A novel inverse design method for super fiber reinforced hydrogel structure based on GAN
Through the GAN-based reverse design method, the problem of insufficient mechanical properties of traditional hydrogels is solved. The multifiber hydrogel structure is predicted by using the generative adversarial network, and efficient design in the fields of drug delivery and flexible electronics is achieved, significantly reducing calculation costs.
Patent Information
- Application Number
- CN202410268949.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-08
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-03-08
AI Technical Summary
Traditional hydrogels lack mechanical strength and toughness in engineering applications, and the existing technology lacks deep learning methods in the structural design of ultrafiber reinforced hydrogels, resulting in high-computing and time-consuming design strategies.
Using a GAN-based reverse design method, the target mechanical properties such as stress and displacement fields are achieved by generating an adversarial network to optimize the fiber distribution by predicting the structure of the hyperfiber hydrogel composite at a given stress or displacement field.
It has huge application potential in the fields of drug delivery and flexible electronics, significantly reducing computing costs, and achieving faster and more accurate composite structural design.
Smart Images

Figure CN118053526B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of super fiber reinforced hydrogel structure design, and in particular to a GAN-based inverse design method for a novel super fiber reinforced hydrogel structure. Background Art
[0002] Hydrogels have attracted great attention in soft robotics, biomedical engineering, and flexible electronic engineering due to their high recoverability, good biocompatibility, and sensitivity to environmental stimuli. However, traditional hydrogels generally have the defects of poor mechanical strength and toughness, which limits their engineering applications.
[0003] To solve the above problems, the prior art improves the mechanical properties of pure hydrogels by integrating different types of fibers into the hydrogel matrix, in order to pursue high strength and toughness and obtain anisotropy in the field of mechanics. Therefore, the prior art proposes the following series of constitutive models to describe the anisotropic deformation of fiber-reinforced hydrogels during swelling:
[0004] Liu Y, Zhang H, Wang J, and Zheng published in "Anisotropic Swelling in Fiber-Reinforced Hydrogels: An Incremental Finite Element Method and Its Applications in Design of Bilayer Structures" a unified intrinsic model of anisotropic fiber-reinforced hydrogels was derived by considering the effect of fiber swelling in the Flory-Rehner theory.
[0005] Liu Y, Zhang H, Zhang J, and Zheng Y published in "Constitutive modeling for polymerhydrogels: A new perspective and applications to anisotropic hydrogels in freeswelling" that an anisotropic constitutive model was derived through the second law of thermodynamics and kinematic constraints between physical and chemical fields.
[0006] Motiwale S, Russell MD, Conroy O, Carruth J, Wancura M, Robinson A, et al. disclose anisotropic elastic behavior of replacement heart valves.
[0007] In addition, Yang Q, Ma L, and Shang J disclosed in "The chemo-mechanical coupling behavior of hydrogels incorporating entanglements of polymer chains" that the chemical-mechanical coupling model was developed by considering the general form of the free energy function of chain entanglement.
[0008] At the same time, it is known from common knowledge in the art that among all fiber types in hydrogel composites, superfibers are a type of supermaterial, which is an artificial structural material used to control and manipulate mechanical properties. Compared with continuous fibers, the mechanical properties of superfibers are derived from the geometric arrangement of their constituent materials, such as multi-directional deformation, adjustable stiffness, negative compressibility, and negative Poisson's ratio. Compared with continuous fibers with the same material parameters, the two types of superfibers produce a two-stage deformation process, during which the superfibers are first straightened and then further stretched. Therefore, superfiber-reinforced hydrogels can enhance the overall mechanical properties by reducing stress concentration after deformation, providing a broader design space for the adjustable mechanical properties of composite materials.
[0009] In addition, the optimal design of hydrogel composites usually relies on the quantitative analysis of mechanical fields such as stress and displacement fields to avoid mechanical failure.
[0010] For example, "A soft pneumatic dexterous gripper with convertible grasping modes.", "Integration of sensing and shape-deforming capabilities for a bioinspired soft robot", and "A comparative study on cylindrical and spherical models in fabrication of bone tissue engineering scaffolds: Finite elements simulation and experiments" respectively disclosed that the stress distribution was designed for the optimized structure of hydrogel-based finger-like grippers, soft robots, and artificial scaffolds.
[0011] Yiming B, Liu T, Nian G, Han Z, Jia Z, and Qu S disclosed in "Mechanics-guided design of shape-morphing composite sheets with hard and soft materials" that finger-like grippers, soft robots, and artificial supports use displacement fields to characterize complex shape transformations for numerically studying the dehydration process of blades.
[0012] It can be seen that traditional structural design strategies, such as topology optimization and trial and error, have the disadvantages of being computationally intensive and time-consuming.
[0013] As another approach, deep learning (DL) methods provide a perspective to explore the mechanical properties of composite materials and reverse design composite structures by learning big data from simulations or experiments. In addition, compared with finite element (FE) methods and molecular dynamics simulations, it is beneficial to solve time-consuming and computationally inefficient problems in complex tasks. Many types of deep learning methods have been applied in solid mechanics. For example, U-Net is used to estimate the local strain field of elastic composites; convolutional neural networks (CNNs) are used to predict the mechanical properties of fiber-reinforced composites. Some researchers have also disclosed data sets obtained through numerical simulations based on multi-scale hydrogel fracture models, and used CNN to predict the fracture behavior of hydrogels; two neural network models, namely generators and discriminators, are trained simultaneously using generative adversarial networks (GANs) for composite materials to generate new structures with target mechanical properties, predict microscale elastic strain fields, and topology optimization.
