy is insufficient. The current model is closed, and a higher-fidelity model is called for prediction. When the global average error When the overall structural accuracy is not up to standard, the weights of each model are adjusted based on the Bayesian algorithm, and the creep strain after fusion is recalculated. When the global maximum error When this occurs, it indicates that there is a prediction blind zone in the current region or working condition. A high-fidelity model is then used to predict the creep strain in that region or working condition. The creep strain data from the prediction blind zone is then fed into a Bayesian algorithm to recalibrate the error standard deviation of the likelihood probability. Optimize the sensitivity of weight updates.

[0048] Furthermore, in the process of optimizing the calculation of posterior weights using error analysis, to avoid frequent model switching or repeated weight adjustments caused by sudden abnormal fluctuations in error, which could lead to oscillations in prediction results and decreased computational efficiency, this invention also employs a smoothing method to process the model's prediction error. The specific processing method is as follows: Using smoothed error The error calculation results are judged, among which... Indicates the current At this moment, for the units in the model Smoothing error after smoothing high-temperature alloys. Indicates the previous moment At this moment, for the units in the model Smoothing error after smoothing high-temperature alloys. Indicates the current At this moment, for the units in the model The original error of high-temperature alloys It is the smoothing coefficient, and .

[0049] Among them, when the high-temperature alloy is in the creep steady state stage, When high-temperature alloys are in the accelerated creep stage, .

[0050] Smoothing methods can be used to process the prediction errors of the model, avoiding abnormal fluctuations in the original error that could trigger frequent model switching. An example illustrates this: Suppose the local relative errors of the medium-fidelity model in the blade element are 8%, 9%, 15%, 10%, and 11% according to the time step. A momentary spike in error occurs at the third time step, causing a single fluctuation at that time step to trigger the call to the high-fidelity model, thus reducing computational efficiency. By setting a smoothing coefficient... Initial smoothing error at the first time step Calculations yielded , , , As a result, the original error of 15% at the third time step was smoothed to 9.56%, avoiding the need to call up a higher-precision model due to a single fluctuation.

[0051] Furthermore, creep strain data from the prediction blind zone can be used to train low-fidelity and medium-fidelity models. An active learning strategy can be adopted to improve the prediction accuracy of low-fidelity and medium-fidelity models. The specific optimization process is as follows: For low-fidelity model parameter optimization: a Gaussian regression surrogate model is used to construct the working conditions. , The mapping relationship between creep rate error and the simulation results of the high-fidelity model is selected in each round, and the Norton power law coefficient (A), stress exponent (n), and activation energy (Q) in the low-fidelity model parameters are adjusted by Bayesian optimization algorithm. When the relative change of the high-fidelity model parameters is <2% or when 50 high-fidelity model calls are reached, the parameter optimization of the low-fidelity model is stopped.

[0052] For parameter optimization of the medium-fidelity model: Real-time recording of dislocation density in key regions of the high-fidelity model. A dynamic dataset is constructed based on local strain rates; an incremental training mode is adopted based on a Long Short-Term Memory (LSTM) neural network, with the input being... The model outputs strain rate; when the prediction confidence is <90%, high-fidelity calculation is triggered and the training set is expanded; after every 5 iterations, the model is evaluated using an independent test set. If the relative change in RMSE between the predicted value and the high-fidelity simulation value or experimental test value of the sample is ≥2% compared to the previous iteration, the parameters are rolled back to prevent overfitting. N is the number of samples in the independent test set. For the first High-fidelity simulation values ​​or experimental test values ​​for a single sample These are the predicted values ​​obtained by LSTM based on a mid-fidelity model.

[0053] During the parameter optimization process of the medium-fidelity model, by setting confidence conditions, the timing of calling the high-fidelity model is triggered, and the prediction accuracy of the medium-fidelity model is improved with the least amount of high-fidelity model data. By evaluating the model with an independent test set and setting the parameter rollback if the RMSE changes by ≥2% compared to the previous time, the optimal weights of the trained model can be guaranteed, the prediction accuracy of the neural network in key areas can be improved, and global parameter optimization can be achieved.

[0054] Furthermore, the model switching process involves calling different models, each with different computational parameters. To ensure the accurate transfer of these parameters during model switching, it is necessary to retain essential data information. For ease of understanding, this implementation, based on the established multi-fidelity model library, provides the following explanation: A. For low-fidelity models Input: Initial boundary conditions Material parameters (volume fraction of precipitate phase) Calculation: Global creep rate field calculation based on Norton's power law Output: High gradient regions are marked to facilitate the identification and calculation of high gradient regions by the medium-fidelity model. Data transmission: Coordinate information is transmitted via JSON file. For medium-fidelity models, the spatial coordinates (x, y, z) help determine the region to be calculated; the time step... This facilitates time synchronization between medium-fidelity and low-fidelity models.

