CN115588470A - Crystal plasticity constitutive parameter calibration method based on two-channel convolutional neural network - Google Patents
Crystal plasticity constitutive parameter calibration method based on two-channel convolutional neural network Download PDFInfo
- Publication number
- CN115588470A CN115588470A CN202211332683.1A CN202211332683A CN115588470A CN 115588470 A CN115588470 A CN 115588470A CN 202211332683 A CN202211332683 A CN 202211332683A CN 115588470 A CN115588470 A CN 115588470A
- Authority
- CN
- China
- Prior art keywords
- groups
- stress
- crystal
- ranges
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/70—Machine learning, data mining or chemometrics
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computer Hardware Design (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Geometry (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Databases & Information Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Crystallography & Structural Chemistry (AREA)
- Chemical & Material Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Complex Calculations (AREA)
Abstract
The invention provides a crystal plasticity constitutive parameter calibration method based on a two-channel convolutional neural network, which comprises the following steps: firstly, establishing a magnesium alloy model by using Neper software, then establishing a magnesium alloy model with a texture by using MTEX, and then establishing a crystal plasticity simulation file by using Python; inputting the crystal plasticity simulation file into DAMASK simulation software for simulation to obtain an IPF (intrinsic stress) -true strain coordinate data and an IPF diagram; then cutting the IPF graph to obtain the IPF S Drawing (1) toDrawing a true stress-true strain curve graph by using true stress-true strain coordinate data; then carrying out normalization processing on constitutive parameters; IPF S The method comprises the steps of taking a graph and a true stress-true strain curve graph as input, taking normalized constitutive parameters as output, and establishing a universal two-channel convolution neural network model for calibrating crystal plastic constitutive parameters through training, wherein the degree of fitting between the parameters calibrated by the model and the parameters measured by experiments is more than 99%; the method provided by the invention has high efficiency and good accuracy, and is superior to the traditional algorithm.
Description
Technical Field
The invention belongs to the field of metal material performance detection, and particularly relates to a crystal plasticity constitutive parameter calibration method based on a dual-channel convolution neural network.
Background
The crystal plasticity is an important calculation tool for crystal material deformation, can simulate the mesoscopic metal deformation process, including crystal grain orientation, crystal grain torsion, texture change and the like, and is increasingly widely applied in the field of material research. However, since crystal plasticity involves many constitutive parameters of the material, the identification and calibration of the parameters are difficult and challenging. The traditional identification method of the crystal plasticity constitutive parameters mainly adopts optimization methods such as a genetic algorithm, a particle swarm algorithm, simulated annealing and the like, the optimization method is long in operation time, low in efficiency and high in trial and error cost, and operators need to have deep understanding and certain use experience on the optimization algorithm. Therefore, a calibration method for the crystal plasticity constitutive parameters needs to be established, so that the workload of an optimization algorithm can be obviously reduced, and the optimization efficiency is improved.
Disclosure of Invention
Based on the above, the invention provides a crystal plasticity constitutive parameter calibration method based on a two-channel convolution neural network, which comprises the following steps:
(1) The magnesium alloy model was established using the Neper software: inputting m groups of parameters x and y into Neper software, wherein x and y respectively represent the random distribution and the crystal grain roundness of crystal grains, the value range of m is 100-100000 and is an integer, the value range of x is 0-1, the value range of y is 0.3-1.0, m groups of magnesium alloy models are obtained and are marked as m groups (x, y), and the model of each group (x, y) is shown as a matrix 1;
matrix 1
i is one of 1, 2 or 3;
(2) A magnesium alloy model with added texture was established using Matlab kit MTEX: setting m groups of parameters d, e, f, g, h and k, wherein the parameters d, e, f, g, h and k respectively represent six different texture strengths of the magnesium alloy in the step (1), the value ranges of d, e, f, g, h and k are all 0-1, and d + e + f + g + h + k =1; inputting m groups of d, e, f, g, h and k parameters and m groups of (x, y) obtained in the step (1) into MTEX, and obtaining m groups of magnesium alloy models (d, e, f, g, h and k) (x, y) with texture added through the Bunge formula operation of MTEX, wherein the m groups of magnesium alloy models are marked as m groups of (d, e, f, g, h and k) (x, y), and data models of each group of (d, e, f, g, h and k) (x, y) are shown as a matrix 2;