[0014] However, there is currently a lack of research published on the use of deep learning to design super-fiber reinforced hydrogel structures. Summary of the invention
[0015] To solve the above problems, the present invention proposes a novel inverse design method of super fiber reinforced hydrogel structure based on GAN. Under the premise of given stress or displacement field, the structure of super fiber hydrogel composite material can be effectively predicted by generative adversarial network to achieve target mechanical properties, such as stress field and displacement field, thus having great application potential in the fields of drug delivery and flexible electronics.
[0016] To achieve the above object, the present invention provides a reverse design method of a novel super fiber reinforced hydrogel structure based on GAN, comprising the following steps:
[0017] S1. Data generation:
[0018] Under chemical and temperature boundary conditions, finite element simulation is performed on the super fiber reinforced hydrogel composite material to obtain the solid mechanics model of the super fiber reinforced hydrogel, so as to obtain the physical field data and fiber distribution data of the super fiber reinforced hydrogel composite material after free expansion, and construct the physical field image data set and fiber distribution image data set respectively;
[0019] S2. Data preprocessing:
[0020] Normalizing the physical field image dataset and the fiber distribution image dataset respectively, and then adjusting the sizes of the images in the physical field image dataset and the fiber distribution image dataset respectively, to construct a dataset for a deep learning model;
[0021] S3. Training Generative Adversarial Network:
[0022] Inputting the physical field information contained in the physical field image data set preprocessed by step S2 into the generative adversarial network, and using the fiber distribution information contained in the fiber distribution image data set preprocessed by step S2 as output data, modifying the parameters between the generator and the discriminator of the generative adversarial network to train the generative adversarial network;
[0023] S4. Verify and use the verified generative adversarial network to design the fiber distribution according to the target physical field.
[0024] Preferably, the physical field described in step S1 includes a stress field and a displacement field.
[0025] Preferably, in step S1, a two-dimensional super fiber reinforced hydrogel composite material model is established, and the free expansion process of the super fiber reinforced hydrogel composite material is caused by simulating the changes in temperature and the concentration of anions and cations ionized by the electrolyzable solute in the external solution, which specifically includes the following steps:
[0026] Assumptions: There is no overlap between the superfibers; the hydrogel composite material is completely swollen; the hydrogel matrix and the superfibers are completely bonded, and the solute in the solution is ionized to produce an ionic solution with an anion-cation concentration ratio of 1:1;
[0027] Under the premise of the established assumptions, the hydrogel matrix is placed in an ionic solution with a set temperature of 320K and a relative ion concentration of 0.03, and then a random number of superfibers with two-dimensional shapes, random distributions and random orientations are embedded to obtain a merged boundary, and it is assumed that the superfibers passing through the merged boundary are cut into segments; then a chemical boundary condition of a relative ion concentration of 0.03 and a temperature boundary condition of 320K are applied to the entire merged boundary. In order to make the hydrogel only swell in the ionic solution, a boundary condition of zero displacement is set at the bottom of the hydrogel.
[0028] Preferably, the superfiber is a linear structural fiber with a sinusoidal structure, which is defined as an isotopic linear elastic solid with a Poisson's ratio and an elastic modulus E, and the total height H and width t of the superfiber in a free state are given the same initial value;
[0029] The shape y of the superfiber is defined by the following sinusoidal function:
[0030]
[0031] Among them, A, ω and They represent the amplitude, angular frequency and phase shift of the sine function respectively.
[0032] Preferably, the process of establishing the hydrogel matrix constitutive model is as follows:
[0033] First, the expansion is simulated by the mixing free energy function based on the Flory-Rehner free energy function, and the chemical field and temperature field are considered to construct the following total mixing free energy density function:
[0034]
[0035] Where T represents temperature, ΔS d represents the entropy change during the expansion process, ΔS s represents the elastic entropy, ΔH mix represents the enthalpy of the hydrogel, t represents the elastic energy density function of the chemical cross-linked chain network, R represents the universal gas constant, f represents the degree of entanglement between the chemical cross-linked chains, N c represents the molar number of chemical cross-linking chains, J = λ1λ2λ3, λ i (i=1,2,3) represents the principal elongation, C s represents the number of moles of solvent molecules in the ionic solution, υ s represents the molar volume of the solvent, N s represents the number of moles of slippable connections between molecular chains, η represents the slip point parameter, and χ represents the dimensionless enthalpy of mixing;
[0036] Then, the total mixing free energy density function is converted into the following principal invariant by Legendre transformation:
[0037]
[0038] Where F represents the deformation gradient tensor, μ represents the chemical potential, and I i (i=1,2,3) represents the main invariant;
[0039] Then, given the initial temperature, chemical potential, and free elongation, the relationship between chemical potential and free elongation is derived:
[0040]
[0041] Then, the chemical potential between the hydrogel matrix and the solvent is assumed to be equal in the initial state, where the chemical potential in the solvent is approximated with the ideal solution assumption by the following equation:
[0042]
[0043] in, is the sum of the initial ionic species concentrations of the solvent;
[0044] Finally, the free elongation λ0 and the initial ion concentration of the solvent are obtained. The relationship between:
[0045]
[0046] The hydrogel matrix of the super fiber reinforced hydrogel composite material was modeled by the total mixing free energy function in formula (2) to characterize the deformation of the hydrogel matrix under chemical-mechanical coupling conditions, and numerical simulation was performed using the finite element software ABAQUS.
[0047] Preferably, step S2 specifically includes the following steps:
[0048] S21, by using the simulation results of the free swelling process of the composite material deposited by the finite element software ABAQUS, the distribution of super fibers in the super fiber reinforced hydrogel composite material before swelling and the physical field distribution of the super fiber reinforced hydrogel composite material after swelling are obtained;
[0049] S22, grayscale processing, and compressing the images of the composite physical field distribution and the fiber distribution into an image with a size of 256×256 pixels;
[0050] S23, normalization processing:
[0051] Normalize in the range [-1,1] using the min-max method:
[0052]
[0053] Among them, z i represents the data after the i-th transformation, x(x1,x2,…x n ) represents the grayscale value in the image.
