1. Introduction

Spray-flame synthesis of nanoparticles, also called flame spray pyrolysis (FSP) has proven to be a powerful process to synthesise countless functional materials for a wide range of practical applications [1–4]. The vast amount of soluble metal-based precursors that are dissolved in combustible liquids or solvent mixtures together with the control of the synthesis route (liquid-to-particle and gas-to-particle conversion) allows for the production of single- and multi-element oxide particles with tailored application-specific characteristics, such as particle size, morphology, composition (surface and bulk), crystallinity, and polydispersity [2–5]. Their reliability, versatility, and scalability make this route towards nanoparticle synthesis very attractive. The spray-flame process can be described by several physico-chemical mechanisms [1,4] that include atomisation of the precursor solution, droplet evaporation, droplet micro-explosion, liquid-phase chemical reactions, precipitation, gas-phase chemical reactions, particle nucleation, surface growth, coagulation, sintering, aggregation, and agglomeration. The interplay among these mechanisms ultimately depends on the reaction environment, that is characterised by a turbulent reacting multi-phase flow field, where temperature, velocity (related to the residence time), and reactant concentration (including fuel-oxidiser ratio) change locally [4]. Although the main mechanisms seem to be already mapped, there is still a lack of complete understanding of how these mechanisms are triggered and interact to better control the final product characteristics and to scale up the process to industrial applications with optimised production outputs.

A deep understanding of the spray-flame synthesis mechanisms can be ultimately acquired based on conjunction between theory, experiments, and simulations [3,4,6]. To this end, a standardised lab-scale reactor for spray flame synthesis has been recently introduced, called SpraySyn burner [3,7]. The geometry of the SpraySyn burner and its proposed standard conditions are designed to facilitate experiments and simulations, as well as the systematic comparison of results among different research groups [7]. The characterisation of the burner and the underlying particle formation processes are being gradually acquired [8–21].

The SpraySyn burner is composed of a central external-mixing twin-fluid nozzle that atomises a flammable precursor solution through a high-momentum oxygen dispersion gas flow. The precursor solution is designed according to the desired synthesised material, solution stability and cost [4,22–24]. Primary and secondary atomisation processes of the precursor solution results in a spray of fine liquid droplets [14,16,17,19]. The spray is ignited and stabilised by a premixed flat methane-oxygen pilot flame [7,17]. An outermost annular co-flowing nitrogen gas shields the inner combusting flow from external fluctuations and assists in the transport of synthesised particles [7,12]. The particle synthesis can follow a gas-to-particle conversion, where nanoparticles nucleate in the gas phase after precursor evaporation. The droplet evaporation can be modelled [20] from the droplet size and velocity [9,14,16], temperature field of the liquid solution [13], and temperature field of the gas [10,15]. Droplet micro-explosions may occur disrupting parent droplets into smaller child droplets, which enhances evaporation [25,26]. Particles grow inside the high-temperature flame region due to coagulation and sintering [4], which is influenced by the gas velocity [12,16,18] and temperature fields [6,10,15]. Then, aggregation and agglomeration mechanisms occur, generating fractal-like structures [6,15,16,21,26]. The particle synthesis can also evolve from a usually undesired liquid-to-particle conversion, where precursor precipitates inside the droplet or at its surface and reacts. After evaporation of the solvent, the final products are large, hollow, or shell-like particles. Multiple particle formation mechanisms can simultaneously occur (e.g., in the presence of a broader droplet size distribution), leading to heterogeneous particle distributions [15–17,21,26].

To better control the spray-flame synthesis process via gas-to-particle conversion and, therefore, the final particle characteristics, a detailed investigation of the turbulent spray flame is paramount. Although the SpraySyn burner is designed to generate a time-averaged symmetric flame, it is instantaneously unsteady, turbulent, and hence inherently three-dimensional. The flow structures (eddies of different length scales) influence the droplet evaporation, flame propagation, particle temperature histories, and the distribution of particles in the flame, which ultimately influence the final particle morphology. Fundamental understanding can be acquired with accurate measurements of the evolution of the 3D flame topology. To this end, tomographic imaging using multi-simultaneous measurements (TIMes) [27] is performed in the present work. This concept has never been applied before for 3D imaging of a spray flame used for nanoparticle synthesis. TIMes is a non-intrusive technique that can reveal the limits of different zones and the spatial correlation of different effects in the entire volume, by simultaneous detection of various signals such as emission from flame radicals (chemiluminescence) or electronically excited atomic ions that have been seeded into selected flow streams. Alkali metal atoms that show intense atomic emission under high temperatures can be seeded into the flow, for example, to analyse the burnt gas regions in the visible wavelength range. TIMes requires a relatively simple and cost-efficient optical setup with multiple 2D sensors allowing spectroscopic combustion diagnostics without the need for expensive laser equipment often used in detailed investigations of particle-synthesis processes [28]. In the present work, we analyse instantaneous and time-averaged 3D fields of natural CH* chemiluminescence (prominent indicator of the flame front) and emission from atomic sodium resulting from NaCl dissolved in the combustible liquid (signal related to the hot burnt spray products) under a variety of operating conditions through the SpraySyn burner.