matrix 2
(3) Using Python to obtain m sets of crystal plastic simulation files: selecting 16 parameters of a crystal plasticity constitutive structure: basal plane slippage critical slitting stress T1, cylindrical surface slippage critical slitting stress T2, conical surface < a > slippage critical slitting stress T3, conical surface < c + a > slippage critical slitting stress T4, basal plane slippage saturated slippage slitting stress T5, cylindrical surface slippage saturated slippage slitting stress T6, conical surface < a > slippage saturated slippage slitting stress T7, conical surface < c + a > slippage saturated slippage slitting stress T8, tensile twin crystal critical slitting stress T9, compressive twin crystal critical slitting stress T10, twin and twin initial hardening modulus T11, slippage and slippage initial hardening modulus T12, slippage and twin initial hardening modulus T13, slippage strain rate sensitive parameter T14, twin strain rate sensitive parameter T15 and slippage hardening parameter T16; t1 ranges from 5 to 80, T2 ranges from 10 to 150, T3 ranges from 20 to 200, T4 ranges from 50 to 500, T5 ranges from 10 to 160, T6 ranges from 20 to 300, T7 ranges from 40 to 400, T8 ranges from 100 to 1000, T9 ranges from 5 to 250, T10 ranges from 10 to 400, T11 ranges from 10 to 500, T12 ranges from 50 to 3000, T13 ranges from 50 to 1200, T14 ranges from 1 to 15, T15 ranges from 5 to 80, and T16 ranges from 0.5 to 10; designing m groups of T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 in the range, adding each group of T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 to the group of magnesium alloy models (d, e, f, g, h, i) (x and y) in the step (2) through Python to obtain a group of crystal plastic simulation files, and finally obtaining m groups of crystal plastic simulation files, wherein each group of crystal plastic simulation files is in a form shown in a matrix 3;
matrix 3
(4) IPF plots and simulated true stress-true strain coordinate data were obtained by simulation: adopting DAMASK as crystal plasticity finite element simulation software, inputting the crystal plasticity simulation file obtained in the step (3) to perform crystal plasticity finite element simulation, inputting m groups of crystal plasticity simulation files in the step (3) into the DAMASK software to perform m groups of simulation, following the crystal plasticity constitutive equation set in the DAMASK simulation, obtaining m groups of simulation results after the simulation is completed, and performing post-processing on the m groups of simulation results by using Python to obtain an IPF (in-plane stress) -true strain coordinate data of m groups of magnesium alloys and m groups of simulated true stress-true strain coordinate data;
the crystal plasticity constitutive equation set is as follows:
n is the initial hardening modulus of the tensile and compressive twins, n is one of T9 and T10,
h 0 the initial hardening modulus values between 3 different slip systems, between slip system and twin system, h 0 Is one of T11, T12 and T13,
q αβ is a coefficient between the different slip systems alpha and beta, q αβ Is one of T14, T15 and T16,
H αβ total hardening modulus;
(5) IPFs plots and true stress-true strain plots were obtained using Python and the crystal plasticity constitutive parameters were normalized: respectively normalizing the m groups of 16 crystal plasticity constitutive parameters T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 in the step (3) to obtain m groups of normalized T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16, drawing the m groups of simulated true stress-true strain coordinate data in the step (4) into m groups of true stress-true strain graphs by using Python, and cutting the m groups of IPF graphs in the step (4) into m groups of standard IPF graphs of standard 50x50 pixels, and marking the m groups of standard IPF graphs as m groups of IPFs graphs; finally obtaining m groups of crystal plasticity data sets, wherein each group of data sets comprises: an IPFs diagram, a true stress-true strain graph and normalized t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16;
the normalization formula is as follows:
W norm is one of t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16 parameters,
w is one of parameters T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16,
W min is the smallest number of the m sets of parameters,
W max maximum number in m sets of parameters;
therefore, an IPFs graph, a true stress-true strain curve graph and a big data sample corresponding to the normalized constitutive parameters are established;
