Determining the three-dimensional atomic structure of a metallic glass
Yao Yang, Jihan Zhou, Fan Zhu, Yakun Yuan, Dillan Chang, Dennis S. Kim, Minh Pham, Arjun Rana, Xuezeng Tian, Yonggang Yao, Stanley Osher, Andreas K. Schmid, Liangbing Hu, Peter Ercius, Jianwei Miao
11 Determining the three-dimensional atomic structure of an amorphous solid
Yao Yang , Jihan Zhou , Fan Zhu , Dillan Chang , Dennis S. Kim , Yakun Yuan , Minh Pham , Arjun Rana , Xuezeng Tian , Yonggang Yao , Stanley Osher , Liangbing Hu , Peter Ercius & Jianwei Miao Department of Physics & Astronomy and California NanoSystems Institute, University of California, Los Angeles, CA 90095, USA. Department of Mathematics, University of California, Los Angeles, CA 90095, USA. Department of Materials Science and Engineering, University of Maryland, College Park, Maryland, 20742, USA. National Center for Electron Microscopy, Molecular Foundry, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA. * These authors contributed equally to this work. † Correspondence and requests for materials should be addressed to J.M. ([email protected]).
Amorphous solids such as glass are ubiquitous in our daily life and have found broad applications ranging from window glass and solar cells to telecommunications and transformer cores . However, due to the lack of long-range order, the three-dimensional (3D) atomic structure of amorphous solids have thus far defied any direct experimental determination . Here, using a multi-component metallic glass as a model, we advance atomic electron tomography to determine its 3D atomic positions and chemical species with a precision of 21 picometer. We quantify the short-range order (SRO) and medium-range order (MRO) of the 3D atomic arrangement. We find that although the 3D atomic packing of the SRO is geometrically disordered, some SROs connect with each other to form crystal-like networks and give rise to MROs, which exhibit translational but no orientational order. We identify four crystal-like MROs face-centred cubic, hexagonal close-packed, body-centered cubic and simple cubic coexisting in the sample, which significantly deviate from the ideal crystal structures. We quantify the size, shape, volume, and structural distortion of these MROs with unprecedented detail. Looking forward, we anticipate this experiment will open the door to determining the 3D atomic coordinates of various amorphous solids, whose impact on non-crystalline solids may be comparable to the first 3D crystal structure solved by x-ray crystallography over a century ago . Since the first discovery in 1960 , metallic glasses have been actively studied for fundamental interest and practical applications . However, due to their disordered internal structure, the 3D atomic arrangement of metallic glasses cannot be determined by crystallography . Over the years, a number of experimental and computational methods have been used to study the metallic glass structure, such as x-ray/neutron diffraction , x-ray absorption fine structure , high-resolution transmission electron microscopy , fluctuation electron microscopy , nanobeam electron diffraction , nuclear magnetic resonance , density functional theory , molecular dynamics simulations and reverse Monte Carlo modelling . Despite all these developments, however, there is currently no experimental method available to directly resolve the 3D atomic structure of metallic glasses. One experimental method that can potentially solve this long-standing problem is atomic electron tomography (AET) . AET combines high-resolution tomographic tilt series with advanced iterative algorithms to determine the 3D atomic positions in materials without assuming crystallinity, which has been applied to image grain boundaries, anti-phase boundaries, stacking faults, dislocations, point defects, chemical order/disorder, atomic-scale ripples, bond distortion and strain tensors with high precision . More recently, 4D (3D + time) AET has been developed to observe crystal nucleation at atomic resolution, showing early stage nucleation results are not consistent with classical nucleation theory . Here, we use a multi-component metallic glass as a model and advance AET to determine the 3D atomic positions in an amorphous solid for the first time. Determining the 3D atomic positions in a multi-component metallic glass
The samples were synthesized by a carbothermal shock technique with a cooling rate as high as 69000 K/s (Supplementary Fig. 1a and Methods), which created high entropy alloy nanoparticles with multi-metal components . The energy-dispersive X-ray spectroscopy data show the nanoparticles are composed of eight elements: Co, Ni, Ru, Rh, Pd, Ag, Ir and Pt (Supplementary Fig. 2). Tomographic tilt series were acquired from six nanoparticles using a scanning transmission electron microscope with an annular dark-field detector (Supplementary Table 1). While most of the nanoparticles show the crystalline structure, particles 1 and 2 exhibit more disordered structures (Supplementary Fig. 3). In this study, we focus on the most disordered nanoparticle (particle 1) among the six, from which a tilt series of 51 high-resolution images was acquired (Supplementary Fig. 4). After pre-processing and image denoising, the tilt series was reconstructed by a newly developed real space iterative algorithm and the 3D atomic positions and chemical species were traced and classified (Supplementary Fig. 5, Methods). Since the image contrast in the 3D reconstruction is based on the atomic number , presently AET is only sensitive enough to classify the eight elements into three different types: Co and Ni as type 1, Ru, Rh, Pd and Ag as type 2, and Ir and Pt as type 3, as the atomic number of the elements in each type is close to each other. After atom classification, we obtained the experimental 3D model of the disordered nanoparticle, consisting of 8794, 8587 and 3472 atoms for type 1, 2 and 3, respectively. To validate the reconstruction, atom tracing and classification procedure, we calculated 51 projections from the experimental atomic model using multislice simulations . Supplementary Fig. 6a and b shows the consistency between the experimental and calculated projections. We then applied the same reconstruction, atom tracing and classification procedure to determine a new 3D atomic model from the 51 multislice projections. By comparing the two models, we confirmed that 96.5% of atoms were correctly identified with a 3D precision of 21 pm (Supplementary Fig. 6c). Figure 1a and Supplementary video 1 show the experimental 3D atomic model of the nanoparticle with type 1, 2 and 3 atoms in green, blue and red, respectively, which exhibits disordered atomic structure. To quantify the disorder, we calculated the local bond orientational order (BOO) parameter of the atoms (Methods). Figure 1b shows the BOO distribution of all the atoms in the nanoparticle, indicating the majority of atoms significantly deviate from the ideal crystal structures. For a comparison, the BOO distribution of the other five nanoparticles is shown in Supplementary Fig. 3h-l. By using a threshold of BOO . 