2. Experimental methods

2.1 Experimental setup for TIMes

In the present work, we measured the chemiluminescence of radical CH*, which is frequently used as an indicator of the flame front due to its presence in close vicinity of the flame front in hydrocarbon flames. The study of this region is relevant in the context of spray-flame synthesis because it is associated with high heat-release and high-temperature regions, which impacts the produced nanoparticles. A trace amount of sodium chloride (NaCl) was dissolved into the injected liquid to act as a tracer for visualising the regions containing hot burnt spray products via emission from atomic Na. The Na signal depends on temperature [29] and therefore is only detected in high-temperature regions (i.e., the burnt side of the spray flame). Considering that the boiling point of the ethanol solution is not affected by the trace amount of NaCl solution and it is equal to that of pure ethanol (about 351 K [30] at atmospheric pressure), negligible signal from Na within the liquid solution jet and atomised droplets are expected due to the local cold temperatures, as measured using two-colour laser-induced fluorescence liquid-phase thermometry in [13]. Therefore, the signal is assumed to represent the gas-phase region in which the hot burnt spray flame products are found. In spray-flame synthesis, these regions are expectedly related to the particle nucleation, surface growth, coagulation, sintering and aggregation steps within the gas-to-particle pathway. Nevertheless, the Na illumination regions must be interpreted with caution, because the Na signal, in addition to exponential dependence on the gas temperature, also depends on the local concentration and equivalence ratio [29]. As the emission of atomic sodium is very intense, only 7.5 mg of NaCl (rock salt) was dissolved in 5 ml of distilled water and 195 ml of absolute ethanol (VWR Chemicals, No. 20821.330), resulting in a diluted solution with a salt concentration of 0.0375 g/l. Na was chosen as a tracer, because the relevant emission peak is spectrally separated from that of the CH* chemiluminescence [27].

Our generic cost-effective multi-camera setup [27,31] was slightly modified for this work. Due to the much smaller size of the SpraySyn burner compared to other burners used in our previous studies, the cameras were shifted nearer to the burner at a distance of about 320 mm to increase the imaging resolution. The cameras were arranged in three inclined levels as a result to accommodate the reduced amount of space, as shown in Fig. 1. Furthermore, the low power of the spray flames (about 1.9 kW for operating condition SF1 [21]) did not cause any heating problems for the cameras. The setup consists of 29 monochrome CCD cameras (Basler acA-645-100gm), each equipped with 12-mm focal-length lenses (Kowa LM12JC) with an aperture opening of f/1.4 arranged in a 168° arc around the burner. The cameras have a sensor of 659×494 pixels with pixels of size 9.9×9.9 µm² (½” Sony ICX414). Two different spectral bandpass filters were mounted on the camera lenses in an alternated pattern, as shown in Fig. 1 by different coloured triangles, with 15 cameras dedicated to the CH* channel and 14 to the Na channel measurement. The CH* emission was separated by BrightLine filters, with a centre wavelength of 433 nm and a full width half maximum (FWHM) of 24 nm, whereas the Na signal was detected by narrowband TECHSPEC filters, with a centre wavelength of 589 nm and a FWHM of 10 nm, as shown in Fig. 3 (see Section 2.3). Different camera exposure times were used for CH* and Na, as presented in Table 1, to account for variations in the signal intensity, and they were optimised based on a trade-off between adequate signal-to-noise ratio and acceptable motion blur, following [27,31]. The image acquisition was simultaneously triggered at 3.5 Hz using a pulse generator (Berkeley Nucleonics, model 577) and simultaneous image acquisition was tested through additional experiments by imaging a stopwatch with all cameras.

 

Fig. 1. Experimental setup for TIMes measurements of the SpraySyn burner with blue and red triangles indicating cameras equipped with CH* and Na filters, respectively.

Download Full Size | PDF

Table 1. Operating conditions of the SpraySyn burner investigated in the present work, with camera exposure times used for the TIMes measurements. The pilot gas flow rate (2 slm of CH₄ + 16 slm of O₂) and sheath gas flow rate of 120 slm were fixed. The global equivalence ratio considers all fuel and oxidizer flows fed to the burner.

View Table | View all tables in this article

2.2 Tomographic reconstruction method

The 3D data analysed in the present work was obtained by computed tomography of various emission signals using an evolutionary reconstruction technique (ERT). The ERT [32] has been applied to reconstruct the 3D chemiluminescence intensity fields of different turbulent and unsteady premixed flames, producing comparable results to state-of-the-art CTC algorithms [31,33]. The ERT has been recently further developed in different aspects for an evolutionary background oriented Schlieren tomography (EBOST) algorithm [34]. Here, the algorithmic and methodological advances that were achieved in the context of the EBOST were used to develop an improved version of the original ERT algorithm [32]. In general, evolutionary methods can be applied to nonlinear problems, which could benefit this project in the future development of an advanced measurement model, for example when considering signal re-absorption in case of OH* detection. Firstly, the basic ERT concepts can be summarized as follows:

A population is defined by individuals, and each individual is associated with a reconstruction domain. The reconstruction domain is discretised into an array of voxels along the x-, y-, and z-directions (here, z starts at the nozzle exit and points upwards), and represents an individual (chromosome) composed of voxels (genes), encoded as a floating-point value. For each individual, a fitness value can be calculated based on an error distance (e.g., L1- or L2-norm), between the reference images (measurements) and the rendered images of its associated domain. Rendering is achieved by volumetrically ray-tracing images of an array of pin-hole cameras (synthetic tomographic setup). The evolutionary process consists of a selection scheme with a bias towards higher fitness values and the application of genetic operators, such as crossover (merging), mutation, annihilation, and filtering, with defined probabilities (mutation rate, annihilation rate and filter rate). The updated ERT version relies on stochastic universal sampling [35] for the selection of individuals that are used to form the new generation. Each selected pair of individuals is merged into an offspring by using a voxel-wise arithmetic average. The population size is constant throughout the evolutionary process and elitism is used to increase the evolutionary pressure.

In addition, our ERT uses a masking method based on a Metropolis algorithm to sample locations where the genetic operators should be applied. This mask is generated with the help of the reference images in the first stage. In the second stage, after the population has evolved for a certain number of generations the mask can be re-created by the best fitting chromosome. For the second stage, either the intensity field of the best chromosome, or the magnitude of its gradient is used interchangeably for sampling. The latter promotes the edges of the object that is reconstructed, for flame emission these correspond to the flame front for example.

