Abstract
We report a synergistic approach of micro-Raman spectroscopic mapping and deep data analysis to study the distribution of crystallographic phases and ferroelastic domains in a defected Al-doped VO2 microcrystal. Bayesian linear unmixing revealed an uneven distribution of the T phase, which is stabilized by the surface defects and uneven local doping that went undetectable by other classical analysis techniques such as PCA and SIMPLISMA. This work demonstrates the impact of information recovery via statistical analysis and full mapping in spectroscopic studies of vanadium dioxide systems, which is commonly substituted by averaging or single point-probing approaches, both of which suffer from information misinterpretation due to low resolving power.
Similar content being viewed by others
Introduction
Vanadium dioxide is a classic example of a strongly-correlated electron system exhibiting a near room temperature uniquely-sharp metal-insulator transition (MIT) that drastically changes the electronic and optical properties of this material. The MIT defines the wide range of existing and potential applications of VO2 in microbolometers1,2, smart windows3, thermochromic devices4, nanoactuators5,6,7, ultra-fast optical switches8, sensors9,10, field-effect transistors11,12,13,14,15 etc. Control over MIT is gained via temperature, elastic stress, electric field and doping. The latter method can be used for tuning of the MIT temperature and for stabilization of a range of crystallographic phases: from the monoclinic M1 and tetragonal R, stable in pure VO2 at ambient pressure, to metastable monoclinic M2 and triclinic T. The Giznburg-Landau phase transition theory describes the lower-symmetry VO2 phases as different structures which can be realized as a result of rutile lattice instabilities at the two non-equivalent R-points of the Brillouin zone16. This phase interdependence substantiates investigation of the low-symmetry phases that may shed light on the controversial question of the origin of MIT in vanadium dioxide17,18,19.
Confocal micro-Raman spectroscopy is a versatile and non-invasive tool that has been extensively used for in situ identification of local specific crystallographic phases of VO2 and their transformations as a function of temperature or mechanical stress. However, its use in VO2 surface mapping has been limited. Only a few groups have utilized micro-Raman mapping of VO2 crystal/film surfaces16, or samples containing VO2/V2O5 nanowires20 to fully unravel the spatial distribution of specific phases. Commonly, simpler approaches such as spectral averaging over a large area or point-probing the sample21,22 (e.g. measuring spectra at selected locations along a VO2 nanobeam length) have been used. Needless to say, micro-Raman spectroscopy offers much richer spectral content that can be used for solving the important question of phase separation (see ref. 23 as an example). However, in many complex multi-phase systems, local conditions may stabilize phases beyond those that exist and have been characterized in the bulk. Hence there is the need for the blind signal unmixing methods that can help to determine the spatial distributions and to deconvolute the spectra of unknown phases. Here we employ a combination of chemometric blind unmixing tools and micro-Raman spectroscopy to investigate the partition of phases in a VO2 microbelt under uneven distribution of stress and dopant concentration.
Results and Discussion
Al-doped VO2 nanostructures were grown on Si/SiO2 substrates according to the protocol described earlier24. A low dopant concentration was used to bring the system into the broad domain of M1/T stability in the temperature-composition phase diagram (see Fig. 1 in ref. 24), where an interesting interplay of these two phases can be observed at room temperature. The growth process involves dissolution of alumina in molten V2O5 microdroplets condensed on the Si/SiO2 substrate25, with the subsequent crystallization of Al-doped VO2 nanostructures from them. The grown nanostructures (nanowires, nanorods, nano and microbelts) are typically single crystals of highest quality, as evidenced by x-ray studies and the sharpness of MIT24. However, occasionally, the local growth conditions (precursor influx, Al concentration, temperature gradient, etc.) favor precipitation of secondary nanocrystal layer on the surface of the previously formed VO2 structures. These newly formed crystals have a rough surface and present a scientific opportunity to model non-ideal VO2 crystals with uneven distribution of stress and dopant concentration. It also has a practical importance in understanding of role of the crystal defects and strain on performance of VO2-based devices.