(6) Training a two-channel convolutional neural network model: subjecting the m sets of IPFs of step (5) to S Taking the graph and a true stress-true strain graph as input of a training dual-channel convolutional neural network model, taking t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16 of the m groups in the step (5) as output of the neural network, and making the dual-channel convolutional neural network model to be randomly divided into p groups of training sets and q groups of verification sets according to a proportion of 5; the two-channel convolution neural network model comprises: 5 layers of alxnet model convolution layer, 3 full connection layers; the convolution layer is as follows: grouping p IPFs in a training set S Inputting the graph and the real stress-real strain curve graph into the convolution layer, abstracting the image into a characteristic graph after convolution operation of the convolution layer, and p groups of IPFs S P groups of characteristic diagrams 1 are obtained after 5 layers of convolution layer processing, p groups of real stress-real strain graphs obtain p groups of characteristic diagrams 2 after 5 layers of convolution layer processing, and the p groups of characteristic diagrams 1 and the p groups of characteristic diagrams 2 are merged and input into full connection layer processing; the full connection layer treatment comprises the following steps: calculating and obtaining p groups of output values through a forward propagation formula containing weights and biases, wherein the weights and the biases are random, the range of the weights is 1-2, the range of the biases is 10-20, and each group of output values is as follows: t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16', then p sets t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16' are calculated and the trainingThe overall error between p groups in the training set t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15, t16; then, recalculating the weight and the bias according to the integral error, then, recalculating by adopting a forward propagation formula again, and performing reciprocating circulation until the integral error is 0.2-1%, and finishing training to obtain a trained dual-channel convolution neural network model; wherein the initial learning rate of the dual-channel convolutional neural network model is set to be 0.00001-0.01, and the learning times are 1000-3000; the overall error of the trained dual-channel convolutional neural network model to the p groups of training sets is 0.2% -1%, and the overall error of the trained dual-channel convolutional neural network model to the q groups of verification sets is 0.4% -1.1%;
the integral error formula is as follows:
p is the number of training set groups,
the sigma is the integral error of the error,
nu is a positive integer with the value of 1-16,
t′ v is one of t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16',
t v is one of t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16,
the integral error, weight and bias calculation formula is as follows:
σ＝W T +b
the sigma is the integral error of the error,
w is a weight value,
t is the transposed symbol and is the symbol,
b is the bias.
Further, the magnesium alloy is one of AZ31, AZ61, AZ91, ZK60 or AM 60;
the invention has the beneficial effects that:
the invention adopts the convolutional neural network in machine learning as a tool for solving the nonlinear problem, utilizes the higher efficiency and the intelligent advantage of the convolutional neural network, borrows the advanced algorithm of the current machine learning, establishes a convolutional neural network model for calibrating the crystal plasticity constitutive parameters based on big data, and effectively changes the problems of low-efficiency acquisition of the current crystal plasticity constitutive parameter calibration, dependence on the optimization algorithm condition and the like.
The method comprises the steps of constructing an IPFs graph, a true stress-true strain curve graph and a big data sample corresponding to a normalized constitutive parameter, using the IPFs graph and the true stress-true strain curve graph as input, using the normalized constitutive parameter graph as output to train a dual-channel convolutional neural network, wherein the error of a trained neural network model to a training set is 0.2% -1%, and the error of a verification set is 0.4% -1.1%; comparing and analyzing the anisotropy coefficient R' and the section reduction rate of the magnesium alloy obtained by the neural network simulation with the anisotropy coefficient R and the section reduction rate of the magnesium alloy obtained by the experiment, wherein the fitting degree of the relevant parameters obtained by the method and the parameters obtained by the experiment is 99.1-99.7%; compared with the traditional methods such as genetic algorithm, particle swarm algorithm, simulated annealing and the like, the method has the advantages that 16 crystal plasticity constitutive parameters can be calibrated at one time, the calibration speed is high, the calibration accuracy is high, the neural network model after the training is completed has high universality, and the method can be suitable for calibrating the crystal plasticity constitutive parameters of the whole magnesium alloy system.
Drawings
Fig. 1 is a diagram of a two-channel convolutional neural network model used in step S6 in embodiment 1.
Detailed description of the invention
The invention is further described with reference to the following specific embodiments and the accompanying drawings.