[0055] B. For medium-fidelity models Input: Coordinates of the high gradient region from low-fidelity output, current stress field, and temperature field. Calculation: The pre-trained Long Short-Term Memory (LSTM) neural network in the mid-fidelity model is used to predict the local creep rate correction term, and the creep rate is calculated. Using the LSTM network, the higher-precision data from the high-fidelity model and pre-collected experimental test data can be used to train the weights and fine-tune the parameters of the mid-fidelity model, improving its simulation efficiency. If the prediction confidence of the mid-fidelity model is <90%, the high-fidelity model calculation is triggered.

[0056] Output: Corrected creep rate Data transfer: Initial values ​​of dislocation density are transferred via HDF5 file. High-fidelity model C. Calculation of the high-fidelity model Input: Initial dislocation density values ​​provided by Medium Fidelity Local microstructure (obtainable through molecular dynamics simulations or electron backscatter diffraction (EBSD)). Calculation: The dislocation density evolution and microstructure evolution characteristics were solved using the crystal plasticity finite element method (CPFEM). Output: Damage variable D and updated dislocation density

[0057] Value passing: returned via PyTorch tensors To the medium fidelity model D. Data Update and Iteration Mid-fidelity model update: The LSTM network is retrained using dislocation density ρ data obtained from the high-fidelity model, which has higher accuracy, to update and improve the accuracy of the mid-fidelity model and reduce potential prediction blind spots. Low-fidelity model update: The corrected creep strain rate is mapped back to the global domain to update and improve the accuracy of the low-fidelity model and supplement any possible prediction blind spots.

[0058] The method established in this embodiment requires only a high-fidelity model for the blade root tenon groove area (approximately 5% of the total model size), a medium-fidelity model for the middle blade area (approximately 45% of the total model size), and a low-fidelity model for the non-load-bearing blade tip area (approximately 50% of the total model size). The number of nodes is... Taking aero-engine blades as an example, the calculation efficiency of different methods for the same model is shown in Table 1.

[0059] Table 1. Comparison of technical efficiency of different methods

[0060] Figure 3 This is a creep-time curve of the CMSX-4 alloy at different temperatures and under different stress conditions at 950°C, representing different temperatures in this embodiment of the invention. The entire model calculation took approximately 4.3 hours, with a deviation of <5% from the experimental data.

[0061] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the processes or functions described in the embodiments of this application are performed entirely or partially. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user equipment, or other programmable device. The computer program or instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; it can also be an optical medium, such as a digital video optical disc (DVD); or it can be a semiconductor medium, such as a solid-state drive (SSD).

[0062] The above embodiments are merely illustrative of the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application. In particular, as long as there is no structural conflict, the various technical features mentioned in the various embodiments can be combined in any way. This application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A dynamic multi-fidelity processing method for predicting creep curves of high-temperature alloys, characterized in that, Includes the following steps: Step 1: Construct a multi-fidelity model library for the creep properties of high-temperature alloys. The multi-fidelity model library includes high-fidelity models, medium-fidelity models, and low-fidelity models. The high-fidelity models adopt a multi-scale modeling strategy, which can predict the creep deformation behavior and microstructure evolution characteristics of high-temperature alloys with different crystal orientations under different temperatures and loads. The medium-fidelity models adopt surrogate models or machine learning models, which can describe the coupled change process between elasticity and creep of high-temperature alloys under a wide range of stress, temperature, and time scales, and can achieve the ability to predict complex, nonlinear, and multi-factor coupled creep behavior. The low-fidelity model uses an empirical formula to describe the deformation behavior of materials in the steady-state creep stage, predicts the creep strain rate and creep strain field distribution of high-temperature alloys under different working conditions, and can quickly estimate the deformation of high-temperature alloys under long-term loads. Step 2: First, use a low-fidelity model to distinguish whether the region where the high-temperature alloy is located is a critical region based on the operating parameters of the high-temperature alloy. These operating parameters include the temperature conditions and stress level of the high-temperature alloy, and the strain gradient of the high-temperature alloy. When the creep strain rate is >3%, the high-temperature alloy is in a critical region. A medium-fidelity model is used to calculate the creep strain rate of the high-temperature alloy at the initial moment; when the strain gradient of the high-temperature alloy... When the creep strain is ≤3%, the high-temperature alloy is in a non-critical region. A low-fidelity model is used to calculate the creep strain rate of the high-temperature alloy at the initial moment. Step 3: Based on the creep strain rate of the high-temperature alloy, select a suitable model from the multi-fidelity model library for creep strain rate simulation iteration; the model selection criterion is: when the creep strain rate of the high-temperature alloy... When the high-temperature alloy is in the accelerated creep stage or critical region, a high-fidelity model is used to recalculate; when Creep strain rate of high-temperature alloys At that time, the high-temperature alloy was in the steady-state creep stage, and the calculation was re-performed using a medium-fidelity model; when the creep strain rate of the high-temperature alloy... At that time, the high-temperature alloy was in the elastic stage or the initial creep stage, and a low-fidelity model was used for calculation. Step 4: As time iterates, based on the creep strain rate at each moment and according to the model selection criteria in Step 3, the model in the multi-fidelity model library is dynamically called in real time to obtain the creep strain change curve of the high-temperature alloy over time.