An optical image of such a microbelt embedded into the growth substrate is shown in Fig. 1a. It is hundreds of microns long, 43 μm wide and ca. 3.2 μm thick, making it a convenient object for micro-Raman mapping. After the growth, the microbeam remains under a substrate-induced axial tensile stress. To minimize the strain energy, a system of ferroelastic domains is usually formed which can be easily accessed with polarized light microscopy (Fig. 1b). The surface of the Al-doped VO2 microbelt is covered with precipitated crystals that form dendritic efflorescence-shaped patterns and a large bowed defect running along the middle of the top crystal facet (blue arrow, Fig. 1a). The defect does not seem to correlate with the ferroelastic domain structure appearing close to the straight domain wall. A closer examination of the deposit with Atomic Force Microscopy (AFM) shows that it is ca. 250 nm tall, reaching 600 nm to 800 nm in the bowed defect (Fig. 2a). The defect itself is comprised of conglomerates of clumped nanocrystals 100 nm to 400 nm in diameter (Fig. 2a, inset). In order to clarify the chemical composition of the deposit, we performed Energy Dispersed X-ray (EDX) analysis of the micro-belt, which revealed a high Al and Si content in the precipitate. The chemical maps of V, Al and Si are displayed in Fig. 2(c–e). These maps illustrate that while vanadium is only found in the microbelt, aluminum is present both in the microbelt and in the substrate, concentrating in the deposit (especially in the bowed defect); in addition, surprisingly, small quantities of silicon were detected in the bowed defect as well. Taking into account that previous study24 of high-quality Al-doped VO2 nanobelts excluded the possibility of Si doping of VO2, we attribute the Si signal exclusively to the deposit. The crystalline form of the deposit is also very different from that of alumina that re-crystalizes from molten vanadia24, which makes it unlikely to be comprised of Al2O3. Furthermore, the deposit is clearly more electrically insulating than the underlying VO2 microbelt, as evidenced by the surface charging during SEM imaging (Fig. 2b). Thus, we conclude that the precipitated nanocrystals might be comprised of heavily-doped VO2, aluminum vanadates/vanadites (AlVO4, AlVO3, etc.) with an admixture of alumosilicates. It is worth noting that in the regions underneath the deposit the microbelt could be more heavily doped with Al than on the average and that the deposit can exert unevenly-distributed local stress on the microbelt surface (see Figure S3 and discussion in Supplemental Information). Both of these factors are known to influence the VO2 crystallographic phase stability at a given temperature. Therefore, spatial mapping of the phase distribution would be particularly helpful in this system.
Micro-Raman measurements were performed in the region of interest of the microbelt, (Fig. 1b, red square) containing the bowed defect using a joint AFM/spectroscopic platform. The diode-pumped solid-state laser (λ = 532 nm) power was minimized to about 20 μW to avoid local overheating and related MIT/oxidation. The laser polarization plane (linear polarization) was set perpendicular to the microbelt length. A 60 μm × 60 μm region was mapped out at a spatial resolution of 100 × 100 pixels, producing a 100 × 100 × 1650 hyperspectral datacube. Visualization of n-dimensional datasets is possible by employing statistical tools, such as principal component analysis (PCA), independent component analysis, k-means clustering, etc.26 However, these and other similar methods yield results that, although statistically correct, are hard to interpret physically. Figure S1 of Supplemental Materials shows PCA decomposition of the Raman dataset into the loading maps and corresponding eigenvectors, revealing the difference between the Raman signals of the oppositely-oriented ferroelastic domains and existence of a separate region (Figure S1, 4) with a unique Raman signature, absent in the optical, AFM or SEM micrographs. However, the spectrum corresponding to this region cannot be easily identified, as PCA eigenvectors are not physically constrained and cannot be directly compared to the Raman spectra of pure phases. Another common chemometric tool used for deconvoluting Raman datacubes is the spectral unmixing via the SIMPLISMA algorithm27,28. SIMPLISMA assumes existence of a chemical variable (e.g. spectral intensity at a certain wavenumber) in the dataset that experiences a major influence from only one of the components (materials) in the mixture. Thus, by identifying such a pure variable, the spectral mixture can be resolved into individual spectra of components (for more details, see SI). As seen in Figure S2, SIMPLISMA easily unmixes the Raman datacube into the VO2 and silicon substrate signals (2 components), but further separation of ferroelastic domain spectra (3 components) and the third phase spectra (4 components) fails. Therefore, here we employed the recently-developed Bayesian Linear Unmixing (BLU) algorithm29,30 to fully unmix the intricate spectroscopic behavior of the microbelt.