Example 1
The method for selecting the AZ31 magnesium alloy and calibrating the AZ31 magnesium alloy crystal plasticity constitutive parameters by establishing the dual-channel convolution neural network comprises the following steps:
s1, establishing an AZ31 magnesium alloy model by using Neper software, inputting 10000 groups of parameters x and y into the Neper software, wherein x and y respectively represent the random distribution and the grain roundness of grains, the value range of x is 0.1-0.8, the value range of y is 0.3-0.9, obtaining 10000 groups of magnesium alloy models which are marked as 10000 groups (x, y), and the model of each group (x, y) is shown as a matrix 1;
matrix 1
i is one of 1, 2 or 3;
s2, establishing an AZ31 magnesium alloy model with the texture added by using a tool box MTEX of Matlab: 10000 groups of d, e, f, g, h, k parameters, d, e, f, g, h, k respectively represent six different texture strengths of the magnesium alloy in the step S1, the value ranges of d, e, f, g, h, k are all 0-1, and d + e + f + g + h + k =1; inputting 10000 groups of d, e, f, g, h, k parameters and 10000 groups (x, y) obtained in the step S1 into MTEX, and obtaining 10000 groups of magnesium alloy models (d, e, f, g, h, k) (x, y) with texture added through the Bunge formula operation of the MTEX, wherein the 10000 groups of magnesium alloy models are marked as 10000 groups (d, e, f, g, h, k) (x, y), and data models of each group (d, e, f, g, h, k) (x, y) are shown as a matrix 2;
matrix 2
S3, obtaining 10000 crystal plastic simulation files by using Python: selecting 16 parameters of a crystal plasticity constitutive structure: basal plane slippage critical slitting stress T1, cylindrical surface slippage critical slitting stress T2, conical surface < a > slippage critical slitting stress T3, conical surface < c + a > slippage critical slitting stress T4, basal plane slippage saturated slippage slitting stress T5, cylindrical surface slippage saturated slippage slitting stress T6, conical surface < a > slippage saturated slippage slitting stress T7, conical surface < c + a > slippage saturated slippage slitting stress T8, tensile twin crystal critical slitting stress T9, compressive twin crystal critical slitting stress T10, twin and twin initial hardening modulus T11, slippage and slippage initial hardening modulus T12, slippage and twin initial hardening modulus T13, slippage strain rate sensitive parameter T14, twin strain rate sensitive parameter T15 and slippage hardening parameter T16; t1 ranges from 5 to 40, T2 ranges from 10 to 100, T3 ranges from 20 to 100, T4 ranges from 50 to 250, T5 ranges from 10 to 100, T6 ranges from 20 to 150, T7 ranges from 40 to 200, T8 ranges from 100 to 500, T9 ranges from 5 to 150, T10 ranges from 10 to 200, T11 ranges from 10 to 250, T12 ranges from 50 to 1500, T13 ranges from 50 to 600, T14 ranges from 1 to 10, T15 ranges from 5 to 40, and T16 ranges from 0.5 to 8; designing 10000 groups of T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 in the range, adding each group of T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 to the group of magnesium alloy models (d, e, f, g, h, i) (x, y) in the step S2 through Python to obtain a group of crystal plasticity simulation files, and finally obtaining 10000 groups of crystal plasticity simulation files, wherein the form of each group of crystal plasticity simulation files is shown as a matrix 3;
matrix 3
S4, obtaining an IPF diagram and simulated true stress-true strain coordinate data through simulation: adopting DAMASK as crystal plasticity finite element simulation software, inputting the crystal plasticity simulation file obtained in the step S3 to simulate the crystal plasticity finite element, inputting 10000 crystal plasticity simulation files in the step S3 into the DAMASK software to simulate 10000 groups of crystals, wherein the DAMASK simulation follows a crystal plasticity constitutive equation set, 10000 groups of simulation results are obtained after the simulation is finished, and the 10000 groups of simulation results are post-processed by Python to obtain an IPF (intrinsic stress) -true strain coordinate data of 10000 groups of magnesium alloys and 10000 groups of simulated true stress-true strain coordinate data;
the crystal plasticity constitutive equation set is as follows:
n is the initial hardening modulus of the tensile and compressive twins, n is one of T9 and T10,
h 0 the initial hardening modulus values between 3 different slip systems, between slip system and twin system, h 0 Is one of T11, T12 and T13,
q αβ is a coefficient between the different slip systems alpha and beta, q αβ Is one of T14, T15 and T16,
H αβ total hardening modulus;
s5, obtaining an IPFs graph and a true stress-true strain curve graph by using Python, and normalizing the crystal plasticity constitutive parameters: respectively normalizing the 10000 sets of 16 crystal plasticity constitutive parameters T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 in the step S3 to obtain 10000 sets of normalized T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16, drawing 10000 sets of simulated true stress-true strain coordinate data in the step S4 into 10000 sets of true stress-true strain graphs by using Python, and cutting the 10000 sets of IPF graphs in the step S4 into 10000 sets of standard IPF graphs of standard 50x50 pixels, which are recorded as 10000 sets of IPFs graphs; 10000 sets of crystal plasticity data sets are finally obtained, wherein each set of data set comprises the following data sets: IPFs diagrams, true stress-true strain graphs, and normalized t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15, t16;
the normalization formula is as follows:
W norm is one of the parameters t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16,