The basic ERT concepts were expanded by including scaled operator hierarchies, as described in [34]. An operator hierarchy is defined as a series of genetic operators for mutation, annihilation, and filtering, respectively. These consider the wide range of scales that are present in the scalar field to be reconstructed. The scaled mutation operator hierarchy consists of operators that mutate a cubic region of voxels. The new value for the voxels in the particular region is sampled from a normal distribution with a mean equal to the average of the central voxel and its nearest neighbours, and a standard deviation sigma that needs to be defined. The edge length of the operators is different and ranges from a single voxel up to, e.g., 15 voxels, which can be dependent on the numbers of voxels in the reconstruction domain in each spatial direction and the expected flame size. Each mutation operator in the set of operators that forms the hierarchy is applied in a random order in the evolution step. The rate at which a mutation operator is applied was assumed to be different for each operator of the hierarchy. Hence the introduction of three different hierarchies, one each for mutation, for annihilation, which sets a defined region of voxels to zero intensity, and for filtering, which applies a 3×3×3 Gaussian stencil to a region of voxels, leads to a large set of potentially tuneable parameters. Manual parameter tuning is out of scope for this set of parameters and therefore a self-adaptive parameter heuristic was implemented. In addition, the mutation standard deviation sigma was added to the self-adaptive set, similar to the meta-learning approach presented in [34].

The updated ERT algorithm was coded in C and CUDA and is GPU accelerated to reduce the run-time of the reconstruction process. A second step to improve the runtime and the performance of the method was to use an island-based genetic algorithm. Each island-population can evolve on a separate GPU device with reduced communication due to an inter-island migration policy, which reduces the memory transfer bottleneck between devices and allows the system to be upscaled to perform reconstructions on larger domain sizes with measurement images that have a higher number of pixels. Encouraged by the positive impact that a migration policy focusing on chromosome diversity [36] had for the EBOST algorithm, the same approach was used here. In Fig. 2, the definition of the chromosomes for the CH* and Na scalar fields is presented. The figure also shows a schematic of the evolution step for an island population of N chromosomes, which is repeated for N_g generations. For simplicity of visualisation, the second part of the chromosome that contains the evolutionary parameters, which are co-evolved in a meta-learning scheme, are omitted.

 

Fig. 2. (a) Definition of the chromosome as a discretised reconstruction domain for the CH* and Na luminescence fields. Each gene of the chromosome represents an intensity value and is treated as a voxel in the volume rendering conducted for fitness evaluation. (b) Evolution step for an island population of N chromosomes.

Download Full Size | PDF

The reconstructions presented in our analysis were performed at the Oculus cluster of the Paderborn Centre for Parallel Computing. For a single reconstruction, a node equipped with two NVIDIA GeForce RTX 2080 GPUs was used and two island-populations were allocated on each device. Convergence of the algorithm was checked by tracking the fitness, here the L2- norm, of the globally best chromosome and the reconstruction process was stopped with the fitness of the currently best solution improving only marginally for several thousand generations (ca. 5% of the total number of generations).

The tomographic reconstruction relies on an accurate calibration of the cameras. The location and orientation (extrinsic parameters) of the cameras in the world coordinate system were determined based on a 3D calibration target [37], while the focal length of the lenses and the principal point (intrinsic parameters) were fixed at 12 mm (datasheet value) and at the centre of the camera sensor, respectively. The calibration method based on a genetic algorithm uses one image of the target per camera and minimises the L2-norm between the acquired image and a rendered view, which is a function of the pin-hole camera parameters and the synthetic calibration target.

For every flame condition in Table 1, a total of 400 images were captured in succession by every camera, simultaneously acquiring the CH* and Na signal at each time instant. Background correction was done by subtracting dark images (without the flame) from the signal images. Negligible light reflections were observed in the images. Due to the fluctuations present in the spray flame, several images with low signal-to-noise ratio were present. These fluctuations in the SpraySyn burner have already been reported in the literature and ascribed to sequences of ignition, combustion, and extinction of the spray flame that are presumably due to a shielding effect of the O₂ dispersion gas preventing ignition of the flammable liquid spray by the pilot flame [17]. The observed fluctuations directly affect the particle trajectory history including the high-temperature residence time, which changes the final synthetised material. For example, the extinguished spray flames (i.e., the lowest intensity images) probably lead to liquid-to-particle formation route due to the colder environment. Therefore, since most of FSP process of interest rely on the gas-to-particle formation route, low-intensity images were sorted out from the analysis. Assuming a flame active time of 50% [17], only the 200 brightest images were kept based on the averaged intensity of all cameras. The tomographic reconstructions of CH* and Na signals were computed based on these brightest instantaneous images. They were cropped to a size of 195×302 pixels (width by height) and resized by a factor of 0.75. The reconstruction domain was discretised into 80×80×160 voxels along x-, y-, and z-directions, respectively, with cubic voxels of size 0.375 mm. The voxel size was selected based on the image resolution and number of viewing directions [31].

2.3 Spectrally-resolved chemiluminescence measurements for experimental validation of reconstructions

Reconstructions will always have some uncertainty and error associated with them that stems from different aspects such as calibration, inversion process, imaging model, discretisation errors and so on. Common practice is to numerically check the accuracy of reconstructions in so-called quantitative phantom studies where synthetic projections of an exactly known field are used to generate a virtual experimental scene, and the original and reconstructed phantoms are then compared. Phantom studies should in fact be performed for any newly developed tomographic algorithm, and an extensive phantom study with the ERT has been previously presented [34]. Nonetheless, phantom studies are also not perfect since the virtual scene is different in many aspects compared to a real experimental one, and in particular incorporating the experimental errors in the synthetic measurements perfectly is not possible. Therefore, experimental validation of reconstructions should also be endeavoured where possible, though it is not a trivial task since ground truth information is not available. Nonetheless, tomography practitioners have attempted this, at most by comparing reconstructions with an independent experimental measurement that is made simultaneously with the tomographic data collection [38–40]. In this work, line-of-sight spectrally-resolved chemiluminescence measurements were obtained to compare with the volumetric reconstructions for experimental validation purposes.

