High-temperature compression rheological stress prediction method based on multi-branch fusion neural network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XI'AN UNIVERSITY OF ARCHITECTURE AND TECHNOLOGY
- Filing Date
- 2026-02-04
- Publication Date
- 2026-06-05
Smart Images
Figure CN122154406A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of interdisciplinary technology of machine learning and materials design, specifically a method for predicting high-temperature compressive rheological stress based on a multi-branch fusion neural network. Background Technology
[0002] In the hot working and microstructure design of metallic materials such as titanium alloys and high-temperature alloys, accurate prediction of the rheological stress curves of materials under different temperatures, strain rates, and strain conditions is an important prerequisite for achieving process optimization, microstructure control, and product quality prediction. During high-temperature compression, materials exhibit a series of competing mechanisms, including strain hardening, dynamic recovery, dynamic recrystallization, and flow softening. This results in a highly nonlinear and complex mapping relationship between rheological stress and multiple variables such as temperature, strain rate, deformation, and microstructure.
[0003] In existing technologies, the prediction of high-temperature compressive rheological stress is mainly based on traditional physical constitutive models and machine learning methods, but both have insurmountable objective problems. Physical constitutive models (such as Arrhenius-type models, Johnson-Cook models, Zener-Hollomon relations, etc.) have fixed formula forms and a limited number of parameters, making it difficult to fully characterize the complex rheological behavior caused by the combined effects of multiple mechanisms such as work hardening, dynamic recovery, and dynamic recrystallization at high temperatures. Moreover, the model parameters are highly sensitive to experimental data, often requiring separate fitting at different temperature and strain rate ranges, resulting in insufficient prediction stability and generalization ability. In addition, most existing rheological stress prediction methods based on artificial neural networks use shallow or medium-depth network structures, which limit the network's expressive power and easily lead to underfitting when dealing with complex nonlinear changes in high-temperature compression curves. At the same time, the training convergence speed of these network models is slow, and they are highly dependent on the amount and distribution of training data. The generalization ability of the models under different deformation conditions is often insufficient, resulting in deviations in key features of the predicted curve, such as peak position, softening degree, and stress level change trend. In summary, existing technologies still have significant shortcomings in terms of expressive power, prediction accuracy, training stability, and consistency across temperature ranges / strain rates, making it difficult to meet the engineering requirements for rheological stress prediction in high-temperature hot processing scenarios. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a high-temperature compressive rheological stress prediction method based on a multi-branch fusion neural network, which nonlinearly characterizes the effects of strain, strain rate and temperature, and couples the three at a high level to achieve accurate prediction of the corresponding stress.
[0005] The technical solution adopted by this invention to solve its technical problem is a high-temperature compressive rheological stress prediction method based on a multi-branch fusion neural network, which includes the following steps: S1: Obtain raw data from high-temperature alloys in high-temperature compression tests, including strain. Absolute temperature T and strain rate A graph showing the relationship between stress σ and stress σ. S2: Preprocess the raw data to obtain the model input feature vector; the preprocessing includes at least performing Min-Max normalization on the temperature T to obtain the normalized temperature. Taking the natural logarithm of the strain rate yields the logarithmic strain rate. ,strain Keep the original value; S3: Construct a multi-branch fusion neural network, which consists of three independent branch subnetworks and a fusion network; S4: Strain Logarithmic strain rate and normalized temperature The inputs are fed into the corresponding sub-networks, and each sub-network is used to learn the nonlinear influence characteristics of the force on the corresponding physical quantity; among them, strain... The branch subnetwork learns the evolution characteristics of work hardening or softening with strain, logarithmic strain rate. Branch subnetworks learn strain rate sensitivity features and normalize temperature. Branch subnetworks learn temperature softening characteristics; S5: Concatenate the output features of the three branch sub-networks into a three-dimensional fusion feature vector; S6: Input the fused feature vector into the fused network, which is used to model the coupling relationship between strain, logarithmic strain rate, and normalized temperature, and output the original rheological stress prediction value. ; S7: Apply a non-negative constraint to the original rheological stress prediction value output by the fusion network to obtain the final predicted true stress value. .
