A stress-strain curve stage-based prediction method based on digital test

Abstract

This invention discloses a staged prediction method for stress-strain curves based on digital experiments, belonging to the fields of electronic engineering and computer science. It includes: designing a data processing module to acquire stress-strain related data and perform data cleaning and classification; designing an offline training module to build three prediction models in stages and achieve model integration; and designing an online testing module to test the test dataset within the integrated model, outputting stress-strain curves and evaluation indicators, forming a closed-loop optimization. This invention can achieve accurate prediction of curves in stress-strain experiments through the fusion of PINN and liquid networks, realizing continuous reconstruction of material properties from the elastic to the plastic stage, significantly reducing the number of tensile and compressive mechanical tests, reducing reliance on high-cost mechanical loading and destructive testing, and is applicable to various materials and working conditions, possessing value for engineering promotion and digital materials research and development.

Description

Technical Field

[0001] This invention belongs to the fields of electronic engineering and computer science, and specifically relates to a staged prediction method for stress-strain curves based on digital experiments. Background Technology

[0002] In material performance evaluation and structural design, traditional methods of obtaining stress and strain rely on mechanical loading tests such as tension or compression. These tests are costly, time-consuming, and difficult to conduct stably under high strain rates or extreme environments. While pure simulation can replace some physical tests, it depends on constitutive models and parameter settings. For example, in numerical simulations of the yielding and hardening stages of metallic materials, excessive model simplification and high parameter uncertainty often lead to significant deviations between predicted curves and actual material responses. Existing numerical predictions mostly employ single-stage fitting methods, achieving predictions by fitting empirical curves or regression models. However, these methods lack physical constraints and struggle to accurately characterize the continuous evolution of the elastic zone—yield point—plastic zone. Significant errors occur, particularly in identifying stage transition points and predicting hardening trends, resulting in insufficient model stability and engineering generalization capabilities. With the development of digital materials experiments and physical information networks, there is an urgent need to establish a new digital experimental framework with staged modeling and dynamic error feedback capabilities to achieve accurate reconstruction of stress-strain curve predictions and closed-loop model optimization. Summary of the Invention

[0003] This invention provides a staged prediction method for stress-strain curves based on digital testing. This method can accurately predict the curves in stress-strain tests by fusing Physics-Informed Neural Networks (PINN) with a liquid network, and realize the continuous reconstruction of material properties from the elastic to the plastic stage. It significantly reduces the number of mechanical tests such as tensile and compressive tests, reduces reliance on high-cost mechanical loading and destructive testing, is applicable to a variety of materials and working conditions, and has value for engineering promotion and digital materials research and development.

[0004] Technical solution of the present invention:

[0005] A staged prediction method for stress-strain curves based on digital experiments includes the following steps:

[0006] Step S110: Obtain stress and strain data, clean and classify the obtained stress and strain data, and divide them into training set and test set according to the ratio;

[0007] Step S120: Based on the training set, construct the elastic stage model, the stage point discrimination model and the plastic stage model respectively, and integrate the three models into a complete stress-strain curve prediction model.

[0008] Step S130: Construct an online testing module to test the prediction model, output stress-strain curves and evaluation indicators, and continuously iterate based on the results to achieve closed-loop optimization.

[0009] A computing device includes: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the method.

[0010] Beneficial effects:

[0011] Compared with existing technologies, this invention can achieve accurate prediction of curves in stress-strain tests by integrating PINN with liquid networks, realize continuous reconstruction of material properties from elastic to plastic stages, significantly reduce the number of mechanical tests such as tensile and compressive tests, reduce reliance on high-cost mechanical loading and destructive testing, and is applicable to a variety of materials and working conditions, with value for engineering promotion and digital material research and development. Attached Figure Description

[0012] Figure 1 This is a flowchart of a staged prediction method for stress-strain curves based on digital experiments according to the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0014] This invention provides a staged prediction method for stress-strain curves based on digital testing, including the design of a data processing module, an offline training module, and an online testing module. By integrating PINN with a liquid network, it achieves accurate prediction of curves in stress-strain tests, realizes continuous reconstruction of material properties from the elastic to the plastic stage, significantly reduces the number of mechanical tests such as tensile and compressive tests, reduces reliance on high-cost mechanical loading and destructive testing, is applicable to a variety of materials and working conditions, and has value for engineering promotion and digital materials research and development.

