1. Introduction

Hyperspectral imaging (HSI) systems are instruments used to capture images, where at each pixel, a spectrum can be extracted [1,2]. These techniques find broad application across diverse fields of scientific investigation, including remote sensing, coastal ocean imaging, and biomedical imaging [3–5]. Taking inspiration from photoluminescence (PL) and Raman spectroscopy, researchers have adopted laser excitation and spectral collection, allowing the measurement of emission spectra across the target’s interest area. This approach facilitates the identification of various properties, such as the distribution of atomic impurities, defect distribution, and strain level analysis [6–8]. To accomplish these applications, precise micrometer-level spatial and sub-nm spectral resolutions are typically required for revealing the chemical and physical characteristics of a sample’s microstructure.

Balancing the design of HSI systems necessitates careful consideration of factors such as range and resolution, both spatial and spectral, as well as data collection and processing speeds, and the resulting file sizes. Given the ongoing development of HSI, various implementations aim to mitigate its limitations to meet the requirements for the specific spatial-spectral sensing application [1,9–17]. Some techniques prioritize non-scanning snapshot measurements [15,18], while others, like the one in this work, rely on sample translation [8,19,20]. To accomplish HSI, a dispersive or spectrally sensitive component is essential for measuring emission spectra, with diffractive elements often offering higher spectral resolution compared to filter-based solutions [21].

When extending HSI from two to three dimensions (2D to 3D), the challenge of balancing these competing parameters becomes even more pronounced. While 3D imaging of materials can be achieved using optical methods such as confocal, structured illumination, and various light sheet microscopy techniques [22–27], these techniques often measure large bands of emission or several discrete spectral bands simultaneously. Acquiring spectral information at each pixel in 3D imaging introduces additional optical complexity and can reduce the signal-to-noise ratio due to the segmentation of collected light into narrow spectral bands.

In contrast to some HSI systems – which use point-scanning (whisk broom), either raster scanned over the sample using scanning mirrors, or via sample translation – we have used line-scanning (push broom) for this instrument [6–8,19,28,29]. Line-scanning systems have some notable advantages over point-scanning. The acquisition of the pixels is parallelized along the length of the spectrometer entrance slit, which allows for either faster acquisition or higher signal collection vs a single point-scanned system using the same power density. The optical setup is simpler than a point-scanned system using scanning mirrors and has one axis which is limited only by the translation stage, whereas scanning mirrors have a single field-of-view (FOV) limited by the objective lens. The developed instrument is also semi-confocal perpendicular to the entrance slit of the microscope, which retains one of the important characteristics of some point-scanning instruments required for 3D imaging – optical sectioning.

Aberrations, such as smile and keystone, are also prevalent in many HSI techniques, causing spectral and spatial blurring in the final dataset [30,31], which is a concern with most spectrometer-based implementations of HSI. 3D HSI suffers from additional aberrations, particularly when sampling within high refractive index (RI) materials. The most significant of these is spherical aberration induced by the RI mismatch between the interface media and the sample [32–34]. This significantly reduces the achievable lateral and axial resolution, and this effect is exacerbated with increased imaging depth. Immersion objective lenses using higher RI media, such as water or oil, can help mitigate these aberrations, but they usually possess shorter working distances, limiting the imaging depth within the material. These objective lenses also typically feature higher numerical apertures (NA), amplifying the spherical aberration caused by RI mismatch. Consequently, immersion can in some cases be less beneficial than using a lower NA objective [35].

Common techniques used to image luminescent defect distributions in materials, such as diamond, include above-bandgap excitation imaging (e.g., the De Beers DiamondView instrument, $\lambda$ $\sim$225 nm) [36], epi-fluorescence imaging at various wavelengths [8], and electron-beam-excited cathodoluminescence (CL) microscopy [29,37,38]. While the latter can potentially provide spectroscopic information depending on the instrument’s capabilities, these methods are limited to visualizing defect distributions, often combining the emission of several luminescent features. Importantly, these techniques are typically restricted to surface analysis. In the case of epi-fluorescence, it may be possible to detect luminescence from greater depths using below-bandgap excitation; however, these setups usually lack optical sectioning capabilities. Point measurement-based PL systems can be extended to imaging or volumetric HSI by scanning across the entire area or volume of interest, but these point-scanning systems often suffer from lengthy acquisition times, limiting their practical applications.