2. The dynamic multi-fidelity processing method for predicting creep curves of high-temperature alloys as described in claim 1, characterized in that, In step 1, the high-fidelity model can be the full-order coupled crystal plastic finite element model (CPFEM), the medium-fidelity model can be the Chaboche model based on the unified creep-plastic constitutive model or the Gaussian process regression (GP) model, and the low-fidelity model can be the Norton power-law creep equation.

3. The dynamic multi-fidelity processing method for predicting creep curves of high-temperature alloys as described in claim 1 or 2, characterized in that, In step 3, a dynamic weighted solution and optimization method is adopted. First, based on the working conditions and regional characteristics of the high-temperature alloy, and the initial weights of each high-fidelity model, the predicted values ​​of each model are calculated. Then, the predicted values ​​of each model are compared with the simulation results of the high-fidelity model or the actual results of the experimental data to calculate the prediction error of each high-fidelity model. Next, based on the Bayesian adaptive weighting algorithm, combined with the error likelihood probability and the initial weights, the posterior weights of each model are calculated to achieve dynamic adjustment of the weights of each high-fidelity model. The predicted values ​​of each model are multiplied by the posterior weights to obtain the optimal creep strain value after fusion. If the posterior weight of a certain model is continuously lower than the threshold and simultaneously meets the creep strain rate condition in the model selection, model weight adjustment is triggered. The model weight adjustment method is as follows:

1. If the weights W of the low-fidelity model LF <0.1, and Creep strain rate of high-temperature alloys At this time, the calculation of the low-fidelity model is turned off, and the weight of the medium-fidelity model is increased; 2. If the weights W of the medium-fidelity model MF <0.1, and the creep strain rate of the high-temperature alloy At that time, increase the weight of the high-fidelity model.

4. The dynamic multi-fidelity processing method for predicting creep curves of high-temperature alloys as described in claim 3, characterized in that, The initial weights assigned to each fidelity model are as follows:

1. When the high-temperature alloy is in the elastic stage or the initial creep stage, and the high-temperature alloy is in a non-critical region, the initial weight W of the low-fidelity model is... LF（0） =0.7, the initial weights W of the medium-fidelity model MF（0） =0.3; 2. When the high-temperature alloy is in the steady-state creep stage or critical region, the initial weight W of the low-fidelity model LF（0） =0.2, the initial weights W of the medium-fidelity model MF（0） =0.7, the initial weights W of the high-fidelity model HF（0） =0.1; 3. When the high-temperature alloy is in the accelerated creep stage and critical region, the initial weight W of the medium-fidelity model MF（0） =0.2, the initial weights W of the high-fidelity model HF（0） =0.

8.

5. The dynamic multi-fidelity processing method for predicting creep curves of high-temperature alloys as described in claim 3, characterized in that, In the process of dynamic weighted summation, error analysis can be used to optimize the calculation of posterior weights, and active learning can be used to optimize the model parameters of low-fidelity and medium-fidelity models.