w is one of parameters T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16,
W min is the smallest number of the 10000 sets of parameters,
W max maximum number in 10000 sets of parameters;
therefore, an IPFs graph, a true stress-true strain curve graph and a big data sample corresponding to the normalized constitutive parameters are established;
s6, training a two-channel convolution neural network model: 10000 IPF groups in step S5 S Taking the graph and a true stress-true strain graph as input of a training dual-channel convolutional neural network model, taking t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16 of the 10000 groups in the step S5 as output of the neural network, manufacturing the dual-channel convolutional neural network model, and randomly dividing the model into 9000 training sets and 1000 verification sets according to a proportion of 9; the two-channel convolution neural network model comprises: 5 layers of alxnet model convolution layer, 3 full connection layers; the convolution layer is as follows: 9000 sets of IPFs in a training set S The graph and the real stress-real strain curve are input into the convolution layer, after convolution operation of the convolution layer, the image is abstracted into a characteristic graph, 9000 groups of IPF S 9000 characteristic diagrams 1 are obtained after 5 layers of convolutional layers are processed, 9000 characteristic diagrams 2 are obtained after 9000 true stress-true strain graphs are processed by 5 layers of convolutional layers, and the 9000 characteristic diagrams 1 and 9000 characteristic diagrams 2 are merged and input into full-connection layer processing; the full connection layer treatment comprises the following steps: 9000 groups of output values are obtained by calculation through a forward propagation formula containing weights and offsets, wherein the weights and the offsets are random, the weight range is 1-2, the offset range is 10-20, and each group of output values is as follows: overall errors between the 9000 sets t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16' and the 9000 sets t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16' in the training set and t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15, t16 in the training set are calculated; then, recalculating the weight and the bias according to the integral error, then, recalculating by adopting a forward propagation formula again, and repeating the calculation until the integral error is 0.2-1%, and finishing the training to obtain a trained dual-channel convolution neural network model Z; the initial learning rate of the dual-channel convolutional neural network model Z is set to be 0.00001, and the learning times are 3000; two-way trainingThe integral error of the per-pass convolutional neural network model Z for 9000 groups of training sets is 0.2% -0.8%, and the integral error of the trained dual-pass convolutional neural network model Z for 1000 groups of verification sets is 0.4% -0.9%;
the overall error formula is as follows:
the sigma is the integral error of the error,
nu is a positive integer and takes the value of 1-16,
t′ v is one of t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16',
t v is one of t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16,
the integral error and weight value and bias calculation formula is as follows:
σ＝W T +b
the sigma is the integral error of the error,
w is a weight value,
t is the transposed symbol, and T is the transposed symbol,
b is an offset.
Example 2
The universality of the two-channel convolution neural network model Z obtained by training in the embodiment 1 is verified through experiments;
s1, selecting a common magnesium alloy in a laboratory: AZ91, shooting an EDSD picture, selecting a magnesium alloy IPF picture, cutting the IPF picture into a standard 50x50 pixel IPFs picture by using Python, obtaining an anisotropy coefficient R and a section reduction rate of the AZ91 magnesium alloy through experiments, then carrying out a uniaxial tension experiment on the material to obtain true stress-true strain coordinate data, and drawing the true stress-true strain coordinate data into a true stress-true strain curve graph through the Python;
s2, inputting the IPFs diagram and the true stress-true strain graph obtained in step S1 into the dual-channel convolutional neural network model Z trained in embodiment 1, outputting a calibrated set of t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15, and t16 by the neural network, and converting the neural network output calibration value into 16 crystal plasticity constitutive parameters by an inverse normalization method: t1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15, T16;
s3, establishing a crystal plasticity simulation file by 16 crystal plasticity constitutive parameters T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 obtained by the neural network calibration in the step S2 and an IPFs diagram obtained by the experiment in the step S1, inputting the crystal plasticity simulation file into DAMASK simulation software to carry out crystal plasticity simulation to obtain a simulation result, and processing the simulation result by Python to obtain a simulated AZ91 magnesium alloy anisotropy coefficient R' and a section reduction rate;
s4, carrying out error analysis on the simulated AZ91 magnesium alloy anisotropy coefficient R' and the section reduction rate obtained in the step S3 and the AZ91 magnesium alloy anisotropy coefficient R and the section reduction rate obtained in the step S1, wherein the results show that: the parameter fitting degree is 99.5-99.7%, so that the accuracy and universality of calibrating the 16 crystal plastic constitutive parameters by adopting the double-channel convolutional neural network are demonstrated.
