Alloy hot deformation prediction system and method based on pre-trained constitutive network model
The alloy hot deformation prediction system based on a pre-trained constitutive network model solves the problems of low prediction accuracy and poor cross-alloy migration ability, and achieves accurate prediction of alloy stress response and process guidance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2026-02-10
- Publication Date
- 2026-07-14
AI Technical Summary
Existing alloy deformation prediction methods are not accurate, cannot be transferred across alloys, have poor extrapolation ability, and cannot predict stress responses outside the training range.
The alloy hot deformation prediction system based on a pre-trained constitutive network model includes modules for data generation, pre-training, experimental data fine-tuning, deformation prediction, and processing condition design. It utilizes an autoregressive encoding and decoding architecture and the FiLM module to learn the hot deformation law of materials and combines it with a small amount of experimental data for fine-tuning, thereby achieving accurate, generalizable, and extrapolable predictions.
It enables accurate prediction of any alloy at different deformation temperatures and strain rates, allows for extrapolation prediction of stress response, and guides alloy design and processing technology design.
Smart Images
Figure CN121683550B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of alloy hot deformation prediction, and in particular to an alloy hot deformation prediction system and method based on a pre-trained constitutive network model. Background Technology
[0002] Traditional constitutive models (Johnson–Cook, Arrhenius, etc.), while concise in formula, rely heavily on assumptions, require extensive experimental data for specific parameter fitting, and exhibit low accuracy in predicting the hot deformation behavior of alloys, making them difficult to generalize across materials. Deep learning methods (CNN, RNN, LSTM, Transformer, etc.) offer superior prediction accuracy, but depend on large amounts of high-quality experimental data, cannot transfer knowledge across alloys, require separate training for each alloy type, have poor extrapolation capabilities, and cannot predict stress responses to strain, strain rate, or temperature outside the training range. Summary of the Invention
[0003] In view of the aforementioned shortcomings of existing technologies, the technical problem to be solved by this invention is that existing alloy deformation prediction methods have low accuracy, existing constitutive models or deep learning methods cannot transfer across alloys, have poor extrapolation capabilities, and cannot predict stress responses outside the range of training processing conditions. This invention provides an alloy hot deformation prediction system and method based on a pre-trained constitutive network model. It trains the basic model with a large amount of synthetic stress-strain data, enabling it to automatically learn the constitutive priors of material hot deformation, and then fine-tunes it with a small amount of experimental data to achieve accurate, generalizable, and extrapolable prediction performance.
[0004] To achieve the above objectives, this invention provides an alloy hot deformation prediction system based on a pre-trained constitutive network model, comprising the following modules:
[0005] The module includes a data generation module, a pre-training module, an experimental data fine-tuning module, a deformation prediction module, and a processing condition design module.
[0006] The data generation module is used to generate large-scale stress-strain curve data based on the Sellars model and its improved models.
[0007] The pre-training module uses an autoregressive encoder-decoder architecture to train based on large-scale stress-strain curve data to learn the thermal deformation law of materials;
[0008] The experimental data fine-tuning module uses a small number of experimental curves to perform physically guided fine-tuning of the model.
[0009] The deformation prediction module is used to predict the true stress-strain curve of an alloy without knowing the temperature, strain rate, or strain range.
[0010] The processing condition design module uses an evolutionary algorithm to inversely deduce the deformation temperature and strain rate from the target stress curve.
[0011] Furthermore, the data generation module iterates through a temperature range of 373-1473K and a strain rate of 0.0001-10 s. -1 And the material parameters in the SCS model Sampling calculations generate millions of training samples, among which For material fitting parameters, The stress sensitivity index, To activate energy, This is the material structure index factor.
[0012] Furthermore, the pre-training module employs an autoregressive GRU encoder and decoder, and uses the FiLM module to conditionally modulate the latent variables to achieve a global representation of the effects of temperature and strain rate.