[0006] Furthermore, in step S4, each of the branch sub-networks is a three-layer feedforward neural network structure, including: First layer: Input x∈R(ε, , Mapped to 64-dimensional hidden feature h1,
[0007] The second layer: Maps h1 to the 32-dimensional hidden feature h2. ; Third layer: Map h2 to the output scalar s. ; in, The first activation function is... To correct the linear unit , , , These are the linear transformation weights of the first, second, and third layers in a single-branch subnetwork, respectively. , , These represent the biases of the first, second, and third layers in a single-branch subnetwork, respectively. , These are the first and second layer implicit features, respectively; This is the output of the subnetwork.
[0008] Furthermore, in S5, specific measures are taken for strain. Logarithmic strain rate and normalized temperature Three sets of parameter-independent branch subnetworks are constructed from three physical inputs to obtain:
[0009] in For strain hardening / softening Evolutionary characteristics; It is a strain rate sensitivity characteristic; For temperature softening features; for the branch feature fusion network, the output features of the three branches are concatenated into a fused feature vector:
[0010] Furthermore, in step S6, the fusion network is a two-layer feedforward neural network structure, including: First fusion layer: Maps the fused feature vector Y∈Z to a 64-dimensional hidden feature u1. ; Second fusion layer: Maps u1 to 32-dimensional hidden feature u2 ; Output layer: Maps u2 to the original rheological stress prediction value. , ; in, The second activation function, the second activation function To correct the linear unit , , , These are the linear transformation weights of the first, second, and third layers in the branch feature fusion network, respectively. , , These represent the biases of the first, second, and third layers in the branch feature fusion network, respectively. , These are the first and second layer hidden features, respectively. This is the output of the fusion network.
[0011] Furthermore, in step S7, the application of the nonnegativity constraint is implemented using the Softplus function, i.e. .
[0012] The beneficial effects of this invention are as follows: By constructing three independent deep branch sub-networks, the nonlinear influence patterns of three key physical variables—strain, strain rate, and temperature—on rheological stress are specifically learned, and the complex coupling relationships among these three are explicitly modeled in a high-level fusion network. This architecture enables the model to completely and accurately characterize the entire rheological behavior from work hardening and peak stress to dynamic recovery and dynamic recrystallization softening. Each branch of this invention is dedicated to learning a specific physical mechanism, effectively separating the change patterns of different mechanisms and avoiding cross-interference between features; the fusion layer corresponds to the real physical process of competition and coupling among multiple mechanisms. This design allows the model to not only achieve high-precision predictions but also learn constitutive relations from data that conform to the laws of materials physics. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of neural network training according to the present invention; Figure 2 It is a schematic diagram of the stress-strain curve; Figure 3 It is a schematic diagram of the stress-strain curve; Figure 4 This is a comparison chart of the stress prediction curve after model training and the stress curve of the test set. Detailed Implementation
[0014] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0015] like Figure 1 As shown, the high-temperature compressive rheological stress prediction method based on a multi-branch fusion neural network of the present invention includes the following steps: S1: Obtain raw data from high-temperature alloys in high-temperature compression tests, including strain. Absolute temperature T and strain rate The relationship between strain and stress σ is plotted; where the raw data refers to the stress-strain curves directly measured through high-temperature compression experiments, including strain (ε), stress (σ), absolute temperature (T), and strain rate. The sequence data, with hundreds of data points per curve, accurately depicts the continuous rheological behavior of materials, providing rich and high-quality samples for training deep learning models.