The experimental setup for recording spectrally-resolved chemiluminescence in the SpraySyn flame constituted mounting the burner on a height-adjustable rack to enable measurements at different heights above the burner exit. Chemiluminescence signal is collected by a 100-mm focal-length UV lens (B. Halle Nachfl.) set with an aperture opening of f/2 followed by a horizontal 50-µm slit before reaching the spectrometer (Princeton Instruments, Acton SP-150). An optical filter can optionally be placed in the beam path between the lens and the slit to block unwanted light and to suppress second-order diffraction artefacts. The spectrometer has a focal length of 150 mm and was operated with 150 or 600 grooves/mm diffraction gratings. A high-sensitivity CCD camera (LaVision, Imager Intense) equipped with an intensifier (LaVision, Intensified Relay Optics) was directly connected at the spectrometer exit port. The chemiluminescence entering the spectrometer was collected from a 31×0.475 mm² region in object space (horizontal and vertical directions, respectively). The horizontal extent was determined by the distance of two illuminating points from an optical fibre placed at the burner, while the vertical size was estimated based on the optical magnification factor. The wavelength-dependent transmission efficiency and detection sensitivity of the optical setup was determined using a broadband laser-driven light source with known spectral emission characteristics (Energetiq Technology, EQ-99X). The wavelength axis was calibrated using the known sharp spectral lines of a PenRay lamp. Spectrally-resolved horizontally-linear measurements were acquired at distinct heights above the burner exit for the different flame operating conditions. The post-processing of raw images and data analysis was done by an in-house MATLAB code.

 

Fig. 3. Chemiluminescence spectra of the SpraySyn flame seeded with NaCl for the standard operating condition SF1 at different heights above the burner and transmission spectra of the applied optical filters. (a) Entire spectra between 210 and 700 nm (150 grooves/mm grating) and (b) higher-resolution spectra around CH* and Na emission peaks (600 grooves/mm grating).

Download Full Size | PDF

Figure 3 shows the spectrally-resolved chemiluminescence of the ethanol SpraySyn flame seeded with NaCl for the standard operating condition SF1 (Table 1) at different heights above the burner z. The overview spectra, Fig. 3(a), were recorded with the coarse diffraction grating (150 grooves/mm), while the higher diffraction grating (600 grooves/mm) allowed a more detailed view of the region of interest around the spectral lines of CH* and Na, Fig. 3(b). Additionally, the transmission spectra of the applied optical filters in the TIMes technique (433 ± 12 nm and 589 ± 5 nm filters) are plotted as transparent areas. They were measured with a UV/VIS absorption spectrometer (Varian Inc., Mod. Cary 400).

Several radical emission peaks can be identified in the chemiluminescence spectra, with the most significant peaks being the OH* emission (around 310 and 287 nm), broader emission from CO₂* (from below 300 nm to above 600 nm [41]), CH* emission (431 and 389 nm) and Na emission (589 nm). The artefact at around 620 nm originates from second-order diffraction of the OH* emission. For the higher spectral resolution measurements of the sodium D-lines, an optical filter (Schott glass, GG420 longpass filter) was placed in front of the spectrometer to block signal below 400 nm and thus to suppress interference with second-order diffraction. Sodium was selected in the present work as a tracer because its signal is strong at a distinct wavelength location not coinciding with other peaks and is close to the best transmission region of the camera lens used in the TIMes method (section 2.1). CH* features comparatively strong emission bands in the visible range, making it suitable for detection by the tomographic camera setup (non-UV) and was hence chosen to represent the flame front.

Radial profiles related to CH* chemiluminescence and Na emission can be computed by integrating the spectra from the emission measurements considering the transmission curves of the corresponding filters used in the TIMes method. The radial profiles have the same horizontal resolution of the spectrally-resolved measurement, and they are spatially integrated along the line-of-sight of the collecting optics. CH* and Na radial profiles are plotted in Fig. 4 for the measured heights above the burner exit at the operating condition Inj4_DG12. Throughout the text, the x'-axis is related to the sensor view and does not coincide with the x-axis of the coordinate system of the reconstruction domain. CH* is mainly located in the upstream region of the flame up to about z = 50 mm, with a maximum at z = 15 mm. The broad CH* emission around z = 0 mm originates from the pilot flame. In contrast, the Na emissions can be detected with the spectrometer-based detection system up to z = 75 mm downstream of the burner, with a maximum emission at z = 30 mm. Compared to CH*, the Na emissions are spread over a broader radial range.

 

Fig. 4. Radial profiles of integrated emission measurements using (a) 433 ± 12 nm filter for CH* chemiluminescence and (b) 589 ± 5 nm filter for Na* emission at the operating condition Inj4_DG12.

Download Full Size | PDF

2.4 SpraySyn burner

The schematics of the SpraySyn burner used (serial number SSB-2018-05) with working gases and liquid precursor solution is presented in Fig. 5. This is the same burner model used previously by Schneider et al. [7]. A liquid solution of a designed precursor in a flammable fluid is fed through the capillary and is dispersed by oxygen gas. The atomised solution produces a spray flame, which is sustained by a laminar premixed methane/oxygen pilot flame, stabilised on top of a porous plate. An inert sheath gas passes through the same plate at the outermost annular region, protecting the flame against environmental influences.

 

Fig. 5. Schematics of the SpraySyn burner with dimensions. Top view picture of the nozzle exit taken with a USB microscope.

