A method for constructing deep learning potential function for magnesium-lithium alloys

Abstract

The present invention discloses a method for constructing a deep learning potential function for magnesium-lithium alloys. The method includes obtaining the lattice constants and elastic constants of magnesium-lithium alloys with different concentrations in a stable state through high-throughput density functional theory calculations; obtaining the surface energy of different crystal planes of magnesium-lithium alloys and the generalized stacking fault energy curves of different slip systems to obtain the energy and atomic force information of all configurations; and using the deep learning software DeePMD-kit to train and test the DFT data set, and fitting a deep learning potential function model for Mg-Li alloys. The present invention adopts the above-mentioned method for constructing a deep learning potential function for magnesium-lithium alloys, and the obtained Mg-Li potential function model is consistent with the DFT calculation results in predicting the trend of basal plane stacking fault energy changing with Li atomic concentration, solving the problem that the classical MEAM potential function cannot accurately predict the trend of basal plane stacking fault energy changing with Li concentration.

Description

Technical Field

[0001] The present invention belongs to the field of molecular simulation computing technology in materials science, and specifically provides a method for constructing a deep learning potential function for magnesium-lithium alloys. Background Art

[0002] Magnesium-lithium alloy is the lightest structural material known to date. The ultra-low density of lithium (0.58g / cm3) indicates that as the lithium content increases, the alloy density will further decrease. Therefore, magnesium-lithium alloy is also widely used in the aerospace field, thereby reducing flight costs. Adding lithium can further improve the plasticity of magnesium alloys. The improvement in the plasticity of magnesium alloys is mainly due to the excitation of non-basal slip systems. For magnesium alloys, adding Li will reduce c / a. Experiments have determined that when the Li content is 5.5wt.%, c / a will decrease from 1.623 to 1.607. The lower c / a will increase the Peierls stress of basal slip. Adding Li can also reduce the nucleation stress of non-basal dislocations, thereby affecting the ratio of basal critical shear stress (CRSS) to non-basal slip CRSS, thereby stimulating non-basal slip systems. Secondly, the addition of lithium will cause changes in the stacking fault energy of the cone surface, promoting the cone surface<c+a> Dislocation decomposition to basal or pyramidal planes<c+a> collinear decomposition of dislocations, both of which promote<c+a> The movement of dislocations further improves the plasticity of the alloy.

[0003] In view of the great potential of magnesium-lithium alloys, many experiments have been conducted to study the deformation process of magnesium-lithium alloys. However, due to the limitations of experimental equipment, the current research perspectives of scholars on magnesium-lithium alloys are still mostly limited to the investigation of the effects of alloying or heat treatment on the microstructure and basic tensile properties, and there are few discussions on the deformation mechanism of Mg-Li alloys. The special microstructure of magnesium-lithium alloys is very sensitive to their deformation behavior. Therefore, the investigation and study of the mechanical properties and deformation failure behavior of magnesium-lithium alloys from a microscopic perspective has very important guiding significance. Large-scale atomic simulations can help supplement experiments and provide new ideas for further understanding the deformation process of magnesium alloys. Among them, molecular dynamics, as an important method to study the behavior and properties of atoms (or molecules) at the microscopic scale, has received more and more attention in the application of magnesium and its alloys in recent years.

[0004] Despite the high demand for large-scale molecular dynamics simulations, little research has been devoted to developing empirical interatomic potentials for Mg–Li alloys. Beauchamp et al. and Canales et al. were the first to develop Mg–Li potentials based on pseudopotential theory to describe pair interactions, but these potentials were primarily used to study the bcc and liquid phases. Furthermore, these pseudopotentials fell far short of the requirements for large-scale simulations. Atomistic simulations based on (semi-)empirical interatomic potentials can handle over one million atoms. Once the interatomic potentials accurately reproduce the fundamental physical properties of the relevant material system, they can become an effective tool for studying deformation mechanisms. Consequently, it was not until 2012 that an accurate interatomic potential for both hcp and bcc Mg–Li binary alloys was developed. This Mg–Li binary potential, constructed based on the second-nearest-neighbor corrected EAM potential, remarkably reproduces the static bulk properties of the Mg–Li alloy.