18 nuclei are situated on or near the surface and only one nucleus is located near the centre of the nanoparticle. To separate the amorphous structure from the crystal nuclei, we focus on the analysis of the atoms with BOO < 0.5 in the following sessions. Figure 1d shows the radial distribution function (RDF) of the amorphous structure, where the splitting of the 2 nd and 3 rd peak was observed with the Gaussian fit (Fig. 1d, inset). The ratios of the 2 nd , 3 rd , 4 th and 5 th to the 1 st peak position were determined to be 1.82, 2.13, 2.74 and 3.62, respectively, which are consistent with those of metallic glasses . The pair distribution functions (PDFs) between type 1, 2 and 3 atoms are shown in Fig. 1e, which consist of 6 pairs - type 11, 12, 13, 22, 23 and 33 atoms. By fitting the Gaussian to the 1 st peaks in the PDFs, we measured the bond lengths of type 11, 22 and 33 atoms to be 2.60, 2.67 and 2.66 Å, respectively, which are consistent with our estimation that the average size of type 2 and 3 atoms is comparable to each other and both are slightly larger than that of type 1 atoms (Methods). While the bond lengths of type 12 (2.65 Å) and type 13 (2.66 Å) agree with the average values of the corresponding atom types, the bond length of type 23 (2.60 Å) is 2.4% shorter than the expected value (Methods), indicating a bond shortening between type 2 and 3 atoms. Furthermore, unlike other five PDFs, the PDF for the type 33 atoms (the yellow curve) exhibits a unique feature with a higher 2 nd peak than the 1 st peak, suggesting that the majority of type 3 atoms are distributed beyond the SRO. The short-range order
To determine the SRO in the metallic glass sample, we used the Voronoi tessellation method to characterize the local atomic arrangement . This method identifies the nearest neighbour atoms around each central atom to form a Voronoi polyhedron, which is designated by a Voronoi index < n , n , n , n > with n i denoting the number of i -edge faces. Figure 2a shows the ten most abundant Voronoi polyhedra in the nanoparticle with a fraction ranging from 7.3% to 3.1%, most of which are geometrically disordered and commonly observed in model metallic glasses such as <0,4,4,3>, <0,4,4,2>, <0,2,8,1> and <0,3,6,3> (Fig. 2b). Although the Voronoi index of <0,4,4,4> can be considered a distorted fcc structure, the 3D atomic positions in these clusters severely deviate from an idea fcc lattice and exhibit a disordered structure (Fig. 2b). Figure 2c shows the local symmetry distribution for all the faces of the Voronoi polyhedra. The 3-, 4-, 5- and 6-fold faces account for 3.2%, 31%, 45.3% and 20.5%, respectively, revealing 5-fold faces are most abundant in the SRO. But only 9.2% of all the Voronoi polyhedra are distorted icosahedra, which include Voronoi indices <0,0,12,0>, <0,1,10,2>, <0,2,8,2> and <0,2,8,1>. This result indicates that most 5-fold faces do not form distorted icosahedra in this metallic glass nanoparticle. From the Voronoi tessellation, we also calculated the coordination number (CN) distribution for the atoms (Fig. 2d), where the average CNs for types 1, 2 and 3 are 10.9, 10.9 and 10.7, respectively. The smaller average CN of type 3 relative to the other two types of atoms can be attributed to the shortening of the type 23 bond (Fig. 1e, Methods), which reduces the space for packing solvent atoms around the type 3 atoms. The bond shortening also implies a stronger interaction or chemical SRO between type 2 and 3 atoms. As a verification, we quantified the chemical SRO using the Warren–Cowley parameter (Methods), confirming that the type 11 and 23 bond are favoured in the metallic glass sample (Supplementary Fig. 7a). The medium-range order
From the PDF of type 33 atoms (Fig. 1e, the yellow curve), we observed that the highest peak is located at 4.76 Å and is 1.22 times higher than the nearest neighbour peak. This result indicates that the majority of type 3 atoms are distributed in the second coordination shell, which is between the first (3.78 Å) and the second minimum (6.09 Å) of the RDF curve (Fig. 1d). According to the efficient cluster packing model , solute atoms are surrounded by solvent atoms to form solute-centred clusters. These solute-centred clusters act as the basic building blocks and are densely packed in 3D space to constitute the MRO of metallic glasses. To quantitatively test this model with experimental data, we first examined the distribution of the type 3 atoms in the second coordination shell. Supplementary Fig. 7b and c shows that 79.8% of type 3 atoms (2769) satisfy this criterion and are fairly uniformly distributed inside the nanoparticle. These type 3 atoms act as solute centres and are surrounded mainly by type 1 and 2 solvent atoms to form atomic clusters. Supplementary Fig. 7d shows the ten most abundant Voronoi polyhedra of the solute-centred clusters. The solute-centred clusters connect with each other by sharing one (a vertex), two (an edge) and three atoms (a face) as well as protrude into each other by sharing four and five atoms (Fig. 3a-e). Figure 3f shows the statistical distribution of the number of the solute-centred cluster pairs, which share from one to five atoms. To locate the MRO, we implemented a breadth-first search algorithm to search for the fcc-, hcp-, bcc-, simple cubic (sc-) and icosahedral-like networks of the solute centres in the metallic glass nanoparticle (Methods). This algorithm enables us to globally search for MROs with a maximum number of solute centres. Each MRO is defined to have five or more solute centres with each solute centre falling within a 0.75 Å radius relative to the perfect fcc, hcp, bcc, sc lattice or icosahedral vertices. We found there are four types of MROs (fcc-, hcp-, bcc- and sc-like), but no icosahedral-like MROs in the nanoparticle (Methods). Figure 3g shows the histogram of the four types of MROs as a function of the size (i.e. the number of solute centres), where the inset illustrates the fraction of the four MRO solute centre atoms. Figure 3h and Supplementary Video 2 show the 3D distribution of the four types of MROs with each having ten solute centres or more. To validate our analysis, we also searched for the MROs with a 1 Å and 0.5 Å radius cut-off. While the total number of the MROs changes as the cut-off radius, the fraction of the four MRO solute centre atoms is consistent among the three different cut-off radii (Supplementary Figs. 8 and 9). Next, we quantified the MROs with a 0.75 Å radius cut-off. Figure 4a and b shows the length and volume distribution of the four types of the MROs in the metallic glass nanoparticle. The average length and volume of the fcc-, hcp-, bcc- and sc-like MROs were measured to be 2.33 , respectively. The high standard deviations indicate a large variation in the length and volume of the MROs. Figure 5a, c, e and g shows four representative fcc-, hcp-, bcc- and sc-like MROs, where the solute-centred clusters exhibit only the translational but no orientational order. To better visualize these MROs, the solute centres are orientated along the fcc, hcp, bcc and sc zone axes (Fig. 5b, d, f and h), showing that the 3D shapes of the MROs are anisotropic and the networks are distorted. We quantified the distortion of the MROs by calculating the partial RDFs of all the fcc-, hcp-, bcc- and sc-like solute centres in the metallic glass nanoparticle with the corresponding maximum peak positions at 4.84, 4.76 4.84 and 3.97 Å, respectively (Fig. 4c). These peak positions represent the average nearest neighbour distances of the solute centres in the four crystal-like MROs and the broadened peaks signify the severe deviation of the MROs from the