Download Full Size | PDF

Flow rates of the pilot gas mixture and the sheath gas (co-flow) were held constant in the present study. A CH₄ flow rate of 2 slm was combined with an O₂ flow rate of 16 slm to produce the premixed pilot flame (fuel/oxidiser equivalence ratio ϕ = 0.25), whereas an N₂ flow with a flow rate of 120 slm was employed as sheath gas. Six operating conditions were considered in the present study by varying the flow rates of the O₂ dispersion gas and the liquid solution (ethanol + trace amount of NaCl) from 9 to 12 slm and from 2 to 4 ml/min, respectively resulting in overall lean flames with a global fuel/oxidiser equivalence ratios ϕ_g ranging from 0.23–0.36 (considering all feed flows). The selected dispersion gas flow rates operate the twin-fluid nozzle under subsonic conditions with cold-flow Reynolds numbers of 6100–8200. All gas flows were metered by mass-flow controllers (Bronkhorst, EL-Flow model) and the liquid solution was injected by a syringe pump (Ayxesis GmbH). The SpraySyn burner was operated without an enclosure at an ambient temperature of about 295 K and atmospheric pressure. Table 1 details the different operating conditions used. The Inj2_DG10 refers to the standard ethanol spray flame condition (SF1) used in [7]. The camera exposure time was varied to account for the intensity differences in the CH* and Na signals. The intensity decreases with increasing dispersion gas flow rate, and it increases with increasing liquid injection flow rate. Generally, the averaged CH* signal is much weaker than that of Na.

3. Burner alignment

The present spray-flame characteristics are very sensitive to minor deviations of the burner geometry, as also frequently observed in many small-scale burners but often overlooked. Small inclinations or asymmetries introduce bias to the measured flow quantities of interest, especially if the measurement technique intrinsically assumes perfect axial symmetry. For accurate and reproducible measurements, the burners need to have equal dimensions and be carefully aligned. The standard alignment procedure of the SpraySyn burner includes comparing features of the present flame with a benchmark under standard operating conditions. The procedure is described in detail in [7] and it will only be briefly explained here.

First, the burner is levelled, and the capillary is centred with the outer nozzle by visual top-view inspection using three micrometre screws at the burner base. Second, an iterative fine adjustment of the spray flame symmetry under operating conditions is performed visually by eye from multiple directions. Then, ethanol spray flame images are captured under the standard condition SF1 by a DSLR camera (Nikon D5300) equipped with a 50-mm focal-length lens (Nikon AF-S 50/1.8G Nikkor) set with an aperture opening of f/1.8. The camera is fixed at a distance of 700 mm from the burner centre with the camera base coinciding with the burner surface plane. The RGB images are acquired using an exposure time of 1/13 s and ISO 100. The image captures the flame radiation transmitted through the embedded Bayer filter of the DSLR camera. A typical instantaneous image can be seen in Fig. 6(c). The brightest intensities in the image roughly correspond to the hottest regions of the flame [10], while dark regions at the centreline close to the exit are related to the cold spray zone. Flame features such as height (location where the intensity along the flame centreline drops down to 10% of the peak intensity downstream), full width at half maximum (FWHM, diameter measured between two points where the intensities are half of the peak value), tilt angle (angle between flame centreline and the vertical) and colour ratio (blue and red colour channels) are quantified and compared with a reference benchmark using a MATLAB script provided in [7]. Large variations of any feature require the capillary and inner nozzle of the SpraySyn burner to be readjusted in an iterative process. The standard alignment procedure states acquiring images from one perspective only, with a suggestion of additional views from other angles around the flame for further checking the flame characteristics.

 

Fig. 6. (a), (b) Flame inclinations $\gamma $ calculated from the time-averaged CH* reconstructions at the SF1 operating condition, as well as the inclination projections on the vertical planes (red lines). Isosurface values in the legend correspond to fractions of the maximum intensity of the field. (c) Typical instantaneous RGB image of the spray flame acquired with a DSLR camera for flame characterisation.

Download Full Size | PDF

Preliminary results from the reconstruction of CH* signal for the SF1 condition following the aforementioned standard procedure show a rather asymmetric and tilted spray flame, as presented in Fig. 6(a). This was the case despite the fact that the flame characterisation results using one viewing direction reasonably agreed with those of the corresponding benchmark. As shown in Fig. 6(a), the computed angle between the projected inclination of the reconstructed flame onto the y-z plane and the vertical is approximately 0°, although the projected inclination angle on the x-z plane is much greater, about 2°. Therefore, if the flame meets the flame characterisation criteria from one view, there is no guarantee that it will satisfy the criteria from another viewing direction. The asymmetry and tilting effects become evident in the 3D reconstructions where multiple views are considered. The 3D tilt angle of the spray flame $\gamma $ can be quantified using the CH* chemiluminescence reconstruction by least-squares fitting a 3D line as shown by the red continuous lines at the centre of the volumes in Fig. 6(a,b). The red lines are calculated based on the intensity peaks within horizontal slices from z = 20 to 43 mm, where the reacting flow is already merged towards the jet axis [7,12,18] and enough signal-to-noise ratio is present. Alternatively, the 3D tilt angle can be roughly estimated from two projections within independent vertical planes as described in Supplement 1. As a trade-off between image acquisition time and alignment accuracy, we propose an improved standard procedure for flame characterisation by analysing flame images from two viewing directions around the burner (preferably perpendicular to each other), to better adjust the investigated flame to the benchmark.

Due to space constraints in the present experimental setup, the flame characterisation was analysed based on two camera views separated by 67° in the x-y plane. After careful realignment of the burner capillary, the computed 3D flame inclination is around 0.3°, with the projected angles deviating minimally from the vertical axes in both the x-z and y-z planes, as shown in Fig. 6(b). The results of the flame characterisation using the improved procedure are summarised in Table 2. The final aligned flame displayed tilt angles that are about 2 to 3 times smaller than that of the benchmark but had a similar colour ratio and was slightly taller and wider. The latter differences seem to be caused by the discrepancies in the flame inclination, since the present flame is much straighter.

Table 2. Results of the flame characterisation for the standard SF1 operating condition by applying the standard alignment procedure discussed in [7] from two viewing directions.

