Method for designing a three-way catalyst using machine learning

Abstract

A method of manufacturing a ternary catalyst for oxygen reduction reaction is disclosed. The method can include constructing a database of catalytic activity of oxygen reduction reaction (ORR) including PtFeCu nanoparticles using a neural network potential (NNP) based on machine learning, determining thermodynamically stable PtFeCu nanoparticles by Monte Carlo calculation, and selecting one type of PtFeCu nanoparticles by analyzing the structure of the PtFeCu nanoparticles.

Description

[0001] Cross-references to related applications

[0002] This application claims priority and benefit to Korean Patent Application No. 10-2021-0037961, filed on March 24, 2021, with the Korean Intellectual Property Office, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This invention relates to a method for designing ternary catalysts using machine learning, and more specifically, to a method for designing PtFeCu catalysts for oxygen reduction reactions in fuel cells. Background Technology

[0004] Proton exchange membrane (PEM) fuel cells show great potential as systems that directly convert renewable fuels into electrochemical energy. Compared to conventional internal combustion engines, they are more environmentally friendly and offer improved efficiency. Developed countries have already commercialized hydrogen-powered vehicles equipped with PEM fuel cells, which offer greater range per refueling compared to lithium-ion powered vehicles.

[0005] PEM fuel cells face several pressing challenges, such as high material costs, significantly slow fuel electrochemical conversion rates, short system durability, and insufficient charging station infrastructure.

[0006] The first two problems are primarily caused by the expensive Pt catalysts used in commercially available PEM fuel cells. Pt catalysts suffer severe structural degradation due to the high overpotential of the electrochemical oxygen reduction reaction (ORR) at the cathode and condensation or dissolution during fuel cell operation. Significant collaborations have been undertaken over the past few decades to develop catalysts capable of reducing overvoltage.

[0007] For example, various nanoscale alloys have been proposed to achieve synergistic effects. Some binary Pt alloys (where M in Pt-M = Cu, Fe, Co, Ni, Y, etc.) have been extensively studied in terms of ORR catalyst performance.

[0008] However, due to factors such as the weak OH binding energy on the surface, PtFe nanoparticles exhibit better ORR catalyst performance than pure Pt. For example, PtCu3 and PtCo demonstrate excellent ORR catalyst performance. However, Pt-M binary nanocatalysts undergo structural decomposition during long-term electrochemical cycling.

[0009] Several studies have been reported on introducing a third component into Pt-M binary nanocatalysts to further modulate the performance and stability of ORR catalysts in acidic media, as well as on using Pt-based ternary nanoparticles for ORR catalyst performance.

[0010] In Pt-based ternary nanoparticles, the stoichiometry of each component is crucial for achieving the optimal oxygen binding energy for ORR catalytic activity. However, the composition and configuration of ternary alloys have not been extensively explored due to the difficulty in using both experimental and computational methods.

[0011] For example, general quantum chemistry-based materials design techniques design materials based on the assumption that the experimental composition and structure are based on bulk structural information. However, actual nanoparticle materials have different volumes, structures, and properties, leading to errors between calculations and experiments.

[0012] Specifically, existing structural calculations for alloy catalysts may not take into account their actual particle size and are time-consuming. Since experimentally synthesizable stable structures can be found by searching millions of structures, the actual prediction of stable structures is limited, and therefore, experimental prediction errors are obviously present in calculations.

[0013] Therefore, in order to overcome the shortcomings of conventional catalyst design methods based on bulk structure information, a computational process based on stable structures is needed. Summary of the Invention

[0014] In a preferred aspect, a method is provided for designing and manufacturing a three-way catalyst for the oxygen reduction reaction, which saves cost and time by using machine learning to find catalysts that can be synthesized experimentally and pre-screening their performance to minimize the number of candidates for experimental testing.

[0015] In one aspect, a method for manufacturing a three-way catalyst for the oxygen reduction reaction is provided. This method may include, in the first step, constructing a database of the catalytic activity of PtFeCu nanoparticles for the oxygen reduction reaction (ORR) using a machine learning-based neural network potential (NNP); in the second step, determining thermodynamically stable PtFeCu nanoparticles through Monte Carlo calculations; and in the third step, selecting a type of PtFeCu nanoparticle for the three-way catalyst by analyzing the structure of the PtFeCu nanoparticles.

[0016] In the first step, a neural network potential can be constructed using machine learning parameters of the atomic interaction energy, which is calculated using density functional theory (DFT).

[0017] The training set for machine learning can consist of local atomic environments with different shapes, sizes, compositions or configurations, where PtFeCu nanoparticles are divided according to their cutoff radius size.

[0018] The atomic local environment may include approximately 100 to 300 cubic octahedral random structures with a particle size less than or equal to approximately 1.5 nm, 10 to 20 cubic octahedral random structures with a particle size less than or equal to approximately 2.0 nm, 100 to 300 truncated octahedron random structures with a particle size less than or equal to approximately 1.1 nm, and 50 to 150 truncated octahedron random structures with a particle size less than or equal to approximately 1.7 nm.

[0019] The catalytic activity of PtFeCu nanoparticles for the oxygen reduction reaction can be calculated using Equation 2.

[0020] [Equation 2]

[0021] ΔG=ΔE+ΔZPE-TΔS-neU

[0022] In Equation 2, ΔG is the free energy of ORR, ΔE is the change in internal energy of the reaction obtained by DFT calculation, ΔZPE and ΔS are the changes in zero-point energy and vibrational entropy, respectively, U is the electrode potential relative to the standard hydrogen electrode, and n is the number of electrons participating in the reaction.