[0015] Figure 1 This is a flowchart illustrating a staged prediction method for stress-strain curves based on digital testing according to the present invention. In the diagram, reference numeral 1 indicates the data processing module, reference numeral 2 indicates the offline training module, and reference numeral 3 indicates the online testing module. An example is a tensile test of AlSi10Mg material manufactured using laser powder bed fusion additive manufacturing; the specific implementation is as follows:

[0016] Step S110: Obtain stress-strain data, clean and classify the acquired stress-strain data, and divide it into training and test sets according to a set ratio. The specific steps are as follows:

[0017] S110-1, based on real tensile tests, obtains time-series stress-strain data and corresponding test operating condition data. The test operating condition data includes tensile strain rate. Structural parameters corresponding to the manufacturing of tensile samples and laser power .

[0018] S110-2, Based on the stress-strain data collected in step S110-1, plot the stress-strain curve, where the horizontal axis is strain and the vertical axis is stress, and obtain a predetermined number of stress-strain curves. Perform cleaning operations on the stress-strain data, including: origin correction, outlier removal, and invalid data segment truncation.

[0019] In one embodiment, the predetermined number may be 178.

[0020] The cleaning process includes:

[0021] To address the issue of the curve origin being offset, each curve is translated as a whole so that the first data point is located at the origin. Stress-strain data points with stress or strain values ​​less than zero after the translation are then deleted.

[0022] To address the issue of significant stress peaks in the elastic phase of the curve, an outlier removal method is employed to process abnormal data, ensuring that stress increases monotonically with strain.

[0023] To address the stress rebound phenomenon that occurs during the unloading process in the plastic stage, the data for the stress rebound stage are truncated to ensure a one-to-one correspondence between strain and stress values.

[0024] S110-3, The stress and strain data cleaned in step S110-2 are classified into training set data and test set data in an 8:2 ratio. The training set data is used for subsequent data calibration and serves as the input for each model, while the test data is used for online testing. Based on the boundary points obtained from the fitting, the training set data is divided into elastic stage data, stage point data, and plastic strain data.

[0025] In the training set, polynomial fitting was performed on the stress-strain data corresponding to strain values ​​less than 0.002 mm for each curve to obtain the elastic modulus of each curve. The boundary between the elastic and plastic stages is defined as the stress-strain data and elastic modulus before the boundary point, which are set as elastic stage data.

[0026] The stress and strain data before the boundary point are calibrated as state 0, and the stress and strain data after the boundary point are calibrated as state 1, and are set as stage point data.

[0027] Stress-strain data after the boundary point were fitted using the first term of the Johnson-Cook model:

[0028] ,

[0029] in, Let A be the stress value, B be the initial yield stress, and B be the strain hardening constant. For equivalent plastic strain, n is the strain hardening exponent. The stress-strain data after the boundary point and the corresponding parameters of each curve are set as plastic strain data.

[0030] In the test set, multiple stress-strain data corresponding to strain values ​​less than 0.002 mm for each curve are input into the test model.

[0031] Step S120: Based on the training set, construct the elastic stage model, the stage-specific discrimination model, and the plastic stage model respectively, and integrate the three models into a complete stress-strain curve prediction model. The specific steps are as follows:

[0032] S120-1, Based on the elastic stage data obtained in step S110-3, construct the elastic stage model. The elastic stage model is divided into the PINN network model based on Hooke's law and the residual compensation network model based on liquid neural network.

[0033] For the PINN network based on Hooke's Law, a GRU-based encoder model and an MLP-based decoder model are constructed. A sliding window is built for elastic stage data of a predetermined number of curves. In one embodiment, the window size of the sliding window can be... The step size can be The temporal stress-strain data within the sliding window is used as input to the GRU network to obtain the outputs of the update gate and the reset gate. The update gate controls the hidden state of the previous time step. How much information is retained in the current state? Reset the hidden state of the door control from the previous moment. How much information is used to calculate candidate states? ;

[0034] The update gate function is shown below:

[0035] ,

[0036] in, To update the gate output, It is the sigmoid activation function. The weight matrix is ​​input to the update gate. The stress-strain data are at time t. The weight matrix from the hidden state to the update gate. Indicates time The hidden state, To update the bias term of the gate.