[0054] Preferably, step S3 specifically includes the following steps:
[0055] S31. Construct a generative adversarial network, which consists of a generator and a discriminator, wherein the generator is a U-Net neural network used to learn the mapping between the super fiber distribution and the physical field, and the discriminator is used to distinguish between the real image in the data set input into the generative adversarial network and the structure image generated by the generator during the training process;
[0056] S32. Optimizing Generative Adversarial Networks:
[0057] The optimization problem of the generative adversarial network is regarded as a minimax problem, where the objective function of the generative adversarial network is defined as:
[0058]
[0059] Among them, P d a t a (m) and P z (z) is the pixel value matrix sampled from the stress field, D represents the discriminator, G represents the generator, represents the expected value of the true sample distribution, represents the expected value of the generated sample distribution, and G(z) represents the generated sample;
[0060] S33. The generator is trained to produce abnormal images, and the discriminator distinguishes the abnormal images from the real images and returns the difference to the generator. The training cycle is repeated until the generator and the discriminator satisfy the Nash equilibrium state, at which time the generative adversarial network algorithm is determined to converge.
[0061] Preferably, the generator described in step S31 includes an encoder and a decoder, and there are jump connections between the image layers;
[0062] The fitting process of the generator is as follows: the encoder extracts important features from the input data, namely the stress field after free expansion, and then uses the convolution kernel to reduce the dimension to obtain feature data. The feature data is then input into the decoder for deconvolution calculation to achieve feature reorganization and achieve the purpose of structure generation. Batch Normalization is used in the fitting process to avoid overfitting of the generator.
[0063] The loss function of the generator is defined as:
[0064] L gen =L gan +λL l1 (9)
[0065] Among them, L gan represents the loss between the generated adversarial network and the array zero point, λ=100 represents the constant coefficient, L l1 It represents the mean absolute error between abnormal images and true images.
[0066] Preferably, the domain function component of the discriminator described in step S31 is a convolutional neural network, wherein the first layer receives the real fiber distribution in the super fiber reinforced hydrogel composite material or the super fiber distribution predicted from the generator from the data set simulated and generated by the finite element software ABAQUS;
[0067] The loss function of the discriminator is defined as
[0068] L dis =L real +L fake (10)
[0069] Among them, L real Represents the loss between the normalized real image of the calculation data set and the identity matrix, L fake represents the loss between the generator’s abnormal images and the zero matrix.
[0070] Preferably, in step S4, the verification of the generative adversarial network is quantitatively evaluated by calculating the mean square error between the real fiber distribution and the fiber distribution predicted by the generative adversarial network and the corresponding physical field respectively;
[0071] The mean square error calculation formula is as follows:
[0072]
[0073] Where M and N represent the width and height of the image respectively, and f′ ij represents the abnormal image generated by the generator, f ij Represents a real image.
[0074] The present invention has the following beneficial effects:
[0075] Under the premise of a given stress or displacement field, the generative adversarial network can be used to effectively predict the structure of superfiber hydrogel composites, which has great application potential in the fields of drug delivery and flexible electronics.
[0076] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0077] Figure 1 It is a flow chart of the reverse design method of the novel super fiber reinforced hydrogel structure based on GAN of the present invention;
[0078] Figure 2 A diagram showing the swelling process of a hydrogel composite material under chemical and temperature boundary conditions according to the inverse design method of a novel super fiber reinforced hydrogel structure based on GAN of the present invention;
[0079] Figure 3 A diagram showing the geometric shape of the fiber used in the superfiber simulation experiment of the present invention;
[0080] Figure 4 This is a fiber stress comparison diagram of the super fiber simulation experiment of the present invention;
[0081] Figure 5 It is a diagram of the parameter estimation of the hydrogel matrix by finite element simulation in the hydrogel matrix simulation experiment of the present invention, and the mechanical field result diagram of the super fiber reinforced hydrogel composite material after free expansion, wherein Figure 5 (a) is the Young's modulus estimation result of the hydrogel matrix; Figure 5 (b) is a schematic diagram of the loading conditions (c0=0.03, T0=320) in the finite element simulation; Figure 5 (c) is the swelling process of the hydrogel matrix and the corresponding mechanical field results; Figure 5 (d) Four representative simulation results of stress (unit: MPa) and displacement (unit: mm) fields of superfiber reinforced hydrogel composites with different superfiber numbers;
[0082] Figure 6 This is a diagram of the architecture of the generative adversarial network described in the present invention;
[0083] Figure 7 This is a performance simulation result diagram of the GAN model based on the stress field according to an embodiment of the present invention; wherein, Figure 7 (a) is the loss function diagram of GAN during training; Figure 7 (b) is the MSE result diagram of GAN between the predicted structure image and the output image in the validation dataset; Figure 7 (c) is the fiber volume fraction V of the predicted structure image f Regression analysis chart of Figure 7 (d) Four representative examples of stress field predicted structural images;
[0084] Figure 8 This is a performance simulation result diagram of the GAN model based on the displacement field according to an embodiment of the present invention; wherein, Figure 8 (a) is the MSE result diagram of GAN between the predicted structure image and the output image in the validation dataset; Figure 8 (b) is the regression analysis diagram of the fiber volume fraction Vf of the predicted structural image; Figure 8 (c) Four representative examples of stress field prediction structures;
[0085] Fig. 9 is a schematic diagram of application example 1 of the present invention; wherein, Fig. 9 (a) is the input image; Fig. 9 (b) is the predicted structure image; Fig. 9(c) is the Mises stress cloud diagram; Fig. 9 (d) is the Mises stress variation diagram of the damage point;
[0086] Fig.10 This is an example diagram of Application Example 1 of the present invention;
[0087] Fig.11 is a schematic diagram of Application Example 2 of the present invention; wherein, Fig.11 (a) is the input image; Fig.11 (b) is the predicted structure image; Fig.11 (c) is the displacement cloud map;
[0088] Fig.12 This is an example diagram of Application Example 2 of the present invention. DETAILED DESCRIPTION
[0089] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions.