[0023] In the second step, the Monte Carlo calculation for each PtFeCu nanoparticle can be made up of n attempts to randomly swap atomic positions, and can be performed in about 10,000 trials at a temperature T (about 0 Kelvin or higher).

[0024] The second step can be performed by mapping the calculated density functional theory (DFT) of PtFeCu nanoparticles to a ternary phase diagram and expressing the Monte Carlo calculation results as an energy convex hull on the ternary phase diagram.

[0025] In the third step, in order to configure PtFeCu nanoparticles including a Pt shell, the Pt content based on the total weight of the PtFeCu nanoparticles can be set to be greater than or equal to about 0.6, in atomic fraction.

[0026] In the third step, the structure of PtFeCu nanoparticles can be analyzed by analyzing the number of Pt, Fe, and Cu atoms in each atomic shell of the PtFeCu nanoparticles.

[0027] In the third step, the chemical stability of the PtFeCu nanoparticles can be evaluated using Equation 7.

[0028] [Equation 7]

[0029] E seg =E(Pt) m Fe n Cu l ) seg -E(Pt m Fen Cu l ) 初始

[0030] In Equation 7, E seg The surface separation energy of alloy components caused by oxygen atom adsorption, E(Pt) m Fe n Cu l ) seg and E(Pt) m Fe n Cu l ) 初始 Pt with or without surface separation m Fe n Cu l The total energy, and m, n and l are the number of Pt, Fe and Cu atoms in the PtFeCu nanoparticles, respectively.

[0031] The method may also include a fourth step, such as experimentally synthesizing the selected PtFeCu nanoparticles and comparing and verifying them.

[0032] The PtFeCu nanoparticles obtained by designing a ternary catalyst for the oxygen reduction reaction can have a composition of Pt. 0.78 Fe 0.09 Cu 0.13 or Pt 0.78 Fe 0.15 Cu 0.07 .

[0033] Methods for manufacturing ternary catalysts for oxygen reduction reactions according to various exemplary embodiments can reduce costs and time by using machine learning to find catalysts that can be synthesized experimentally and screening their performance in advance to minimize the number of candidates for experimental testing.

[0034] Specifically, the methods for manufacturing ternary catalysts for oxygen reduction reactions according to various exemplary embodiments can be used to search for superior materials at speeds 6 million times or more faster than existing quantum mechanics calculations and at least 100 times faster than existing computational simulation methods. In the system catalyst, optimal compositions that can improve catalyst performance and durability can be obtained, and experimental roadmaps can be established based on candidates expected to have excellent durability and performance obtained through calculations, so that research can be carried out efficiently.

[0035] Other aspects of the invention are disclosed below. Attached Figure Description

[0036] Figure 1 The exemplary ternary alloy configuration search, theoretical prediction, and experimental verification of ORR are shown.

[0037] Figures 2A-2E The configuration space of an exemplary ternary PtFeCu nanoparticle with a size of 2.0 nm is shown in the atomic fraction range of 0.6 < Pt < 1. Specifically, Figure 2A This shows the number of atoms in the primary shell; Figure 2B This shows the number of atoms in the secondary shell; and Figure 2C The number of atoms in the (C) tertiary shell of truncated ternary PtFeCu nanoparticles with a particle size of 2.0 nm is shown. Furthermore, Figure 2D The atomic fraction of Pt is shown. Figure 2E The atomic fraction of Fe is shown, and Figure 2F The atomic fraction of Cu is shown.

[0038] Figures 3A-3F The thermodynamic ternary plot and spectral properties of Pt-based ternary nanoparticles are shown. Specifically, Figure 3A An exemplary ternary graph is shown; Figure 3B The convex hull points of ternary PtFeCu nanoparticles (2.0 nm) in the atomic fraction range of 0.6 < Pt < 1 are shown; Figure 3C The XRD pattern is shown; Figure 3D The structural information of representative compositions of PtFe and PtFeCu nanoparticles is shown; Figure 3E The STEM-EDS mapping is shown, and Figure 3F The outline of the STEM-EDS line is shown. Figure 3B In the diagram, orange stars represent synthetic compositions, while... Figure 3C In the middle, the dashed line represents the (111) peak of Pt (JCPDS No.04-0802).

[0039] Figures 4A-4D The theoretical predictions and electrochemical properties of pure (Pt), binary (PtFe), and ternary (PtFeCu) nanoparticles are shown. Specifically, Figure 4A An exemplary catalytically active volcano is shown, calculated for ORR using the d-band center energy of Pt; Figure 4B The elemental distribution in the subshell (secondary shell) of 2.0 nm PtFe and PtFeCu nanoparticles is shown; Figure 4C The CV is shown; and Figure 4D The LSV curve of ORR is shown.

[0040] Figures 5A-5D The electrochemical and chemical stability of pure (Pt), binary (PtFe), and ternary (PtFeCu) nanoparticles are shown. Specifically, Figure 5A The separation energies of each catalyst in the DFT calculations are shown. Figure 5BThe atomic fractions of the primary and secondary shells of PtFe, PtFe-high-Cu-low, and PtFe-low-Cu-high 4 nm nanoparticles are shown. Figure 5C ECSA is shown; while Figure 5D The mass activity of AST30k at 0.9V (relative to RHE) is shown. Detailed Implementation

[0041] The advantages and features of this disclosure, as well as the methods of implementing this disclosure, can be more readily understood by referring to the following detailed description and accompanying drawings of preferred embodiments. However, this disclosure may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. Terms as defined in commonly used dictionaries may not be interpreted ideally or exaggeratingly unless explicitly defined. Throughout this specification, unless explicitly stated otherwise, the word "comprise" and variations such as "comprises" or "comprising" are to be understood as implying inclusion of the stated elements but not exclusion of any other elements.