ideal crystal structures. Compared with the other three partial RDFs, the partial RDF of the sc-like MRO has two peaks and the ratio of the 2 nd to the 1 st peak position is about √2 (Fig. 4c, the purple curve), which corresponds to the ratio of the diagonal to the side length of a square. The shorter nearest neighbour distance of the sc-like MROs relative to the other three crystal-like networks and the appearance of the two peaks in the partial RDF indicate that the sc-like solute-centred clusters are more closely connected with their neighbours. Figure 4d shows the distribution of sharing one, two, three, four and five atoms between neighbouring solute-centred clusters for the four types of the MRO, confirming that the solute-centred clusters in sc-like MROs tend to share more atoms with their neighbours than those in other types of MROs. Our quantitative analysis of the SRO and MRO in a multi-component metallic glass confirm the general framework of the efficient cluster packing model , that is, solute-centred clusters are densely packed in 3D space to give rise to the MRO. However, we observed that the shortening of a solute and solvent bond contributes to the formation of the solute-centred clusters. We revealed that fcc-, hcp-, bcc- and sc-like MROs co-exist in the multi-component metallic glass and there are no icosahedral-like MROs. By quantifying their length, volume and 3D structure, we found that the MROs not only have a large variation in length and volume, but also significantly deviate from the ideal crystal structures (Fig. 4c). These results indicate that the MROs in real metallic glasses are more complicated than previously thought. Furthermore, as the size of MROs is comparable to that of shear transformation zones in metallic glasses , AET could also be applied to determine the 3D atomic structure of shear transformation zones and link the structure and properties of metallic glasses . Outlook
Over the last century, crystallography has been broadly applied to determine the 3D atomic structure of crystalline samples . The quantitative 3D structural information has been fundamental to the development of many scientific fields. However, for amorphous solids, their 3D structure has been primarily inferred from experimental data, where the statistical structural information can be obtained but the 3D local structure is lost . This qualitative approach has hindered our fundamental understanding of the 3D structure of amorphous solids and related phenomena such as the crystal-amorphous phase transition and the glass transition . Here, we demonstrate the ability to directly determine the 3D atomic structure of an amorphous solid using AET, which enables us to quantitatively analyse the SRO and MRO at the single-atom level. Although we focus on a metallic glass nanoparticle in this study, this method is generally applicable to different sample geometry such as thin films and extended objects (Methods). Furthermore computer simulations have indicated that AET can also be used to resolve the 3D atomic positions of silicate glasses . Thus, we anticipate that this work will open the door to determining the 3D structure of a wide range of amorphous solids. Furthermore, by annealing amorphous solid samples at different times and temperatures (below and above the transition temperature), 4D AET can be applied to reveal their atomic structure as a function of time and temperature , which would not only enable us to capture the amorphous-crystalline phase transition and interface dynamics, but may also shed light on the glass transition at the atomic scale . References Zallen, R.
The Physics of Amorphous Solids (Wiley, 1998). 2.
Elliott, S. R.
Physics of Amorphous Materials (Harlow, Essex, England: Longman Scientific & Technical, 1990). 3.
Zachariasen, W. H. The atomic arrangement in glass.
J. Am. Chem. Soc. , 3841-3851 (1932). 4. Bernal, J. D. & Mason, J. Packing of spheres: co-ordination of randomly packed spheres.
Nature , 910-911 (1960). 5.
Scott, G. D. Packing of equal spheres.
Nature , 908-909 (1960). 6.
Finney, J. L. Random packings and structure of simple liquids. 1. Geometry of random close packing.
Proc. Roy Soc. Lond. Ser. A - Math. Phys. Sci. ,479-493 (1970). Nelson, D. R. & Spaepen, F. Polytetrahedral order in condensed matter.
Solid State Phys. , 1-90 (1989). 8. Miracle, D. B. A structural model for metallic glasses.
Nat. Mater. , 697-702 (2004). 9. Sheng, H. W., Luo, W. K., Alamgir, F. M., Bai, J. M. & Ma, E. Atomic packing and short-to-medium-range order in metallic glasses.
Nature , 419-425 (2006). 10.
Miracle, D. B., Egami, T., Flores, K. M. & Kelton, K. F. Structural aspects of metallic glasses.
MRS Bull. , 629-634 (2007). 11. Cheng, Y. Q. & Ma, E. Atomic-level structure and structure-property relationship in metallic glasses.
Prog. Mater. Sci. , 379-473 (2011). 12. Miracle, D. B. A Physical Model for Metallic Glass Structures: An Introduction and Update.
JOM , 846–855(2012). 13. Bragg, W. L. The structure of some crystals as indicated by their diffraction of X-rays.
Proc. R. Soc. London Ser . A , 248-277 (1913). 14. Klement Jun, W., Willens, R. H. & Duwez, P. Non-crystalline structure in solidified gold–silicon alloys.
Nature , 869-870 (1960). 15.
Greer, A. L. Metallic glasses.
Science , 1947-1953 (1995). 16.
Peker, A. & Johnson, W. L. A highly processable metallic glass: Zr
41. 2 Ti
13. 8 Cu
12. 5 Ni
10. 0 Be
22. 5 . Appl. Phys. Lett. , 2342-2344 (1993). 17. Inoue, A. Stabilization of metallic supercooled liquid and bulk amorphous alloys.
Acta Mater. , 279-306 (2000). 18. Wang, W. H., Dong, C. & Shek. C. H. Bulk metallic glasses.
Mater. Sci. Eng.: R: Rep. , 45-89 (2004). 19. Chen, M.W. A brief overview of bulk metallic glasses.
NPG Asia Mater . , 82-90 (2011). 20. Crystallography at 100.
Science , 1091-1116 (2014). 21.
Egami, T. & Billinge, S. J. L.
Underneath the Bragg Peaks: Structural Analysis of Complex Materials. (Pergamon Materials Series, v.7, 2003). Kelton, K. F. et al. First X-ray scattering studies on electrostatically levitated metallic liquids: demonstrated influence of local icosahedral order on the nucleation barrier.
Phys. Rev. Lett. , 195504 (2003). 23. Zhong, L., Wang, J., Sheng, H., Zhang, Z. & Mao, S. X. Formation of monatomic metallic glasses through ultrafast liquid quenching.
Nature , 177-180 (2014). 24.
Hwang, J. et al. Nanoscale structure and structural relaxation in Zr Cu A l5 bulk metallic glass. Phys. Rev. Lett. , 195505 (2012). 25.
Hirata, A. et al. Direct observation of local atomic order in a metallic glass.
Nat. Mater. , 28-33 (2011). 26. Pekin, T.C. et al. Direct measurement of nanostructural change during in situ deformation of a bulk metallic glass.
Nat. Commun. , 2445 (2019). 27. Tang, X. P., Geyer, U., Busch, R., Johnson, W. L. & Wu, Y. Diffusion mechanisms in metallic supercooled liquids and glasses.
Nature , 160-162 (1999). 28.
Sachdev, S. & Nelson, D. R. Order in metallic glasses and icosahedral crystals.
Phys. Rev. B , 4592-4606 (1985). 29. Tang, C., & Harrowell, P. Anomalously slow crystal growth of the glass-forming alloy CuZr.
Nat. Mater. , 507-511 (2013). 30. Cheng, Y. Q., Ma, E. & Sheng, H. W. Atomic level structure in multicomponent bulk metallic glass.
Phys. Rev. Lett. , 245501 (2009). 31.
Hu, Y. C., Li, F. X., Li, M. Z., Bai, H. Y. & Wang, W. H. Five-fold symmetry as indicator of dynamic arrest in metallic glass-forming liquids.