[0090] It should be noted that the terms "include", "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or devices.
[0091] Like reference numerals and letters denote similar items in the following drawings, and thus, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0092] like Figure 1 As shown, the inverse design method of the novel super fiber reinforced hydrogel structure based on GAN includes the following steps:
[0093] S1. Data generation:
[0094] Under chemical and temperature boundary conditions, finite element simulation is performed on the super fiber reinforced hydrogel composite material to obtain the solid mechanics model of the super fiber reinforced hydrogel, so as to obtain the physical field data and fiber distribution data of the super fiber reinforced hydrogel composite material after free expansion, and construct the physical field image data set and fiber distribution image data set respectively;
[0095] The physical field described in step S1 includes a stress field and a displacement field.
[0096] like Figure 2 As shown, in step S1, a two-dimensional super fiber reinforced hydrogel composite material model is established, and the free expansion process of the super fiber reinforced hydrogel composite material is caused by simulating the changes in the temperature and the concentration of anions and cations (such as sodium chloride, calcium chloride solution, ferric nitrate, etc., and this method is not limited to fixing a certain anion and cation) ionized by the electrolyzable solute in the external solution, which specifically includes the following steps:
[0097] Assumptions: There is no overlap between the superfibers; the hydrogel composite material is completely swollen; the hydrogel matrix and the superfibers are completely bonded, and the solute in the solution is ionized to produce an ionic solution with an anion-cation concentration ratio of 1:1;
[0098] Under the premise of the established assumptions, the hydrogel matrix is placed in an ionic solution with a set temperature of 320K and a relative ion concentration of 0.03, and then a random number of superfibers with two-dimensional shapes, random distributions and random orientations are embedded to obtain a merged boundary, and it is assumed that the superfibers passing through the merged boundary are cut into segments; then a chemical boundary condition of a relative ion concentration of 0.03 and a temperature boundary condition of 320K are applied to the entire merged boundary. In order to make the hydrogel only swell in the ionic solution, a boundary condition of zero displacement is set at the bottom edge of the hydrogel.
[0099] According to the inverse design method of the novel super fiber reinforced hydrogel structure based on GAN, it is characterized in that: the super fiber is a linear structural fiber with a sinusoidal structure, which is defined as a linear elastic solid with a Poisson's ratio and an elastic modulus E, and the total height H and width t of the super fiber in a free state are given the same initial value;
[0100] The shape y of the superfiber is defined by the following sinusoidal function:
[0101]
[0102] Among them, A, ω and They represent the amplitude, angular frequency and phase shift of the sine function respectively.
[0103] Super fiber simulation experiment
[0104] In this simulation experiment, two types of super fibers with periods of 7.5 and 15 are used. The two types of super fibers are defined as having different angular frequencies, for example, ω1=3.75π and ω2=7.5π.
[0105] like Figure 3 and Figure 4 As shown, both types of superfibers produce a two-stage deformation process, during which the fibers are first straightened and then further stretched. The transition behavior of long-period superfibers (period is 15) appears earlier than that of short-period superfibers (period is 7.5). In addition, under the same displacement boundary conditions, short-period superfibers bear less load than long-period superfibers. This proves that the mechanical properties of superfibers in hydrogel composites have a flexible design space
[0106] The process of establishing the hydrogel matrix constitutive model is as follows:
[0107] First, the expansion is simulated by the mixing free energy function based on the Flory-Rehner free energy function, and the chemical field and temperature field are considered to construct the following total mixing free energy density function:
[0108]
[0109] Where T represents temperature, ΔS d represents the entropy change during the expansion process, ΔS s represents the elastic entropy, ΔH mix represents the enthalpy of the hydrogel, t represents the elastic energy density function of the chemical cross-linked chain network, R represents the universal gas constant, f represents the entanglement between the chemical cross-linked chains, N c represents the molar number of chemical cross-linking chains, J = λ1λ2λ3, λ i (i=1,2,3) represents the principal elongation, C s represents the number of moles of solvent molecules in the ionic solution, υ s represents the molar volume of the solvent, N s represents the number of moles of slippable connections between molecular chains, η represents the slip point parameter, and χ represents the dimensionless enthalpy of mixing;
[0110] Then, the total mixing free energy density function is converted into the following principal invariant by Legendre transformation:
[0111]
[0112] Where F represents the deformation gradient tensor, μ represents the chemical potential, and I i (i=1,2,3) represents the main variation;
[0113] Then, given the initial temperature, chemical potential μ0 and free elongation λ0, the relationship between chemical potential and free elongation is derived:
[0114]
[0115] Then, the chemical potential between the hydrogel matrix and the solvent is assumed to be equal in the initial state, where the chemical potential in the solvent is approximated with the ideal solution assumption by the following equation:
[0116]
[0117] in, is the sum of the initial ionic species concentrations of the solvent;
[0118] Finally, the free elongation λ0 and the initial ion concentration of the solvent are obtained. The relationship between:
[0119]
[0120] The hydrogel matrix of the super fiber reinforced hydrogel composite material was modeled by the total mixing free energy function in formula (2) to characterize the deformation of the hydrogel matrix under chemical-mechanical coupling conditions, and numerical simulation was performed using the finite element software ABAQUS.