This paper presents the development of a volumetric HSI system for PL and Raman scattering measurements of high refractive index (RI) materials. We have developed a line-scanning-based hyperspectral imaging microscope that captures luminescence emission spectra for diamond volumes up to 400 mm³ with micrometer-level spatial resolution and spectral coverage from visible to near-infrared regions (400-1100 nm) with up to 0.25 nm spectral resolution. The rapid sampling rate of up to 6 minutes per 545 million data points for this HSI system enables practical applications in volumetric hyperspectral imaging. The system exhibits low distortion and chromatic aberration across the multi-dimensional spatial-spectral dataset, facilitating the non-destructive assessment of defect distributions in diamond. We have applied this HSI to natural and lab-grown diamond samples, demonstrating its capability to non-destructively reveal impurity concentration distribution between mineral growth planes and microstructures within the bulk high RI materials.

2. Background

Minerals such as diamond and colored gemstones (e.g., corundum [sapphire and ruby], emerald, spinel, etc.) have famously been used for millennia in jewelry. Diamond is a wide-bandgap semiconductor ($\sim$5.5 eV), and can host a range of different point defects, many that are associated with color (termed color centers). The successful synthesis of diamond using either the chemical vapour deposition (CVD) or high pressure, high temperature (HPHT) methods has improved its availability and led to intensive research into tailoring its properties by defect engineering through controlled growth or treatment [39,40]. Most recently, quantum technology applications have been explored that exploit the electronic spin dynamics of color centers consisting of vacancies (V) and elements such as nitrogen (NV), silicon (SiV), germanium (GeV), and tin (SnV) [41–44].

Diamond and colored gemstones can possess fluorescent defects, with characteristic luminescence spectra that may be detected between the deep ultraviolet (UV) to near infrared (NIR), allowing their identification [29,39,40,45–49]. High spectral resolution is necessary for the positive identification of these features as their spacing can be as narrow as 0.2 nm. This is of importance to the jewelry industry, as the presence and absence of certain defects are the basis for identification criteria for mineral species and may extend to the determination of the growth method (natural vs laboratory-grown) and treatments [40,46,47]. The distributions of these features can be particularly valuable as they may relate back to changes in growth conditions (e.g., temperature, pressure, fluid or gaseous composition, internal morphology and crystallographic growth direction), that affect the defect incorporation, producing a unique map of a sample’s formation. This information is of interest to geologists, gemologists as well as scientists who aim to reproducibly create materials with homogeneity requirements, or samples with a patterned design (e.g., layers).

Surface HSI of photoluminescent defects (often referred to as PL mapping) has shown promise as a technique to analyze distributions of spectrally identified features [6,8,19,28]. 2D and 3D datasets are created by collecting emission spectra for each pixel in an image and can be used to extract defect maps by fitting the characteristic peaks of specific defects. Studies have shown this method being applied to both natural and laboratory-grown diamonds, as well as colored gemstones [6,8,19,28]. One of the key features of HSI, and exploited in this work, is that it can allow for the non-destructive analysis of defect distributions within a sample. Traditionally, the only way to access such information would be to cut and polish to reveal the interior section of interest, which is destructive.