Nat. Commun. , 8310 (2015). 32. Scott, M. C. et al. Electron tomography at 2.4-angstrom resolution.
Nature , 444-447 (2012). 33.
Miao, J., Ercius, P. & Billinge, S. J. L. Atomic electron tomography: 3D structures without crystals.
Science , aaf2157 (2016). Chen, C.-C. et al . Three-dimensional imaging of dislocations in a nanoparticle at atomic resolution.
Nature , 74-77 (2013). 35.
Goris, B. et al . Three-Dimensional Elemental Mapping at the Atomic Scale in Bimetallic Nanocrystals.
Nano Lett. , 4236-4241 (2013). 36. Xu, R. et al . Three-dimensional coordinates of individual atoms in materials revealed by electron tomography.
Nat. Mater. , 1099-1103 (2015). 37. Goris, B. et al . Measuring lattice strain in three dimensions through electron microscopy.
Nano Lett. , 6996-7001 (2015). 38. Haberfehlner, G. et al . Formation of bimetallic clusters in superfluid helium nanodroplets analysed by atomic resolution electron tomography.
Nat. Commun. , 8779 (2015). 39. Yang, Y. et al . Deciphering chemical order/disorder and material properties at the single-atom level.
Nature , 75-79 (2017). 40.
Tian, X. et al. Correlating the three-dimensional atomic defects and electronic properties of two-dimensional transition metal dichalcogenides.
Nat. Mater. (2020). https://doi.org/10.1038/s41563-020-0636-5. 41.
Zhou, J. et al. Observing crystal nucleation in four dimensions using atomic electron tomography.
Nature , 500-503 (2019). 42.
Yao, Y. et al. Carbothermal shock synthesis of high-entropy-alloy nanoparticles.
Science , 1489-1494 (2018). 43.
Pryor, A., Ophus, C. & Miao, J. A streaming multi-GPU implementation of image simulation algorithms for scanning transmission electron microscopy.
Adv. Struct. Chem. Imag. , 15 (2017). 44. Steinhardt, P. J., Nelson, D. R. & Ronchetti, M. Bond-orientational order in liquids and glasses.
Phys. Rev. B , 784-805 (1983). 45. Liu, X. J. et al. Metallic liquids and glasses: atomic order and global packing.
Phys. Rev. Lett. , 155501 (2010). Spaepen, F. A microscopic mechanism for steady state inhomogeneous flow in metallic glasses.
Acta Metall. , 407-415 (1977). 47. Argon A. S. Plastic-deformation in metallic glasses.
Acta Metall. , 47–58 (1979). 48. Johnson, W. L. Thermodynamic and kinetic aspects of the crystal to glass transformation in metallic materials.
Prog. Mater. Sci. , 81-134 (1986). 49. Debenedetti, P. G. & Stillinger, F. H. Supercooled liquids and the glass transition.
Nature , 259-267 (2001). 50.
Zhu, C. et al. Towards three-dimensional structural determination of amorphous materials at atomic resolution.
Phys. Rev. B , 100201 (2013). Acknowledgements
This work was primarily supported by STROBE: A National Science Foundation Science & Technology Center under Grant No. DMR 1548924. This work was also supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, Division of Materials Sciences and Engineering under Award No. DE-SC0010378 and the NSF DMREF program under Award No. DMR-1437263. J.M. acknowledges partial support from an Army Research Office MURI grant on Ab-Initio Solid-State Quantum Materials: Design, Production and Characterization at the Atomic Scale. The ADF-STEM imaging with TEAM 0.5 was performed at the Molecular Foundry, which is supported by the Office of Science, Office of Basic Energy Sciences of the U.S. DOE under Contract No. DE-AC02—05CH11231. Figures and Figure legends
Figure 1 | Determining the 3D atomic structure of a multi-component metallic glass with AET . a , Experimental 3D atomic model of the metallic glass nanoparticle with a diameter of 9 nm. b , The BOO distribution of all the atoms in the nanoparticle, indicating the majority of atoms significantly deviate from the ideal crystal structures. c , The distribution of the atoms with BOO d , The RDF of the atoms with BOO < 0.5. The ratios of the 2 nd , 3 rd , 4 th and 5 th to the 1 st peak position ( R /R , R /R , R /R and R /R ) are 1.82, 2.13, 2.74 and 3.62, respectively. The inset shows the splitting of the 2 nd and 3 rd peaks with the Gaussian fit. e , The PDFs between type 1, 2 and 3 atoms, consisting of 6 pairs - type 11, 12, 13, 22, 23 and 33 atoms. The PDF for the type 33 atoms (the yellow curve) shows a unique feature with a higher 2 nd peak than the 1 st peak. Figure 2 | The short-range order of the metallic glass nanoparticle . a , Ten most abundant Voronoi polyhedra in the nanoparticle. b , Six representative Voronoi polyhedra containing type 3 central atoms, where <0,4,4,3>, <0,4,4,2>, <0,2,8,1> and <0,3,6,3> are the four highest fraction Voronoi indices, <0,4,4,4> shows a severely distorted polyhedron, and <0,0,12,0> represents an icosahedron. c , The 3-, 4-, 5- and 6-fold face distribution for all Voronoi polyhedra, where the 5-fold faces are the most abundant (45.3%). d , The coordination number (CN) distribution for type 1, 2 and 3 atoms. The average CNs for type 1, 2 and 3 atoms are 10.9, 10.9 and 10.7, respectively. Figure 3 | The connectivity and distribution of the MROs in the metallic glass nanoparticle . a-e , Representative pairs of the solute-centred clusters that are connected with each other by sharing one, two, three, four and five atoms, respectively, where the central atom of each cluster is indicated by a large red sphere. f , Statistical distribution of the number of the solute-centred cluster pairs, which share from one to five atoms. g , Histogram of the four types of MROs fcc- (in blue), hcp- (in red), bcc- (in green) and sc-like (in purple) as a function of the size (i.e. the number of solute centres). The total number of fcc-, hcp-, bcc- and sc-like MROs is 86, 84, 30 and 29, respectively. The inset shows the fraction of the four MRO solute centre atoms. h , Distribution of the four types of the MROs in the metallic glass nanoparticle. To better visualize the networks, only those with ten solute centre atoms or more are shown. Figure 4 | Quantitative characterization of the MROs.
The length ( a ) and volume ( b ) distribution of the four types of the MROs in the metallic glass nanoparticle, where the length was measured along the longest direction of each MRO. c, Partial
RDFs of the fcc-, hcp-, bcc- and sc-like solute centres in the metallic glass nanoparticle, where the maximum peak positions are located at 4.84, 4.76 4.84 and 3.97 Å, respectively. Compared with the other three partial RDFs, the partial RDF of the sc-like solute centres (the purple curve) shows two peaks with the ratio of the 2 nd to the 1 st peak position about √2 . a , Distribution of sharing one, two, three, four and five atoms between neighbouring solute-centred clusters for the four types of the MROs. Figure 5 | 3D atomic packing of four representative MROs.