[0005] The 2NN-MEAM potential developed by Kim et al. fairly well reproduces various fundamental physical properties of both materials at 0 K, but its reliability at finite temperatures (except for the heat of mixing at 1000 K) has not been verified. Furthermore, its prediction of the variation of the basal plane stacking fault energy with lithium content does not conform to the trends reported in the literature. Using first-principles calculations and experimental data, Karewar et al. developed a thermodynamically reliable concentration-dependent embedded atom method (CD-EAM) interatomic potential. This potential predicts various properties of magnesium-lithium alloys, such as lattice parameter, heat of mixing, stacking fault energy, and bulk modulus, and establishes temperature and concentration dependencies that agree well with literature data. Most importantly, it reproduces the experimental Mg–Li binary phase diagram. This CD-EAM potential is derived from the pure elemental EAM potentials of Li and Mg. Nogaret et al. and Ghazisaeidi et al. showed that this pure Mg potential well predicts the stacking fault energy of the cone I slip system, but the dislocation core structure differs from density functional theory (DFT) results. At the same time, this potential function cannot predict the stable stacking fault energy of the cone II slip system. These shortcomings greatly limit the wide application of this type of potential function.

[0006] Density functional theory calculation method is currently recognized as an important theoretical method for studying atomic structure and electronic properties. Density functional theory method can accurately predict the geometric structure, elastic properties and mechanical parameters such as stacking fault energy of metal materials. Moitra et al. used first principles to calculate the effect of different solute elements on the stacking fault energy of magnesium alloys. The results showed that the unstable stacking fault energy γ of the cone II of pure magnesium is USF =239mJ / m 2 , can be reduced to 235mJ / m by adding lithium 2, thereby reducing the nucleation resistance of conical dislocations. This demonstrates the close connection between atomic-scale dislocations and macroscopic mechanical properties in alloys. However, due to spatial scale limitations, density functional theory can only simulate the motion of tens to hundreds of atoms. Simulating dislocations in metals requires a far larger number of atoms than density functional theory can achieve. Therefore, it is necessary to construct a high-precision potential function suitable for large-scale atomistic simulations.

[0007] With the continuous development of artificial intelligence and computer technology, machine learning potential functions have become a hot topic in the field of computational materials. Potential functions fitted using machine learning offer advantages such as high accuracy and low computational complexity, and are expected to address the inaccurate description of interatomic interactions in certain complex systems by classical potential functions. Machine learning potential functions establish interatomic interactions by fitting the potential energy surface (PES) of metals, unconstrained by the precise functional form required by classical potential functions. The construction of machine learning potential functions involves: (i) selecting appropriate descriptors to describe the local atomic environment; (ii) using first-principles methods to construct an energy and force database (training set) containing diverse atomic structures; and (iii) applying regression algorithms (such as neural networks and gradient boosting trees) to optimize the parameters in the machine learning framework to achieve the best match between the trained potential function and the training set. However, precisely because they are regression algorithms, machine learning potential functions are not suitable for extrapolation to structures that differ significantly from the training set. At the same time, it is important to emphasize that classical interatomic potential functions, with their fixed functional form and finite parameters, have inherent limitations in accurately fitting material properties; the fitting results depend on the target properties and the regression algorithm.

[0008] In recent years, there has been a significant body of work on machine learning potential functions. For example, the machine learning potential function fitted by Zuo et al. yielded poor predictions for the stable stacking fault energy (SFE) of fcc Ni and Cu. This is primarily due to the fact that the training set does not include information about stacking fault structures. Furthermore, while the self-supervised learning method used effectively samples close crystal configurations, it does not generate stacking faults or vacancies. Therefore, the potential function trained by Zuo is not suitable for simulating plastic deformation processes. Kobayashi et al. developed a neural network potential (NNP) for Al-Mg-Si alloys. By incorporating intermetallic compounds, solute-solute interactions, and interface characteristics into the training set, they achieved excellent predictions for edge and screw dislocation structures, solute-dislocation interactions, and precipitation behavior. Maresca et al. fitted Gaussian approximate potentials (GAPs) for Fe and W based on the basic data required to describe dislocations, and found that the double-kink nucleation process controls the plastic deformation of bcc metals. These works highlight the promise of machine learning potential functions in the field of computational materials. Recently, Professor Curtin's team at the Swiss Federal Institute of Technology in Lausanne constructed a dataset of 443 structures using first-principles calculations and, using a neural network algorithm, successfully fitted a molecular dynamics potential function for metallic magnesium. This neural network potential function outperformed traditional embedded-atom potentials in predicting both the crystal elastic constants and stacking fault energies. However, their potential function exhibited significant fluctuations in its predictions of the basal plane generalized stacking fault energy curve and equilibrium energy, suggesting that further improvement in model stability is needed. Summary of the Invention