The BLU method considers data Y as a linear combination of position-independent endmembers, M, with respective relative abundances, A, corrupted by additive Gaussian noise N: Y = MA + N. This method incorporates several built-in constraints that allow physical interpretation of results: the non-negativity (Mi ≥ 0, Ai ≥ 0), full additivity and sum-to-one (∑Ai = 1) constraints for both the endmembers and the abundance coefficients. To solve the blind unmixing problem, the BLU algorithm estimates the initial projection of endmembers in a dimensionality reduced subspace (PCA) via N-FINDR – a geometrical method that searches for a simplex of maximum volume that can be inscribed within the hyperspectral data set using a simple non-linear inversion. The endmember abundance priors as well as noise variance priors are then chosen by a multivariate Gaussian distribution, where the posterior distribution is calculated based on endmember independence using Markov Chain Monte Carlo that generates asymptotically distributed samples probed by Gibbs sampling strategy. Due to non-negativity of the resulting endmembers M and normalization of abundances, the spectrum at each location is decomposed into a linear combination of spectra of individual components in corresponding proportions. The number of spectral components must be provided by the researcher and can be estimated using PCA or by under- and oversampling. A detailed description and testing of BLU against PCA, k-means and other statistical methods can be found elsewhere31,32. The PCA maps (Figure S1) suggest that there are 4 to 5 components, corresponding to the Si substrate, 2 types of ferroelastic domains and 1 to 2 other phase regions. Bayesian unmixing of the datacube into 4 components is shown in Fig. 3. As can be seen, a proper separation of the 2 ferroelastic domains (Fig. 3: 1 to 2), substrate (Fig. 3: 3) and a third phase (Fig. 3: 4) is achieved. The corresponding endmembers (unmixed spectra of pure components) can be directly compared to the spectra of individual materials and are easily identifiable. Endmembers 1 and 2 correspond to the Raman spectra of the M1 phase ferroelastic domains, as identified by their shapes and spatial distribution. They mostly differ by the relative intensity of the V-motion modes33 at 195 cm−1 and 224 cm−1, the former decreasing in intensity, the latter staying constant. Both of these modes have Ag symmetry and a non-circular dependence of intensity on the polarization angle20. Endmember 3 shows one strong peak at 521 cm−1 originating from the silicon substrate. The most intriguing fourth component, localized in the vicinity of the bowed defect, can now be tentatively identified by comparing its spectrum to the spectra of the known VO2 crystallographic phases.
Figure 1b presents a temperature dependence of the averaged Raman spectra recorded from a free-standing single-crystal nanobelt released and collected from the same substrate as the studied microbelt and having similar doping level (2‰ to 3‰). At room temperature the nanobelt’s phase can be clearly identified as T and a gradual transition to the M1 phase is observed upon cooling to −131 °C. The spectral signatures distinguishing these two phases at room temperature are: a) appearance of a T-phase peak at 584 cm−1, b) bifurcation of the M1 peak at 390 cm−1 into two T-peaks at 383 cm−1 and 401 cm−1, c) appearance of a small T-peak at 443 cm−1 and d) a bathochromic shift of an M1-peak from 143 cm−1 to 126 cm−1. Hence, the fourth Bayesian component can be tentatively identified as the VO2 T-phase. Recalling that the M1-T transition is gradual (unlike other phase transitions in VO2 that are sharp)24, characterized by a continuous distortion of the monoclinic lattice to the triclinic, we expect to see some intermixture of signatures. Indeed, the lower-frequency peak of component 4 appears not at 126 cm−1, but at 140 cm−1 and the peaks of component 2 around 600 cm−1 and 400 cm−1 appear to be slightly bifurcated, which is similar to the case of the nanobelt spectrum recorded at -51 °C (Fig. 1b), when transition to the M1 phase was not yet completed. As discussed before, this presents an additional challenge for the spectral unmixing algorithms, as the pure component spectrum can be very similar to a linear combination of the spectra of other components. Several tokens in Fig. 3 indicate that the unmixing may not be optimal: the Si signal (component 3) has non-zero intensity