[0013] Furthermore, the experimental data fine-tuning module only updates the FiLM and decoder layers of the pre-trained model, freezing the encoder body to maintain the constitutive prior and adapt to the experimental data of the target material.
[0014] Furthermore, the deformation prediction module is set to perform stress extrapolation prediction within the unseen temperature range and the unseen strain range, while maintaining the true hardening-softening behavior and recrystallization characteristics.
[0015] Furthermore, the processing condition design module searches for temperature and strain rate based on the particle swarm optimization algorithm to achieve optimal matching between the predicted stress curve and the target stress curve in the sense of mean square error.
[0016] In a preferred embodiment of the present invention, an alloy hot deformation prediction method based on a pre-trained constitutive network model is provided, and an alloy hot deformation prediction system based on a pre-trained constitutive network model is provided, comprising the following steps:
[0017] Based on the Sellars model and with the addition of a recrystallization softening term, the hardening and soft behavior of different alloys were simulated to generate stress-strain curve data;
[0018] The generated stress-strain curve data were interpolated to 20 strain points, and then the stress data, deformation temperature and strain rate were Z-score normalized.
[0019] An autoregressive encoder-decoder network model trained on large-scale stress-strain data is used to obtain a deformation pre-trained model.
[0020] For experimental data of a specific alloy, freeze the main encoder part of the pre-trained model, update only the FiLM module and decoder, and fine-tune the model to enable the model to quickly adapt to the target alloy.
[0021] The fine-tuned model is used to predict stress-strain curves for a specific alloy at arbitrary processing temperatures and strain rates.
[0022] By combining the fine-tuned model with the particle swarm optimization algorithm, the processing conditions can be solved in reverse according to the required stress-strain curve, thus realizing the intelligent design of the thermal deformation path.
[0023] Furthermore, based on the Sellars model and by incorporating a recrystallization softening term, the hardening and softening behaviors of different alloys were simulated. Specifically, based on the Sellars model...
[0024]
[0025] And add recrystallization softening agent
[0026]
[0027] Simulate the hardening and softening behavior of different alloys; among which Indicates strain, stress, and strain rate; Here, T is the gas constant, and T is the temperature. Represents material constants. It is the softening index. It is the softening rate factor.
[0028] Furthermore, regarding the parameters in the constitutive formula Sampling was performed, with sampling ranges of (5e) -3 ~1e -1 ),(1~12),(1e 4 ~2e 5 ),(1e 1 ~1e 20 After sampling, a fifth-order polynomial was used for fitting to avoid oscillations and underfitting; simultaneously, the actual hot working range was traversed, with a temperature range of 373~1473 K, a step size of 1 K, and a strain rate of 0.001~10 s. -1 There are 36 discrete points in total. Each condition is sampled 20 times. Finally, outlier curves are removed, resulting in approximately 1.2 million curves.
[0029] Furthermore, the generated stress-strain curve data is interpolated to 20 strain points, and then the stress data, deformation temperature, and strain rate are Z-score normalized, as shown in the following formula.
[0030]
[0031] Where X represents the stress, deformation temperature, and strain rate matrix. and Let represent the mean and variance of stress, deformation temperature, and strain rate, respectively. This represents the standardized data.
[0032] Compared with existing inventions, the present invention has the following advantages:
[0033] 1. Pre-trained constitutive network models can accurately predict the stress response of any alloy at different deformation temperatures and strain rates.
[0034] 2. Pre-trained constitutive network models can perform stress extrapolation prediction of deformation temperature, strain rate, and strain;
[0035] 3. Pre-trained constitutive network models can visualize the deformation mechanism of alloys and guide alloy design;
[0036] The pre-trained constitutive network model combined with the evolutionary algorithm can infer the alloy deformation temperature and strain rate from the target stress curve, thus guiding the design of the processing technology.
[0037] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description
[0038] Figure 1 This is a diagram of the DeformPTM network architecture according to a preferred embodiment of the present invention;
[0039] Figure 2 This is a schematic diagram illustrating the predicted performance of DeformPTM in BCC steel, FCC high-temperature alloys, and HCP magnesium alloys according to a preferred embodiment of the present invention.