3. Experimental details

3.1 System design

The configuration of the instrument is depicted in Fig. 1. The instrument uses a single excitation wavelength, 405 nm diode laser at 50 mW (IB-405-B, PicoQuant), where elliptical laser output is circularized by a pair of anamorphic prisms (PS875-A, Thorlabs), and is directed onto a Powell lens (30$^{\circ }$ for 3 mm beam, Laser Line Optics). A pair of achromatic doublet lenses (AC508-075-A-ML, and AC508-300-A-ML, Thorlabs) expands the beam to fill the aperture of the objective lens, whilst relaying the generated line. A band pass (BP) filter (FF01-400/12-25, Semrock), spectrally cleans the output from the 405 nm excitation laser, and the laser light is reflected into the objective lens by a dichroic mirror (Di03-R405-t3-25x36, Semrock) which is focused into the sample mounted to an X-Y-Z linear motor translation stage (AI-LM-3000XY, and AI-BSD-1500-Z, Alio). An extended apochromatic corrected (400–1000 nm) 10X 0.4 NA objective lens (UPLXAPO10X, Olympus) was selected due to its broad band chromatic correction, magnification, and NA. Emission collected by the objective lens transmits through the dichroic and a long pass (LP) filter (LP02-407RU-25, Semrock) to block the reflected and scattered excitation light, and a tube lens (SWTLU-C, Olympus) focuses the emission onto a motorized entrance slit of the spectrometer with a 22 mm entrance height.

 

Fig. 1. A schematic diagram of the developed instrument showing the key components. The 405 nm laser is circularized by a pair of anamorphic prisms, before a Powell lens (30$^{\circ }$ fan angle for 3 mm beam), a pair of achromatic lenses (Lens 1–75 mm, Lens 2–300 mm) and the objective lens (10X 0.4 NA) focus a line to the sample plane on an X-Y-Z stage. Luminescence is collected by the objective lens and passes through the dichroic mirror and long pass filter into a tube lens. This is focused onto the entrance slit (10 $\mathrm{\mu}$m) of an IsoPlane 320 imaging spectrograph, which disperses the line onto a 2D cooled CMOS camera.

Download Full Size | PDF

The developed instrument exploits the recent development of aberration corrected Schmidt-Czerny-Turner spectrometers (IsoPlane-320, Princeton Instruments), which uses a reflective Schmidt corrector plate to mitigate or eliminate the effect of many aberrations at the detection plane, which manifests as effects such as smile and keystone. This enables a much larger field of view (FOV) at the detection plane of the spectrometer and – coupled with a full frame small pixel (3.75 $\mathrm{\mu}$m) CMOS camera – large sample areas can be measured with high spectral resolution and range. A choice of diffraction gratings are available (150 lines/mm 500 nm blazed, 300 lines/mm 500 nm blazed, 2400 lines/mm holographic grating), and the spectrally dispersed emission is focused onto a Peltier cooled full frame CMOS camera (QHY600PRO SBFL, QHYCCD) with USB and fiber optic data transfer. The microscope frame is a custom openFrame upright microscope (openFrame, Cairn Research), with the benefit of multiple entrance and exit ports.

Spatially, the developed instrument can measure 2.2 mm in the X-axis, the Y-axis is limited by the stage travel at 30 mm, and the Z-axis range is limited by the objective working distance. In comparison to other line-scanning based HSI systems, this instrument can use the full height of the slit, as opposed to the typical 3 mm of an uncorrected Czerny-Turner system, which significantly increases the fixed FOV in the non-scanned axis. This is a >7x improvement over other hyperspectral line-scanning instruments, and allows for a significant speed increase for the same magnification, whilst preserving spatial resolution [8]. The lateral resolution of the instrument is $\sim$1–3 $\mathrm{\mu}$m, although this depends on sample geometry, imaging depth, and spatial sampling. The axial resolution has been estimated experimentally as $\sim$35 $\mathrm{\mu}$m (not presented here), to a depth of atleast 1.5 mm inside the sample, based on linear features known as striations in CVD-grown diamond [50]. We note that the estimation of axial resolution is challenging, and will decrease with increasing depth into the sample. It depends on depth, spatial sampling, and material’s index of refraction, and the axial resolution will be lower than its theoretical estimation due to the spherical aberration. Traditional methods of measuring axial resolution are unsuitable as the diffraction-limited object must be in high RI media, which for a solid is non-trivial. Therefore, we are limited to existing small objects within a sample, resulting in an upper limit for the resolution.