on the microbelt itself and the corresponding endmember has several small peaks originating from VO2; component 3 map appears to have contributions from both types of the ferroelastic domains; component 2 has pronounced peaks at 588 cm−1 and 397 cm−1. In addition, the average abundance of the fourth component in the region of its highest intensity is only about 50%. Thus, in order to ensure the effectiveness of the BLU, we performed deconvolution of the datacube into five components. As can be seen (Fig. 4), this allowed complete separation of the ferroelastic domain signals and a better highlighting of component 4 by increasing its average abundance to ca. 80%. The ferroelastic domain signal of component 2 lost the 397 cm−1 mode and its 594 cm−1 peak decreased in intensity to the level typically observed in the M1 phase. The component 3 spectrum became completely void of VO2 peaks and has zero intensity on the microbelt. The mixed Si-VO2 signal was adjudicated into a separate component 5, whose intensity is around 90% on the microbelt edges. Further increase in the number of components is not feasible, as it brings about a physically-meaningless splitting of the existing components.
Interestingly, comparison of the component 4 maps to the AFM and SEM images indicates that although the T-phase region must be pinned by the bowed defect, it is in no way identical in shape with the later. The bowed defect concaves toward the top of the image, whereas the T-region convexes in the same direction (Fig. 4,4). Moreover, the T-region features a secondary bow that curves in the opposite direction (to the bottom of the image), which has no direct correspondence in the topographic images. We speculate that the T phase stabilization in this second bow is not driven by the local stress exerted by the precipitate on the microbelt surface, but by the local uneven Al doping inside the microbelt.
Conclusions
To summarize, we have studied distribution of crystallographic and ferroelastic phases in a defected Al-doped VO2 microcrystal employing a synergistic conjunction of micro-Raman and Bayesian linear unmixing algorithm. The deep data analysis allowed discriminating an uneven distribution of the T phase, partly correlated with the surface defects. Stabilization of the T phase is likely due to the defect-induced stress and local uneven doping. This work demonstrates the importance of full mapping in hyperspectral micro-Raman studies of vanadium dioxide systems, which is commonly substituted with averaging or single point-probing approaches that suffer from low resolving power.
From a more general point of view, full data acquisition is required in order to increase the scientific output of the existing materials characterization techniques. However, that brings about the challenge of data analysis, visualization and interpretation. The big data and deep data statistical approaches can solve this problem by unraveling the information on the functional properties present in the raw data without losing unsought-for signatures of serendipitous phenomena. The deep data methods also allow for deconvolution of concurrent processes and for ascribing physical models to the statistically-relevant behaviors. The continuing progress in information and computer technologies suggests that the usage of these methods and their applicability to the areas of physics, chemistry and materials science will significantly expand in the nearest future, boosting the progress in materials design.
Additional Information
How to cite this article: Strelcov, E. et al. Local coexistence of VO2 phases revealed by deep data analysis. Sci. Rep. 6, 29216; doi: 10.1038/srep29216 (2016).
References
Zerov, V. Y. et al. Features of the operation of a bolometer based on a vanadium dioxide film in atemperature interval that includes a phase transition. J. Opt. Technol . 66, 387 (1999).
Chen, C., Yi, X., Zhao, X. & Xiong, B. Characterizations of VO2-based uncooled microbolometer linear array. Sensors and Actuators A: Physical 90, 212–214, doi: 10.1016/s0924-4247(01)00495-2 (2001).
Gao, Y. et al. Nanoceramic VO2 thermochromic smart glass: A review on progress in solution processing. Nano Energy 1, 221–246, doi: 10.1016/j.nanoen.2011.12.002 (2012).
Li, S. Y., Niklasson, G. A. & Granqvist, C. G. Thermochromic fenestration with VO2-based materials: Three challenges and how they can be met. Thin Solid Films 520, 3823–3828, doi: 10.1016/j.tsf.2011.10.053 (2012).