[0037] The reset gate function is shown below:

[0038] ,

[0039] in, To reset the output of the door, The weight matrix is ​​input to the reset gate. The weight matrix from the hidden state to the reset gate. To reset the door's bias;

[0040] Based on current input Generate candidate states using the reset historical information:

[0041] ,

[0042] in, Candidate state The hyperbolic tangent activation function is used. The weight matrix is ​​the input to the gate. The weight matrix is ​​the hidden state to the gate. Represents the Hadamard product. For the gate's bias term;

[0043] The final output state is shown below:

[0044] ,

[0045] in, It is the hidden state at time t, and the output of the current state. To update the gate output, Represents the Hadamard product. This indicates the output of the previous state. The retention weight represents historical information. In the candidate hidden state, Introduce weights to candidate states;

[0046] The result The output is The experimental operating condition data corresponding to the curve for each window. , , Input into the MLP model architecture, output elastic modulus ;

[0047] When training a neural network, a loss function is constructed. :

[0048] ,

[0049] in, For the predicted elastic modulus The mean square error function of E, where E is the elastic modulus of the actual elastic phase data. for The weight value is 1. For the physical loss function of the elastic phase, for The weight value is 1. It can be as follows:

[0050] ,

[0051] in, This represents the actual stress value during the elastic stage. This represents the actual strain value during the elastic stage. The function, which is the predicted elastic modulus, conforms to Hooke's Law.

[0052] For residual compensation networks based on liquid neural networks, an encoder with a liquid neural network architecture and a decoder with an MLP architecture are built. A predetermined number of curve data points are transformed based on the following formula:

[0053] ,

[0054] ,

[0055] in, These are the residual stress values ​​from the elastic stage data input into the residual compensation network. These are the actual stress values ​​from the elastic phase data. These are the elastic stage data strain values ​​input into the residual compensation network. These are the actual strain values ​​from the elastic phase data. Construct a sliding window, for example, with a window size of... Step size is The temporal stress-strain data within the sliding window is input into the liquid neural network, and the update formula of the liquid neural network is as follows:

[0056] ,

[0057] The above formula is discretized using the Euler method. This represents the hidden state vector at time n+1. Represents the hidden state at time n. Represents the discrete time step. Represents the circular connection weight matrix. Represents an element-wise nonlinear activation function. Represents the input connection weight matrix. Represents the external input at time n. This represents a time-dependent constant.

[0058] The liquid neural network cell unit LTCcell is updated using the above formula. The LTCcells are stacked, and through residual connections, they output the encoded results. These encoded results are then compared with the experimental operating condition data of the curve corresponding to each sliding window. , , The subsequent sliding window (e.g., number) The strain value corresponding to point ) is used as input to the MLP model architecture, and the stress values ​​corresponding to the same number of subsequent points are output. During training, the loss function As shown below:

[0059] ,

[0060] During training, after passing a preset threshold (e.g.) After 10 epochs, the loss function Instead of lowering the rate, a teacher strategy is introduced, using a certain probability (e.g., probability). Insert model predictions into the actual values. In this process, hybrid training is achieved until the model... It will not decrease further.

[0061] S120-2, Based on the elastic stage data and segmented point data obtained in step S110-3, a segmented point discrimination model is constructed. This model is a GRU neural network guided by prior knowledge. The prior knowledge is that the strain value corresponding to the boundary between the elastic and plastic stages of common metallic materials is usually between 0.002mm and 0.008mm. For a predetermined number of stress-strain curves (e.g., the 178 curves mentioned above) containing strain values ​​between 0.002mm and 0.008mm, a sliding window is constructed, the window size of which can be... Step size is The temporal stress-strain data within each sliding window are used as input to the GRU neural network. The GRU neural network outputs a staged state identifier of 0 or 1, where 0 represents the elastic stage and 1 represents the plastic stage.

[0062] S120-3, Based on the plastic strain data obtained in step S110-3, construct the plastic stage model. The plastic stage model is divided into the PINN network model based on the Johnson-Cook model and the residual compensation network model based on the liquid neural network.