[0042] Unless otherwise stated, all figures, values ​​and / or expressions relating to the amounts of ingredients, reaction conditions, polymer compositions and formulations used herein should be understood to be modified in all cases by the term “about”, as these figures are approximations by nature, and these approximations reflect in particular the various measurement uncertainties encountered in obtaining these values.

[0043] Furthermore, unless specifically stated or obvious from the context, as used herein, the term “about” is understood to mean within the normal tolerance range in the field, such as within 2 standard deviations of the mean. “About” can be understood as within 10%, 9%, 8%, 7%, 6%, 5%, 4%, 3%, 2%, 1%, 0.5%, 0.1%, 0.05%, or 0.01% of the stated value. Unless the context clearly indicates otherwise, all numerical values ​​provided herein are modified by the term “about”.

[0044] In this specification, when a range is described for a variable, it will be understood that the variable includes all values ​​described within the stated range, including the endpoints. For example, the range “5 to 10” will be understood to include any subranges such as 6 to 10, 7 to 10, 6 to 9, 7 to 9, etc., as well as individual values ​​of 5, 6, 7, 8, 9, and 10, and will also be understood to include any values ​​between valid integers within the stated range, such as 5.5, 6.5, 7.5, 5.5 to 8.5, 6.5 to 9, etc. Similarly, for example, the range “10% to 30%” will be understood to include subranges such as 10% to 15%, 12% to 18%, 20% to 30%, etc., as well as all integers including values ​​up to 30% such as 10%, 11%, 12%, 13%, etc., and will also be understood to include any values ​​between valid integers within the stated range, such as 10.5%, 15.5%, 25.5%, etc.

[0045] Furthermore, there are no particular restrictions on the number of times each step is repeated or the process conditions, as long as they do not impair the purpose of this invention.

[0046] Furthermore, unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0047] In one aspect, a method for manufacturing a ternary catalyst for the oxygen reduction reaction is provided, which may include a first step of constructing a database of the oxygen reduction reaction (ORR) catalytic activity of PtFeCu nanoparticles, a second step of determining the composition of thermodynamically stable PtFeCu nanoparticles, and a third step of analyzing the structure of PtFeCu nanoparticles to select a stable structure.

[0048] In the first step, a database of the oxygen reduction reaction (ORR) catalytic activity of PtFeCu nanoparticles can be constructed using machine learning-based neural network potentials (NNPs).

[0049] To effectively obtain the neural network potential (NNP) of ternary PtFeCu nanoparticles, shape effects such as morphology and size alloying effects such as composition and configuration can be separated, such as... Figure 1 As shown.

[0050] A database can be constructed to simulate the configuration space of PtFeCu nanoparticles. Shape and alloying effects can be varied depending on the cutoff radius. For larger nanoparticles, their atomic environment can be replicated and applied to all nanoparticles using local symmetry functions, but smaller nanoparticles may be exposed to significantly different local environments. This can be tuned by changing the cutoff radius, taking into account the local atomic environment. However, information asymmetry arises when the cutoff radius is small, and computational costs increase dramatically when it is large. For example, the cutoff radius could be set to... To ensure the atomic environment of the nearest neighbor atoms.

[0051] In other words, the training set for machine learning consists of atomic local environments, in which PtFeCu nanoparticles can be divided according to their cutoff radius and may have different morphologies, sizes, compositions, or configurations.

[0052] For example, an atomic local environment may include about 100 to 300 cubic octahedral random structures with a particle size less than or equal to about 1.5 nm, 10 to 20 cubic octahedral random structures with a particle size less than or equal to about 2.0 nm, 100 to 300 truncated octahedral random structures with a particle size less than or equal to about 1.1 nm, and 50 to 150 truncated octahedral random structures with a particle size less than or equal to about 1.7 nm.

[0053] Neural network potentials (NNPs) can be constructed using machine learning parameters of atomic interaction energies calculated by density functional theory (DFT).

[0054] For example, Kohn-Sham density function theory (DFT) calculations applied in VASP (Vienna Ab-initio Simulation Package) can be appropriately used (G. Kresse and J. Furthmuller, Physical Review B, 1996, 54, 11169; J. Hafner, Journal of Computational Chemistry, 2008, 29, 2044-2078).

[0055] For the interaction between the nucleus and the electron, the projector-augmented wave (PAW) pseudopotential can be used (PEBlochl, Physical Review B, 1994, 50, 17953).

[0056] All of Shen Lüjiu's DFT equations can be derived using approximately 10... -5 eV and the treaty The energy and force conversion (convergence) calculations were performed, and the basic plane wave was extended to a cutoff energy of approximately 520 eV.

[0057] The exchange correlation function can be computed using the generalized gradient approximation (GGA) of the Perdew-Burke-Ernzerhof (PBE) or the revised Perdew-Burke-Ernzerhof (RPBE) (JP Perdew, K. Burke and M. Ernzerhof, Physical Review Letters, 1996, 77, 3865; B. Hammer, LB Hansen and JK Physical Review B, 1999, 59, 7413.

[0058] The magnetism of PtFeCu can be obtained by considering spin polarization and van der Waals (vdW) interactions in the DFT-D3 method and initializing different magnetic moments of Fe (S. Grimme, J. Antony, S. Ehrlich and H. Krieg, The Journal of Chemical Physics, 2010, 132, 154104).