[0009] In order to solve the problem that the classic magnesium-lithium potential function is inaccurate in predicting the trend of basal plane stacking fault energy with lithium content, the present invention targets a training data set containing a series of samples obtained by VASP calculation, and constructs a high-precision and low-cost deep learning potential function for magnesium-lithium alloys based on a deep learning method. This method involves the development of an interatomic interaction model in metal materials. It takes Mg-Li alloy as the research object, trains and learns the structure and energy characteristic database based on a deep learning framework, and fits the molecular dynamics potential function with first-principles accuracy.

[0010] A method for constructing a deep learning potential function for magnesium-lithium alloys comprises the following steps:

[0011] Step 1: Use Materials Studio to build a binary magnesium-lithium alloy model. By randomly replacing Mg atoms, magnesium-lithium alloy models with different Li contents are formed. VASP is used to optimize the initial crystal structure of the material to obtain the lattice constant of the material.

[0012] Step 2: Use the lattice constants obtained in step 1 to build a model, and obtain the elastic constants C of the magnesium-lithium alloy by applying a deformation matrix to the unit cell. 11、C 12 、C 13 、C 33 and C 44 ; After the strain is applied, the change of the total energy of the system with the strain component can be expressed as:

[0013]

[0014] Where e is the strain component in different directions, and the strain energy formula is used to fit the quadratic coefficient, where the quadratic coefficient is the elastic constant of the crystal or a linear combination of elastic constants. The strain energy formula is as follows:

[0015]

[0016] Where ΔE is the change in crystal energy after the strain is applied, V is the equilibrium volume of the crystal, δ is the strain, and C 11 、C 12 、C 13 、C 33 and C 44 are the independent components of the elastic constants.

[0017] Step 3: Use the lattice constants obtained in step 1 to build a model. Use VASP to fully release the cut slab model for structural optimization and calculate the surface energy on different surfaces of the magnesium-lithium alloy. Generally, the following formula is used to calculate the surface energy:

[0018]

[0019] Where E slab is the energy of the slab model after a certain crystal plane of the block model is cut, E bulk is the energy of a single atom in the bulk, i.e., the energy of the bulk divided by the number of atoms, where n is the number of atoms in the slab model, and A is the surface area of ​​the slab model. The denominator is divided by 2 to calculate the surface energy of a single surface.

[0020] Step 4: Use the lattice constants obtained in step 1 to build a model. Use VASP to calculate the generalized stacking fault energy curves of different slip systems of magnesium-lithium alloy by sliding some atoms on the slip plane along the slip plane with different Burgers vectors. The generalized stacking fault energy curve is usually calculated using the formula:

[0021]

[0022] Where E SFis the energy of the supercell containing the stacking fault, E0 is the energy of the perfect crystal, and A is the area of ​​the slip plane. The upper half of the supercell is rigidly shifted by a certain Burgers vector. By shifting different distances, the corresponding stacking fault energy is calculated according to formula (8). Finally, a generalized stacking fault energy curve can be obtained with the shift distance as the horizontal axis and the stacking fault energy as the vertical axis.

[0023] Step 5: Use the energy and atomic forces corresponding to different configurations in the OUTCAR calculation results of the above steps as a training data set, and use the deep learning software DeePMD-kit to train and learn different interatomic forces to obtain the magnesium-lithium alloy potential energy model. The training of the model is mainly controlled by the input file xx.json. The more atomic types, the more complex the model, and the slower the training. For the magnesium-lithium binary alloy system, the "type" in the "descriptor" setting is set to "se_e2_a", which mainly sets the descriptor type; the "sel" parameter is set to [64,64], which controls the upper limit of the number of atoms within the rcut; the "rcut" parameter mainly affects the atomic interactions within the truncation radius. Through multiple attempts, it was found that the parameter rcut has a significant impact on the deep learning results. For the magnesium-lithium binary system, setting rcut to 14 is a relatively reasonable value.