[0121] Simulation experiment of hydrogel swelling
[0122] In this simulation experiment, the stretching process of the standard dog-bone sample of the hydrogel matrix was simulated by the finite element software ABAQUS, and based on the user subroutine UHYPER of ABAQUS, the Person correlation coefficient between the numerical simulation and the experimental results was 0.997, indicating that the user subroutine has satisfactory consistency in quantitatively describing the constitutive model of the hydrogel matrix.
[0123] Meanwhile, the superfiber reinforced hydrogel composites underwent a swelling process under constant chemical and thermal boundary conditions, where the horizontal span L of the superfiber was 30 mm and the thickness was 1 mm. As for the superfibers, it was assumed that the Young's modulus of the isotopic elastic fibers was 9 MPa, the Poisson's ratio was 0.3, and they were randomly distributed in the hydrogel matrix without overlap. In order to generate identifiable stress or displacement field results by the adversarial network, two different shapes of superfibers with periods of 15 and 7.5 were designed, respectively, and the total number of superfibers ranged from 5 to 20. The average Young's modulus of the hydrogel matrix was approximated to 900 kPa by the Dog-bone model tensile test. Other initial simulation constants are listed in Table 1.
[0124] Table 1 Input parameters of the hydrogel matrix in the UHYPER subroutine
[0125]
[0126] A 100mm×100mm super fiber reinforced hydrogel composite was placed at the bottom of the boundary with fixed boundary conditions. The swelling process of the super fiber reinforced hydrogel composite was driven by a temperature drop from 320K to 300K and a dilute ion concentration of 0.03 to 0.015. Finite element analysis was performed using CPE4H units, and the validity of the simulation results of the swelling process was verified by testing the unit size and number. The mechanical field of the swollen super fiber reinforced hydrogel composite was then obtained using UHYPER in ABAQUS.
[0127] like Figure 5 As shown in the figure, compared with pure hydrogel materials, the stress concentration of hydrogel composites has changed due to the integration of superfibers. The stress concentration of superfiber-reinforced hydrogel composites usually appears on the concave side of the middle of the superfiber, unlike the continuous fiber-reinforced hydrogel composites that produce stress concentration at the fiber ends. The number of stress concentration areas depends on the shape of the superfiber, and the location of the stress concentration area is closely related to the distribution of the superfiber. The displacement range of the two vertical sides of the superfiber-reinforced hydrogel composite is large, from 0 at the fixed point to a maximum of 46.5 mm, close to 46% strain. In addition, the displacements of different superfiber distributions in the hydrogel composite on the opposite side of the fixed side also show obvious diversity.
[0128] According to the simulation results, the deformation behavior of the specific fiber reinforced hydrogel can be adjusted by rationally designing the specific fiber distribution. The adjustment of the stress field provides a wide range of design space for material function and structural design. For example, the stress concentration of super fiber reinforced hydrogel composites usually leads to material failure problems, which can be optimized by super fiber distribution; the displacement field of super fiber reinforced hydrogel composites can be reorganized by the imported super fibers to adapt to the desired application.
[0129] S2. Data preprocessing:
[0130] Normalizing the physical field image dataset and the fiber distribution image dataset respectively, and then adjusting the sizes of the images in the physical field image dataset and the fiber distribution image dataset respectively, to construct a dataset for a deep learning model;
[0131] Step S2 specifically includes the following steps:
[0132] S21, by using the simulation results of the free swelling process of the composite material deposited by the finite element software ABAQUS, the distribution of super fibers in the super fiber reinforced hydrogel composite material before swelling and the physical field distribution of the super fiber reinforced hydrogel composite material after swelling are obtained;
[0133] S22, grayscale processing, and compressing the images of the composite physical field distribution and the fiber distribution into an image with a size of 256×256 pixels;
[0134] S23, normalization processing:
[0135] Normalize in the range [-1,1] using the min-max method:
[0136]
[0137] Among them, z i represents the data after the i-th transformation, x(x1,x2,…x n ) represents the grayscale value in the image.
[0138] S3. Training Generative Adversarial Network:
[0139] Inputting the physical field information contained in the physical field image data set preprocessed by step S2 into the generative adversarial network, and using the fiber distribution information contained in the fiber distribution image data set preprocessed by step S2 as output data, modifying the parameters between the generator and the discriminator of the generative adversarial network to train the generative adversarial network;
[0140] Step S3 specifically includes the following steps:
[0141] S31, construct Figure 6 The generative adversarial network shown in the figure consists of a generator and a discriminator, where the generator is a U-Net neural network used to learn the mapping between the super fiber distribution and the physical field, and the discriminator is used to distinguish between the real images in the data set input to the generative adversarial network and the structural images generated by the generator during the training process;
[0142] The generator described in step S31 includes an encoder and a decoder, and there are jump connections between the image layers;
[0143] The fitting process of the generator is as follows: the encoder extracts important features from the input data, namely the stress field after free expansion, and then uses the convolution kernel to reduce the dimension to obtain feature data. The feature data is then input into the decoder for deconvolution calculation to achieve feature reorganization and achieve the purpose of structure generation. Batch Normalization is used in the fitting process to avoid overfitting of the generator.
[0144] The loss function of the generator is defined as:
[0145] L gen =L gan +λL l1 (9)
[0146] Among them, L ganrepresents the loss between the generated adversarial network and the array zero point, λ=100 represents the constant coefficient, L l1 It represents the mean absolute error between abnormal images and true images.