To sum up: the invention applies the convolution neural network used in the macroscopic field to the microscopic crystal structure parameter calibration field, the obtained double-channel convolution neural network can accurately and synchronously calibrate 16 parameters of the crystal plasticity constitutive structure at one time, and the calibration speed and the calibration accuracy are superior to those of the calibration method disclosed in the prior art. In addition, the successfully trained neural network model has universality, is applicable to a magnesium alloy system, can accurately and synchronously calibrate the crystal plasticity constitutive parameters of various magnesium alloys of the magnesium alloy system at one time, does not need to repeatedly establish a model database and repeatedly train a convolutional neural network, and simplifies the calculation process.
Claims (2)
1. A crystal plasticity constitutive parameter calibration method based on a two-channel convolution neural network is characterized by comprising the following steps: it comprises the following steps:
(1) The magnesium alloy model was established using the Neper software: inputting m groups of parameters x and y into Neper software, wherein x and y respectively represent the random distribution and the crystal grain roundness of crystal grains, the value range of m is 100-100000 and is an integer, the value range of x is 0-1, the value range of y is 0.3-1.0, m groups of magnesium alloy models are obtained and are marked as m groups (x, y), and the model of each group (x, y) is shown as a matrix 1;
matrix 1
i is one of 1, 2 or 3;
(2) A magnesium alloy model with added texture was established using Matlab kit MTEX: setting m groups of d, e, f, g, h and k parameters, d, e, f, g, h and k respectively representing six different texture strengths of the magnesium alloy in the step (1), wherein the numeric area of d, e, f, g, h and k is 0-1, and d + e + f + g + h + k =1; inputting m groups of d, e, f, g, h and k parameters and m groups of (x, y) obtained in the step (1) into MTEX, and obtaining m groups of magnesium alloy models (d, e, f, g, h and k) (x, y) with texture added through the Bunge formula operation of MTEX, wherein the m groups of magnesium alloy models are marked as m groups of (d, e, f, g, h and k) (x, y), and data models of each group of (d, e, f, g, h and k) (x, y) are shown as a matrix 2;
matrix 2
(3) Using Python to obtain m sets of crystal plastic simulation files: selecting 16 parameters of a crystal plasticity constitutive structure: basal plane slippage critical slitting stress T1, cylindrical surface slippage critical slitting stress T2, conical surface < a > slippage critical slitting stress T3, conical surface < c + a > slippage critical slitting stress T4, basal plane slippage saturated slippage slitting stress T5, cylindrical surface slippage saturated slippage slitting stress T6, conical surface < a > slippage saturated slippage slitting stress T7, conical surface < c + a > slippage saturated slippage slitting stress T8, stretching twin crystal critical slitting stress T9, compressing twin crystal critical slitting stress T10, twin and twin initial hardening modulus T11, slippage and slippage initial hardening modulus T12, slippage and twin initial hardening modulus T13, sliding strain rate sensitive parameter T14, twin strain rate sensitive parameter T15 and sliding hardening parameter T16; t1 ranges from 5 to 80, T2 ranges from 10 to 150, T3 ranges from 20 to 200, T4 ranges from 50 to 500, T5 ranges from 10 to 160, T6 ranges from 20 to 300, T7 ranges from 40 to 400, T8 ranges from 100 to 1000, T9 ranges from 5 to 250, T10 ranges from 10 to 400, T11 ranges from 10 to 500, T12 ranges from 50 to 3000, T13 ranges from 50 to 1200, T14 ranges from 1 to 15, T15 ranges from 5 to 80, and T16 ranges from 0.5 to 10; designing m groups of T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 in the range, adding each group of T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 to the group of magnesium alloy models (d, e, f, g, h, i) (x and y) in the step (2) through Python to obtain a group of crystal plastic simulation files, and finally obtaining m groups of crystal plastic simulation files, wherein each group of crystal plastic simulation files is in a form shown in a matrix 3;
matrix 3
(4) IPF plots and simulated true stress-true strain coordinate data were obtained by simulation: adopting DAMASK as crystal plasticity finite element simulation software, inputting the crystal plasticity simulation file obtained in the step (3) to perform crystal plasticity finite element simulation, inputting m groups of crystal plasticity simulation files in the step (3) into the DAMASK software to perform m groups of simulation, following the crystal plasticity constitutive equation set in the DAMASK simulation, obtaining m groups of simulation results after the simulation is completed, and performing post-processing on the m groups of simulation results by using Python to obtain an IPF (in-plane stress) -true strain coordinate data of m groups of magnesium alloys and m groups of simulated true stress-true strain coordinate data;