[0016] S2: Preprocess the raw data to obtain the model input feature vector; the preprocessing includes at least performing Min-Max normalization on the temperature T to obtain the normalized temperature. Taking the natural logarithm of the strain rate yields the logarithmic strain rate. ,strain Keep the original values; among them, the temperature is normalized using the Min–Max method, and the normalization formula is:
[0017] in The normalized temperature. For temperature, This is the minimum temperature. The maximum temperature is represented by . Normalization is used to linearly map the temperature to the interval [0,1] to preserve the relative variation trend between different experimental conditions.
[0018] strain rate To avoid training instability due to excessively large scales when the scale exceeds a certain order of magnitude, the natural logarithm of the strain rate is taken, as shown in the following formula:
[0019] in For strain rate, This is the logarithmic strain rate after taking the natural logarithm. Strain It is the most important physical variable describing the deformation state of a material; its absolute value has a clear physical meaning, and its influence remains consistent under different temperatures and strain rates. Therefore, this invention selects... Instead of normalizing the values, we use their original values as input to the neural network.
[0020] S3: Construct a multi-branch fusion neural network, which consists of three independent branch subnetworks and a fusion network; S4: Strain Logarithmic strain rate and normalized temperature The inputs are fed into the corresponding sub-networks, and each sub-network is used to learn the nonlinear influence characteristics of the force on the corresponding physical quantity; among them, strain... The branch subnetwork learns the evolution characteristics of work hardening or softening with strain, logarithmic strain rate. Branch subnetworks learn strain rate sensitivity features and normalize temperature. The branch subnetwork learns temperature softening characteristics; each branch subnetwork is a three-layer feedforward neural network structure, including: First layer: Input x∈R(ε, , Mapped to 64-dimensional hidden feature h1,
[0021] The second layer: Maps h1 to the 32-dimensional hidden feature h2. ; Third layer: Map h2 to the output scalar s. ; in, The first activation function is... To correct the linear unit , , , These are the linear transformation weights of the first, second, and third layers in a single-branch subnetwork, respectively. , , These represent the biases of the first, second, and third layers in a single-branch subnetwork, respectively. , These are the first and second layer implicit features, respectively; This is the output of the sub-network. Each branch is a small deep network responsible for extracting high-level, abstract features from a single physical input. Its three-layer structure (64→32→1) is designed to progressively refine information: the first layer performs basic feature transformation and expansion; the second layer performs feature compression and integration; and the third layer outputs a scalar feature(s) that represents the impact of the physical mechanism. Modified linear units are used as activation functions to introduce non-linearity and alleviate the vanishing gradient problem in deep networks, ensuring the network can learn complex mappings.
[0022] S5: Concatenate the output features of the three branch subnetworks into a three-dimensional fused feature vector; respectively targeting strain Logarithmic strain rate and normalized temperature Three sets of parameter-independent branch subnetworks are constructed from three physical inputs to obtain:
[0023] in For strain hardening / softening Evolutionary characteristics; It is a strain rate sensitivity characteristic; For temperature softening features; for the branch feature fusion network, the output features of the three branches are concatenated into a fused feature vector:
[0024] Scalars output from the three branches , , The features are concatenated along their dimensions to form a new 3D vector Z. This operation is equivalent to placing the three learned independent physical mechanism representations in the same space, preparing for subsequent coupled modeling.
[0025] S6: Input the fused feature vector into the fused network, which is used to model the coupling relationship between strain, logarithmic strain rate, and normalized temperature, and output the original rheological stress prediction value. ; The fusion network is a two-layer feedforward neural network structure, including: First fusion layer: Maps the fused feature vector Y∈Z to a 64-dimensional hidden feature u1. ; Second fusion layer: Maps u1 to 32-dimensional hidden feature u2 ; Output layer: Maps u2 to the original rheological stress prediction value. , ; in, The second activation function, the second activation function To correct the linear unit , , , These are the linear transformation weights of the first, second, and third layers in the branch feature fusion network, respectively. , , These represent the biases of the first, second, and third layers in the branch feature fusion network, respectively. , These are the first and second layer hidden features, respectively. This is the output of the fusion network.