[0147] The domain function component of the discriminator described in step S31 is a convolutional neural network, wherein the first layer receives the real fiber distribution in the super fiber reinforced hydrogel composite material or the super fiber distribution predicted from the generator from the data set simulated and generated by the finite element software ABAQUS;
[0148] The loss function of the discriminator is defined as
[0149] L dis =L real +L fake (10)
[0150] Among them, L real Represents the loss between the normalized real image of the calculation data set and the identity matrix, L fake represents the loss between the generator’s abnormal images and the zero matrix.
[0151] S32. Optimizing Generative Adversarial Networks:
[0152] The optimization problem of the generative adversarial network is regarded as a minimax problem, where the objective function of the generative adversarial network is defined as:
[0153]
[0154] Among them, P data (m) and P z (z) is the pixel value matrix sampled from the stress field, D represents the discriminator, G represents the generator, represents the expected value of the true sample distribution, represents the expected value of the generated sample distribution, and G(z) represents the generated sample;
[0155] S33. The generator is trained to produce abnormal images, and the discriminator distinguishes the abnormal images from the real images and returns the difference to the generator. The training cycle is repeated until the generator and the discriminator satisfy the Nash equilibrium state, at which time the generative adversarial network algorithm is determined to converge.
[0156] S4. Verify and use the verified generative adversarial network to design the fiber distribution according to the target physical field.
[0157] In step S4, the verification of the generative adversarial network is quantitatively evaluated by calculating the mean square error between the real fiber distribution and the fiber distribution predicted by the generative adversarial network and the corresponding physical field;
[0158] The mean square error calculation formula is as follows:
[0159]
[0160] Where M and N represent the width and height of the image respectively, and f ij ' represents the abnormal image generated by the generator, f ij Represents a real image.
[0161] Example
[0162] The dataset of the generative adversarial network (GAN) in this embodiment consists of finite element simulation images for the free expansion of super fiber reinforced hydrogel composites with random fiber numbers and distribution positions. More specifically, a total of 1800 image pairs, including the physical field distribution after expansion as the input image and the fiber distribution as the output image, are aggregated to train the generative adversarial network, of which 80% are randomly divided into a training set and the remaining 20% are divided into a validation set.
[0163] In this embodiment, GAN is trained on an NVIDIA P2000 GPU for 10,000 training iterations, with 4 training samples in each cycle. In addition, the generator in GAN uses an Adam optimizer with a learning rate of 4.0E-4, decaying to 1.0E-8, and the discriminator applies a stochastic gradient descent (SGD) optimizer with a learning rate of 4.0E-4 and no decay.
[0164] Table 2 GAN layer parameters
[0165]
[0166]
[0167] like Figure 7 As shown in Figure 1, the loss function in GAN changes from 0 to 10,000 iterations. In the figure, the generator G and the discriminator D continue to compete until the GAN model converges, where the loss functions of the two components decrease with the training iterations and stabilize at the minimum value. Four representative prediction samples of the generator G are inserted in different iterations. Figure 7 (a) provides a more accurate prediction process for composite materials compared to the true structures in the validation set.
[0168] GAN quantitative analysis based on stress field
[0169] All predicted super-fiber-reinforced hydrogel composite structures show relatively small prediction errors, with the overall mean square error (MSE) less than 0.1 (and most of them less than 0.04).
[0170] The fiber volume fraction V defined by the fiber pixels in the structure image predicted by GAN and the fiber volume fraction V defined by the super fiber reinforced hydrogel composite pixels f Perform regression analysis on Figure 7 The x-axis in corresponds to the corresponding output image in the validation set, that is Figure 7 (c) The y-axis shows that the relative error distribution of the GAN prediction is less than 9%, and the regression coefficient R2 of the y=x linear fit between the structure image and the corresponding output image in the validation set is 0.957, indicating that the predicted structure is highly correlated with the target structure.
[0171] Figure 7 (d) Four representative examples with two different fiber periods illustrate the comparison between the predicted structures and the corresponding structures in the validation set. It can be seen that GAN accurately predicts the key information of the superfiber, including period, position, and direction. The superfiber with a long period of 15 is fully predicted with a smooth curve and no extra pixels.
[0172] Quantitative analysis of GAN based on displacement field
[0173] like Figure 8 As shown, the predicted structure is highly similar to the true structure with a relative error less than 0.04, mostly concentrated at 0.01 with a frequency of 35%. The fiber volume fraction is defined as the ratio of fiber pixels to composite pixels in the structure image predicted based on the displacement field and the corresponding GAN (x-axis). Output images in the validation dataset (y-axis).
[0174] The relative error distribution of GAN prediction is less than 17.5%. Through linear regression analysis, the regression coefficient R2 between the predicted output structure image and the verification output image is 0.912, showing a high correlation, indicating that the predicted structure is closely related to the target structure.
[0175] like Figure 8 As shown in (c), using four representative examples with two different fiber periods, the results show that GAN can accurately predict the key information of superfibers, including period, position, and orientation. When using displacement fields to design superfiber-reinforced hydrogel microstructures, the GAN model can well predict the structure of superfibers with lower V f Position and period of fibers in the microstructure of superfibers.
[0176] Application Example 1
[0177] like Fig. 9As shown, the super fiber reinforced hydrogel structure designed by the inverse design method of the novel super fiber reinforced hydrogel structure based on GAN described in the present invention is used in drug transport. As an effective way to deliver drugs to the target destination in the human body, the hydrogel carrier can prevent the drug from dissolving before reaching the focus, and use the fiber to predict the destruction position of the hydrogel carrier for drug delivery, so as to release the loaded drug after destruction by environmental stimulation.