[0040] Figure 3 This is a schematic diagram of process extrapolation and strain extrapolation prediction of HCP magnesium alloy by DeformPTM according to a preferred embodiment of the present invention.
[0041] Figure 4 This is a schematic diagram comparing the predictions of DeformPTM and traditional constitutive models, ANN, and SVM models for different alloys according to a preferred embodiment of the present invention.
[0042] Figure 5 This is a schematic diagram of the DeformPTM guiding process design, which is a preferred embodiment of the present invention. Detailed Implementation
[0043] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0044] In the following description, specific details, such as particular internal procedures and techniques, are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will appreciate that the invention may be practiced in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of the invention with unnecessary detail.
[0045] This invention provides an alloy hot deformation prediction system based on a pre-trained constitutive network model, such as... Figure 1 As shown, it includes the following modules:
[0046] The module includes a data generation module, a pre-training module, an experimental data fine-tuning module, a deformation prediction module, and a processing condition design module.
[0047] The data generation module is used to generate large-scale stress-strain curve data based on the Sellar model and its improved models;
[0048] The decoding architecture is trained based on large-scale stress-strain curve data to learn the thermal deformation law of materials;
[0049] The experimental data fine-tuning module uses a small number of experimental curves to perform physically guided fine-tuning of the model.
[0050] The deformation prediction module is used to predict the true stress-strain curve of an alloy without knowing the temperature, strain rate, or strain range.
[0051] The processing condition design module uses an evolutionary algorithm to inversely deduce the deformation temperature and strain rate from the target stress curve.
[0052] Furthermore, the data generation module iterates through a temperature range of 373-1473K and a strain rate of 0.0001-10s. -1 And the material parameters in the SCS model Sampling calculations generate millions of training samples, among which For material fitting parameters, The stress sensitivity index, To activate energy, This is the material structure index factor.
[0053] Furthermore, the pre-training module employs an autoregressive GRU encoder and decoder, and uses the FiLM module to conditionally modulate the latent variables to achieve a global representation of the effects of temperature and strain rate.
[0054] Furthermore, the experimental data fine-tuning module only updates the FiLM and decoder layers of the pre-trained model, freezing the encoder body to maintain the constitutive prior and adapt to the experimental data of the target material.
[0055] Furthermore, the deformation prediction module can perform stress extrapolation prediction within the unseen temperature range and the unseen strain range, preserving the true hardening-softening behavior and recrystallization characteristics.
[0056] Furthermore, the processing condition design module searches for deformation temperature and strain rate based on the particle swarm optimization algorithm, so that the predicted stress curve and the target stress curve are optimally matched in the sense of mean square error.
[0057] In a preferred embodiment of the present invention, an alloy hot deformation prediction method based on a pre-trained constitutive network model is provided, and an alloy hot deformation prediction system based on a pre-trained constitutive network model is provided, comprising the following steps:
[0058] Step 1: Based on the Sellars model and with the addition of a recrystallization softening term, simulate the hardening and softening behavior of different alloys to generate stress-strain curve data;
[0059] Step 2: Interpolate the generated stress-strain curve data to 20 strain points, and then perform Z-score normalization on the stress data, deformation temperature and strain rate.
[0060] Step 3: Based on the large-scale stress and strain data, an autoregressive encoder-decoder network model is trained and constructed to obtain the deformation pre-trained model;
[0061] Step 4: For experimental data of a specific alloy, freeze the encoder part of the pre-trained model, update only the FiLM module and decoder, and fine-tune the model to enable the model to quickly adapt to the target alloy.
[0062] Step 5: Use the fine-tuned model to predict the stress-strain curves for a specific alloy at arbitrary processing temperatures and strain rates.
[0063] Step 6: By combining the fine-tuned model with the particle swarm optimization algorithm, the processing conditions can be solved in reverse according to the required stress-strain curve, thereby realizing the intelligent design of the thermal deformation path.