the crystal plasticity constitutive equation set is as follows:
n is the initial hardening modulus of the tensile and compressive twins, n is one of T9 and T10,
q αβ is a coefficient between the different slip systems alpha and beta, q αβ Is one of T14, T15 and T16,
H αβ total hardening modulus;
(5) IPFs plots and true stress-true strain plots were obtained using Python and the crystal plasticity constitutive parameters were normalized: respectively normalizing the m groups of 16 crystal plasticity constitutive parameters T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 in the step (3) to obtain m groups of normalized T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16, drawing the m groups of simulated true stress-true strain coordinate data in the step (4) into m groups of true stress-true strain graphs by using Python, and cutting the m groups of IPF graphs in the step (4) into m groups of standard IPF graphs of standard 50x50 pixels, and marking the m groups of standard IPF graphs as m groups of IPFs graphs; finally obtaining m groups of crystal plasticity data sets, wherein each group of data sets comprises: IPFs diagrams, true stress-true strain graphs, and normalized t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15, t16;
the normalization formula is as follows:
W norm is one of T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 parameters, W is one of T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15 and T16 parameters,
W min is the smallest number of the m sets of parameters,
W max in m sets of parametersMaximum number;
therefore, an IPFs graph, a true stress-true strain curve graph and a big data sample corresponding to the normalized constitutive parameters are established;
(6) Training a two-channel convolutional neural network model: subjecting the m sets of IPFs of step (5) to S Taking the graph and a true stress-true strain graph as input of a training dual-channel convolutional neural network model, taking t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16 of the m groups in the step (5) as output of the neural network, manufacturing the dual-channel convolutional neural network model, and randomly dividing the model into p training sets and q verification sets according to a proportion of 5-30; the two-channel convolution neural network model comprises: 5 layers of alxnet model convolution layer, 3 full connection layers; the convolution layer is as follows: grouping p IPFs in a training set S Inputting the graph and the real stress-real strain curve graph into the convolution layer, abstracting the image into a characteristic graph after convolution operation of the convolution layer, and p groups of IPFs S P groups of characteristic diagrams 1 are obtained after 5 layers of convolution layer processing, p groups of real stress-real strain graphs obtain p groups of characteristic diagrams 2 after 5 layers of convolution layer processing, and the p groups of characteristic diagrams 1 and the p groups of characteristic diagrams 2 are merged and input into full connection layer processing; the full connection layer treatment comprises the following steps: p groups of output values are obtained through calculation of a forward propagation formula containing weights and biases, wherein the weights and the biases are random, the range of the weights is 1-2, the range of the biases is 10-20, and each group of output values is as follows: the overall error between the p groups t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16' is then calculated, and the p groups t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16' and the p groups t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15, t16 in the training set; then, recalculating the weight and the bias according to the integral error, then, recalculating by adopting a forward propagation formula again, and performing reciprocating circulation until the integral error is 0.2-1%, and finishing training to obtain a trained dual-channel convolution neural network model; wherein an initial learning rate of the two-channel convolutional neural network model is set to0.00001-0.01, and the learning times are 1000-3000; the overall error of the trained dual-channel convolutional neural network model to the p groups of training sets is 0.2% -1%, and the overall error of the trained dual-channel convolutional neural network model to the q groups of verification sets is 0.4% -1.1%;
the overall error formula is as follows:
p is the number of training set groups,
the sigma is the overall error of the signal,
nu is a positive integer and takes the value of 1-16,
t′ v is one of t1', t2', t3', t4', t5', t6', t7', t8', t9', t10', t11', t12', t13', t14', t15', t16',
t v is one of t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14, t15 and t16,
the integral error and weight value and bias calculation formula is as follows:
σ＝W T +b
the sigma is the overall error of the signal,
w is a weight value,
t is the transposed symbol and is the symbol,
b is an offset.