[0059] The calculation of nanoparticles can be performed using the Γ-point scheme, which guarantees... The vacuum space is used to ignore the interaction between periodic images.

[0060] The composition and configuration of the nanoparticles can be randomly generated in the bulk (0.975 bulk lattice) using a squeezed lattice parameter to incorporate a squeezed effect optimized for nanoparticles in vacuum (Z. Huang, P. Thomson, and S. Di, Journal of Physics and Chemistry of Solids, 2007, 68, 530-535). Each nanoparticle can be calculated as a single point.

[0061] Adsorption energy (E) of the adsorbed material ads It can be calculated using Equation 1.

[0062] [Equation 1]

[0063] E ads =E NP+吸附质 -E NP -E 吸附质

[0064] Among them, E NP+吸附质 E NP and E吸附质 These are the energy of nanoparticles and adsorbates, and the total energy of nanoparticles and adsorbates, respectively.

[0065] The free energy (ΔG) of ORR was calculated using Equation 2 to confirm the thermodynamic potential of the spontaneous ORR reaction.

[0066] [Equation 2]

[0067] ΔG=ΔE+ΔZPE-TΔS-neU

[0068] Where ΔE is the change in internal energy of the reaction obtained by DFT calculation, ΔZPE and ΔS are the changes in zero-point energy and vibrational entropy, respectively, U is the electrode potential relative to the standard hydrogen electrode, and n is the number of electrons participating in the reaction.

[0069] The free energy diagram was plotted at pH=0, and the relevant ORR mechanism of equations 3 to 6 can be considered.

[0070] [Equation 3]

[0071] *+O2+H + +e - →*OOH

[0072] [Equation 4]

[0073] *OOH+H + +e - →*O+H2O

[0074] [Equation 5]

[0075] *O+H + +e - →*OH

[0076] [Equation 6]

[0077] *OH+H + +e - →*+H2O

[0078] The reaction free energy of ORR can be calculated from the experimental values ​​of the reaction O2 + 2H2 → 2H2O, at 298.15 K and 0.035 bar: ΔG = -4.92 eV (JK). J. Rossmeisl, A. Logadottir, L. Lindqvist, J.R. Kitchin, T. Bligaard and H. Jonsson, The Journal of Physical Chemistry B, 2004, 108, 17886-17892.

[0079] All oxygen intermediates can be adsorbed on the (111) surface of the nanoparticle model, which is oxygen passivated at the edges and vertices, and the most common sites appearing on the surface are selected as adsorption sites (R. Jinnouchi, KKTSuzuki and Y. Morimoto, Catalysis Today, 2016, 262, 100-109; G.-F. Wei and Z.-P. Liu, Physical Chemistry Chemical Physics, 2013, 15, 18555-18561).

[0080] Neural network potentials (NNPs) can be constructed using the Atomic Simulation Environment (ASE) and the Atomic Machine Learning Package (AMP) (A. Khorshidi and A.A. Peterson, Computer Physics Communications, 2016, 207, 310-324; A.H. Larsen, J.J. Mortensen, J. Blomqvist, E.E. Astelli, R. Christensen, M. Dulak, J. Friis, M.G. Roves, B. Hammer and C. Hargus, Journal of Physics: Condensed Matter, 2017, 29, 273002).

[0081] Using the Gaussian descriptor (G) with radial and angular symmetric functions proposed by Behler. 2 and G 4 (J. Behler and M. Parrinello, Physical Review Letters, 2007, 98, 146401). Local symmetry is thought to exist in... Within the range. G 2 and G 4 The various parameter sets can be used to obtain a total of 108 symmetric functions as unique vectors for each atom species, which are used in the input layer of the neural network.

[0082] The activation function between hidden layers can be hyperbolic tangent, and the potential can be trained using the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm until the root mean square error (RMSE) is less than 1 meV·atom. -1(CG. Broyden, IMA Journal of Applied Mathematics, 1970, 6, 76-90; R. Fletcher, The Computer Journal, 1970, 13, 317-322; D. Goldfarb, Mathematics of Computation, 1970, 24, 23-26).

[0083] In order to maintain 10 meV· atoms -1 While testing RMSE, the potential was effectively trained, and different hyperparameters of the number of nodes and hidden layers were tested at 40-30-20-10 and 15-15.

[0084] In the second step, thermodynamically stable PtFeCu nanoparticles can be identified or selected through Monte Carlo calculations.

[0085] To search for thermodynamically stable configurations as a function of alloy composition, Monte Carlo (MC) simulations can be performed in a large canonical ensemble scheme.

[0086] Each MC step consists of n attempts to swap atomic positions, which are randomly selected according to Metropolis's algorithm (B. Han, A. Van der Ven, G. Ceder, and B.-J. Hwang, Physical Review B, 2005, 72, 205409). All simulations were performed on 10,000 trials at temperature T (0 Kelvin or higher), neglecting thermal effects.

[0087] In the third step, structural analysis can be used to identify or select a type of PtFeCu nanoparticle that may be suitable for a ternary catalyst.

[0088] Structural analysis of PtFeCu nanoparticles can be performed by analyzing the number of Pt, Fe, and Cu atoms in each atomic shell of the PtFeCu nanoparticles.

[0089] Furthermore, in the third step, the chemical stability of the ternary PtFeCu nanoparticles can be calculated using Equation 7 by determining the surface segregation energy (Ei) of the alloy composition caused by oxygen atom adsorption. seg To evaluate and select nanoparticles.