Spectral resolution and range are identical to any point-scanning technique while using the same spectrometer. When using the 150 lines/mm grating, the spectral range of this system is 711 nm, and the central spectral resolution of the system (using a 10 $\mathrm{\mu}$m entrance slit) is $\sim$0.25 nm. This spectral range and resolution are an improvement over most Czerny-Turner spectrometer systems due to the Schmidt correction plate, and reduction in off-axis light from the low f-number illumination onto the entrance slit. Due to data collection and storage challenges, the data acquisition parameters are carefully tuned to achieve reasonable file sizes and rate of data acquisition, with sufficient resolution. It should be noted that the absolute spectral sensitivity of the instrument has not been calibrated.

The RI mismatch-induced spherical aberration is limited in this instrument by using a comparatively low NA objective lens, a property that is typical of low magnification lenses. High chromatic correction is important for this application due to the extended wavelength range possible, and the aberrations will compound. In this design, the 0.4 NA is enough to provide sufficient optical sectioning for the semi-confocal measurement without inducing significant spherical aberration. We tested several different NA objective lenses in the instrument, and found that at an NA of 0.4, the aberrations only became notable below $\sim$2 mm at the voxel sizes we were interested in sampling at. These aberrations could be limited through techniques such as adaptive optics, however they are more challenging to implement in line scanning systems, and the benefits are smaller for low NA systems typical of low magnification, high FOV instruments that we were aiming to develop. Reducing the NA below 0.4 presented a detrimental effect to the axial resolution, which was deemed undesirable for the application. In line-scanning based HSI systems the smile and keystone can be noticeable, but the aberration correction in the dispersed path allows for these effects to be strongly mitigated. To the best of our knowledge, there are currently no commercial HSI instruments that exploit the aberration correction of a Schmidt-Czerny-Turner spectrometer.

3.2 Method

A 3D HSI scan consists of acquiring data at a single Y spatial position, where the camera collects a 16-bit image containing the X spatial data from along the emission line, and spectral data on the other camera axis. The sample is then translated by the pixel size of the camera (native 3.76 $\mathrm{\mu}$m), which is typically binned, multiplied by the magnification of the objective lens. This ensures square pixels for image reconstruction. Acquisition continues for the full length of the sample, generating a 3D data cube of X-Y-$\lambda$. The 150 l/mm grating is used in all presented examples. The sample is then translated by Z, and the process repeats until a full X-Y-Z-$\lambda$ data hypercube is collected. This Z-step size is set based on real space and then is converted to sample space by multiplying by the RI of the material, and all values given in this work are internal sample space sizes. Typical camera integration times are 0.2 s per image when using fiber transfer, but this depends on sample geometry, signal strength, camera binning, and grating. The entire instrument is controlled by a custom LabVIEW interface. Image processing is conducted using ImageJ (FIJI), ICY, and a custom MATLAB script. Data sets are either reduced to single wavelength/defect greyscale X-Y-Z images or false color X-Y-Z images where the RGB color space represent different detected wavelengths/emitting species, and can be viewed either as 2D image stacks, GIFs or movie files. Full 3D manipulation of the data was achieved using ICY.

To restrict the analysis to only that originating from a specified defect, a custom written MATLAB script is used to find the mean value of two points at the extreme ends of a section of the emission spectrum, typically its zero-phonon line (ZPL), and then calculate the area of a trapezium. The entire spectrum is then summed, and the two calculated values are subtracted, which results in a value for the area under the curve with a linear background subtracted. Through careful selection of the emission window, specific overlapping features can be avoided, and an example is shown in Figure S1, where a background spectrum has been removed. A threshold value is used to eliminate signal from saturated regions, or sample contamination (e.g. dust) which may be brightly emitting under the selected laser excitation. This process is conducted to generate every image in the stack.