[0024] The present invention adopts the above-mentioned method for constructing a deep learning magnesium-lithium alloy potential function, which has the following advantages:

[0025] (1) The present invention obtains the corresponding relationship between different configurations and energies through first-principles calculations, obtains a database that can characterize the physical properties of magnesium-lithium alloys at low concentrations, and further constructs a deep learning potential function for magnesium-lithium alloys;

[0026] (2) The present invention solves the problem that the classical MEAM potential function cannot accurately predict the trend of basal plane stacking fault energy with concentration, and at the same time makes up for the shortcoming of low efficiency of first-principles calculation, achieving a calculation efficiency close to that of the empirical potential function;

[0027] (3) The present invention provides a theoretical model and necessary foundation for in-depth research on the evolution of the microstructure of magnesium-lithium alloys under plastic deformation, and has obvious advantages and broad application prospects in the design and application of new magnesium-lithium alloys.

[0028] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Flowchart for constructing a deep learning potential function for magnesium-lithium alloys in an embodiment of the present invention.

[0030] Figure 2 This is the deep learning potential function learning result of magnesium-lithium alloy in an embodiment of the present invention. Figure 2(a) Energy prediction error of the deep learning magnesium-lithium alloy potential function; Figure 2 (b) The prediction error of atomic forces in different directions of deep learning magnesium-lithium alloy potential function.

[0031] Figure 3 This is the prediction result of different physical properties of magnesium-lithium alloy deep learning potential function in the embodiment of the present invention.

[0032] Figure 4 Magnesium-lithium alloy base and Slip system, Figure 4 (a) and Figure 4 (b) is the result of calculation when the Li content is 2 at%, Figure 4 (c) and Figure 4 (d) is the result calculated when the Li content is 4 at%, Figure 4 The middle dashed line is the calculation result of the MEAM potential function, the black solid line is the DFT calculation result; the light red line is the calculation result of the potential function of thirty neural networks magnesium-lithium alloy.

[0033] Figure 5 Magnesium-lithium alloy cylinder and cone Slip system, Figure 5 (a) and Figure 5 (b) is the result of calculation when the Li content is 2 at%, Figure 5 (c) and Figure 5 (d) is the result calculated when the Li content is 4 at%, Figure 5 The middle dashed line is the calculation result of the MEAM potential function, the black solid line is the DFT calculation result; the light red line is the calculation result of thirty neural network magnesium-lithium alloy functions.

[0034] Figure 6 Magnesium-lithium alloy cone Ⅰ and cone II Slip system, Figure 6 (a) and Figure 6 (b) is the result of calculation when the Li content is 2 at%, Figure 6 (c) and Figure 6 (d) is the result calculated when the Li content is 4 at%, Figure 6 The middle dashed line is the calculation result of the MEAM potential function, the black solid line is the DFT calculation result; the light red line is the calculation result of the potential function of thirty neural networks for magnesium-lithium alloy. DETAILED DESCRIPTION

[0035] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0036] Example 1

[0037] Step 1: Use Materials Studio to build a binary magnesium-lithium alloy model. By randomly replacing Mg atoms, magnesium-lithium alloy models with different Li contents are formed. VASP is used to optimize the initial crystal structure of the material to obtain the lattice constant of the material.

[0038] Step 2: Use the lattice constants obtained in step 1 to build a model, and obtain the elastic constants C of the magnesium-lithium alloy by applying a deformation matrix to the unit cell. 11 、C 12 、C 13 、C 33 and C 44 ; After the strain is applied, the change of the total energy of the system with the strain component can be expressed as:

[0039]

[0040] Where e is the strain component in different directions, and the strain energy formula is used to fit the quadratic coefficient, where the quadratic coefficient is the elastic constant of the crystal or a linear combination of elastic constants. The strain energy formula is as follows:

[0041]

[0042] Where ΔE is the change in crystal energy after the strain is applied, V is the equilibrium volume of the crystal, δ is the strain, and C 11 、C 12 、C 13 、C 33 and C 44 are the independent components of the elastic constants.