Representative fcc- ( a ), hcp- ( c ), bcc- ( e ) and sc-like ( g ) MROs, consisting of 13, 10, 11 and 25 solute centres (large red spheres), respectively, where the solute-centred clusters exhibit only the translational but no orientational order. The solute centres are oriented along the fcc ( b ), hcp ( d ), bcc ( f ) and sc ( h ) zone axes, showing the 3D shapes of the MROs are anisotropic and the networks significantly deviate from the ideal crystal structures. METHODS
Sample preparation.
The multi-component metallic nanoparticle samples were synthesized using the thermal shock procedures published elsewhere . Individual metal salts (chlorides or their hydrate forms) were dissolved in ethanol at a concentration of 0.05 mol/L. After complete dissolving with hydrochloric 0 acid, the individual salt precursor solutions with different cations were mixed and sonicated for 30 minutes. The homogenously mixed precursor solution was loaded onto the carbon substrates (reduced graphene oxide) and heated to a temperature as high as 1,763 K for 55 milliseconds (Supplementary Fig. 1). The sample was suspended on a trench and connected with copper electrodes by silver paste for both heating and effective cooling as a giant heat sink. The thermal shock synthesis was triggered by electric Joule heating in an argon-filled glovebox using a Keithley 2425 SourceMeter where the high temperature and duration can be effectively controlled by tuning the input power and duration. The temperature of this process was monitored by a time-resolved spectrometer. The max cooling rate was estimated to be 69,000 K/s at the cooling stage (Supplementary Fig. 1). The resulting nanoparticles on reduced graphene oxide were dispersed in ethanol with sonication. After deposited on to 5-nm-thick silicon nitride membranes, the nanoparticles were baked at 100 °C for 12 hours in vacuum to eliminate any hydrocarbon contamination. Data acquisition.
A set of tomographic tilt series were acquired from six nanoparticles using the TEAM 0.5 microscope with the TEAM stage . Images were collected at 200 kV in ADF-STEM mode (Supplementary Table 1). To minimize sample drift, four sequential images per tilt angle were measured with a dwell time of 3 μs. To monitor any potential damage induced by the electron beam, we took 0° projection images before, during and after the acquisition of each tilt series and ensured that no noticeable structural change was observed for the six particles. The total electron dose of each tilt series was estimated to be between 7 e - /Å and 9.5 e - /Å (Supplementary Table 1). Image pre-processing and denoising.
For each experimental tilt series, we performed the following procedure for image post-processing and denoising. i) Image registration. At each tilt angle, we used the first image as a reference and calculated normalized cross-correlation between the reference and the other three images using a step size of 0.1 pixel . These four images were aligned and summed to form an experimental image at that tilt angle. ii) Scan distortion correction . Two steps were used to correct the scan distortion for the experimental images. First, a set of low-magnification images were taken from nanoparticles and their positions were fitted with a Gaussian. Based on the geometric relation of the nanoparticles at different angles, the scan coil directions were calibrated to be perpendicular and equal in strength. Second, six high-magnification images were taken from a multi-component metallic nanoparticle and scan distortion parameters were estimated by minimizing the mean squared error of the common line of the six images. These scan distortion parameters were applied to the experimental images. iii) Image denoising. The experimental images contain mixed Poisson and Gaussian noise and were denoised by the block-matching and 3D filtering (BM3D) algorithm , which has been demonstrated to be effective in reducing noise in AET . The BM3D denoising parameters were optimized by the following three steps. First, Poisson and Gaussian noise level were estimated from the experimental tilt series. Second, several images were simulated based on a model nanoparticle, which has a similar size and elemental distribution as those of an experimental image. The same level of Poisson and Gaussian noise was added to the simulated images. Third, these noisy images were denoised by BM3D with different parameters. The denoising parameters corresponding to the largest cross-correlation coefficient between the denoised and the original images were chosen and applied to denoise the experimental images. 1 iv) Background subtraction and alignment. After denoising, a 2D mask was defined from each experimental image, which is slightly larger than the size of the nanoparticle. The background inside the mask was estimated by the discrete Laplacian in Matlab. After background subtraction, each tilt series of the experimental images were scaled and aligned by the centre of mass and common line methods . The REal Space Iterative REconstruction (RESIRE) algorithm.
After post-processing and denoising, the experimental images were reconstructed by the RESIRE algorithm. The algorithm iteratively minimizes an error function defined by, 𝜀 𝜃 (𝑂) = 12 ∑|Π 𝜃 (𝑂){𝑥, 𝑦} − 𝑏 𝜃 {𝑥, 𝑦}| (1) where 𝜀 𝜃 (𝑂) is an error function of a 3D object ( 𝑂 ) at tilt angle 𝜃 , Π 𝜃 (𝑂) projects 𝑂 to generate a 2D image at angle 𝜃 , 𝑏 𝜃 is the experimental image at angle 𝜃 , and {𝑥, 𝑦} is the coordinates. The minimization is solved via the gradient descent, ∇𝜀 𝜃 (𝑂){𝑢, 𝑣, 𝑤} = Π 𝜃 (𝑂){𝑥, 𝑦} − 𝑏 𝜃 {𝑥, 𝑦} 𝑤ℎ𝑒𝑟𝑒 [𝑢𝑣𝑤] = 𝑅 𝜃 [𝑥𝑦𝑧] 𝑓𝑜𝑟 𝑠𝑜𝑚𝑒 𝑧 (2) where ∇ represents the gradient and 𝑅 𝜃 is the rotation matrix at tilt angle θ , which transforms coordinates {𝑥, 𝑦, 𝑧} to {𝑢, 𝑣, 𝑤} . Supplementary Fig. 5 shows the schematic diagram of the iterative process, where the j th iteration consists of the following four steps. i) A set of images are calculated from the 3D object of the j th iteration using a Fourier method. The 3D object is first padded with zeros by properly choosing an oversampling ratio . Applying the fast Fourier transform to the zero-padded object generates a 3D array in reciprocal space, from which a series of 2D Fourier slices are obtained at different tilt angles. These 2D Fourier slices are inverted to a set of images via the inverse Fourier transform. ii) The error function defined in equation 1 is calculated between the computed and experimental images. iii) The gradient of the error function is computed for every voxel using equation 2. iv) The 3D object of the (j+1) th iteration is updated by, 𝑂 𝑗+1 = 𝑂 𝑗 − ∆𝑛𝑁 ∑ ∇𝜀 𝜃 (𝑂 𝑗 ) 𝜃 (3) where ∆ is the step size ( ∆ = 2 was chosen for the reconstruction of our experimental data), n is the number of images and N is the dimension of each image ( N × N ). 𝑂 𝑗+1 {𝑢, 𝑣, 𝑤} is used as an input for the (j+1) th iteration. The convergence of the algorithm is monitored by the R-factor, 𝑅 = 1𝑛 ∑ ∑ |Π 𝜃 (𝑂){𝑥, 𝑦} − 𝑏 𝜃 {𝑥, 𝑦}| 𝑥,𝑦 ∑ |𝑏 𝜃 {𝑥, 𝑦}| 𝑥,𝑦𝜃 . (4) Usually, after several hundreds of iterations, the algorithm converges to a high-quality 3D reconstruction from a limited number of images. Both our numerical simulation and experimental results have indicated that RESIRE outperforms other iterative tomographic algorithms such as generalized Fourier iterative reconstruction and simultaneous iterative reconstruction technique . By avoiding iterating between real 2 and reciprocal space, RESIRE can be applied to general sample geometry such as thin films and extended objects. The details of the RESIRE algorithm will be reported in a follow-up paper. For each aligned and scaled experimental tilt series, we first ran RESIRE for 200 iterations. From the initial 3D reconstruction, angular refinement and spatial alignment were applied iteratively until there was no further improvement. Next, the background of each experimental image was re-evaluated and re-subtracted. Using the refined experimental images and tilt angles, we ran another 200 iterations of RESIRE to obtain the final 3D reconstruction of each experimental tilt series (Supplementary Table 1). Determination of 3D atomic coordinates and species.