Two diamond samples were selected to demonstrate the instrument’s capabilities analyzing the interior of high RI materials. All measurements were conducted at room temperature, though low temperature data could be collected using a cryostat. Sample details can be found in the supplementary information.

4. Results and discussion

4.1 Sample 1 - natural diamond

Sample 1 is a tetrahexahedroid natural diamond that has been partially cut and polished into a 0.09 ct cuboid shape (1.9 $\times$ 1.7 $\times$ 1.4 mm), leaving small portions of natural resorbed {110} rough faces intact. Two separate 3D data stacks that traversed the full sample depth along one axis were collected, probing from two opposing sides; here we present one example. The shown stack imaged a 2.17 $\times$ 2.00 $\times$ 1.20 mm volume with a 1.5 $\times$ 1.5 $\times$ 12.0 $\mathrm{\mu}$m voxel size (2 $\times$ 2 pixel binning). Spectra were collected over 411 – 759 nm, achieving a$\sim$0.7 nm spectral resolution. Every other data slice was removed to achieve square aspect pixels. This measurement took 23 hours to complete, whilst a comparative scan using a commercially available instrument would have taken $\sim$10 days to complete. Videos showing the full 3D distribution of the defects can be viewed online in Visualization 1.

Figure. 2(a) are false color 2D XY, YZ, and XZ slices from within the sample, where blue is associated with emission from the N3 ($\text {N3V}^\text {0}$) defect (ZPL = 415 nm), and green with a broad band emission centered on 625 nm with unknown origin. A truncated growth pattern consisting of concentric rectangles is observed, consistent with octahedral growth. The layers are growth horizons, with the changing defect concentrations indicating variations in the growth environment or diamond-forming fluid composition during growth. The YZ cross section also shows evidence of the diamond’s octahedral growth [38]. Growth structures in the XZ image are more subtle, though a green growth zone can be noted running at an angle through the sample. Figure 2(b) shows a slice through the full 3D data set, showing internal growth structure. An approximate <111> plane is shown in the supplementary Fig. S2 (c), showing how this technology can allow for specific growth planes to be identified. Averaged spectra from two regions of interest (ROI) in the sample are shown in Fig. 2(c), where the slight differences in the intensity of the unidentified broad band around 625 nm are shown. These subtle spectral differences are responsible for most of the spatial emission variation that reveals the diamond growth structure in the false color images. Minor artifacts relating to the probing direction resulted in shadows along the map edges with lower spatial information, and vertical intensity related features due to the non-uniform excitation source provided by the Powell lens.

 

Fig. 2. False color images of natural diamond sample 1 were acquired using the 10X 0.4 NA objective lens using 405 nm excitation, where blue represents the N3 defect, and green an unknown broad emission band centered at 625 nm. The scale bar is 0.33 mm and the images in (a) are XY, YZ, and XZ images from a 3D data stack from within the sample. Image (b) shows an arbitrary plane inside the sample. The graph in (c) shows two different ROIs; ROI 1 and ROI 2 are regions that are represented as primarily blue or green in the maps, respectively. N3 center fluorescence at room temperature consists of a ZPL at 415 nm, along with broad vibronic structure that extends till 600 nm, R denotes the diamond Raman peak, and the unknown broad band centered at 625 nm is shown in the inset.

Download Full Size | PDF

4.2 Sample 2 - CVD-grown diamond

Synthesis details for Sample 2 are included in the supplementary materials. This emerald cut sample has been fashioned so that the macroscopic <100> growth direction lies approximately perpendicular to the table facet. The total imaging volume was 2.00 $\times$ 4.85 $\times$ 3.36 mm, probing through the table facet (Z-axis $\sim$parallel to growth direction). A 3.0 $\times$ 3.0 $\times$ 48.4 $\mathrm{\mu}$m voxel size (4 x 4 pixel binning) was used, which is comparatively coarse to reduce the overall data collection time and storage requirements. The axial resolution is much worse than the lateral resolution, and becomes worse with depth, so better sampling laterally was chosen due to the potential structures of interest being more apparent laterally. The spectral range was set to 425 – 746 nm, with a resolution of $\sim$ 0.7 nm. An integration time of 0.75 s was used per slice, leading to a 24-hour total acquisition time.