[0026] This network receives the concatenated feature vector Z and its function is to learn the interaction between the three mechanism features. Its two-layer structure (64→32) is designed similarly to the branch sub-network, aiming to further nonlinearly transform and fuse the coupled features. Finally, a raw stress prediction value is generated through a linear output layer without an activation function. .
[0027] S7: Apply a non-negative constraint to the original rheological stress prediction value output by the fusion network to obtain the final predicted true stress value. The application of nonnegativity constraints is achieved through the Softplus function, i.e. . It is the direct mathematical output of the model and can theoretically be any real number. However, rheological stress must physically be non-negative. The Softplus function is used to... Mapping to a suitable physical space yields the final predicted value. .
[0028] Example
[0029] This embodiment uses a certain type of nickel-based superalloy as an example.
[0030] Data preparation and preprocessing
[0031] First, see Figure 2 and Figure 3 Eight complete true stress-true strain curves of the alloy were obtained from high-temperature compression tests, denoted as curves 1 to 8. Each curve corresponds to a specific set of temperature and strain rate conditions, covering typical rheological behaviors from work hardening, dynamic recovery to dynamic recrystallization. The original data points include strain. Absolute temperature T and strain rate With stress .
[0032] Next, all data undergoes uniform preprocessing to generate the model input features: The temperature (T) is normalized using the Min-Max method to obtain the normalized temperature. The formula is in and These are the minimum and maximum temperatures in the dataset.
[0033] strain rate Taking the natural logarithm, we obtain the logarithmic strain rate. To smooth out its numerical distribution.
[0034] strain Maintaining the original physical values as input preserves the continuous physical meaning and evolutionary patterns, which is beneficial for the model to learn the inherent hardening / softening rules and enhance generalization ability. After preprocessing, the feature vector of each data sample is... , , The corresponding monitoring target is the stress measured at that point. .
[0035] Constructing a multi-branch fusion neural network model
[0036] Construct three independent branch subnetworks, each receiving strain. Logarithmic strain rate, and normalized temperature As input, each branch is designed as a three-layer feedforward structure. Through its internal weights and bias parameters, and using a modified linear unit activation function for nonlinear transformation, each branch ultimately outputs a scalar feature that characterizes the influence of the corresponding physical mechanism. , , This design allows each branch to specialize and deeply learn the complex nonlinear mapping relationship between a single physical variable and stress.
[0037] A fusion network is constructed to concatenate the output scalar features of the three branches into a three-dimensional comprehensive feature vector. This vector is then input into a two-layer feedforward fusion network. The role of the fusion network is to learn and model the interaction and coupling effects among strain, strain rate, and temperature features, ultimately outputting a raw stress prediction value. .
[0038] Apply physical constraints and modify the original values output by the fusion network using the Softplus function. Processing is performed to ensure the final predicted rheological stress value. It is always positive, which is consistent with physical facts.
[0039] Model training
[0040] Based on the typical training pattern for verifying the generalization ability of the model, the dataset was divided into a training set (curves 1, 2, 3, 4, 6, 7, 8) and an independent test set (curve 5). The constructed multi-branch fusion neural network was trained using the training set data. During training, the backpropagation algorithm was used to optimize all parameters in the network with the goal of minimizing the mean square error between the predicted stress and the actual stress.
[0041] Model Application and Advantages
[0042] Once trained, the model can be used for prediction. When it is necessary to predict the rheological stress of the alloy under a new set of hot working conditions, simply preprocess these conditions, then input the resulting feature vectors into the trained model, and the model will automatically output a complete stress prediction curve from the initial deformation to the specified strain.
[0043] In this process, the technical advantages of the present invention are specifically demonstrated, see [link to relevant documentation]. Figure 4 Even when tested using curve 5, an independent condition that was not involved in the training, the model was still able to predict the complete rheological curve with high accuracy, including the accurate initial hardening rate, peak stress, and subsequent dynamic recrystallization softening trend. This fully demonstrates that the multi-branch fusion structure proposed in this invention can effectively learn universal physical mechanisms, rather than simply memorizing training data, thus possessing powerful cross-condition generalization prediction capabilities and providing a reliable tool for dealing with complex and variable thermal processing conditions in practical engineering.