From each final
3D reconstruction, the atomic coordinates and chemical species were identified using the following procedure . i) Each 3D reconstruction was upsampled by a factor of 3 using the spline interpolation, from which all the local maxima were identified. Starting from the highest intensity peak, polynomial fitting was performed on a 0.8 Å × 0.8 Å × 0.8 Å (7 × 7 × 7 voxel) volume around each local maximum to locate the peak position. If the distance between the fitted peak position and existing potential atom positions is larger than or equal to 2 Å, it was listed as a potential atom. After repeating this step for all the local maxima, a list of potential atom positions was obtained. ii) From the list of the potential atoms, manual checking was performed to correct any unidentified or misidentified atoms due to the broadened local intensity peaks from multiple atoms. iii) A K-mean clustering method was used to classify three types of atoms and non-atoms (Co and Ni as type 1, Ru, Rh, Pd and Ag as type 2, and Ir and Pt as type 3) based on the integrated intensity of a 0.8 Å × 0.8 Å × 0.8 Å volume around each potential atom position. An initial atomic model with 3D atomic coordinates and chemical species was determined from each 3D reconstruction. iv) Due to the missing wedge problem and noise in the experimental images, there is local intensity variation in each 3D reconstruction. A local reclassification was iteratively performed to refine the type 1, 2 and 3 atoms. Each atom was defined as the centre of a 10-Å-radius sphere. The average intensity distribution of type 1, 2 and 3 atoms was computed within the sphere. The L norm of the intensity distribution between the centre atom and the average type 1, 2 and 3 atom was calculated. The centre atom was assigned to the type with the smallest L norm. The procedure was iteratively repeated until there are no further changes. Refinement of 3D atomic coordinates and species.
The 3D atomic coordinates were refined by minimizing the error between the calculated and measured images using the gradient descent . Each atom was first fit with a 3D Gaussian function with a height 𝐻 and a width 𝐵 ′ , where 𝐻 and 𝐵 ′ were considered the same for the same type of atoms. A 3D atomic model was obtained by, 𝑂{𝑥, 𝑦, 𝑧} = ∑ 𝐻 𝑖 exp [− |𝑥 − 𝑥 𝑖 | + |𝑦 − 𝑦 𝑖 | + |𝑧 − 𝑧 𝑖 | 𝐵 𝑖′ ] |𝑥−𝑥 𝑖 |,|𝑦−𝑦 𝑖 |,|𝑧−𝑧 𝑖 |≤𝜌 (5) where 𝑥 𝑖 , 𝑦 𝑖 , 𝑧 𝑖 , 𝐻 𝑖 and 𝐵 𝑖′ are the coordinates, height and standard deviation of the i th atom, respectively, and 𝜌 is a cut-o ff size of the 3D Gaussian function. From the 3D atomic model, a set of projection images were computed at different tilt angle θ by, 3 Π 𝜃 (𝑂){𝑢, 𝑣} = ∑ 𝐻 𝑖 exp [− |𝑢 − 𝑢 𝑖 | + |𝑣 − 𝑣 𝑖 | 𝐵 𝑖′ ] |𝑢−𝑢 𝑖 |,|𝑣−𝑣 𝑖 |≤𝜌 ∑ 𝑒𝑥𝑝 [− |𝑤 − 𝑤 𝑖 | 𝐵 𝑖′ ] |𝑤−𝑤 𝑖 |≤𝜌 where [𝑢 𝑖 𝑣 𝑖 𝑤 𝑖 ] = 𝑅 𝜃 [𝑥 𝑖 𝑦 𝑖 𝑧 𝑖 ] . (6) Substituting equation (6) into (1), an error function was calculated, from which the gradient descent method was used to search for the optimal atomic position at the (j+1) th iteration, {𝑥 𝑖 , 𝑦 𝑖 , 𝑧 𝑖 } 𝑗+1 = {𝑥 𝑖 , 𝑦 𝑖 , 𝑧 𝑖 } 𝑗 − ∆ ∑[Π 𝜃 (𝑂){𝑢, 𝑣} − 𝑏 𝜃 {𝑢, 𝑣}]∇ 𝑖 [Π 𝜃 (𝑂){𝑢, 𝑣}] 𝜃 (7) Where ∇ 𝑖 is the spatial gradient operator with respect to the atomic position (𝑥 𝑖 , 𝑦 𝑖 , 𝑧 𝑖 ) . The iterative refinement process was terminated when the L norm error could not be further reduced. The local bond orientational order (BOO) parameter . The local BOO parameter was calculated from the 3D atomic model of each nanoparticle using a method described elsewhere . The Q and Q order parameters were computed up to the second shell with a shell radius set by the first valley in the RDF curve of the 3D atomic model. The local BOO parameter was normalized between 0 and 1, where 0 means Q = Q = 0 and 1 represents a perfect fcc crystal structure. Fig. 1b and Supplementary Fig. 3g-l show the distribution of the local BOO parameter of all the atoms in particles 1-6. Based on the BOO distribution of a Cu Zr metallic glass structure obtained from molecular dynamics simulations (Supplementary Fig. 3m), we chose BOO = 0.5 as a cut-off to separate crystal nuclei from amorphous structure. The radial distribution function (RDF) and pair distribution function (PDF).
The RDF was calculated for the 3D atomic model of each nanoparticle using the following procedure. i) The distance of all atom pairs in each 3D atomic model was computed and binned into a histogram. ii) The number of atom pairs in each bin was normalized with respect to the volume of the spherical shell corresponding to each bin. iii) The histogram was scaled so that the RDF approaches one for large separations. After calculating the RDF for each nanoparticle, the first valley of the RDF was used to determine the local BOO parameter (Fig. 1b and Supplementary Fig. 3g-l). By choosing the atoms in the metallic glass nanoparticle (particle 1) with BOO < 0.5, we applied the above procedure to plot the RDF (Fig. 1d). For type 1, 2 and 3 atoms, we identified six sets of atoms pairs (type 11, 12, 13, 22, 23 and 33) in the nanoparticle. For each set of atom pairs, we used the above procedure to calculate the PDF shown in Fig. 1e.