Figure. 3(a) represents the total summed luminescence of the top XY plane parallel to the sample’s table facet, approximating the growth structure near the end of synthesis. The emission from the sample was complex, with different emitters becoming more prevalent at different depths. Luminescence was detected from defects commonly observed in CVD-grown diamonds such as the structurally unidentified defect with a ZPL at 468 nm, H3 ($\text {N}_{2}\text {V}^{0}$, ZPL = 503.2 nm), the neutral nitrogen-vacancy defect ($\text {NV}^\text {0}$, ZPL = 575 nm), and the negatively charged silicon-vacancy defect ($\text {SiV}^\text {-}$, 737 nm). Raman signals from $\text {sp}^\text {3}$- (diamond) and $\text {sp}^\text {2}$- (non-diamond) bonded carbon was also detected at 428 and $\sim$432 nm, respectively, along with a broad luminescence feature attributed to $\text {sp}^\text {2}$ carbon.

 

Fig. 3. The images shown in (a) – (d) are XY slices at the table of the CVD-grown diamond Sample 2 extracted from a 3D data stack (scale bar = 1 mm), demonstrating different data processing conditions and what they can reveal regarding defect distributions. Image (a) shows the summed total intensity of the emission. Image (b) shows the emission from a 4.5 nm wide band centered on the $\text {sp}^\text {2}$-bonded carbon Raman feature. (c) is a false color image taken for three emission bands, with blue = 468 nm defect, green =$\text {NV}^\text {0}$, and red = $\text {SiV}^\text {-}$, and has spectral overlap. (d) is a false color image using the same color scale as (c) but contains light only from the associated defect. (e) is a YZ slice. The bulk sample growth direction was along the Z-axis, going from right to left in (e). Unlike (a) – (d), the pixels in (e) are not square; the full image is 3.36 mm in Z (horizontal dimension) and 4.85 mm in Y. (e) was analyzed in the same way as (d).

Download Full Size | PDF

 

Fig. 4. All images show the same false color representation where green is the 468 nm defect, and red is the $\text {SiV}^\text {-}$ defect, and the scale bar is 1 mm in (a). Images in (a) are XY, XZ, and YZ segments of the 3D data set collected with the 10$\times$ 0.4 NA objective lens under 405 nm excitation, where the red arrow shows the location of the XY plane in the YZ and XZ images. The graph in (b) shows the averaged emission spectra from two locations in image (c). The image in (c) is a higher resolution XY image within the diamond present in the 3D stack from (a), shown by the blue arrows in the XZ and YZ images, where etching pits can be seen. The scale bar in (c) is 0.25 mm.

Download Full Size | PDF

Many of the luminescence bands were broadened and overlapped as the measurements were taken at room temperature. This highlights the advantages of this technique, as the broad spectral range measured allows for the conclusive detection of these defects. High spatial resolution and sampling was necessary as some of the structures were < 20 $\mathrm{\mu}$m in size, providing detailed information about the distribution of these defects as the growth conditions were varied. Different data analysis methods were tested to expose the defect distributions, with selected examples shown in Fig. 3(a) – (e). Videos of the full 3D data stack are available online in Visualization 2, Visualization 3, Visualization 4.