[0043] Step 3: Use the lattice constants obtained in step 1 to build a model. Use VASP to fully release the cut slab model for structural optimization and calculate the surface energy on different surfaces of the magnesium-lithium alloy. Generally, the following formula is used to calculate the surface energy:

[0044]

[0045] Where E slab is the energy of the slab model after a certain crystal plane of the block model is cut, E bulk is the energy of a single atom in the bulk, i.e., the energy of the bulk divided by the number of atoms, where n is the number of atoms in the slab model, and A is the surface area of ​​the slab model. The denominator is divided by 2 to calculate the surface energy of a single surface.

[0046] Step 4: Use the lattice constants obtained in step 1 to build a model. Use VASP to calculate the generalized stacking fault energy curves of different slip systems of magnesium-lithium alloy by sliding some atoms on the slip plane along the slip plane with different Burgers vectors. The generalized stacking fault energy curve is usually calculated using the formula:

[0047]

[0048] Where F SF is the energy of the supercell containing the stacking fault, E0 is the energy of the perfect crystal, and A is the area of ​​the slip plane. The upper half of the supercell is rigidly shifted by a certain Burgers vector. By shifting different distances, the corresponding stacking fault energy is calculated according to formula (8). Finally, a generalized stacking fault energy curve can be obtained with the shift distance as the horizontal axis and the stacking fault energy as the vertical axis.

[0049] According to the present invention, step 1 is first performed to calculate the lattice constant of pure magnesium. A 4×3×2 supercell structure containing 48 atoms is selected. By randomly replacing Mg atoms, a binary magnesium-lithium alloy model with different lithium concentrations is established. The lattice constant of the magnesium-lithium alloy is obtained by fitting the BM curve.

[0050] Then proceed to step 2, use VASP software to apply different strain matrices to the magnesium-lithium alloy for structural optimization, and set the strain components to -0.02, -0.015, -0.01, -0.005, 0, 0.005, 0.01, 0.015 and 0.02 in sequence, and obtain the elastic constants of the magnesium-lithium alloy by fitting the energy-volume curve through the formula.

[0051] In step 3, Materials Studio was used to build models of different surfaces of the magnesium-lithium alloy. The cut slab model was fully opened for structural optimization and the surface energy of different surfaces of the magnesium-lithium alloy was calculated.

[0052] In step 4, use Materials Studio to build models of different slip systems of magnesium-lithium alloy. Each slip system model contains 120 atoms, ensuring that the slip planes are perpendicular to the z-axis and filling the upper and lower ends of the z-axis. The generalized stacking fault energy curve is calculated by moving the atoms above the slip plane to different distances, assuming a vacuum layer and atoms are only allowed to relax in the z-axis direction.

[0053] In step 5, the energy, atomic forces, and potential information for all configurations obtained in steps 1-5 are used as a database. Based on a deep learning framework, the structure and energy feature database is trained to fit the molecular dynamics potential function with first-principles accuracy. Model training is primarily controlled by the input file xx.json. The more atomic types and the more complex the model, the slower the training. For the magnesium-lithium binary alloy system, the "descriptor" setting is "type" set to "se_e2_a", which mainly sets the descriptor type; the "sel" parameter is set to [64,64], which controls the upper limit of the number of atoms within the rcut; the "rcut" parameter mainly affects the atomic interactions within the cutoff radius. Through multiple experiments, it was found that the rcut parameter has a significant impact on the deep learning results. For the magnesium-lithium binary system, an rcut setting of 14 is a relatively reasonable value.

[0054] Using the deep learning magnesium-lithium alloy potential function developed in this embodiment, calculations show that:

[0055] First, the deep learning magnesium-lithium alloy developed in this embodiment can well repeat the reference values ​​of energy and atomic force on the training set and test set, such as Figure 2 、 3 shown.

[0056] Second, the physical properties of magnesium-lithium alloys calculated using the deep learning potential function developed in this embodiment, the existing classical MEAM potential function, and the first principle calculation are shown in Tables 1 and Figure 4 The results show that the potential function model of the present invention predicts the alloy system better than the classic MEAM potential function results, is closer to the DFT calculation results, and has less volatility. The potential function model meets the expected results.