View Table | View all tables in this article

4. Results and discussion

Instantaneous and averaged 3D reconstructed fields under SF1 condition are presented in Fig. 7 with blue representing the flame front and red the hot spray flame products. Coexistence of signals from both the reacting zones and hot burnt spray products is represented by the magenta regions (blue and red combined). Horizontal slices from both signals at different heights above the burner are also shown in Fig. 7 using a similar colour scheme, for signals that are normalised within the slice for better visualisation. The instantaneous 3D field shows that the flame front structures are observed to be more constricted around the central vertical axis, following the dispersion of the atomised spray. The instantaneous structures of the hot spray products are on the other hand much larger and undulated. The expansion of the burnt gases due to heat release enhances the observed characteristics of the Na structures, as explained in the particle image velocimetry study of [12]. Under combustion, the local release of chemical energy from exothermic reactions between a fuel mixture and an oxidiser (here referred to as heat release) generates hot burnt products, which have lower density causing the gas to expand, increasing the fluid velocity and thus turbulent fluctuations. Some motion blur is inevitably present in the reconstructions (originating from the imaging), particularly in the CH* field due to the higher exposure times needed to capture the weaker CH* chemiluminescence signals compared to Na emission. A diagonal streak starting at the nozzle exit is visible in the instantaneous CH* field, which possibly indicates burning droplets. Similar structures were observed from broadband chemiluminescence images by other work [10]. The instantaneous reconstructed field in Fig. 7 is additionally presented as a rotational visualisation movie (Visualization 1) for illustration of the flame geometry detail. The 3D fields show that the occurrence of Na signal generally starts at further downstream locations compared to those of the CH* signal (approximately at z = 4 mm in averaged reconstructed field of the SF1 condition). This trend is expected, because the detected signal from atomic Na emission is expected to occur downstream of the flame front, following the upward motion of the main jet stream.

 

Fig. 7. Instantaneous (t = t₁) and time-averaged 3D reconstructed fields of CH* and Na signals at the standard SF1 operating condition using TIMes. Normalised intensities within horizontal slices are presented at selected heights above the burner.

Download Full Size | PDF

The CH* slice at z = 0.5 mm in Fig. 7 (with a thickness of ∼1 mm) represents the closest cross-section to the burner surface of the lifted pilot flame. The analysis of this region is very important, because the pre-mixed pilot gas enters the porous matrix through a defined annular region with inner and outer diameters of 8 and 15 mm, respectively. The pilot gas mixture exits the matrix through an annular region with an inner diameter of 6 mm (i.e., outer diameter of the nozzle inset) and an outer diameter that depends on the dynamic balance between the pilot and the sheath gas flow rates (revisit Fig. 5). Assuming an empirical threshold to obtain the inner diameter of the annular pilot flame in the CH* slice equals to 6 mm, we can quantitatively assess the pilot gas outer diameter immediately downstream the matrix, which is 25.1 mm for a pilot gas flow rate of 18 slm (2 slm of CH₄ + 16 slm of O₂) and a sheath gas flow rate of 120 slm at 295 K temperature and atmospheric pressure. The computed diameter agrees well with the broadband flame chemiluminescence picture taken using the DSLR camera, Fig. 8(c) and can be used as the outer boundary of the pilot gas inlet in computational simulations of the SpraySyn burner starting at the burner surface. The signal emitted from the pilot flame within the slice at z = 0.5 mm can also be used for checking the integrity of the porous matrix since extensive use of the burner or its operation under non-standard conditions can deteriorate it [7]. The good spatial homogeneity and roundness of the present pilot gas signal indicates appropriate matrix structure of the burner mode used in this work. The instantaneous horizontal slices at z = 15 mm and 40 mm in Fig. 7 reveal cross-sections of detailed 3D structures that could not be directly inferred from the raw camera images, due to the line-of-sight integration nature of the measurements. The structures evolve from asymmetric dual-peak distributions towards symmetric bell-shaped distribution at downstream positions.

 

Fig. 8. (a), (b) Comparison of the full width at half maxima FWHM of CH* and Na temporally-averaged signals at z = 40 mm for various operating conditions computed from the spectrally-resolved measurements (spectrometer) with those computed from the TIMes reconstructions based on vertical planes (TIMes plane) and projections from the 3D data (TIMes proj.). (c) Typical projection image computed from the CH* reconstruction.

Download Full Size | PDF

TIMes measurements were validated by comparing the full width at half maximum of CH* and Na reconstructions with spectrally-resolved chemiluminescence measurements (setup described in section 2.3). Results at z = 40 mm for all investigated operating conditions are presented in Fig. 8(a,b). The height z = 40 mm was chosen because both signals coexist at this location with sufficient signal-to-noise ratios and their intensity distributions are relatively axisymmetric. The FWHM points from TIMes are averaged over 36 FWHMs obtained in two different ways. One way is to compute the FWHMs from vertical slices positioned at the burner centreline that are extracted from the CH* and Na averaged fields at different rotation angles about the z-axis (every 5° within a half-circle arc). The other way to compute the FWHM values is from projection images of the 3D reconstruction based on the optical parameters of the spectrally-resolved detection system (Fig. 8(c)) . The latter approach was introduced to mimic the line-of-sight integration effect through the collecting lens optics of the present spectrometer measurements. Error bars in the plots of Fig. 8 correspond to one standard deviation from the 36 FWHM values reflecting the spatial fluctuations due to the asymmetric time-averaged 3D field. The results show good agreement between the TIMes reconstructions and the spectrally-resolved measurements, and similar trends for different operating conditions can be seen. Differences between reconstructed cross-sectional data and line-of-sight integrated measurements are expected and they are found to be more pronounced for Na emission. This seems to be caused by the presence of Na signal over a wider location, enhancing the line-of-sight integration effect. Additionally, the asymmetries of the present spray flames enhance the differences between the measurements from those techniques.