[0064] The above steps will be explained in detail below.
[0065] Step 1: Based on the Sellars model and by adding a recrystallization softening term, simulate the hardening and softening behavior of different alloys to generate stress-strain curve data; specifically, based on the Sellars model...
[0066]
[0067] And add recrystallization softening agent
[0068]
[0069] in Indicates strain, stress, and strain rate; is the gas constant, and T is the temperature. This represents the material constant. It is the softening index. It is the softening rate factor.
[0070] Simulate the hardening and softening behavior of different alloys. Consider the parameters in the constitutive formula. Sampling was performed, with sampling ranges of (5e) -3 ~1e -1 ),(1~12),(1e 4 ~2e 5 ),(1e 1 ~1e 20 ).in For material fitting parameters, The stress sensitivity index, To activate energy, This represents the material structure index factor. A fifth-order polynomial was used for fitting after sampling to avoid oscillations and underfitting. Simultaneously, the actual hot working range was traversed, with a temperature range of 373–1473 K (in 1 K increments) and a strain rate range of 0.001–10 s⁻¹. -1 (A total of 36 discrete points) The formula was used 20 times under different material parameters for each condition. Finally, outlier curves were removed, resulting in approximately 1.2 million curves.
[0071] Step 2: For each of the 120w stress-strain curves obtained from the calculation, use cubic spline interpolation to reach 20 strain points. Then, perform Z-score normalization on the stress data of the curves, the corresponding deformation temperature and strain rate of each curve, as shown in the following formula:
[0072]
[0073] Where X represents the stress, deformation temperature, and strain rate matrix. and Let represent the mean and variance of stress, deformation temperature, and strain rate, respectively. This represents the standardized data.
[0074] Step 3: Normalize the simulated stress and strain and the corresponding deformation temperature and strain rate Data is input into the network model DeformPTM for training: DeformPTM uses an autoregressive encoder-decoder network architecture oriented towards deformation constitutive model to generate complete stress-strain curves point by point.
[0075] This invention employs an autoregressive encoder-decoder network oriented towards deformation constitutive models, which consists of the following components:
[0076] 1) Input data structure
[0077] Each sample contains a stress sequence: Process conditions: temperature T, strain rate Finally, the data is input into the encoder via splicing.
[0078] 2) Encoder
[0079] The encoder employs a GRU autoregressive network model to incorporate "historical stress" Current process conditions "As input, it is mapped to a high-dimensional hidden state z by the encoder, and then z is mapped to two latent variables: (Latent variables in the stress sequence) (Latent variables of process conditions)
[0080] 3) FiLM conditional modulator
[0081] The input is fed into the FiLM (Feature-wise Linear Modulation) network to generate the scaling factor. Translation factor And it acts on the hidden state of the decoder GRU at each strain point. :
[0082]
[0083] The FiLM module can perform feature-level linear modulation by utilizing latent variables from process condition data. The resulting scaling factor Translation factor During network model training, the curve shape at different temperatures and strain rates is learned by controlling the latent variables mapped from the stress data. To achieve control of the entire curve by process conditions.
[0084] 4) Decoder
[0085] The decoder also uses an autoregressive network, with the input being the predicted stress and latent variables from the previous time step. Latent variables output by FiLM The output is the next stress value. A complete stress-strain curve is generated point-by-point. The entire generation process simulates the material's true "loading history dependence".
[0086] 5) The training loss function of the DeformPTM pre-trained model is:
[0087]
[0088] in The term represents the stress reconstruction error. To reconstruct errors based on processing conditions, the model learns both "curve shape" and "physical condition mapping" simultaneously.