[0063] For the PINN network based on the Johnson-Cook model, construct an encoder model with the same GRU architecture and a decoder model with the same MLP architecture as in step S120-1. For a predetermined number of curves in the elastic data stage, construct a window size of [missing information]. Step size is A sliding window is used, and the temporal stress-strain data within the sliding window is used as input to the encoder model of the GRU architecture. The encoder model performs encoding, and the obtained data is then processed. The output is The experimental operating condition data corresponding to the curve for each window. , , Each window has several subsequent (e.g., number) The strain value corresponding to point ) As input to the MLP model architecture, it outputs the stress values ​​corresponding to the same number of subsequent points. and the parameters of each prediction curve, including The initial yield stress, It is the strain hardening constant. The strain hardening index;

[0064] When training a neural network, a loss function is constructed. :

[0065] ,

[0066] in, For the predicted stress value Compared with the actual stress value The mean square error function, for The weight value is 1. Let physical loss function be used during the plastic stage. for The weight value is 1. It can be as follows:

[0067] ,

[0068] in, True stress value To predict the initial yield stress, To predict the strain hardening constant, To predict the strain hardening exponent, For true equivalent plastic strain, this function conforms to the Johnson-Cook model;

[0069] For the residual compensation network based on liquid neural network, a predetermined number of curve data are transformed based on the following formula:

[0070] ,

[0071] ,

[0072] in, These are the residual stress values ​​from the plastic stage data input into the residual compensation network. This is the actual stress value during the plastic stage. These are the predicted stress values ​​of the PINN network based on the Johnson-Cook model during the plastic stage. This is the true strain value during the plastic stage. These are the residual strain values ​​from the plastic stage data input into the residual compensation network.

[0073] To achieve the same architecture as the elastic phase, update the liquid neural network cell unit (LTCcell), stack the LTCcells, and output the encoded results through residual connections. Compare the encoded results with the experimental operating condition data of the curve corresponding to each sliding window. , , The subsequent windows (e.g., the number is...) The strain value corresponding to a given point is used as input to the MLP model architecture, and the stress values ​​corresponding to the same number of subsequent points are output. During training, the loss function As shown below:

[0074] ,

[0075] During training, after passing a preset threshold (e.g.) After 10 epochs, the loss function Instead of lowering the rate, teacher strategies are introduced, using a certain probability, for example... Insert the model predictions into In this process, hybrid training is achieved until the model... It will not decrease further.

[0076] S120-4, Integrate the elastic stage model obtained in step S120-1, the phase point model obtained in step S120-2, and the plastic stage model obtained in step S120-3 to form a complete prediction model;

[0077] The specific integration process and data flow are as follows:

[0078] First, the stress and strain data after cleaning are input into the phased point discrimination model constructed in step S120-2, and the current material state is judged by the GRU neural network guided by prior knowledge.

[0079] If the stage state identifier output by the discrimination model is 0, then elastic stage prediction is performed: the data flows to the elastic stage model constructed in step S120-1, and the stress and strain data are synchronously input to the PINN network model based on Hooke's law and the residual compensation network model based on liquid neural network. The PINN network calculates the physical reference value, the residual compensation network calculates the nonlinear deviation, and the two are added together to output the elastic stress prediction result.

[0080] If the stage state indicator output by the discrimination model changes from 0 to 1, then the plastic stage prediction is performed: the data flows to the plastic stage model constructed in step S120-3, the stress and strain data are input into the PINN network model based on the Johnson-Cook model and the corresponding residual compensation network model, and the stress and strain data at the end of the elastic stage are used as the initial boundary conditions to introduce into the plastic calculation. Finally, the plastic stress prediction result is obtained by adding the outputs of the PINN network and the residual compensation network of the plastic stage.