[0044] The embodiments described herein are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for predicting high-temperature compressive rheological stress based on a multi-branch fusion neural network, characterized in that, Includes the following steps: S1: Obtain raw data from high-temperature alloys in high-temperature compression tests, including strain. Absolute temperature T and strain rate With stress Relationship graph; S2: Preprocess the raw data to obtain the model input feature vector; the preprocessing includes at least performing Min-Max normalization on the temperature T to obtain the normalized temperature. Taking the natural logarithm of the strain rate yields the logarithmic strain rate. ,strain Keep the original value; S3: Construct a multi-branch fusion neural network, which consists of three independent branch subnetworks and a fusion network; S4: Strain Logarithmic strain rate and normalized temperature The inputs are fed into the corresponding sub-networks, and each sub-network is used to learn the nonlinear influence characteristics of the force on the corresponding physical quantity; among them, strain... The branch subnetwork learns the evolution characteristics of work hardening or softening with strain, logarithmic strain rate. Branch subnetworks learn strain rate sensitivity features and normalize temperature. Branch subnetworks learn temperature softening characteristics; S5: Concatenate the output features of the three branch sub-networks into a three-dimensional fusion feature vector; S6: Input the fused feature vector into the fused network, which is used to model the coupling relationship between strain, logarithmic strain rate, and normalized temperature, and output the original rheological stress prediction value. ; S7: Apply a non-negative constraint to the original rheological stress prediction value output by the fusion network to obtain the final predicted true stress value. .
2. The high-temperature compressive rheological stress prediction method based on a multi-branch fusion neural network according to claim 1, characterized in that, In step S4, each of the branch sub-networks is a three-layer feedforward neural network structure, including: First layer: Input x∈R(ε, , Mapped to 64-dimensional hidden feature h1, The second layer: Maps h1 to the 32-dimensional hidden feature h2. ; Third layer: Map h2 to the output scalar s. ; in, The first activation function is... To correct the linear unit , , , These are the linear transformation weights of the first, second, and third layers in a single-branch subnetwork, respectively. , , These represent the biases of the first, second, and third layers in a single-branch subnetwork, respectively. , These are the first and second layer implicit features, respectively; This is the output of the subnetwork.
3. The high-temperature compressive rheological stress prediction method based on a multi-branch fusion neural network according to claim 2, characterized in that, In S5, strain is addressed separately. Logarithmic strain rate and normalized temperature Three sets of parameter-independent branch subnetworks are constructed from three physical inputs to obtain: in For strain hardening / softening Evolutionary characteristics; It is a strain rate sensitivity characteristic; For temperature softening features; the output features of the three branches are concatenated into a three-dimensional fused feature vector: 。 4. The high-temperature compressive rheological stress prediction method based on a multi-branch fusion neural network according to claim 3, characterized in that, In step S6, the fusion network is a two-layer feedforward neural network structure, including: First fusion layer: Maps the fused feature vector Y∈Z to a 64-dimensional hidden feature u1. ； Second fusion layer: Maps u1 to 32-dimensional hidden feature u2 ； Output layer: Maps u2 to the original rheological stress prediction value. , ； in, The second activation function, the second activation function To correct the linear unit , , , These are the linear transformation weights of the first, second, and third layers in the branch feature fusion network, respectively. , , These represent the biases of the first, second, and third layers in the branch feature fusion network, respectively. , These are the first and second layer hidden features, respectively. This is the output of the fusion network.
5. The high-temperature compressive rheological stress prediction method based on a multi-branch fusion neural network according to claim 4, characterized in that, In step S7, the application of nonnegativity constraints is achieved through the Softplus function, i.e. .