[0089] In the process of DeformPTM predicting stress point by strain point, the predicted value of the stress at the current strain point can be obtained by the following formula:
[0090]
[0091] Step 4: After the model has finished training on the simulated stress-strain curve data in Step 3, freeze the main encoder part of the model. When the system is used for a real alloy, only a small number of stress-strain experimental curves of the target alloy need to be input to update the FiLM module and decoder, allowing the model to quickly adapt to the target alloy. Experiments show that only 4-8 experimental curves are needed to obtain high-accuracy predictions. The fine-tuning process follows:
[0092]
[0093] After fine-tuning, DeformPTM's deformation priors are adapted to the deformation behavior of the target alloy, enabling DeformPTM to accurately predict the deformation behavior of the target alloy, such as... Figure 2 As shown, the DeformPTM and SCS models are compared under varying deformation temperatures (523–1423 K) and strain rates (0.001–10 s⁻¹). -1 Under the given conditions, the predicted and experimentally true stress-strain curves for BCC low-alloy steel (S40CVS), FCC nickel-based superalloy (FGH98), and HCP magnesium alloy (AQ80) are presented. Furthermore, the mean square error of the DeformPTM and SCS models for S40CVS, FGH98, and AQ80 is given, grouped by strain rate. Figure 4 The predictive performance of DeformPTM, Sellars constitutive model, artificial neural network (ANN), and support vector machine (SVM) models on steel (S40CVS, MnCrNiCo), magnesium alloys (AQ80, AZ31, ZK61), aluminum alloys (Al6063), titanium alloys (TC11), and nickel-based superalloys (FGH98, FGH4113A) is demonstrated.
[0094] Step 5: Use the fine-tuned model to perform stress-strain curve prediction for arbitrary processing temperatures and strain rates. The fine-tuned model can perform stress-strain curve prediction for arbitrary processing temperatures and strain rates, stress extrapolation prediction for high strain ranges (e.g., 0.4-0.6), and multi-material prediction (steel, magnesium, aluminum, and nickel-based superalloys, etc.). Figure 3 As shown, the DeformPTM and SCS models are compared on HCP AQ80 magnesium alloy at a strain rate of 0.001 s⁻¹. -1 0.01 s -1 0.1 s -1 and 1 s -1 The results show the in-distribution processing predictions under the given conditions and the extrapolation results for strain ranges beyond the training window. DeformPTM maintains physically consistent softening behavior outside the training strain range (extrapolated from 0.40 to 0.60, while training only covers 0.05–0.40), while the SCS model shows significant bias.
[0095] Step 6: Based on the required stress response (stress-strain curve) of the alloy during processing, the processing conditions are deduced to achieve intelligent design of the hot deformation path, such as... Figure 5 This paper demonstrates the reverse process condition design based on DeformPTM and its experimental verification. Two target flow curves are presented: one representing high deformation resistance and limited softening (Target 1), and the other representing low deformation resistance and steady-state flow characteristics (Target 2). By coupling DeformPTM with particle swarm optimization, the curves for Target 1 and Target 2 are reverse-optimized, and the results are verified experimentally. The obtained target stress curve is basically consistent with the experimentally measured stress curve.
[0096] This step is used to inversely solve for the processing conditions (temperature T and strain rate) based on the stress-strain curve of the desired target alloy. This enables intelligent design of thermal deformation paths. The core idea is to use a finely tuned DeformPTM as a surrogate model, and through an optimization algorithm, search the input space to make the generated stress-strain response approximate the expected target curve.
[0097] 1) Construction of target stress-strain curve
[0098] First, construct the target stress-strain trajectory based on actual needs. The trajectory can be manually set according to the desired organizational evolution (such as strong DRX, weak DRX);
[0099] In FGH4113A powder superalloy, this invention selected two typical target curves: a high stress / weak dynamic recrystallization (weak DRX) target curve, characterized by significant work hardening, relatively high peak stress, and a relatively slow decreasing trend; and a low stress / strong dynamic recrystallization (strong DRX) target curve, characterized by lower peak stress, faster softening rate, and lower stable flow stress.