[0081] Step S130: Construct an online testing module to test the prediction model, output stress-strain curves and evaluation indicators, and continuously iterate based on the results to achieve closed-loop optimization. The specific steps are as follows:

[0082] S130-1 inputs the stress-strain data of multiple points corresponding to strain values ​​less than 0.002mm for each curve in the test set into the prediction model to achieve the prediction output of the entire stress-strain curve;

[0083] S130-2: Obtain the actual stress-strain curve using real data from the test set. Compare and evaluate the actual stress-strain curve with the predicted stress-strain curve to construct an overall evaluation index. Utilize multiple indicators such as the coefficient of determination (COD) and root mean square error (RMSE) for evaluation. If the COD and RMSE meet the preset evaluation thresholds, for example, the COD... Root mean square error index If the model meets the requirements, it is determined that the model does not meet the requirements. If it does not meet the requirements, the hyperparameters of the prediction model are adjusted and each model is retrained until the prediction model integration is completed and the performance requirements are met.

[0084] In summary, this invention discloses a staged prediction method for stress-strain curves based on digital experiments. This method involves three parts: a data processing module, an offline training module, and an online testing module. By integrating PINN with a liquid network, it achieves accurate prediction of curves in stress-strain experiments, enabling continuous reconstruction of material properties from the elastic to the plastic stage. This significantly reduces the number of mechanical tests such as tensile and compressive stress tests, reduces reliance on high-cost mechanical loading and destructive testing, and is applicable to various materials and working conditions. It has value for engineering promotion and digital materials research and development.

[0085] The present invention also provides a computing device, comprising: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the method described thereon.

[0086] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0087] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered to fall within the protection scope of the present invention.

Claims

1. A staged prediction method for stress-strain curves based on digital experiments, characterized in that, Includes the following steps: Step S110: Obtain stress and strain data, clean and classify the obtained stress and strain data, and divide them into training set and test set according to the ratio; Step S120: Based on the training set, construct the elastic stage model, the stage point discrimination model and the plastic stage model respectively, and integrate the three models into a complete stress-strain curve prediction model. Step S130: Construct an online testing module to test the prediction model, output stress-strain curves and evaluation indicators, and continuously iterate based on the results to achieve closed-loop optimization.

2. The method for staged prediction of stress-strain curves based on digital experiments according to claim 1, characterized in that, Step S110 includes: Step S110-1: Obtain time-series stress-strain data and corresponding test operation condition data based on actual tensile tests. The test operation condition data includes tensile strain rate. Structural parameters corresponding to the manufacturing of tensile samples and laser power ; Step S110-2: Based on the stress-strain data collected in step S110-1, plot the stress-strain curve, where the horizontal axis is strain and the vertical axis is stress. Obtain a predetermined number of stress-strain curves. Perform cleaning operations on the stress-strain data, including: origin correction, outlier removal, and invalid data segment truncation. Step S110-3: The stress and strain data cleaned in step S110-2 are classified into training set data and test set data in an 8:2 ratio. The training set data is used for subsequent data calibration and serves as the input for each model, while the test set data is used for online testing. Based on the boundary points obtained from the fitting, the training set data is divided into elastic stage data, stage point data, and plastic strain data.

3. The method for staged prediction of stress-strain curves based on digital experiments according to claim 2, characterized in that, Step S110-3 includes: In the training set, polynomial fitting was performed on the stress-strain data corresponding to strain values ​​less than 0.002 mm for each curve to obtain the elastic modulus of each curve. The boundary between the elastic and plastic stages is defined as the elastic stage data, and the stress-strain data and elastic modulus before the boundary are set as elastic stage data. The stress and strain data before the boundary point are calibrated as state 0, and the stress and strain data after the boundary point are calibrated as state 1, and are set as stage point data. Stress-strain data after the boundary point were fitted using the first term of the Johnson-Cook model: ， in, Let A be the stress value, B be the initial yield stress, and B be the strain hardening constant. For equivalent plastic strain, n is the strain hardening exponent. The fitted stress-strain data after the boundary point and the corresponding parameters of each curve are set as plastic strain data. In the test set, multiple stress-strain data corresponding to strain values ​​less than 0.002 mm for each curve are input into the prediction model.