Voronoi tessellation and the coordination number (CN).
The analysis of Voronoi tessellation was performed by following the procedure published elsewhere , where the surface atoms of the nanoparticle were excluded. To reduce the effect of the experimental and reconstruction error on Voronoi tessellation, those surfaces with areas less than 1% of the total surface area of each Voronoi polyhedron were removed . From the Voronoi tessellation, each polyhedron is designated by a Voronoi index 〈𝑛 , 𝑛 , 𝑛 , 𝑛 , ⋯ 〉 with 𝑛 𝑖 denoting the number of i -edge faces and the CN was calculated by ∑ 𝑛 𝑖𝑖 . Measurement of the bond length and quantification of the chemical SRO.
By fitting a Gaussian to the 1 st peak of the PDFs (Fig. 1e), we determined the type 11, 22, 33, 12, 13 and 23 bond length to be 2.60, 2.67, 2.66, 2.65, 2.66 and 2.60 Å, respectively. The type 11, 22 and 33 bond length are consistent with the metallic radii of type 1, 2 and 3 atoms . But the type 23 bond is 2.4% shorter than the average length of 4 the type 22 and 33 bond, indicating a stronger interaction between type 2 and 3 atoms than other atom pairs. To quantify the chemical SRO, we computed the Warren–Cowley parameter ( 𝛼 𝑙𝑚 ), 𝛼 𝑙𝑚 = 1 − 𝑍 𝑙𝑚 𝜒 𝑚 𝑍 𝑙 (8) where 𝑙, 𝑚 = 1, 2 or 3 , 𝑍 𝑙𝑚 is the partial CN of type 𝑚 atoms around type 𝑙 atoms, 𝜒 𝑚 is the fraction of type 𝑚 atoms, and 𝑍 𝑙 is the total CN around type 𝑙 atoms. After excluding the surface atoms, we estimated 𝜒 , 𝜒 and 𝜒 to be 39.16%, 42.76% and 18.08%, respectively. Using the partial CNs (Supplementary Fig. 7a), we calculated 𝛼 = -0.12, 𝛼 = 0.06, 𝛼 = 0.12, 𝛼 = -0.01, 𝛼 = 0.02, 𝛼 = -0.03, 𝛼 = 0.04, 𝛼 = -0.08, and 𝛼 = 0.11, indicating that both the type 11 and 23 bond are favoured in the metallic glass sample. The favouring of the type 23 bond is also consistent with the shortening of the type 23 bond (Fig. 1e). Determination of the MROs.
The MROs were identified using the following procedure. First, 2769 solute centres were chosen from a total of 3472 type 3 atoms based on a criterion that the solute centres must be distributed within the second coordination shell, which is between the first (3.78 Å) and the second minimum (6.09 Å) of the RDF curve (Fig. 1d). Second, a breadth-first-search algorithm was used to search for the possible fcc-, hcp-, bcc- or sc-like MROs from the 2769 solute centres. Each possible MRO must satisfy the following three criteria: i) there are at least five or more solute centres; ii) every solute centre must fall within a 0.75 Å radius from its fitted lattice vector; and iii) every solute centre must be within a medium range order distance (between 3.74 Å and 6.05 Å) to at least one other solute centre. Third, after identifying all the possible MROs, they were sorted by number of solute centres to generate a possible network queue. If there was a tie, the average error of fitting the solute centres into the lattice vectors was used to break the tie. Fourth, starting from the largest possible network, it was classified as an MRO if none of the solute centres in the network was already occupied by another MRO. If some solute centres were already occupied, then those solute centres were removed from the possible network, refitted while also accounting for connectivity, and then added back into the queue of possible networks. This process was repeated until the queue was depleted and every solute centre belonged to one and only one MRO. All the classified network with five or more solute centres were counted as the final MROs (Figs. 3g, h and 5). Finally, to corroborate our analysis, we repeated the above steps with a 1 Å and 0.5 Å radius cut-off and the corresponding final MROs are shown in Supplementary Figs. 8 and 9, respectively. An attempt was also made in searching for icosahedral-like MROs. The breadth-first-search algorithm was used to find the possible MROs that fall within a 0.75 Å radius from the 12 vertices of an icosahedron. Because the icosahedron cannot be periodically packed in three dimensions, only the nearest neighbour vertices were searched, making the largest possible MRO have 13 solute centres (1 central solute centre plus 12 nearest neighbours). After performing the search, the resulting possible MROs have a mean value of 3.9, meaning on average each solute centre is connected to only 3 others when constrained to an icosahedron within the second coordination shell. Furthermore, although the largest possible MROs have 7 solute centres, none of these solute centres form five-fold symmetry. We also repeated this analysis with a 1 Å radius cut-off. The mean value of solute centres becomes 4.5, the largest 5 possible MROs have 8 solute centres, and there are 19 five-fold symmetries. But these numbers are substantially less than those of other MROs (Supplementary Fig. 8). 51.
Xu, Z., Sun, H., Zhao, X. & Gao, C. Ultrastrong fibers assembled from giant graphene oxide sheets.
Adv. Mater. , 188-193 (2013). 52. Ercius, P., Boese, M., Duden, T. & Dahmen, U. Operation of TEAM I in a user environment at NCEM.
Microsc. Microanal. , 676-683 (2012). 53. Lewis, J. P. Fast normalized cross-correlation.
Vis. Interface , 120-123 (1995). 54. Dabov, K., Foi, A., Katkovnik, V. & Egiazarian, K. Image denoising by sparse 3-D transform-domain collaborative filtering.
IEEE Trans. Image Process . , 2080-2095 (2007). 55. Miao, J., Sayre, D. & Chapman, H. N. Phase retrieval from the magnitude of the Fourier transform of non-periodic objects,
J. Opt. Soc. Am. A , 1662-1669 (1998). 56. Pryor, A. et al. GENFIRE: a generalized Fourier iterative reconstruction algorithm for high-resolution 3D imaging.
Sci. Rep. , 10409 (2017). 57. Gilbert, P. Iterative methods for the three-dimensional reconstruction of an object from projections.
J. Theor. Biol. , 105-117 (1972). 58. Rogers, S. S., Waigh, T. A., Zhao, X. & Lu, J. R. Precise particle tracking against a complicated background: polynomial fitting with Gaussian weight.
Phys. Biol. , 220–227 (2007). 59. Lloyd, S. Least squares quantization in PCM.
IEEE Trans. Infor. Theory , 129-137 (1982). 60. Lechner, W. & Dellago, C. Accurate determination of crystal structures based on averaged local bond order parameters.
J. Chem. Phys. , 114707 (2008). 61.
Yu, H. B. & Samwer, K. Atomic mechanism of internal friction in a model metallic glass.
Phys. Rev. B , 144201 (2014). 62. Greenwood, N. N. & Earnshaw, Alan.
Chemistry of the Elements
Warren, B. E.
X-ray diffraction (New York: Dover Publications Inc. 1990). 64.
Cowley, J. M. X ‐ Ray Measurement of order in single crystals of Cu Au.