[0090] [Equation 7]

[0091] E seg =E(Pt) m Fe nCu l ) seg -E(Pt m Fe n Cu l ) 初始

[0092] In Equation 7, E seg The surface separation energy of alloy components caused by oxygen atom adsorption, E(Pt) m Fe n Cu l ) seg and E(Pt) m Fe n Cu l ) 初始 Pt with or without surface separation m Fe n Cu l The total energy is given by m, n, and 1, where m, n, and 1 represent the number of Pt, Fe, and Cu atoms in the PtFeCu nanoparticles, respectively. In this paper, the lower the separation energy, the easier the separation occurs.

[0093] Example

[0094] The process of selecting PtFeCu nanoparticles through the second and third steps will be described below with reference to specific embodiments.

[0095] First, 2.0 nm ternary PtFeCu nanoparticles with octahedral structure were searched for in the atomic fraction range of 0.6 < Pt < 1 using parameterized NNP. To form PtFeCu nanoparticles including a Pt shell, the Pt content was set to 0.6 or greater as an atomic fraction based on the total weight of the PtFeCu nanoparticles.

[0096] The atomic fraction is the ratio of the number of each atom to the total number of all atoms. For example, the atomic fraction of Pt can be calculated by (the number of Pt atoms) / (the number of Pt, Fe, and Cu atoms).

[0097] As shown in Figure 2, the thermodynamically most stable composition among various alloy compositions was confirmed by MC simulation along the Pt composition line.

[0098] like Figure 2A-2C As shown, each alloying element tends to occupy a specific shell.

[0099] Pt atoms can be in the outermost (first) shell of the nanoparticle, while Fe atoms can be in the second shell, and Cu atoms can be more or less randomly dispersed in the inner shell of the nanoparticle.

[0100] The Pt in the ternary PtFeCu nanoparticles is subjected to compression deformation by smaller Fe atoms. Figure 2D-2E It shows the relationship with Figure 2A-2C The same trend.

[0101] Furthermore, these results are consistent with experimental observations of typical Pt-based alloy catalysts.

[0102] In configuration analysis, the Pt content should be greater than or equal to the theoretical prediction (0.6 atomic fraction) to form ternary nanoparticles including a Pt shell.

[0103] This means that ternary nanoparticles have subtle interactions that may not be detectable by thermodynamic surface energy alone. Pt has the lowest surface energy among the three elements.

[0104] Furthermore, Fe is more likely to be present in the daughter shell, and the three elements compete for space in the core.

[0105] like Figures 3A-3B As shown, the DFT of the alloy PtFeCu nanoparticles was calculated to plot a ternary phase diagram, where the ground-state structure is identified by the energy convex hull in the composition range of 0.6 < Pt < 1.0 atomic fraction.

[0106] Most of the inclusion spots in Pt compositions with an atomic fraction greater than 0.8 indicate the structure of the Pt shell. In Pt compositions with an atomic fraction less than 0.8, Fe and Cu are distributed together in the primary shell.

[0107] like Figure 2D-2F As shown, since the initial compositions of Cu and Fe are 0.80 atomic fractions and 0.73 atomic fractions respectively, Cu is more likely to exist in the primary shell than Fe.

[0108] To predict the catalytic performance of ternary PtFeCu nanoparticles for ORR, three different Pt-FeCu nanoparticle compositions were selected. 0.82 Fe 0.18 (PtFe), Pt 0.82 Fe 0.12 Cu 0.06 (PtFe 高 Cu 低 ) and Pt 0.8 Fe 0.08 Cu 0.12 (PtFe 低 Cu 高 ).like Figure 3D As shown, the selected composition has a Pt surface structure and a thermodynamically stable energy convex hull.

[0109] As the Cu content increases, Cu may be located in the outermost shell, limiting the composition of Cu.

[0110] Optionally, in the fourth step, the selected PtFeCu nanoparticles are experimentally synthesized and then compared and verified with calculated values.

[0111] Figure 3C X-ray diffraction patterns of synthesized PtFe and PtFeCu catalysts and commercial Pt / C are shown.

[0112] All samples exhibited the same face-centered cubic (fcc) structure as bulk Pt. Furthermore, no phase separation was observed in any region. Additionally, the (111) peak of the PtFe and PtFeCu catalysts was shifted at a higher angle than that of the Pt / C peak.

[0113] This means that relatively small Fe and Cu atoms are incorporated into the Pt lattice, leading to compressive deformation. Based on the Scherrer equation, the Pt... 0.83 Fe 0.17 (PtFe), Pt 0.78 Fe 0.15 Cu 0.07 (PtFe 高 Cu 低 ) and Pt 0.78 Fe 0.09 Cu 0.13 (PtFe 低 Cu 高 The crystal sizes of each sample were 2.1 nm, 2.2 nm, and 2.3 nm, respectively.

[0114] All samples were highly dispersed on the carbon support and showed a uniform particle size (less than or equal to 3 nm).

[0115] like Figure 3D-3E As shown, the elemental distribution was analyzed for STEM-EDS mapping and line profiles were scanned at a point resolution of 0.08 nm to examine the compositional information of PtFe and PtFeCu nanoparticles.

[0116] For PtFe nanoparticles, Fe-K signaling was observed in the core region and Pt-M signaling was observed throughout the nanoparticles.

[0117] Based on the difference between the two profiles, it can be predicted that the PtFe catalyst consists of about 1 to 2 Pt shells (about 0.3 nm to 0.5 nm) on the catalyst surface.