[0100] 2) Definition of the search space for processing parameters
[0101] The processing conditions are represented as a two-dimensional search vector:
[0102]
[0103] 3) The loss function is defined as follows:
[0104]
[0105] The objective of PSO is:
[0106]
[0107] 4) Particle Swarm Optimization Process
[0108] The particle swarm parameters are set as follows:
[0109] Number of particles: 20
[0110] Maximum number of iterations: 30
[0111] Inertia weight:
[0112] Cognitive / social factors:
[0113] The operation process is as follows:
[0114] Population initialization: Multiple sets of temperature and strain rate were randomly sampled as initial values for PSO particles.
[0115] Predicted stress curves: for each group , Input DeformPTM to generate
[0116] Calculate the particle error MSE: compare it with the target curve to obtain the fitness.
[0117] Update example speed and position
[0118] Use the standard PSO update mechanism:
[0119]
[0120]
[0121] Repeat the iteration until convergence.
[0122] The following specific embodiments will illustrate the steps of the alloy hot deformation prediction method based on a pre-trained constitutive network model according to the present invention.
[0123] Example 1: Generation and Processing of Thermal Deformation Data
[0124] 1.1 Generation of Thermal Deformation Data
[0125] In this embodiment, to enable the model to learn the thermal deformation behavior of different alloy materials under high-temperature processing conditions, a basic constitutive stress-strain relationship is first constructed based on the Sellars model:
[0126]
[0127] To simulate the dynamic recrystallization (DRX), work hardening, and softening behavior of actual high-temperature alloys during hot deformation, a softening term is added to the above model:
[0128]
[0129] Among them, critical strain With temperature change, softening coefficient It is related to the recrystallization rate of the material.
[0130] To ensure that the model covers a wide range of alloy systems, the constitutive model parameters are sampled as follows:
[0131]
[0132] After sampling, the parameters are fitted using a fifth-order polynomial to avoid curve oscillation or underfitting, thereby ensuring that the generated stress-strain curve has physical rationality.
[0133] 1.2 Processing Condition Traversal
[0134] Typical hot working conditions were investigated, with the temperature range set at 373-1473 K (in 1 K increments) and the strain rate at 0.001-10 s⁻¹. -1 (A total of 36 discrete points), with each condition sampled 20 times. After removing outlier curves, approximately 1.2 million stress-strain curves were obtained, forming the pre-training dataset for the model.
[0135] 1.3 Data Preprocessing
[0136] All curves were interpolated to 20 uniform strain points; then Z-score normalization was performed.
[0137]
[0138] in and These are the global mean and standard deviation, respectively.
[0139] Example 2: Construction and training of the pre-trained constitutive model DeformPTM
[0140] 2.1 The input data structure includes a stress sequence: Processing conditions, temperature strain rate Input is fed into the encoder via splicing.
[0141] 2.2 The encoder adopts a GRU autoregressive network model, which incorporates "historical stress" Current process conditions "As input, it is mapped to a high-dimensional hidden state z by the encoder, and then z is mapped to two latent variables: (Latent variables in the stress sequence) (Latent variables of process conditions)
[0142] 2.3 FiLM Conditional Modulation Module
[0143] The input is fed into the FiLM (Feature-wise Linear Modulation) network to generate the scaling factor. Translation factor And it acts on the hidden state of the decoder GRU at each strain point. :
[0144]
[0145] The FiLM module can perform feature-level linear modulation by utilizing latent variables from process condition data. The resulting scaling factor Translation factor During network model training, the curve shape at different temperatures and strain rates is learned by controlling the latent variables mapped from the stress data. To achieve control of the entire curve by process conditions.
[0146] 2.4 The decoder also uses an autoregressive network, with the input being the predicted stress and latent variables from the previous time step. Latent variables output by FiLM The output is the next stress value. A complete stress-strain curve is generated point-by-point. The entire generation process simulates the material's true "loading history dependence".
[0147] 2.5 The training loss function of the DeformPTM pre-trained model is
[0148]
[0149] in The term represents the stress reconstruction error. To reconstruct errors based on processing conditions, the model learns both "curve shape" and "physical condition mapping" simultaneously.