The bulk of this emission originates from $\text {sp}^\text {2}$-bonded carbon fluorescence outlining tabular blocky structures, with some very bright spots originating from sample contamination (e.g., dust). These tabular structures may be flat hillocks that are converging together [51]. Figure. 3(b) shows the summed intensity from the $\text {sp}^\text {2}$-bonded carbon Raman feature, which is weakly shown in ROI 1 at 431 nm in supplementary Fig. S3. Although the concentration of $\text {sp}^\text {2}$-bonded carbon is higher at the interfaces of the tabular structures, Fig. 3(b) reveals an underlying irregular network that extends further. This coincides with regions that have dislocation bundles as observed using above band-gap energy fluorescence imaging (shown in supplementary Fig. S4), i.e., the crystallinity of the diamond was compromised, leading to $\text {sp}^\text {2}$-bonded carbon. Several spectral bands are used to give an impression of the spatial variation for the different emitting species in Fig. 3(c), and these are bands centered at the 468 nm defect ZPL (2.5 nm wide band, blue), which includes $\text {sp}^\text {2}$ luminescence, $\text {NV}^\text {0}$ defect centered on the ZPL (7 nm wide, green), which includes emission from 468 nm defect and $\text {sp}^\text {2}$, and $\text {SiV}^\text {-}$ defect centered on the ZPL (11.5 nm wide, red) which includes some $\text {NV}^\text {0}$ emission.

To mitigate the effect of overlapping spectral bands, images were generated considering only the areas of the ZPLs for the 468 nm (includes the first 3 phonon replicas), $\text {NV}^\text {0}$, and $\text {SiV}^\text {-}$ defects using the MATLAB script described in Section 3. Blue, green and red colors were assigned to these defects, respectively, with the brightness scaling with the peak area of the ZPL through the complete data volume. Figure 3(d) shows an XY slice of the top layer of the 3D stack, shown by the red arrows, while Fig. 3(e) shows the depth profile of a section of the stack (YZ slice). Due to the significantly lower intensity of $\text {NV}^\text {0}$ in the top layer of the sample relative to the substrate, the assigned green color is not visible in (d). Broadly, the 468 nm and $\text {SiV}^\text {-}$ defects follow the same spatial distribution, dropping in intensity in the regions where $\text {sp}^\text {2}$-bonded carbon is detected in Fig. 3(b). The YZ slice in Fig. 3(e) shows that the substrate used for growth was dominated by emission from $\text {NV}^\text {0}$ centers. Growth, that here traversed from right to left, was layered, illustrating variations in growth conditions as well as the two interruptions described in the supplementary materials.

It is interesting to compare this to the behavior of the 468 nm and $\text {SiV}^\text {-}$ defects for the XY slice shown in Fig. 4(d). Figure 4(d) demonstrates two different growth regions – a smooth region consistent with step-flow growth (top), and another one dominated by the converging hillocks surrounded by $\text {sp}^\text {2}$-bonded carbon (bottom). The former smooth region shows that the 468 nm and $\text {SiV}^\text {-}$ defects are similarly distributed, but it was not possible to compare their emission in the latter region due to the interference of the bright emission from the $\text {sp}^\text {2}$-bonded carbon. This illustrates the effect of the microscopic growth morphology on the defect distributions.

Additional data was collected using a higher spatial sampling to further investigate the finely spaced layers within the sample revealed by the YZ slice in Fig. 3(e). A 2.10 $\times$ 2.10 $\times$ 0.67 mm region of the sample was measured at a starting depth of $\sim$1 mm from the table facet, with a voxel size of 1.50 $\times$ 1.50 $\times$ 19.36 $\mathrm{\mu}$m, using 4 $\times$ 4 pixel binning. A video of the full 3D data set can be accessed online in Visualization 5. Figure 4(a) shows three different perspectives of the 3D data set, XY, XZ, YZ, revealing $\sim$ 5 layers in the XZ and YZ images. Green and red colors were assigned to ZPLs of the 468 nm and $\text {SiV}^\text {-}$ defects, respectively. Representative spectra are plotted in Fig. 4(b). For the XY slice (representing a growth horizon), the $\text {SiV}^\text {-}$ defect possesses significant spatial variation, yet the 468 nm defect appears to be homogeneously distributed. $\text {SiV}^\text {-}$ is known to preferentially incorporate into specific growth sectors, and monochromatic images for each contribution are included in the supplementary Fig. S6. This behavior contrasts with that observed in Fig. 4(d), where both defects are more uniformly distributed with contrast only provided by absorptive features.