[0178] In this application example, microfibers were arranged in a hexagonal shape as a reinforcing phase embedded in a hydrogel matrix to simulate multi-point drug release problems. Fig. 9 In (a), the sinusoidal stress cloud is considered to simulate the condition of drug release from the center of the super fiber reinforced hydrogel carrier. The initial stress field is set to a sinusoidal shape, and the Max Mises value is located around the fiber. Then the target stress field is used as the input of the GAN model to predict the structure. After the predicted structure is obtained by the GAN model, the FE method is used to simulate Fig. 9 (b) shows the structure, and the Fig. 9 (c) shows the Mises stress cloud diagram. Finite element calculations found that the maximum stress in the predicted structure is located at the center of the super fiber, and the maximum stress is 2.791MPa. At the same time, there is also a stress concentration area on the concave side of the super fiber. Fig. 9 As shown in (d), the stress value increases with the increase of expansion time. At the same time, the rate of change of stress is very sensitive to changes in the environment, indicating that the function of drug release can be achieved in different environments. Similarly, based on the deep learning model, the following can be designed: Fig.10 Other fiber-reinforced hydrogel carriers are shown to achieve single-point and multi-point drug release problems.
[0179] Application Example 2
[0180] like Fig.11 and Fig.12 As shown, the super fiber reinforced hydrogel structure designed by the inverse design method of the novel super fiber reinforced hydrogel structure based on GAN described in the present invention is used in a flexible sensor. Another significant application of the super fiber reinforced hydrogel composite material is to generate a special electronic signal of deformation anisotropy by controlling the fiber distribution during the expansion process. Due to its excellent biocompatibility and adjustable deformation mode, it has been widely used as a sensing component in the fields of wearable strain sensors, bionic electronic skins and pressure sensors. In the above application, the resistance response characteristics of the super fiber reinforced hydrogel composite material are changed by the form of tunable distribution of fibers after the hydrogel absorbs water, so that the hydrogel can sense the mechanical movement of the human body and monitor physiological behaviors such as sweating.
[0181] In addition, the GAN model performs directional structural design of super fiber reinforced hydrogel composites under the target displacement field, making the hydrogel sensing performance highly adjustable, making it potentially valuable in multi-scenario adaptive interaction between humans and machines. Similar to the design strategy of the stress field, the target displacement field is adopted to obtain a hydrogel composite with super fiber distribution through GAN. Figure 5 (d), the target displacement field is used as an important goal for designing the superfiber distribution of hydrogel composites. By configuring the number of fibers and their distribution, microstructures with anisotropic displacement fields can be obtained to design flexible sensor devices that meet the target mechanical properties.
[0182] Fig.11 The two different displacement fields predetermined in (a) constrain the horizontal and vertical displacements and are then used as inputs to the GAN model to obtain the predicted structure based on the target displacement field and simulated by the FE method. Fig.11 (b) shows the structure, and the Fig.11 (c) shows the displacement cloud. Finite element calculations show that the predicted structure produces the same displacement distribution as the design, indicating that for a predetermined displacement field, the GAN model can better design the corresponding fiber-reinforced hydrogel microstructure. Based on this, different displacement field distributions can be designed according to the deformation modes of stretching, bending, torsion, etc. of different parts of the human body, thereby designing different biosensors to generate electrical signals to monitor the health of the human body. When the deformation strain field of the human body is the same as the deformation strain field of the preform, other fiber-reinforced hydrogel flexible sensors can be designed based on GAN to monitor the health of the human body.
[0183] Therefore, the present invention adopts the above-mentioned GAN-based novel inverse design method of super-fiber reinforced hydrogel structure. Using the GAN model, an end-to-end mapping between mechanical properties and super-fiber distribution can be obtained, so that the GAN model can be used to predict the fiber distribution required for the target mechanical properties under a given mechanical field (including stress field and displacement field), while significantly reducing the computational cost.
[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.
Claims
1. A new GAN-based inverse design method for super fiber reinforced hydrogel structure, characterized by: The following steps are involved: S1. Data generation: Under chemical and temperature boundary conditions, finite element simulation is performed on the super fiber reinforced hydrogel composite material to obtain the solid mechanics model of the super fiber reinforced hydrogel, so as to obtain the physical field data and fiber distribution data of the super fiber reinforced hydrogel composite material after free expansion, and construct the physical field image data set and fiber distribution image data set respectively; In step S1, a two-dimensional super fiber reinforced hydrogel composite material model is established, and the free expansion process of the super fiber reinforced hydrogel composite material is caused by simulating the changes in temperature and the concentration of anions and cations ionized by the electrolyzable solute in the external solution, which specifically includes the following steps: Assumptions: There is no overlap between the superfibers; the hydrogel composite material is completely swollen; the hydrogel matrix and the superfibers are completely bonded, and the solute in the solution is ionized to produce an ionic solution with an anion-cation concentration ratio of 1:1; Under the premise of the established assumptions, the hydrogel matrix is placed in an ionic solution with a set temperature of 320K and a relative ion concentration of 0.03, and then a random number of superfibers with two-dimensional shapes, random distribution and random orientation are embedded to obtain a merged boundary, and it is assumed that the superfibers passing through the merged boundary are cut into segments; then a chemical boundary condition of a relative ion concentration of 0.03 and a temperature boundary condition of 320K are applied to the entire merged boundary, and a boundary condition of zero displacement is set at the bottom edge of the hydrogel to make the hydrogel only swell in the ionic solution; The superfiber is a linear structural fiber with a sinusoidal structure, which is defined as a linear elastic solid with Poisson's ratio and elastic modulus E. The total height H and width t of the superfiber in the free state are given the same initial value; Ultrafiber shape It is defined by the following sine function: (1) in, , and They represent the amplitude, angular frequency and phase shift of the sine function respectively; The process of establishing the hydrogel matrix constitutive model is as follows: First, the expansion is simulated by the mixing free energy function based on the Flory-Rehner free energy function, and the chemical field and temperature field are considered to construct the following total mixing free energy density function: (2) in, Indicates