[0150] Example 3: Reverse Design Module for Processing Conditions
[0151] In this embodiment, the finely tuned DeformPTM is used as a surrogate model, and combined with the particle swarm optimization (PSO) algorithm, the heat treatment conditions are solved in reverse based on the desired stress-strain curve.
[0152] 3.1 Construction of target stress-strain curve
[0153] Taking FGH4113A powder superalloy as an example, based on actual application requirements, this invention selects two typical objectives:
[0154] This invention selects two typical target curves: a high-stress / weak dynamic recrystallization (weak DRX) target curve, characterized by significant work hardening, relatively high peak stress, and a relatively slow decreasing trend; and a low-stress / strong dynamic recrystallization (strong DRX) target curve, characterized by lower peak stress, faster softening rate, and lower stable flow stress.
[0155] 3.2 Definition of the Search Space for Processing Parameters
[0156] The processing conditions, temperature and strain rate, are represented as a two-dimensional search vector:
[0157]
[0158] The search range for temperature is (1173 K~1423 K), and the search range for strain rate is... .
[0159] 3.3 Definition of Loss Function
[0160] For candidate processing conditions
[0161]
[0162] The model generates a prediction curve, and its error function is:
[0163]
[0164] PSO seeks the processing conditions that minimize this error:
[0165]
[0166] 3.4 Particle Swarm Optimization Process
[0167] The PSO parameters are set to a particle count of 20, a maximum number of iterations of 30, and an inertia weight. Cognitive and social factors
[0168] The iterative update method is as follows:
[0169]
[0170]
[0171] Repeat the above steps until convergence.
[0172] 3.5 Verification of Inversion Results
[0173] Through five independent optimization experiments, the present invention yielded the following stable processing parameters:
[0174] Weak DRX target:
[0175]
[0176] Strong DRX target:
[0177]
[0178] The above conditions were further verified using actual hot compression experiments:
[0179] Weak DRX corresponds to less recrystallization and significant work hardening, consistent with the target curve.
[0180] Strong DRX corresponds to complete recrystallization and rapid softening, which also matches the target curve.
[0181] The above embodiments demonstrate that the reverse design module of the present invention can accurately reconstruct the desired tissue evolution path.
[0182] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. An alloy hot deformation prediction system based on a pre-trained constitutive network model, characterized in that, Includes the following modules: The module includes a data generation module, a pre-training module, an experimental data fine-tuning module, a deformation prediction module, and a processing condition design module. The data generation module is used to generate large-scale stress-strain curve data based on the Sellars model and its improved models. Specifically, it simulates the hardening and softening behavior of different alloys at different deformation temperatures and strain rates using the Sellars model with the addition of a recrystallization softening term, thereby generating stress-strain curve data. ， And add recrystallization softening agent ， Simulate the hardening and softening behavior of different alloys, among which Indicates strain, stress, and strain rate; Here, T is the gas constant, and T is the temperature. Represents material constants. It is the softening index. It is the softening rate factor; The data generation module iterates through a temperature range of 373-1473K and a strain rate of 0.0001-10 s. -1 and the Sellars model Sampling calculations generate millions of training samples, among which For material fitting parameters, The stress sensitivity index, To activate energy, This refers to the material structure index factor, specifically including parameters in the constitutive formula. Sampling was performed, with sampling ranges of (5e) -3 ~1e -1 ),(1~12),(1e 4 ~2e 5 ),(1e 1 ~1e 20 After sampling, a fifth-order polynomial was used for fitting to avoid oscillations and underfitting; simultaneously, the actual hot working range was traversed, with a temperature range of 373~1473 K, a step size of 1 K, and a strain rate of 0.001~10 s⁻¹. -1 There are 36 discrete points in total. Each condition is sampled 20 times. Finally, outlier curves are removed to obtain 1.2 million curves. The pre-training module uses an autoregressive encoding and decoding architecture to train based on large-scale stress-strain curve data to learn the thermal deformation law of the material. The experimental data fine-tuning module uses a small number of experimental curves to perform physically guided fine-tuning of the model. The deformation prediction module is used to predict the true stress-strain curve of the alloy without knowing the temperature, strain rate, or strain range. The processing condition design module uses an evolutionary algorithm to inversely deduce the deformation temperature and strain rate from the target stress curve.