Figure 4(c) shows a higher resolution XY slice within the stack illustrated in Fig. 4(a) from the blue arrows, with 0.75 $\mathrm{\mu}$m pixels using 2 $\times$ 2 binning, and is a digitally magnified region of (supplementary Fig. S5) which covers a 2.1 $\times$ 2.1 mm, revealing several square-shaped structures. The same color assignments to defects were maintained. The position of this slice coincides with the interface between two growth layers, where the growth had been intentionally interrupted (described in the supplementary materials). The square-shaped structures are approximately 100-200 $\mathrm{\mu}$m in size, resemble the pyramidal or flat hillocks that have been reported to form during epitaxial CVD diamond growth on <100>-oriented substrates (as used for Sample 2) [51,52]. These 3D growth structures have four-fold symmetry with <111> lateral facets, which may have macro-steps running along the <110> directions and can range in size from $\sim$10 – 500 $\mathrm{\mu}$m, depending on growth conditions and duration. These structures will develop into either a flat or pyramidal hillock [52].

Since analysis was conducted at a depth of 1 mm, the axial resolution is estimated to be $\sim$ 100 $\mathrm{\mu}$m and it is impossible to clearly distinguish whether the hillocks are flat or pyramidal. The hillocks appear to extend $\leq$ 100 $\mathrm{\mu}$m along the bulk growth direction (Z axis) and were subsequently overgrown. The impurity and defect uptake differs for the different crystallographic growth directions of the hillock faces and edges, resulting in the luminescence center distribution observed in Fig. 4(c). The pyramid hillocks’ peaks are truncated and topped by unepitaxial diamond crystallites, whereas the flat hillocks are dominated by <100> square facets. Both hillock types are thought to originate from threading dislocations that form at the interface between the substrate and new growth layer, created at inclusions at the substrate surface. These inclusions act as localized nucleation centers which enhance the local deposition rate, resulting in the formation of <111> micro-facets.

The ability to non-destructively measure the growth microstructure at the surface of the diamond as well as within the bulk is one of the key advantages of the 3D HSI technique. Alternatively, the sample would have had to be analyzed between growth steps. Cutting and polishing down to the intermediate interface after completing the multi-layered sample synthesis would likely destroy the evidence of growth hillock formation. Consequently, this is a powerful tool to understand how growth progressed during synthesis and draw conclusions relating to the effect of growth condition variations.

5. Conclusions and future outlook

Luminescence from different depths of natural and CVD-grown diamonds have been successfully measured using a newly developed high-speed line-scanning HSI system for 3D imaging of high RI materials, capable of high resolution and range, both spectral and spatial, under conditions that are not practical with other commonly used imaging techniques. Analysis of collected HSI data has non-destructively provided information regarding the distribution of defects and internal growth morphology of a natural diamond, and the spatially varying incorporation of several defects in a multi-layered CVD-grown diamond.

The flexibility of the instrument design means that further improvements could readily be made to tailor it to potential applications. The acquisition speed could be increased by moving to high powered multimode diode laser sources, and a camera with a less restrictive frame rate would also increase the minimum acquisition time. Selecting a coarser spectral and spatial sampling could increase the signal to noise and limit data storage requirements. Through laser and grating optimization, this instrument could also be employed for high resolution 3D Raman HSI with applications in inclusion identification and strain mapping. In future we anticipate adding a galvanometer scanning mirror to replace the Powell lens would provide a more uniform line illumination.

The developed 3D HSI instrument is a powerful analytical tool for material scientists, geologists and gemologists alike, enabling a detailed interpretation of the growth progression for transparent high RI materials. This could also be applied to the detection of artificial treatments, which can modify the distribution and relative concentrations of defects in potentially recognizable manners. The non-destructive nature of this technique is especially attractive when analyzing high value materials such as gemstones. Further studies with this system will demonstrate its full potential.