[0118] Furthermore, regardless of composition, PtFeCu catalysts all possess a core-shell structure.

[0119] The core-shell structure is expected to form during synthesis through successive acid and heat treatments. Furthermore, Cu atoms are more abundant on the exterior of the nanoparticles than Fe atoms. Figure 3F ).

[0120] The experimentally synthesized PtFeCu nanoparticles exhibited similar structural conditions, such as particle size, composition, and elemental distribution. In other words, the experimental results were consistent with the computational model system.

[0121] For example, PtFe catalysts can be synthesized using a simple ultrasound-assisted polyol method.

[0122] First, 2.6 mmol of Pt(acac)2, 3.9 mmol of Fe(acac)3 and 1.35 g of thermally graphitized Ketjen black were dispersed in 100 mL of argon-purified ethylene glycol (EG) at 600 J (1200 °C).

[0123] The precursor dispersion was irradiated with a solid-state horn-type ultrasonic generator (tip diameter: 13 mm, amp. 40%, VCX-750, Sonic & Materials, Inc.). The ultrasonic reaction was carried out at 150°C or higher for 4 hours.

[0124] The resulting dark slurry was then sieved using a membrane filter (0.4 μm pore size, Advantec Toyo Kaisha, Ltd.).

[0125] The obtained sample was washed several times with excess ethanol and deionized water to remove residual water and EG.

[0126] The obtained sample sections were dried overnight in an oven at 80°C.

[0127] Then, the prepared sample was placed in an alumina crucible and annealed at 400°C for 2 hours in a mixture of H2 / Ar (v / v% = 4 / 96).

[0128] The annealed sample was dispersed in a mixture of ethanol and 0.1 M HClO4 at a ratio of 1 / 4 (v / v%), and then acid-treated twice at 94 °C to remove unwanted residues such as FeO. x .

[0129] Except for the composition and ratio of the metal precursors, ternary PtFeCu catalysts were synthesized using the same procedure as described above. All samples were synthesized in batches of 2 g.

[0130] The characteristics of the ternary PtFeCu catalyst were measured as follows.

[0131] X-ray diffraction (XRD, Bucker D2 PHASER XE, Cu kα) was used. )Measure its crystal structure.

[0132] The elemental composition and Pt loading of the samples were measured and averaged using an elemental analyzer (FlashEA 1112, Thermo Finnigan) and an inductively coupled plasma atomic emission spectrometer (ICP-AES, OPTIMA 4300DV).

[0133] The Pt loading of the catalyst film in the rotating disk electrode (RDE) was evaluated using an X-ray fluorescence analyzer (XRF, Hofriba, MESA-50).

[0134] The particle size and alloy element distribution of the samples were obtained by field emission transmission electron microscopy (FM-TEM, FEI, Talos F200X, 200kV) and Cs-corrected FE-TEM (FEI, Titancubed G260-300, 300kV).

[0135] In addition, the electrochemical characteristics of the prepared ternary PtFeCu catalyst were measured as follows.

[0136] Disperse 10 mg of powder in deionized water and isopropanol (IPA) (v / v% = 4:1).

[0137] Subsequently, 10 μL of ionomer dispersion (FSS-2, ASAHIGLASS Co., Ltd.) was added to the catalyst dispersion, and then ultrasonically treated with a bath apparatus until a uniform catalyst ink was formed.

[0138] Subsequently, 13.3 μL of catalyst ink was placed on a glassy carbon rotating disk electrode (RDE, 5.0 mm disk outer diameter, 12.0 mm outer diameter PTFE shield, active area: 0.196 cm). 2 On the Pine Research Instrumentation.

[0139] The Pt loading of glassy carbon was fixed at 20.4 μg·cm⁻¹. -2 .

[0140] Electrochemical measurements were performed using a three-electrode cell system, which consisted of a catalyst-coated RDE as the working electrode, a reversible hydrogen electrode (RHE, Gaskatel GmbH) as the reference electrode, and a Pt wire as the counter electrode.

[0141] Prior to measurement, the catalyst-coated RDE was washed for 300 cycles in N2-saturated 0.1M HClO4 at potentials ranging from 0.03V to 1.1V (relative to RHE).

[0142] In solution at 20 mVs -1 The scanning rate was used to record the cyclic voltammetry (CV) for each sample.

[0143] In O2-saturated 0.1M HClO4 at 10 mVs -1 Linear scan voltametry (LSV) for ORR was measured for each sample at a scan rate of 1600 rpm over a potential range of 0.0 V to 1.1 V (relative to RHE).

[0144] In addition to a 0.1M HClO4 solution saturated with N2, LSV curves were also measured under the same conditions to remove background current.

[0145] IR compensation was performed by measuring the impedance at 0.7V, 0.8V, and 0.9V during ORR catalytic performance.

[0146] Accelerated stress testing (AST) was performed by applying square wave potential cycles between 0.6V (3s) and 0.95V (3s) at 30,000 (30k) according to the U.S. Department of Energy (DOE) electrocatalyst protocol.

[0147] By mapping four nanoparticles (such as Pt, PtFe, PtFe) 高 Cu 低 and PtFe 低 Cu 高 The ORR free energy diagram of PtFe was used to evaluate the catalyst performance. 高 Cu 低 >PtFe>PPtFe 低 Cu 高 The properties of four nanoparticles were obtained in the order of Pt with overvoltages of 0.31 eV, 0.33 eV, 0.37 eV and 0.45 eV.