2. The alloy hot deformation prediction system based on a pre-trained constitutive network model as described in claim 1, characterized in that, The pre-training module employs an autoregressive GRU encoder and decoder, and uses the FiLM module to conditionally modulate the latent variables to achieve a global representation of the effects of temperature and strain rate.
3. The alloy hot deformation prediction system based on a pre-trained constitutive network model as described in claim 1, characterized in that, The experimental data fine-tuning module only updates the FiLM and decoder layers, and freezes the encoder body to maintain the constitutive prior and adapt to the experimental data of the target material.
4. The alloy hot deformation prediction system based on a pre-trained constitutive network model as described in claim 1, characterized in that, The deformation prediction module performs stress extrapolation prediction in the absence of temperature, strain rate range and strain range, thus preserving the true hardening-softening behavior and recrystallization characteristics.
5. The alloy hot deformation prediction system based on a pre-trained constitutive network model as described in claim 1, characterized in that, The processing condition design module searches for deformation temperature and strain rate based on the particle swarm optimization algorithm, so that the predicted stress curve and the target stress curve are optimally matched in the sense of mean square error.
6. A method for predicting hot deformation of alloys based on a pre-trained constitutive network model, characterized in that, An alloy hot deformation prediction system based on a pre-trained constitutive network model includes the following steps: Based on the Sellars model with the addition of a recrystallization softening term, the hardening and softening behavior of different alloys under different deformation temperatures and strain rates was simulated, generating stress-strain curve data. Specifically, based on the Sellars model... ， And add recrystallization softening agent ， Simulate the hardening and softening behavior of different alloys, among which Indicates strain, stress, and strain rate; Here, T is the gas constant, and T is the temperature. Represents material constants. softening index It is the softening rate factor; The generated stress-strain curve data were interpolated to 20 strain points, and then the stress data, deformation temperature and strain rate were Z-score normalized. An autoregressive encoder-decoder network model trained on large-scale stress-strain data is used to obtain a deformation pre-trained model. For experimental data of a specific alloy, freeze the main encoder part of the pre-trained model, update only the FiLM module and decoder, and fine-tune the model to enable the model to quickly adapt to the target alloy. Predict stress-strain curves for a specific alloy at arbitrary processing temperatures and strain rates using a finely tuned model. By combining the fine-tuned model with the particle swarm optimization algorithm, the processing conditions are solved in reverse according to the required stress-strain curve, thereby realizing the intelligent design of the thermal deformation path.
7. The alloy hot deformation prediction method based on a pre-trained constitutive network model as described in claim 6, characterized in that, For the parameters in the constitutive formula Sampling was performed, with sampling ranges of (5e) -3 ~1e -1 ),(1~12),(1e 4 ~2e 5 ),(1e 1 ~1e 20 After sampling, a fifth-order polynomial was used for fitting to avoid oscillations and underfitting; simultaneously, the actual hot working range was traversed, with a temperature range of 373~1473 K, a step size of 1 K, and a strain rate of 0.001~10 s⁻¹. -1 There are 36 discrete points in total. Each condition is sampled 20 times. Finally, outlier curves are removed, resulting in 1.2 million curves.
8. The alloy hot deformation prediction method based on a pre-trained constitutive network model as described in claim 6, characterized in that, The generated stress-strain curve data is interpolated to 20 strain points, and then the stress data, deformation temperature, and strain rate are Z-score normalized, as shown in the following formula. ， Where X represents the stress, deformation temperature, and strain rate matrix. and Let represent the mean and variance of stress, deformation temperature, and strain rate, respectively. This represents the standardized data.