Centreline profiles of temporally and spatially (circumferential direction) averaged reconstructed fields of CH* and Na signal using 15 and 14 cameras, respectively, are presented in Fig. 9(a) for the standard operating condition SF1. Profiles are normalised by their maximum. Additional CH* statistics from 60 instantaneous reconstructions are computed based on reconstructed signals acquired by all 29 cameras (all equipped with bandpass spectral filters for CH*), to provide data with the best reconstruction quality of our generic multi-camera setup. Error bars shown in the plot represents the computed uncertainties of the CH* field from 29 cameras, incorporating the flame fluctuations over time and space. These uncertainties are only plotted for CH* to avoid overcrowding the graph. The uncertainty in the averaged intensity ($\delta \bar{I}$) can be roughly estimated as [42]:

 

Fig. 9. (a) Centreline intensity along the centreline of the temporally and circumferentially averaged fields of CH* and Na, and (b) radial profiles of CH* and Na intensity fields for the standard SF1 operating condition with uncertainties. Intensities are normalised by the maximum of the corresponding 3D field.

Download Full Size | PDF

The CH* and Na centreline profiles in Fig. 9 show intensity values increasing up to a maximum before decreasing monotonically at downstream locations. The spray-flame height computed from the CH* signal is approximately 45 mm, which is smaller than the height observed from broadband chemiluminescence (Table 2). On average, the most intense reaction zone of the flame coincides with the highest intensity of CH* approximately at z = 17 mm, while the most intense signal of Na emission at z = 37 mm is related to a combination of high concentrations of burnt spray-flame products with a high-temperature region. Good agreement between the averaged centreline profiles of CH* chemiluminescence reconstructed from 15 and 29 camera measurements is observed indicating that an adequate resolution and number of cameras were employed in the present TIMes, as discussed in detail previously [31].

Corresponding radial profiles of CH* and Na averaged fields for selected heights above the burner are presented in Fig. 9(b). For both fields, the radial profiles evolve from an off-centred peak close to the burner surface, due to the liquid spray and cold dispersion gas, towards bell-shaped distributions for downstream locations. The radial profiles of CH* obtained from reconstructions using 15 and 29 cameras are similar to one another, except for minor differences at z = 10 mm. Close to the burner exit, the CH* averaged structure of the spray flame shows the smallest scales due to the atomisation region and strong asymmetry. Therefore, the structure reconstructed using 15 cameras was not as sharp and defined as that reconstructed using 29 cameras in that region. It is well known that a better resolution can be achieved by increasing the number of cameras [27,31], nevertheless the present TIMes technique required the other 14 cameras for simultaneous acquisition of Na emission.

Instantaneous and averaged 3D reconstructed fields of Na and CH* signals for all investigated operating conditions are presented in Fig. 10. Only half of the volume is plotted for better visualisation of overlapping regions. Isosurface values shown in the legend of the figure correspond to fractions of the maximum intensity (e.g., 0.05 CH* refers to 5% of the maximum intensity of this field). The instantaneous fields show elongated smooth CH* structures (largely due to motion blur from the higher exposure time in the imaging stage), while Na structures are generally wider and contain a variety of spatial scales. The instantaneous fields show higher noise levels (around 3% of the maximum intensity) than the averaged fields, as expected. The averaged fields are asymmetric, probably as a result of the single dispersion inlet pipe of the present burner and/or asymmetric inner nozzle. Since the burner is aligned based on the standard operating condition SF1 (Inj2_DG10), large deviations from the standard flow conditions can cause changes in the flame inclination and symmetry. For example, the Inj2_DG12 and Inj4_DG12 cases are remarkably more asymmetric than Inj2_DG10, where the dispersion gas flow rate of the former cases was 20% higher than the standard case. However, although the present spray flames are asymmetric near the burner exit, they become more symmetric further downstream as a result of the flame spray development. Doubling the liquid injection flow rate and keeping the flow rate of the dispersion gas close to the standard case (i.e., cases Inj4_DG9 and Inj4_DG10) shows negligible changes in the flame asymmetry.

 

Fig. 10. Instantaneous (t = t₁, top row) and time-averaged (bottom row) 3D reconstructed fields of Na and CH* signal intensities visualised by isosurfaces for the investigated operating conditions of the SpraySyn flame using TIMes. Isosurface values in the legend correspond to fractions of the maximum intensity of each field.

Download Full Size | PDF

The isosurfaces of averaged reconstructions for both signals in Fig. 10 show vertical prolate-ellipsoid-like shapes of the spray flames. The shapes from Na signal are generally bigger and offset to upstream locations compared to those of CH*. The latter characteristic is enhanced by the buoyancy effect, assisting the upward motion of the hot burnt spray products in a cooler surrounding, as discussed in [12]. Similar spray-flame morphology can be inferred from high-temperature zones in the time-averaged thermocouple measurements of [10] assuming symmetry about the central axis. The Na signal decreases in the downstream direction due to a combined effect of reduced temperature [10,15] and dilution of the burnt spray-flame products through mixing with entrained sheath gas as the jet flow spreads [12,18]. The spray-flame height from the averaged CH* fields varies from about 37 to 45 mm according to the operating condition. The heights computed from Na signal are longer, but not reported here because the signal expands further downstream of the investigated volume.

Overall, the spray flame becomes shorter and narrower with increasing dispersion gas flow rate. This can be explained using information from previous studies that utilised a dual phase-Doppler anemometer (PDA) [14,19] to show that the size distribution of atomised droplets in the two-fluid nozzle shifts towards smaller droplets when the ratio between the dispersion gas and liquid injection flow rate increases. This leads to a decrease in the time needed for the droplet full evaporation (according to the diameter-square law [20]). Regarding the weak CH* signal that is related to the pilot flame, it cannot be detected in the cases with liquid injection rate of 4 slm, due to the shorter exposure time used during image capture to compensate for the significantly brighter spray flame.