[0148] This prediction is consistent with the d-band center energy of Pt, such as Figure 4A As shown. Due to the optimal binding energy of the oxygen intermediate, PtFe in the nanoparticles 高 Cu 低 Located at the top of a volcano. The activity of Pt-based alloy nanocatalysts can be controlled by a strain field created by mixing nanoparticles with elements of different sizes. A key function of this technique is finding an appropriate strain to obtain the optimal binding energy. Therefore, controlling the spatial distribution of elements is crucial for tuning catalytic activity.

[0149] like Figure 4B As shown, PtFe and PtFe were calculated. 高 Cu 低 and PtFe 低 Cu 高The elemental distribution within the secondary shell is shown. Compressive strain is primarily determined by the Fe content in the secondary shell. Within the secondary shell, variations in Pt content have little effect on Cu substitution for Fe. Therefore, a lower Cu loading can optimize bond strength.

[0150] To verify this theoretical prediction, the catalytic performance of the nanoparticles was measured using a typical three-electrode cell in an acidic medium. Figure 4C The periodic voltage-current plots of the sample in N2-saturated 0.1M HClO4 and the periodic voltage-current plots of Pt / C are shown.

[0151] PtFe 高 Cu 低 and PtFe 低 Cu 高 A copper dissolution peak was observed at 0.7 V in the first cycle, but the peak disappeared immediately. This provides evidence of Cu dealloying in the outermost shell.

[0152] The electrochemical surface area (ECSA) of the samples was evaluated by integrating the charge of the hydrogen desorption peak in the potential range of 0.03 V to 0.4 V (relative to RHE). Samples: Pt / C, PtFe, PtFe 高 Cu 低 and PtFe 低 Cu 高 The calculated ECSA values ​​were 84.1, 102.6, 86.7, and 67.4 m, respectively. 2 g -1 .

[0153] Linear sweep voltammograms (LSVs) of the samples against the ORR were measured at 1600 rpm in an O2-saturated 0.1 M HClO4 electrolyte. Compared to Pt / C, both PtFe and PtFeCu catalysts showed higher onset potentials and higher half-wave potentials (E0). 1 / 2 ).

[0154] like Figure 4D As shown, PtFe 高 Cu 低 It exhibits the highest mass activity (0.67 m at 0.9 V). 2 g -1 This mass activity is the mass activity of Pt / C (0.21m). 2 g -1 The ORR performance is 3.2 times that of the standard PtFe binary catalyst. Based on electrochemical measurements, the ORR performance can be improved when Cu is appropriately added to the PtFe binary catalyst.

[0155] Figure 4 shows that both calculations and experiments demonstrate the improved activity of the ternary alloys. However, these alloys are difficult to use practically and commercially without guaranteeing long-term stability. Therefore, the stability of the ternary alloy nanoparticles was evaluated.

[0156] When calculating model nanoparticles PtFe and PtFe with the same particle size 高 Cu 低 and PtFe 低 Cu 高 When the electrochemical dissolution potential was calculated, the results were 0.96V, 0.93V, and 0.91V, respectively.

[0157] It is estimated that both binary (PtFe) and ternary (PtFeCu) nanoparticles have higher dissolution potentials than pure 2.0 nm Pt nanoparticles (relative to SHE 0.83 V).

[0158] PtFe binary catalysts exhibit the highest electrochemical stability. However, since alloy nanocatalysts inevitably undergo chemical reactions on their surfaces, their durability depends on more than just the electrochemical environment.

[0159] In fact, three-dimensional transition metals such as Fe and Cu can be easily separated due to their oxygen absorption capacity. Therefore, although metals absorb oxygen intermediates during ORR, it is necessary to determine whether the metals maintain the integrity of their surface structure.

[0160] To examine surface separation caused by oxygen adsorbates, a passivation model can be hypothesized, where oxygen is poisoned at the edges or apex of the nanoparticles due to strong oxygen binding. Then, the alloying elements (Fe and Cu) of PtFe binary nanoparticles and PtFeCu ternary nanoparticles were investigated to separate into four surface regions (

[111] , edge (

[111] X

[111] ), and edge (

[111] X

[100] ),

[100] ).

[0161] like Figure 5A As shown, surface separation occurs in the order of difficulty:

[100] >

[111] > edge (

[111] x

[100] ) > edge (

[111] x

[111] ). The

[100] and

[111] surfaces exhibit stronger surface separation resistance than the edge (

[111] x

[100] ) and edge (

[111] x

[111] ). The

[111] surface is very important for ORR, but remains relatively strong in terms of surface separation. However, Fe, which has weaker surface separation resistance, may reduce long-term durability. On the other hand, Cu exhibits high surface separation resistance, and therefore, PtFe... 高 Cu 低 and PtFe 低 Cu 高The Fe separation energies for the

[111] facet (partial) are -1.07 and -0.96 eV, respectively, while the Cu separation energies are -0.46 and -0.58 eV, respectively, for example, about half the Fe affinity. Figure 5A As shown, relatively low surface separation energies of Cu atoms were also observed at other locations. This implies that the high Cu composition in the PtFeCu ternary alloy can improve chemical stability due to Cu replacing Fe in the daughter shell.

[0162] However, such as PtFe 低 Cu 高 (Pt 0.8 Fe 0.08 Cu 0.12 As shown in the figure, when the Cu concentration is limited, Cu atoms may easily reside on the outermost surface of the ternary catalyst as the Cu composition increases. Therefore, the Pt surface structure may be disrupted due to the high Cu content through the dealloying process of Cu on the surface. Figure 2A ).