4. The method for staged prediction of stress-strain curves based on digital experiments according to claim 3, characterized in that, Step S120 includes: Step S120-1: Based on the elastic stage data obtained in step S110-3, construct the elastic stage model. The elastic stage model is divided into the PINN network model based on Hooke's Law and the residual compensation network model based on the liquid neural network. Step S120-2: Based on the stage point data obtained in step S110-3, a stage point discrimination model is constructed. This model is a GRU neural network guided by prior knowledge. The prior knowledge is that the strain value corresponding to the boundary between the elastic and plastic stages of the metallic material is distributed in the range of 0.002 mm–0.008 mm. For a predetermined number of stress-strain curves with strain values ​​in the range of 0.002 mm–0.008 mm, a sliding window is constructed. The time-series stress-strain data in each sliding window is used as the input of the GRU neural network. The GRU neural network outputs a stage state identifier of 0 or 1, where 0 represents the elastic stage and 1 represents the plastic stage. Step S120-3: Based on the plastic strain data obtained in step S110-3, construct a plastic stage model. The plastic stage model is divided into a PINN network model based on the Johnson-Cook model and a residual compensation network model based on a liquid neural network. Step S120-4: Integrate the elastic stage model obtained in step S120-1, the stage point model obtained in step S120-2, and the plastic stage model obtained in step S120-3 to form a complete prediction model. The specific integration process and data flow are as follows: First, the stress and strain data after cleaning are input into the phased point discrimination model constructed in step S120-2, and the current material state is judged by the GRU neural network guided by prior knowledge. If the stage state identifier output by the discrimination model is 0, then elastic stage prediction is performed: the data flows to the elastic stage model constructed in step S120-1, and the stress and strain data are synchronously input to the PINN network model based on Hooke's law and the residual compensation network model based on liquid neural network. The PINN network calculates the physical reference value, the residual compensation network calculates the nonlinear deviation, and the two are added together to output the elastic stress prediction result. If the stage state indicator output by the discrimination model changes from 0 to 1, then the plastic stage prediction is performed: the data flows to the plastic stage model constructed in step S120-3, the stress and strain data are input into the PINN network model based on the Johnson-Cook model and the corresponding residual compensation network model, and the stress and strain data at the end of the elastic stage are used as the initial boundary conditions to introduce into the plastic calculation. Finally, the plastic stress prediction result is obtained by adding the outputs of the PINN network and the residual compensation network of the plastic stage.

5. The method for staged prediction of stress-strain curves based on digital experiments according to claim 3, characterized in that, Step S120-1 includes: For the PINN network based on Hooke's Law, a GRU-based encoder model and an MLP-based decoder model are constructed. A sliding window is built for the elastic stage data of a predetermined number of curves. The temporal stress-strain data within the sliding window is used as the input to the GRU network to obtain the outputs of the update gate and the reset gate. The update gate controls the hidden state of the previous time step. How much information is retained in the current state? Reset the hidden state of the door control from the previous moment. How much information is used to calculate candidate states? ; The result The output is Experimental operating condition data corresponding to the curve of each sliding window , , Input into the MLP model architecture, output elastic modulus ; For residual compensation networks based on liquid neural networks, an encoder with a liquid neural network architecture and a decoder with an MLP architecture are built. Based on the following formula, a predetermined number of curve data points are transformed to obtain strain values. Stress value : ， ， in, For the predicted elastic modulus, These are the residual stress values ​​from the elastic stage data input into the residual compensation network. These are the actual stress values ​​from the elastic phase data. These are the elastic stage data strain values ​​input into the residual compensation network. These are the actual strain values ​​from the elastic phase data; A sliding window is constructed, and the temporal stress-strain data within the sliding window is input into the liquid neural network. The update formula of the liquid neural network is as follows: ， The above formula is discretized using the Euler method. This represents the hidden state vector at time n+1. Represents the hidden state at time n. Represents the discrete time step. Represents the circular connection weight matrix. Represents an element-wise nonlinear activation function. Represents the input connection weight matrix. Represents the external input at time n. Represents a time-dependent constant; The liquid neural network cell unit LTCcell is updated using the above formula. The LTCcells are stacked, and through residual connections, they output the encoded results. These encoded results are then compared with the experimental operating condition data of the curve corresponding to each sliding window. , , The strain values ​​corresponding to several points after the window are used as input to the MLP model architecture, and the stress values ​​corresponding to the same number of subsequent points are output. .