temperature, represents the entropy change during the expansion process, represents the elastic entropy, represents the enthalpy of the hydrogel, t represents the elastic energy density function of the chemical cross-linked chain network, is the universal gas constant, represents the degree of entanglement between chemically cross-linked chains, represents the number of moles of chemical cross-linking chains, , represents the principal extension, represents the molar number of solvent molecules in the ionic solution, represents the molar volume of the solvent, It indicates the number of moles of slippable connections between molecular chains. represents the sliding point parameter, represents the dimensionless enthalpy of mixing; Then, the total mixing free energy density function is converted into the following principal invariant by Legendre transformation: (3) in, represents the deformation gradient tensor, represents the chemical potential, Describe the main changes; Then, given the initial temperature, chemical potential and free elongation Under the premise of , the relationship between chemical potential and free elongation is derived: (4) Then, the chemical potential between the hydrogel matrix and the solvent is assumed to be equal in the initial state, where the chemical potential in the solvent is approximated with the ideal solution assumption by the following equation: (5) in, is the sum of the initial ionic species concentrations of the solvent; Finally, the free elongation is obtained and the initial solvent ion concentration The relationship between: (6) The hydrogel matrix of the super fiber reinforced hydrogel composite material was modeled by the total mixing free energy function in formula (2) to characterize the deformation of the hydrogel matrix under chemical-mechanical coupling conditions, and numerical simulation was performed using the finite element software ABAQUS; S2. Data preprocessing: Normalizing the physical field image dataset and the fiber distribution image dataset respectively, and then adjusting the sizes of the images in the physical field image dataset and the fiber distribution image dataset respectively, to construct a dataset for a deep learning model; S3. Training Generative Adversarial Network: Inputting the physical field information contained in the physical field image data set preprocessed by step S2 into the generative adversarial network, and using the fiber distribution information contained in the fiber distribution image data set preprocessed by step S2 as output data, modifying the parameters between the generator and the discriminator of the generative adversarial network to train the generative adversarial network; S4. Verify and use the verified generative adversarial network to design the fiber distribution according to the target physical field.
2. The inverse design method of the novel super fiber reinforced hydrogel structure based on GAN according to claim 1, characterized in that: The physical field described in step S1 includes a stress field and a displacement field.
3. The inverse design method of the novel super fiber reinforced hydrogel structure based on GAN according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21, by using the simulation results of the free swelling process of the composite material deposited by the finite element software ABAQUS, the distribution of super fibers in the super fiber reinforced hydrogel composite material before swelling and the physical field distribution of the super fiber reinforced hydrogel composite material after swelling are obtained; S22, grayscale processing, and compressing the images of the composite physical field distribution and the fiber distribution into an image with a size of 256×256 pixels; S23, normalization processing: Normalize in the range [-1,1] using the min-max method: (7) in, Indicates The transformed data, Represents the grayscale value of the image in the dataset.
4. The inverse design method of the novel super fiber reinforced hydrogel structure based on GAN according to claim 3 is characterized in that: Step S3 specifically includes the following steps: S31. Construct a generative adversarial network, which consists of a generator and a discriminator, wherein the generator is a U-Net neural network used to learn the mapping between the super fiber distribution and the physical field, and the discriminator is used to distinguish between the real image in the data set input into the generative adversarial network and the structure image generated by the generator during the training process; S32. Optimizing Generative Adversarial Networks: The optimization problem of the generative adversarial network is regarded as a minimax problem, where the objective function of the generative adversarial network is defined as: (8) in, and is the matrix of pixel values sampled from the stress field, represents the discriminator, represents a generator, represents the expected value of the true sample distribution, represents the expected value of the generated sample distribution, represents the generated samples; S33. The generator is trained to produce abnormal images, and the discriminator distinguishes the abnormal images from the real images and returns the difference to the generator. The training cycle is repeated until the generator and the discriminator satisfy the Nash equilibrium state, at which time the generative adversarial network algorithm is determined to converge.
5. The inverse design method of the novel super fiber reinforced hydrogel structure based on GAN according to claim 4, characterized in that: The generator described in step S31 includes an encoder and a decoder, and there are jump connections between the image layers; The fitting process of the generator is as follows: the encoder extracts important features from the input data, namely the stress field after free expansion, and then uses the convolution kernel to reduce the dimension to obtain feature data. The feature data is then input into the decoder for deconvolution calculation to achieve feature reorganization and achieve the purpose of structure generation. Batch Normalization is used in the fitting process to avoid overfitting of the generator. The loss function of the generator is defined as: (9) in, represents the loss between the generated adversarial network and the zero point of the array, λ=100 represents the constant coefficient, It represents the mean absolute error between abnormal images and true images.
6. The inverse design method of the novel super fiber reinforced hydrogel structure based on GAN according to claim 5, characterized in that: The domain function component of the discriminator described in step S31 is a convolutional neural network, wherein the first layer receives the real fiber distribution in the super fiber reinforced hydrogel composite material or the super fiber distribution predicted from the generator from the data set simulated and generated by the finite element software ABAQUS; The loss function of the discriminator is defined as (10) in, Represents the loss between the normalized real image of the calculation data set and the identity matrix, represents the loss between the generator’s abnormal images and the zero matrix.
7. The inverse design method of the novel super fiber reinforced hydrogel structure based on GAN according to claim 6, characterized in that: In step S4, the verification of the generative adversarial network is quantitatively evaluated by calculating the mean square error between the real fiber distribution and the fiber distribution predicted by the generative adversarial network and the corresponding physical field; The mean square error calculation formula is as follows: (11) Where M and N represent the width and height of the image respectively. represents the abnormal image generated by the generator, Represents a real image.