[0163] In addition, the distribution of alloying elements was analyzed to infer the stability of PtFeCu with different compositions. Figure 5B Since the subsurface elements are separated into the outermost shell, the elemental distribution in the primary and secondary shells was calculated. The atomic fraction was obtained by averaging ten initial configurations of randomly generated 4.0 nm nanoparticles in a 9,000 MC step simulation. The Fe composition in both the primary and secondary shells decreased with increasing Cu content. This result indicates that the introduction of Cu can improve catalyst stability.

[0164] Since the Cu in the outermost shell is highly likely to be de-alloyed, the primary and secondary shells contain low Fe content PtFe. 低 Cu 高 Nanoparticles are resistant to electrochemical cycling. The improved durability achieved through the formation of these ternary alloys can be demonstrated through durability testing.

[0165] AST was performed by cycling spherical potentials of 0.6V for 3 seconds and 0.95V for 3 seconds. Figures 5C-5D ECSA and MA at 0.9V before and after AST30k were compared.

[0166] The Pt / C and PtFe catalysts were confirmed to have similar durability. The significant decrease in ECSA (-42.1%) of the PtFe catalyst was due to particle growth (from 2.95 nm to 4.56 nm) during AST 30 k. However, both PtFeCu catalysts only grew by 0.3 nm during the experiment.

[0167] Additionally, PtFe occurred after AST 30k. The change is 17mV, PtFe 高 Cu 低 of The change is 4mV, PtFe 低 Cu 高 of The change was -4 mV. These results indicate that the addition of Cu improves the durability of the PtFe catalyst.

[0168] In addition, PtFe 高 Cu 低 The MA of the sample was three times that of the Pt / C sample, and it exceeded the DOE's 2020 target (0.44 mg) around 30 kDa. Pt -1 ()( Figure 5D ).

[0169] According to various exemplary embodiments of the present invention, the alloying ratio of PtFeCu can be optimized by systematically examining the previously reported beneficial effects of Cu, which adjusts the strain energy of Pt on the surface and prevents Fe from separating into the active surface of the catalyst, and also employs a configuration space search algorithm based on machine learning.

[0170] Therefore, PtFe 低 Cu 高 The best catalyst performance can be obtained in ternary composition samples.

[0171] While the invention has been described in conjunction with what is now considered to be exemplary embodiments, it should be understood that the invention is not limited to the disclosed embodiments, but rather is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method for manufacturing a three-way catalyst for oxygen reduction reaction, comprising the following steps: A database of catalytic activity of oxygen reduction reaction (ORR) for PtFeCu nanoparticles was constructed using machine learning-based neural network potentials. Thermodynamically stable PtFeCu nanoparticles were determined using Monte Carlo calculations; and By analyzing the structure of the PtFeCu nanoparticles, a type of PtFeCu nanoparticles was selected for the ternary catalyst. The training set used for machine learning includes atomic local environments with different morphologies, sizes, compositions, or configurations, wherein the PtFeCu nanoparticles are divided according to their cutoff radius size; and The atomic local environment includes 100 to 300 cubic octahedral random structures with a particle size less than or equal to 1.5 nm, 10 to 20 cubic octahedral random structures with a particle size less than or equal to 2 nm, 100 to 300 truncated octahedral random structures with a particle size less than or equal to 1.1 nm, and 50 to 150 truncated octahedral random structures with a particle size less than or equal to 1.7 nm. The catalytic activity of the PtFeCu nanoparticles for the oxygen reduction reaction was calculated using Equation 2: [Equation 2] In Equation 2, is the free energy of ORR, is the change in the internal energy of the reaction obtained by DFT calculations, and respectively are the zero point energy and the change in vibrational entropy, U is the electrode potential relative to the standard hydrogen electrode, and n is the number of electrons participating in the reaction; The thermodynamically stable PtFeCu nanoparticles are determined by mapping the calculated density functional theory (DFT) of PtFeCu nanoparticles to a ternary phase diagram and expressing the Monte Carlo calculation results as the energy convex hull on the ternary phase diagram; and The chemical stability of the PtFeCu nanoparticles was evaluated using Equation 7: [Equation 7] In Equation 7, E seg is the surface segregation energy of the alloy component caused by the adsorption of oxygen atoms, E( Pt m Fe n Cu l ) seg and E( Pt m Fe n Cu l ) 初始 Pt with or without surface separation m Fe n Cu l The total energy, and m, n, and l represent the number of Pt, Fe, and Cu atoms in the PtFeCu nanoparticles, respectively.

2. The method of claim 1, wherein the neural network potential is constructed using machine learning parameters of atomic interaction energy, wherein the atomic interaction energy is calculated using density functional theory (DFT).

3. The method of claim 1, wherein the Monte Carlo calculation for each PtFeCu nanoparticle includes n attempts to randomly swap atomic positions.

4. The method of claim 1, wherein, in order to configure the PtFeCu nanoparticles comprising a Pt shell, the Pt content based on the total weight of the PtFeCu nanoparticles is set to be greater than or equal to 0.6 atomic fractions.

5. The method according to claim 1, wherein the structure of the PtFeCu nanoparticles is analyzed by analyzing the number of Pt, Fe, and Cu atoms in each atomic shell of the PtFeCu nanoparticles.

6. The method according to claim 1, further comprising the following step: The selected PtFeCu nanoparticles were synthesized.

7. The method of claim 1, wherein the PtFeCu nanoparticle produced by the method is Pt 0.78 Fe 0.09 Cu 0.13 or Pt 0.78 Fe 0.15 Cu 0.07 .

Images