J. Appl. Phys. , 24-29 (1950). 65. Lee, C. Y. An algorithm for path connections and its applications.
IEEE Trans. Electron. Comput.
EC-10 , 346-365 (1961). 66.
Larsen, P. M., Schmidt, S. & Schiøtz, J. Robust structural identification via polyhedral template matching.
Model. Simul. Mater. Sci. Eng. , 055007 (2016). 6 Supplementary Table and Figures
Supplementary Table 1 | AET data collection, processing, reconstruction, refinement and statistics.
Particle 1 Particle 2 Particle 3 Particle 4 Particle 5 Particle 6
Data collection and processing
Voltage (kV) 200 200 200 200 200 200 Convergence semi-angle (mrad) 25 25 25 25 25 25 Probe size (Å) 0.8 0.8 0.8 0.8 0.8 0.8 Detector inner angle (mrad) 38 38 38 38 38 38 Detector outer angle (mrad) 190 190 190 190 190 190 Depth of focus (nm) 8 8 8 8 8 8 Pixel size (Å) 0.347 0.347 0.347 0.347 0.347 0.347 +69.5 +63.4 +63.4 +69.4 +66.4 +72.0 Total electron dose (10 e/Å ) 9.5 7 7.6 7.6 7.6 7.7 Reconstruction
Algorithm RESIRE RESIRE RESIRE RESIRE RESIRE RESIRE Oversampling ratio 4 4 4 4 4 4 Number of iterations 200 200 200 200 200 200
Refinement R (%) a b ) Type 1 atoms 36.4 65.2 48.8 39.6 60.0 40.4 Type 2 atoms 38.2 42.4 38.4 41.6 45.6 42.4 Type 3 atoms 35.5 32.8 34.0 35.6 40.0 36.0 Statistics a The R -factor is defined as equation (5) in ref. 39. b The R-factor is defined in equation 4 in Methods. Supplementary Fig. 1 | Temperature change during the rapid thermal shock heating process . The average and maximum temperatures (black and blue dots, respectively) were measured by a high-speed Phantom Miro M110 camera (Vision Research). The cooling rate for the average and maximum temperature curves was determined to be 51000 K/s (i.e. the slope of the red line) and 69000 K/s (i.e. the slope of the green line), respectively. Supplementary Fig. 2 | Elemental mapping of multi-component metallic nanoparticles. a,
Low-resolution
ADF-STEM image of the nanoparticles. Energy-dispersive X-ray spectroscopy images show the distribution of Ni ( b ), Co ( c ), Ru ( d ), Rh ( e ), Pd ( f ), Ag ( g ), Ir ( h ) and Pt ( i ). k , The spectrum of all the elements distributed in the images ( b-i ), where cps stands for counts per second. Scale bar, 20 nm. Supplementary Fig. 3 | Analysis of six multi-component metallic nanoparticles. a-f , Representative ADF-STEM images of particles 1-6, respectively. Scale bar, 2 nm. g - l , The BOO distribution of the atoms in particles 1-6, showing the structural transition of the nanoparticles from disorder to order, where the red lines correspond to BOO = 0.5. According to the BOO, particles 1 and 2 are more disordered than other four particles. m , The BOO distribution of a Cu Zr metallic glass obtained from molecular dynamics simulations as a reference, from which we chose BOO = 0.5 as a cut-off to separate crystal nuclei from amorphous structure. Supplementary Fig. 4 | Tomographic tilt series from the most disordered nanoparticle (particle 1) among the six.
51 ADF-STEM images of the nanoparticle with a tilt range from −69.4° to +69.4° with each tilt angle shown at top right of each panel. The images were denoised by BM3D (Methods). The total electron dose of the tilt series was estimated to be 9.5 × 10 e/Å . Scale bar, 2 nm. Supplementary Fig. 5 | Schematic diagram of the RESIRE algorithm . From the experimental projections, RESIRE minimizes the L2-norm error metric using gradient descent. The j th iteration of the algorithm consists of the following steps. i) RESIRE computes a set of images from the 3D object of the (j-1) th iteration (middle left). As the algorithm is not sensitive to the initial input, a random or empty 3D object can be used as the input of the 1 st iteration. ii) The difference between the computed and corresponding experimental images is calculated (middle right), from which an error metric is defined to monitor the convergence of the algorithm. iii) The difference is back projected to real space, yielding the gradient of the 3D reconstruction (bottom right). iv) The 3D reconstruction of the j th iteration is updated by combining the gradient with the reconstruction of the (j-1) th iteration (bottom left), where constraints such as positivity and support can be enforced. Usually, after several hundreds of iterations, RESIRE converges to a 3D reconstruction from a limited number of images. Supplementary Fig. 6 | Validation of the experimental 3D atomic model using multislice simulations.
Comparison between a representative experimental ( a ) and a multislice ADF-STEM image ( b ). To account for the source size and incoherent effects, the multislice image was convolved with a Gaussian function. c , Histogram of the root mean square deviation of the atomic positions between the experimental atomic model and that obtained from 51 multislice images. 96.5% of the atoms are identical between the two models with a root-mean-square deviation of 21 pm. Scale bar, 2 nm. Supplementary Figure 7 | The partial CNs, the distribution of the solute centres, and the Voronoi polyhedra of the solute-centred clusters. a , The partial CNs around type 1, 2 and 3 atoms. b , 3D distribution of the 2769 solute centres (red dots), which are between the first (3.78 Å) and the second minimum (6.09 Å) of the RDF curve (Fig. 1d). The yellow shades show the locations of the crystal nuclei in the nanoparticle. c , Radial distribution of the solute centre density, obtained by dividing the number of the solute centres in each 2-Å-thick shell by the volume of the shell. One crystal nucleus near the centre and 18 nuclei on or near the surface of the nanoparticle affect the distribution of the solute centres in these regions. d , Ten most abundant Voronoi polyhedra of the solute-centred clusters. Supplementary Fig. 8 | Identification of MROs with a 1 Å radius cut-off. a , Histogram of the four types of MROs fcc- (in blue), hcp- (in red), bcc- (in green) and sc-like (in purple) as a function of the size (i.e. the number of solute centres). b , The fraction of the four MRO solute centre atoms. Representative fcc- ( c ), hcp- ( e ), bcc- ( g ) and sc-like ( i ) MROs, consisting of 30, 27, 17 and 30 solute centres (large red spheres), respectively. The solute centres are orientated along the fcc ( d ), hcp ( f ), bcc ( h ) and sc ( j ) zone axes. Supplementary Fig. 9 | Identification of MROs with a 0.5 Å radius cut-off. a , Histogram of the four types of MROs fcc- (in blue), hcp- (in red), bcc- (in green) and sc-like (in purple) as a function of the size. b , The fraction of the four MRO solute centre atoms. Representative fcc- ( c ), hcp- ( e ), bcc- ( g ) and sc-like ( i ) MROs, consisting of 12, 9, 8 and 17 solute centres (large red spheres), respectively. The solute centres are orientated along the fcc ( d ), hcp ( f ), bcc ( h ) and sc ( jj