6. The method for staged prediction of stress-strain curves based on digital experiments according to claim 5, characterized in that, Step S120-1 includes: constructing a loss function for the PINN network based on Hooke's Law. : ， in, For the predicted elastic modulus The mean square error function of E, where E is the elastic modulus of the actual elastic phase data. for The weight value, For the physical loss function of the elastic phase, for The weight value, As shown below: ， in, This represents the actual stress value during the elastic stage. This represents the actual strain value during the elastic stage. The function, which is the predicted elastic modulus, conforms to Hooke's Law. For residual compensation networks based on liquid neural networks, during training, the loss function... As shown below: ， During training, after a preset threshold number of epochs, the loss function... Instead of lowering the value, a teacher strategy is introduced to insert the model's predicted value into the model with a certain probability. In this process, hybrid training is achieved until the model... It will not decrease further.

7. The method for staged prediction of stress-strain curves based on digital experiments according to claim 5, characterized in that, Step S120-3 includes: for the PINN network based on the Johnson-Cook model, constructing an encoder model with the same GRU architecture as in step S120-1, and a decoder model with an MLP architecture; constructing a sliding window for the elastic data stages of a predetermined number of curves; using the temporal stress-strain data within the sliding window as input to the encoder model with the GRU architecture; and encoding the obtained data through the encoder model. The output is The experimental operating condition data corresponding to the curve for each window. , , The strain values ​​corresponding to several points after each window As input to the MLP model architecture, it outputs the stress values ​​corresponding to the same number of subsequent points. and the parameters of each prediction curve, including The initial yield stress, It is the strain hardening constant. The strain hardening index; For the residual compensation network based on liquid neural network, the stress value is obtained by transforming a predetermined number of curve data according to the following formula. strain value : ， ， in, These are the residual stress values ​​from the plastic stage data input into the residual compensation network. This is the actual stress value during the plastic stage. These are the predicted stress values ​​of the PINN network based on the Johnson-Cook model during the plastic stage. This is the true strain value during the plastic stage. These are the strain values ​​from the plastic stage input into the residual compensation network; The architecture is the same as in the elastic phase. The liquid neural network cell unit (LTCcell) is updated, stacked, and connected via residuals to output the encoded results. These results are then compared with the experimental operating condition data of the curve corresponding to each sliding window. , , The strain values ​​corresponding to several points after the window are used as input to the MLP model architecture, and the stress values ​​corresponding to the same number of subsequent points are output. .

8. The method for staged prediction of stress-strain curves based on digital experiments according to claim 7, characterized in that, For the PINN network based on the Johnson-Cook model, a loss function is constructed. : ， in, For the predicted stress value Compared with the actual stress value The mean square error function, for The weight value, This is the physical loss function during the plastic stage. for The weight value, As shown below: ， in, True stress value To predict the initial yield stress, To predict the strain hardening constant, To predict the strain hardening exponent, For true equivalent plastic strain, this function conforms to the Johnson-Cook model; For residual compensation networks based on liquid neural networks, the loss function is... As shown below: ， in, It is the residual stress value predicted by the residual compensation network; During training, after a preset threshold number of epochs, the loss function... Instead of lowering the value, a teacher strategy is introduced to insert the model's predicted value into the model with a certain probability. In this process, hybrid training is achieved until the model... It will not decrease further.

9. The method for staged prediction of stress-strain curves based on digital experiments according to claim 8, characterized in that, Step S130 includes: Step S130-1: Input multiple stress-strain data points corresponding to strain values ​​less than 0.002 mm in each stress-strain curve of the test set into the prediction model to realize the prediction output of the entire stress-strain curve; Step S130-2: Obtain the corresponding true stress-strain curves based on real experimental data in the test set, and compare and evaluate the true stress-strain curves with the predicted stress-strain curves to construct an overall performance evaluation index system, using the coefficient of determination. The root mean square error (RMSE) is used to comprehensively evaluate the model's predictive performance. If the coefficient of determination and the RMSE meet the preset evaluation thresholds, the predictive model is deemed to meet the requirements. If the preset evaluation thresholds are not met, the hyperparameters of the predictive model are adjusted and each model is retrained until the predictive model integration is completed and the performance requirements are met.

10. A computing device, characterized in that, include: At least one processor and a memory storing program instructions; When the program instructions are read and executed by the processor, the computing device performs the method as described in any one of claims 1-9.

Images