6. The dynamic multi-fidelity processing method for predicting creep curves of high-temperature alloys as described in claim 5, characterized in that, The method of optimizing the calculation of posterior weights using error analysis involves the following steps: A. Real-time monitoring of the local normalized error and global error of the fused creep strain, wherein the local normalized error... ,in Representing the units in the model High-temperature alloys in Local normalization error at time t, Representing the units in the model High-temperature alloys in At time i, the creep strain is predicted using the i-model; This refers to units retrieved from real experimental data. High-temperature alloys in The actual creep strain value at time t; If no real test values ​​are available, a high-fidelity model prediction unit is used. High-temperature alloys in Creep strain at a given moment; This represents the maximum absolute error of prediction allowed by the predefined model i; The global error includes the global average error. and global maximum error Global average error Global maximum error , where N represents the total number of high-temperature alloy units in the model; B. Based on the error calculation results and the magnitude of the creep strain rate, formulate different model adjustment strategies: When local normalization error If the creep strain rate meets the selection criteria of the current model, the current model weights are maintained, the model is used again, and the simulation proceeds to the next time step. When local normalization error If the creep strain rate does not meet the selection criteria of the current model, the model accuracy is insufficient. The current model is closed, and a higher-fidelity model is called for prediction. When the global average error When the overall structural accuracy is not up to standard, the weights of each model are adjusted based on the Bayesian algorithm, and the creep strain after fusion is recalculated. When the global maximum error When this occurs, it indicates that there is a prediction blind zone in the current region or working condition. A high-fidelity model is then used to predict the creep strain in that region or working condition. The creep strain data from the prediction blind zone is then fed into a Bayesian algorithm to recalibrate the error standard deviation of the likelihood probability. Optimize the sensitivity of weight updates.

7. The dynamic multi-fidelity processing method for predicting creep curves of high-temperature alloys as described in claim 6, characterized in that, In the process of optimizing the calculation of posterior weights using error analysis, a smoothing method is used to process the prediction error of the model. Specifically, the smoothed error is used... The error calculation results are judged, among which... Indicates the current At this moment, for the units in the model Smoothing error after smoothing high-temperature alloys. Indicates the previous moment At this moment, for the units in the model Smoothing error after smoothing high-temperature alloys. Indicates the current At this moment, for the units in the model The original error of high-temperature alloys It is the smoothing coefficient, and .

8. The dynamic multi-fidelity processing method for predicting creep curves of high-temperature alloys as described in claim 5, characterized in that, An active learning strategy is employed to improve the prediction accuracy of low-fidelity and medium-fidelity models. The specific optimization process is as follows: For low-fidelity model parameter optimization: a Gaussian regression surrogate model is used to construct the working conditions. , The mapping relationship between creep rate error and the high-fidelity model simulation results are used to select the 10 sample points with the largest variance in each round. The Norton power law coefficient A, stress exponent n, and activation energy Q in the low-fidelity model parameters are adjusted by Bayesian optimization algorithm. When the relative change of the high-fidelity model parameters is less than 2% or when there are 50 calls to the high-fidelity model, stop the parameter optimization of the low-fidelity model. For parameter optimization of the medium-fidelity model: Real-time recording of dislocation density in key regions of the high-fidelity model. A dynamic dataset is constructed based on local strain rates; an incremental training mode is adopted based on a Long Short-Term Memory (LSTM) neural network, with the input being... The model outputs creep strain rate; when the prediction confidence is <90%, high-fidelity calculation is triggered and the training set is expanded; after every 5 iterations, the model is evaluated using an independent test set. If the relative change in RMSE between the predicted value and the high-fidelity simulation value or experimental test value of the sample is ≥2% compared to the previous iteration, the parameters are rolled back to prevent overfitting. N is the number of samples in the independent test set. For the first High-fidelity simulation values ​​or experimental test values ​​for a single sample These are the predicted values ​​obtained by LSTM based on a mid-fidelity model.

9. A dynamic multifidelity device for predicting the creep curve of high-temperature alloys, characterized in that, The dynamic multi-fidelity processing method for predicting the creep curve of high-temperature alloys, as described in any one of claims 1-8, can be used to perform data storage, model calculation, and output: The test data storage and retrieval module is used to store and retrieve test and simulation data of the creep performance of high-temperature alloys under various operating conditions; A multi-fidelity model library is used to store and run low-fidelity, medium-fidelity, and high-fidelity models, and output simulation results such as creep strain rate, creep strain, and dislocation density based on the simulation results of each model. The creep strain output module dynamically calls models from the multi-fidelity model library in real time based on the creep strain rate at each moment to obtain the creep strain curve of the high-temperature alloy over time.

10. The dynamic multifidelity device for predicting the creep curve of a high-temperature alloy as described in claim 9, characterized in that, It also includes a dynamic weighted solution and optimization module, which can dynamically call the prediction results of each high-fidelity model according to the model calling and switching criteria, and compare the prediction results of each high-fidelity model with the simulation results of the high-fidelity model or the actual results of the experimental data. Based on the comparison results, according to its own stored dynamic weighted solution and optimization strategy, it performs weighted fusion on the simulation results calculated by each model to obtain the optimal creep strain value after fusion, and feeds it back to the creep strain output module.

Images