Tselev, A. et al. Electromechanical Actuation and Current-Induced Metastable States in Suspended Single-Crystalline VO2 Nanoplatelets. Nano Letters 11, 3065–3073, doi: 10.1021/nl200493k (2011).
Wang, K. et al. Performance Limits of Microactuation with Vanadium Dioxide as a Solid Engine. ACS Nano 7, 2266–2272, doi: 10.1021/nn305419e (2013).
Liu, K. et al. Giant-Amplitude, High-Work Density Microactuators with Phase Transition Activated Nanolayer Bimorphs. Nano Letters 12, 6302–6308, doi: 10.1021/nl303405g (2012).
Rini, M. et al. Optical switching in VO2 films by below-gap excitation. Applied Physics Letters 92, doi: 10.1063/1.2921784 (2008).
Strelcov, E., Lilach, Y. & Kolmakov, A. Gas Sensor Based on Metal−Insulator Transition in VO2 Nanowire Thermistor. Nano Letters 9, 2322–2326, doi: 10.1021/nl900676n (2009).
Byon, J. W. et al. Electrothermally Induced Highly Responsive and Highly Selective Vanadium Oxide Hydrogen Sensor Based on Metal–Insulator Transition. The Journal of Physical Chemistry C 116, 226–230, doi: 10.1021/jp2080989 (2011).
Sengupta, S. et al. Field-effect modulation of conductance in VO2 nanobeam transistors with HfO2 as the gate dielectric. Applied Physics Letters 99, doi: 10.1063/1.3624896 (2011).
Liu, K. et al. Dense Electron System from Gate-Controlled Surface Metal–Insulator Transition. Nano Letters 12, 6272–6277, doi: 10.1021/nl303379t (2012).
Ko, C. & Ramanathan, S. Observation of electric field-assisted phase transition in thin film vanadium oxide in a metal-oxide-semiconductor device geometry. Applied Physics Letters 93, doi: 10.1063/1.3050464 (2008).
Hormoz, S. & Ramanathan, S. Limits on vanadium oxide Mott metal–insulator transition field-effect transistors. Solid-State Electronics 54, 654–659, doi: 10.1016/j.sse.2010.01.006 (2010).
Nakano, M. et al. Collective bulk carrier delocalization driven by electrostatic surface charge accumulation. Nature 487, 459–462, doi: 10.1038/nature11296 (2012).
Tselev, A. et al. Symmetry Relationship and Strain-Induced Transitions between Insulating M1 and M2 and Metallic R phases of Vanadium Dioxide. Nano Letters 10, 4409–4416, doi: 10.1021/nl1020443 (2010).
Park, J. H. et al. Measurement of a solid-state triple point at the metal-insulator transition in VO2 . Nature 500, 431–434, doi: 10.1038/nature12425 (2013).
Wei, J., Wang, Z., Chen, W. & Cobden, D. H. New aspects of the metal-insulator transition in single-domain vanadium dioxide nanobeams. Nat Nano 4, 420–424 (2009).
Budai, J. D. et al. Metallization of vanadium dioxide driven by large phonon entropy. Nature 515, 535–539, doi: 10.1038/nature13865 (2014).
Chou, J. Y., Lensch-Falk, J. L., Hemesath, E. R. & Lauhon, L. J. Vanadium oxide nanowire phase and orientation analyzed by Raman spectroscopy. J. Appl. Phys. 105, doi: 10.1063/1.3075763 (2009).
Zhang, S., Chou, J. Y. & Lauhon, L. J. Direct Correlation of Structural Domain Formation with the Metal Insulator Transition in a VO2 Nanobeam. Nano Letters 9, 4527–4532, doi: 10.1021/nl9028973 (2009).
Chang, S. J. et al. In situ probing of doping- and stress-mediated phase transitions in a single-crystalline VO2 nanobeam by spatially resolved Raman spectroscopy. Nanoscale 6, 8068–8074, doi: 10.1039/c4nr01118j (2014).
Tselev, A. et al. Mapping internal structure of coal by confocal micro-Raman spectroscopy and scanning microwave microscopy. Fuel 126, 32–37, doi: 10.1016/j.fuel.2014.02.029 (2014).