[0057] Third, the deep learning magnesium-lithium alloy potential function developed in this embodiment, the existing classical MEAM potential function and the generalized stacking fault energy curve of the magnesium-lithium alloy basal slip system calculated by first principles are used. The results are as follows Figure 5 As shown, the potential function model of the present invention predicts the alloy system better than the classic MEAM potential function results, is closer to the DFT calculation results, and has less volatility. The potential function model meets the expected results.

[0058] Fourth, the deep learning magnesium-lithium alloy potential function developed in this embodiment, the existing classical MEAM potential function and the generalized stacking fault energy curve of the magnesium-lithium alloy non-basal slip system calculated by first principles are used. The results are as follows: Figure 5As shown, the potential function model MLP2 of the present invention predicts the alloy system better than the classic MEAM potential function results, is closer to the DFT calculation results, and has less volatility, and the potential function model meets the expected results. In the table, we also provide the deep learning potential function model MLP3. Unlike MLP2, the rcut is set to 10 during the MLP3 training process. By comparison, when calculating γsf(I2), the MLP3 calculation results are significantly greater than the first-principles calculation results. Therefore, we believe that rcut has a significant impact on the deep learning results and set rcut to 14 in the magnesium-lithium binary system.

[0059] Table 1

[0060]

[0061]

[0062] It can be seen that the deep learning magnesium-lithium alloy potential function model developed in the present invention can accurately describe the lattice constant, elastic constant, surface energy and generalized stacking fault energy curves of different slip systems of nickel-magnesium-lithium alloys, and is superior to the existing classic MEAM potential function.

Claims

1. A method for constructing a deep learning potential function for magnesium-lithium alloys, characterized in that: The following steps are involved: Step 1: Use Materials Studio to establish a binary magnesium-lithium alloy model. By randomly replacing Mg atoms, magnesium-lithium alloy models with different Li contents are formed. VASP is used to optimize the initial crystal structure of the material to obtain the lattice constant of the material. Step 2: Use the lattice constants obtained in step 1 to build a model, and obtain the elastic constants C of the magnesium-lithium alloy by applying a deformation matrix to the unit cell. 11 、C 12 、C 13 、C 33 and C 44 ; Step 3: Use the lattice constants obtained in step 1 to build a model, use VASP to fully release the cut slab model for structural optimization, and calculate the surface energy of different crystal planes of magnesium-lithium alloy. Step 4: Use the lattice constants obtained in step 1 to build a model, and use VASP to calculate the generalized stacking fault energy curves of different slip systems of magnesium-lithium alloy by sliding some atoms on the slip plane along the slip plane with different Burgers vectors; Step 5: The energy and atomic forces corresponding to different configurations in the OUTCAR calculation results of the above steps are used as a training data set, and the deep learning software DeePMD-kit is used to train and learn different interatomic forces to obtain the magnesium-lithium alloy potential energy model; The change of the total energy of the system after strain with the strain component is expressed as: Among them, e i , e j They are the strain components in different directions, C ij is the elastic constant in different directions, and the strain energy formula is used to fit the quadratic coefficient, where the quadratic coefficient is the elastic constant of the crystal or a linear combination of the elastic constants. The strain energy formula is as follows: Where ΔE is the change in crystal energy after the strain is applied, V is the equilibrium volume of the crystal, δ is the strain, and C 11 、C 12 、C 13 、C 33 and C 44 are the independent components of the elastic constants.

2. The method for constructing a deep learning potential function for magnesium-lithium alloy according to claim 1, characterized in that: The following formula is used to calculate surface energy: Where E slab is the energy of the slab model after a certain crystal plane of the block model is cut, E bulk It is the energy of a single atom in the bulk, that is, the energy of the bulk divided by the number of atoms, n is the number of atoms in the slab model, and A is the surface area of ​​the slab model; the purpose of dividing the denominator by 2 is to calculate the surface energy of a single surface.

3. The method for constructing a deep learning potential function for magnesium-lithium alloy according to claim 1, characterized in that: The generalized stacking fault energy curve is calculated using the formula: Where E SF is the energy of the supercell containing the stacking fault, E0 is the energy of the perfect crystal, and A is the area of ​​the slip plane. The upper half of the supercell is rigidly moved by a certain Burgers vector. By moving different distances, the corresponding stacking fault energy is calculated according to formula (8), and finally a generalized stacking fault energy curve is obtained with the moving distance as the horizontal coordinate and the stacking fault energy as the vertical coordinate.

Images