The spatial relationship between CH* and Na fields can be investigated using 3D local cross-correlation of the reconstructed signals. This was carried out by computing the local Pearson’s correlation coefficient r between the 3D CH* and Na intensity fields within a cubic region of 7×7×7 voxel as follows

3D fields of local correlation coefficient for the investigated operating conditions are plotted in Fig. 11 as half-sided isosurfaces of 0.5 with colour-coded caps. The instantaneous field was computed from the same instant presented in Fig. 10. The 3D local correlation coefficient from instantaneous fields shows complex 3D structures with a wide range of scales typical of turbulent reacting flows. The structure observed from the averaged field is smooth and shows low correlation values at the centreline close to the burner exit, due to the presence of the liquid spray and cold dispersion gas, and high values downstream, due to the coexistence of flame front and hot burnt spray-flame products. At these locations, the atomised droplets are more dispersed, due to droplet evaporation and jet spreading, enabling single-droplet combustion (i.e., most of the droplets surrounded by individual flames), as discussed in [19]. Close to the burner exit, a high coefficient value is observed in the form of a hollow structure probably as a result of external and internal group combustion at the surroundings of the liquid spray jet, where most of the droplets inside the jet core hardly evaporate and the droplet clouds at the spray boundaries burn [19]. No correlation is observed at the pilot region, as expected, because no NaCl seeding was used in that stream. Although, these plots give a general idea on how the signals change with respect to one another, it is important to note that r only considers the linear relationship between the two signals. Therefore, lower values do not necessary mean that the two fields are not locally correlated, but they could be correlated in a non-linear fashion.

 

Fig. 11. 3D local cross-correlation coefficient between the instantaneous (t = t₁) and time-averaged CH* and Na reconstructed signals under different operating conditions of the SpraySyn flame.

Download Full Size | PDF

Joint probability distribution functions (JPDFs) of CH* and Na intensity fields within regions are employed to better understand the spray-flame morphology. Normalised JPDF from instantaneous and averaged fields of CH* and Na at the standard operating condition SF1 for different height intervals are presented in Fig. 12. The CH* and Na intensity fields were normalised by the maximum intensities of the corresponding fields beforehand. The instantaneous case plotted in the first row of the figure corresponds to the data shown in Fig. 7. Prior to computing the JPDF, binary masks were applied to reduce the influence of the background as aforementioned. For visualisation purposes, the JPDF plots of each region are normalised to their maximum values for each height interval. Results very close to the burner exit and above 50 mm height are not plotted, because CH* signals were not detected together with Na signals in these regions. Due to the turbulent nature of the spray flame, JPDFs of instantaneous cases are very different from each other for the different instances in time. Generally, higher intensity values of CH* chemiluminescence coexist with lower values of Na emission, while higher values of Na coincide with lower ones of CH* (i.e., higher occurrences close to the horizontal and vertical axes, respectively). This is expected since the Na signal is detected at the burnt side, whereas the CH* signal is maximum at the flame front.

 

Fig. 12. Normalised JPDFs for the normalised CH* and Na signals from instantaneous and time-averaged reconstructions at different height intervals for the standard SpraySyn operating condition SF1.

Download Full Size | PDF

Regarding the averaged results, the JPDFs of CH* and Na intensity in Fig. 12 exhibit a different behaviour. The distribution near the burner exit (10 mm < z < 20 mm) is very spread out, but it evolves towards a curved distribution downstream. At the downstream locations, both fields are more correlated due to single-droplet combustion within an environment of mixed gases and diluted droplet number, as aforementioned. Since the Na signal is the most dominant downstream compared to CH*, peaks of the JPDFs are clustered at higher intensity values of Na and lower intensity values of CH*.

5. Conclusions

Tomographic imaging using multi-simultaneous measurements (TIMes) was successfully performed on the SpraySyn burner for different operating conditions to reconstruct concurrent instantaneous 3D fields of CH* chemiluminescence (an indicator of the flame front) and Na emission from vaporised rock salt that was dissolved in the liquid solution (related to the hot burnt spray-flame products). Preliminary reconstructions of the spray-flame morphology showed large asymmetry and 3D inclination, which was greatly reduced after improved alignment procedure using two frontal viewing directions. The TIMes reconstructions were experimentally validated against spectrally-resolved chemiluminescence measurements. Comparing the 15-camera reconstructions of CH* with 29-camera reconstructions (almost double the number of measurements expectedly provides better results) showed a good agreement, putting more confidence in the accuracy of the TIMes results.

Instantaneous and time-averaged reconstructed fields were analysed to characterise the 3D structures of the spray flames which are designed for nanoparticle synthesis within the SpraySyn burner. The reconstructed fields revealed structural detail with regions of mixing and interaction between the precursor-containing spray stream and flame front. CH* structures were generally more constricted along the centre axis, whereas Na structures were much larger. The averaged reconstructions showed that the investigated spray flames were always asymmetric near the burner exit where aerosolisation and droplet evaporation begins, and this might impact the quality of the final synthesised material. However, the flames became more symmetric at downstream locations. Isosurface visualisations were employed to characterise the 3D overlapping regions of CH* and Na, while 3D fields of local correlation coefficient and joint probability distributions were computed to study the spatial relationship between both signals. Horizontal slices closest to the burner exit, extracted from the 3D CH* data confirmed the integrity of the porous matrix used and could be used to determine the outer diameter and shape of the pilot flame, which provides an accurate inlet boundary dimension for future simulations of the SpraySyn burner starting downstream of the porous matrix.

Centreline and radial profiles based on temporally and spatially averaged fields of CH* and Na at the standard spray-flame condition revealed how the intensity fields evolve towards bell-shaped radial distributions at downstream locations. The computed heights and widths of the spray flame decreased as the dispersion gas flow rate increased, probably due to differences in the size distribution of atomised liquid solution droplets, which agrees with the results of previous studies. Furthermore, the data containing 3D reconstructed fields and corresponding profiles can be used to validate numerical simulations and to better understand the nanoparticle synthesis within the SpraySyn burner. However, comparison of data between numerical simulations and experiments is not trivial, as the former usually assumes perfect settings (i.e., perfectly aligned flame and perfect geometry of the burner), and the latter involves a flame that would be randomly titled to some degree. The burner was carefully aligned to provide reconstruction of the standard SpraySyn condition SF1, and uncertainties based on temporal and spatial fluctuations in the experiment were provided. This makes the 3D dataset easier to compare with simulations.