Strelcov, E. et al. Doping-Based Stabilization of the M2 Phase in Free-Standing VO2 Nanostructures at Room Temperature. Nano Letters 12, 6198–6205, doi: 10.1021/nl303065h (2012).
Strelcov, E., Davydov, A. V., Lanke, U., Watts, C. & Kolmakov, A. In Situ Monitoring of the Growth, Intermediate Phase Transformations and Templating of Single Crystal VO2 Nanowires and Nanoplatelets. ACS Nano 5, 3373–3384, doi: 10.1021/nn2007089 (2011).
Belianinov, A. et al. Big data and deep data in scanning and electron microscopies: deriving functionality from multidimensional data sets. Adv Struct Chem Imag 1, 1–25, doi: 10.1186/s40679-015-0006-6 (2015).
Windig, W. & Guilment, J. Interactive self-modeling mixture analysis. Analytical Chemistry 63, 1425–1432, doi: 10.1021/ac00014a016 (1991).
Windig, W., Heckler, C. E., Agblevor, F. A. & Evans, R. J. Self-modeling mixture analysis of categorized pyrolysis mass spectral data with the SIMPLISMA approach. Chemometrics and Intelligent Laboratory Systems 14, 195–207, doi: 10.1016/0169-7439(92)80104-C (1992).
Dobigeon, N., Moussaoui, S., Coulon, M., Tourneret, J. Y. & Hero, A. O. Joint Bayesian Endmember Extraction and Linear Unmixing for Hyperspectral Imagery. IEEE Transactions on Signal Processing 57, 4355–4368, doi: 10.1109/tsp.2009.2025797 (2009).
Dobigeon, N. & Brun, N. Spectral mixture analysis of EELS spectrum-images. Ultramicroscopy 120, 25–34, doi: 10.1016/j.ultramic.2012.05.006 (2012).
Strelcov, E. et al. Deep Data Analysis of Conductive Phenomena on Complex Oxide Interfaces: Physics from Data Mining. ACS Nano 8, 6449–6457, doi: 10.1021/nn502029b (2014).
Strelcov, E., Belianinov, A., Hsie, Y.-H., Y.-H., C. & Kalinin, S. V. Constraining data mining with physical models: voltage- and oxygen pressure-dependent transport in multiferroic nanostructures. Nano Letters, doi: 10.1021/acs.nanolett.5b02472 (2015).
Marini, C. et al. Optical properties of V1−xCrxO2 compounds under high pressure. Phys. Rev. B J1 - PRB 77, doi: 10.1103/PhysRevB.77.235111 (2008).
Acknowledgements
This research was conducted at the Center for Nanophase Materials Sciences, which is sponsored at Oak Ridge National Laboratory by the Scientific User Facilities Division, Office of Basic Energy Sciences, U.S. Department of Energy. ES acknowledges support under the Cooperative Research Agreement between the University of Maryland and the National Institute of Standards and Technology Center for Nanoscale Science and Technology, Award 70NANB10H193, through the University of Maryland.
Author information
Authors and Affiliations
Contributions
E.S. and A.K. have synthesized the sample, E.S. and A.I. performed measurements, A.B. adapted the BLU analysis codes, E.S. performed the analysis and wrote the manuscript, S.V.K. and A.T. supervised the research. All authors discussed the results and commented on the manuscript.
Ethics declarations
Competing interests
The authors declare no competing financial interests.
Electronic supplementary material
Rights and permissions
This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/
About this article
Cite this article
Strelcov, E., Ievlev, A., Belianinov, A. et al. Local coexistence of VO2 phases revealed by deep data analysis. Sci Rep 6, 29216 (2016). https://doi.org/10.1038/srep29216
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/srep29216
This article is cited by
-
Automated Interpretation and Extraction of Topographic Information from Time of Flight Secondary Ion Mass Spectrometry Data
Scientific Reports (2017)
-
Synthesis and thermochromic property studies on W doped VO2 films fabricated by sol-gel method
Scientific Reports (2017)
-
In situ nanomechanical behaviour of coexisting insulating and metallic domains in VO2 microbeams
Journal of Materials Science (2017)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.