2. The method for calibrating the crystal plasticity constitutive parameters based on the dual-channel convolutional neural network as claimed in claim 1, wherein: the magnesium alloy is one of AZ31, AZ61, AZ91, ZK60 or AM 60.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211332683.1A CN115588470A (en) | 2022-10-28 | 2022-10-28 | Crystal plasticity constitutive parameter calibration method based on two-channel convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211332683.1A CN115588470A (en) | 2022-10-28 | 2022-10-28 | Crystal plasticity constitutive parameter calibration method based on two-channel convolutional neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115588470A true CN115588470A (en) | 2023-01-10 |
Family
ID=84782078
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211332683.1A Pending CN115588470A (en) | 2022-10-28 | 2022-10-28 | Crystal plasticity constitutive parameter calibration method based on two-channel convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115588470A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117877642A (en) * | 2024-01-16 | 2024-04-12 | 华中科技大学 | Weld joint crystal plastic model constitutive parameter calibration method and system |
-
2022
- 2022-10-28 CN CN202211332683.1A patent/CN115588470A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117877642A (en) * | 2024-01-16 | 2024-04-12 | 华中科技大学 | Weld joint crystal plastic model constitutive parameter calibration method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110084221B (en) | Serialized human face key point detection method with relay supervision based on deep learning | |
| US20210132904A1 (en) | Apparatus and methods for neural network operations supporting floating point numbers of short bit length | |
| Yamanaka et al. | Deep neural network approach to estimate biaxial stress-strain curves of sheet metals | |
| Sun et al. | Fast implementation of DeLong’s algorithm for comparing the areas under correlated receiver operating characteristic curves | |
| CN109933925B (en) | Method for predicting stamping forming performance of metal plate | |
| CN108985335B (en) | Integrated learning prediction method for irradiation swelling of nuclear reactor cladding material | |
| CN112052951B (en) | Pruning neural network method, system, equipment and readable storage medium | |
| CN100507460C (en) | Dynamic soft measuring and form establishing method base pulse response formwork and parameter optumization | |
| WO2014141794A1 (en) | Stress-strain relationship simulation method, spring back prediction method, and spring back analyzing device | |
| CN111539071B (en) | Method and system for establishing differential plate crystal plasticity constitutive model and electronic equipment | |
| CN105046277A (en) | Robust mechanism research method of characteristic significance in image quality evaluation | |
| CN115588470A (en) | Crystal plasticity constitutive parameter calibration method based on two-channel convolutional neural network | |
| CN112560967B (en) | Multi-source remote sensing image classification method, storage medium and computing device | |
| CN107292029A (en) | A kind of determination method that sheet forming technological parameter is predicted based on forming defects | |
| CN114330042A (en) | Program load spectrum compiling method and system based on SN curve and storage medium | |
| CN110647990A (en) | Cutting method of deep convolutional neural network model based on grey correlation analysis | |
| JP2000275154A (en) | Method for simulating relationship between stress and distortion | |
| Koenuma et al. | Estimation of texture-dependent stress-strain curve and r-value of aluminum alloy sheet using deep learning | |
| CN106372348A (en) | Vector fitting model order reduction method based on error control in linear system | |
| CN110750926A (en) | Particle swarm algorithm-based high-speed tensile curve processing and predicting method | |
| CN112149351B (en) | Microwave circuit physical dimension estimation method based on deep learning | |
| CN106682328B (en) | Vibration deformation measurement calculation method for vertical high-rise structure vibration isolation system | |
| CN115620836A (en) | Optimization method for constitutive parameters of mixed hardened material model | |
| CN115310472A (en) | Nuclear pulse peak sequence-based one-dimensional convolution neural network nuclide identification method | |
| Nadarajah et al. | Characterizations of the Weibull and uniform distributions using record values |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |