1 Introduction

Raman scattering is a valuable spectroscopic approach to gaining a noninvasive, label-free insight into the chemical composition of a sample. The Raman scattering spectrum contains a molecular fingerprint, because it results from vibrational resonances of the molecules under investigation. Although commonly applied spontaneous Raman scattering [1] gives access to the full spectral Raman response, this technique suffers from a low Raman scattering cross section, resulting in long acquisition times in the order of hundreds of milliseconds per spectrum, rendering the observation of dynamic spectral changes difficult to achieve [2]. With the development of new laser sources, coherent Raman scattering (CRS) techniques, like femtosecond stimulated Raman scattering (FSRS) [3–5], gained a lot of attention during the last decade. In CRS, at least two synchronized laser pulses cause a stimulated excitation of one or more Raman transitions, such that the efficiency of the signal generation is coherently enhanced. In the case of FSRS, a broad-bandwidth ultrashort (femtosecond) pulse acts as the probe field, in order to excite the full Raman scattering spectrum at once, such that FSRS is now an established technique for ultrafast spectroscopic investigations [6, 7]. The femtosecond probe pulse duration allows for ultrafast pump–probe experiments, like the time-resolved vibrational studies of, e.g., classical chemical reactions and excited-state reaction dynamics [8, 9], or structural dynamics of molecules [10, 11]. In FSRS, the acquisition time of the signal is limited by the read-out time of the detector of the spectrometer, which is in the order of milliseconds. In this temporal region, the noise in the Raman signals is dominated by laser power fluctuations (including pulse-to-pulse fluctuations and shot noise) and the read-out noise of the detector, while thermal noise is negligible. In order to acquire a sufficient signal-to-noise (S/N) ratio, the FSRS spectra need to be averaged over multiple shots, resulting in acquisition times of up to minutes [9]. FSRS has also been applied in proof-of-principle experiments to hyperspectral imaging [12, 13], where the Raman scattering spectrum from a raster-scanned sample serves as a contrast for microscopic imaging. With a pixel dwell time in the order of hundreds of milliseconds per pixel, the acquisition of a single image amounts to minutes, whereby the integration or averaging time of the spectra is dominating. FSRS-based imaging would thus greatly benefit from a faster scheme, especially if non-static samples are observed.

The dispersive Fourier transformation (DFT) [14] is a method to measure the spectrum of a single pulse. DFT overcomes the lower limit of the acquisition time of common spectrometers by temporally dispersing the signal by means of a huge amount of chromatic dispersion, such that the individual frequency components can be resolved in real time by a single fast photodetector. The DFT technique is extensively implemented in experiments capturing rare events [15], measuring noise [16] or dynamical processes [17]. In another approach, called serial time-encoded amplified microscopy (STEAM) [18], the wavelength-to-time transformation via chromatic dispersion has been applied to high-speed imaging. Herein, the spatial information was mapped to individual spectral components, which were recorded via a variant of DFT, called amplified DFT, meaning an active DFT. In 2008, amplified DFT has also been successfully applied to the observation of rapidly changing stimulated Raman scattering signals [19]. Note, that in this case, the generated Raman scattering signals needed to be amplified by a fiber-based Raman amplifier, which severely limits the detectable optical bandwidth, necessary for yielding the full Raman spectrum. Also, spontaneous Raman scattering signals were generated by a picosecond pulse and detected by using the wavelength-to-time transformation in a single-mode fiber and a time-correlated photon counting system [20]. However, it seems that none of the Raman scattering-related work was continued, as we have found no further studies within the Raman scattering community so far.

In this work, DFT will be used for the acquisition of FSRS signals in order to overcome the limitations in acquisition speed of commonly applied spectrometers in FSRS. Consequently, this broad-bandwidth CRS technique is expected to be as fast as state-of-the-art narrow-bandwidth photodetector-based CRS techniques (like coherent anti-Stokes Raman scattering [21] or stimulated Raman scattering [22]), but additionally does yield the full spectral Raman response on a similar timescale. Combined with a fast raster-scanning system, DFT-FSRS would allow the acquisition of hyperspectral datasets, i.e., a full Raman scattering spectrum per pixel, with pixel dwell times in the order of microseconds, which is, as discussed before, orders of magnitudes faster, than it was feasible before. In this way, the spectroscopic contrast is not limited to one specific line within the spectrum, but all inherent spectral information is acquired at once, in order to ensure a high chemical selectivity.

2 Experiment

The experimental setup for FSRS was based on our previously reported FSRS related work [23] and is schematically shown in Fig. 1. The narrow-bandwidth pump pulse and the broad-bandwidth femtosecond probe pulse for the coherent excitation of Raman scattering were provided by a home-made master oscillator power-amplifier system, operating at a repetition frequency of 1 MHz, and a subsequent supercontinuum generation. The supercontinuum probe pulse contained about 2 nJ of pulse energy (2 mW average power), covering a spectral range from 750 to 950 nm. The dispersion of the optical components in the beam path of the probe pulse needed to be compensated for by a prism-based pulse compressor. Together with the pump pulse with up to 40 nJ pulse energy (40 mW average power) at a central wavelength of 1035 nm, the probe pulse covered the relevant spectral region ranging from 700 to 3600 cm\(^{-1}\). As FSRS induces an alteration of the broad-bandwidth femtosecond probe pulse in the form of a stimulated Raman loss at covered Raman resonances, this FSRS signal was acquired as the relative change between two probe pulse spectra, the first measured as a reference and the second with the pump pulse turned on. An acousto-optical modulator was used in order to perform the rapid switching of the pump pulses. Its frequency was locked to half the repetition frequency of the laser system, i.e., 500 kHz, such that every second pump pulse was directed onto the sample. Note that the spectral resolution of FSRS is given by the spectral width of the pump pulse. In these proof-of-principle measurements, the pump pulse exhibited a spectral width of 2.1 THz, equal to 70 cm\(^{-1}\), which was suited to resolve the coarsely spaced lines in the presented spectra. Typical Raman linewidths are well below this value, such that future experiments need to be conducted with more appropriate pump pulse parameters. The pump and probe pulses were spatially and temporally overlapped by a dichroic mirror (DM) and focused into the sample with a microscope objective (MO, \(20{\times }\) magnification). A second MO was used to collect the transmitted probe pulse, which had to be analyzed spectrally.

The subsequent DFT setup (Fig. 1) consisted of three parts: firstly, the femtosecond probe pulse needed to be pre-chirped temporally with a 10-cm-long piece of highly dispersive H-ZF13-glass, in order to avoid nonlinearities occurring in the following fiber-stretcher part. Otherwise, the long propagation distance inside this fiber, combined with a high peak power of the probe pulse, would have resulted in an alteration of the spectral composition of the probe pulse due to self-phase modulation. The major amount of chromatic dispersion, necessary for the DFT, was introduced by a single-mode fiber (SMF, supplier: iXFiber, type: IXF-SMF-780-4-125) with a length of 1.75 km, which stretched the probe pulse to a duration of about 45 ns and consequently accomplished the wavelength-to-time transformation. Lastly, the temporal signal and thus the FSRS signal were measured with a fast photodetector (PD, Femto Messtechnik GmbH, model: HSA-X-S-1G4-SI-FC) and recorded with a sampling oscilloscope (Teledyne LeCroy, model: WavePro 725Zi-A 2.5 GHz). Although the spectral signal is read out with a single detector, as, e.g., in Fourier transform infrared spectroscopy, the spectral intensity is not measured by evaluating the coherence, i.e., the interference pattern of the light in the time domain, but the dispersed spectral components of the probe pulse are sampled directly by the oscilloscope.

Fig. 1
figure 1

Schematic experimental DFT-FSRS setup (for details, see text). DM dichroic mirror, MO microscope objective, F spectral filter, FM flip mirror, Spec CCD spectrometer, GB H-ZF13 glass block, SMF single-mode fiber, PD photodetector

Full size image

As a first step, the DFT system needed to be calibrated, regarding the wavelength-dependent temporal response. The probe pulse was spectrally filtered to narrow-bandwidth pulses at various central wavelengths between 740 and 970 nm. The temporal read-out of the photodetector was triggered to the laser output, so that relative temporal shifts, resulting from the wavelength-dependent group velocity of the dispersive stretcher (glass block and fiber), became apparent. The timing jitter of the oscilloscope was specified to be below 1 ps, which was two orders of magnitude below the resolvable temporal features of the impulse response of the photodetector (compare with Fig. 2b). The measured temporal position as a function of central wavelength is shown in Fig. 2a as red dots. The error bars of this measurement are not depicted, as the uncertainty of the evaluation of the temporal (\({\le }{\pm }100\) ps) and spectral positions (\({\le }{\pm }1.5\) nm) was negligible and would not be visible within the plot. The measurement matches the theoretical blue curve, which was calculated assuming a fiber length of 1.75 km and a second- and third-order dispersion, determined in Ref. [16], of \(4.030\times 10^{-2}\,\hbox {ps}^2{/}\hbox {m}\) and \(2.344\times 10^{-5}\,\hbox {ps}^3{/}\hbox {m}\), respectively.

Fig. 2
figure 2

Calibration measurement: a wavelength-dependent temporal delay (red dots) and the corresponding theoretical curve (blue line), b temporal response of the photodetector on an ultrashort pulse

Full size image

In order to estimate the spectral resolution of the DFT setup, the impulse response of the photodetector was acquired with the oscilloscope, by measuring the temporal response on a femtosecond pulse. As shown in Fig. 2b, the full-width at half-maximum (FWHM) response was 370 ps, which was in accordance with the specifications. Taking the wavelength calibration into account, this resulted in a spectral resolution of 3.78 nm at a wavelength of 750 and 8.34 nm at a wavelength of 950 nm, corresponding to 67 and 92 cm\(^{-1}\), respectively. As already discussed, this performance was just sufficient to resolve the coarsely spaced signatures in the presented spectra. Accepting a loss of signal strength, the spectral resolution could be improved by simply introducing an even higher amount of dispersion onto the probe pulse, or by utilizing a photodetector and oscilloscope, both with a larger measurement bandwidth. The application of, e.g., commercially available detectors and oscilloscopes with bandwidths beyond 20 GHz would easily result in an increased spectral resolution by one order of magnitude. The sensitivity of the DFT system was limited by the intrinsic losses, in both the glass block (Fresnel reflections at the uncoated surfaces) and the fiber (ranging from 1.3 dB/km in the long wavelength region around 950 nm, to 4.3 dB/km in the short wavelength region around 750 nm), together with coupling losses. The application of aberration and chromatically corrected objectives, not only for collecting the light coming from the sample, but also for coupling it into the single-mode fiber, would allow for an overall higher detection efficiency. In the present case, the overall transmission amounted to below 6%, which was still sufficient to make use of the linear dynamic range of the photodetector over the spectral region of interest, ranging from 750 to 950 nm. The oscilloscope reading of the photodetector exhibited a read-out noise amplitude of around 1 mV and a nonlinear response above 1.5 V, which saturated at about 2 V. The thermal noise contribution was negligible on the nanosecond timescale. As will be seen later, the dominant noise contribution was related to laser power fluctuations, i.e., shot noise, as the Raman loss signal is imprinted on a huge background and relative measurements between two successive pulses have to be performed.

Fig. 3
figure 3

Probe pulse spectrum measured with a DFT, b a CCD-based spectrometer, each time as a reference (pump off, blue line) and with the pump pulse turned on (red line). c Relative pump-induced amplitude changes (FSRS spectra) for a single-shot (red line) and a 20-time-averaged trace (orange-colored line), acquired both with DFT, or with a CCD-based spectrometer (green line). d Corresponding spontaneous Raman scattering spectrum. The shown measurements were performed on a sample of Cinnamomum cassia, significant Raman peaks are marked by vertical lines, and the total acquisition time is given for each method. a DFT single shot (\(1\,\upmu \hbox {s}\) per spectrum), b CCD spectrometer (1 ms per spectrum)

Full size image

The capabilities of the DFT technique allow a single-shot retrieval of the spectral composition, which is shown in Fig. 3a for the transmitted probe pulse, depicted as the blue line. Here, the temporal trace with a duration of about 45 ns was mapped on the wavelength axis, using the calibration from Fig. 2. The red line shows the probe pulse spectrum, after its interaction with a pump pulse of 40 nJ pulse energy in the sample, which was Cinnamomum cassia. This liquid sample shows a very strong and thus readily visible Raman line at 1625 cm\(^{-1}\). The relative changes between the probe pulse spectra, recorded with the pump turned off and on (i.e., the ratio of the measured spectral intensities of the blue and red lines), represent the FSRS spectrum. As only relative intensity changes are relevant, the wavelength-dependent sensitivity of the detection system does not need to be accounted for. The FSRS spectrum of single-shot measurements (1-\(\upmu\)s acquisition time per spectrum) is shown in Fig. 3c as a red line. The S/N ratio, here and in the following, was determined as the ratio between the maximum signal amplitude at 1625 cm\(^{-1}\) and the root-mean-square (rms) deviation in the silent region around 2500 cm\(^{-1}\), where no Raman signal was apparent. The S/N ratio of 8 just allowed for the Raman signal of the dominant peak to be visible. The rms noise amplitude of the relative Raman loss spectra of 2.5% in this measurement corresponded to an absolute amplitude of about 8 mV and is an order of magnitude higher than the read-out noise of the detection system and thus indicated the dominance of laser power fluctuations in the form of shot noise.

As a reference, the spontaneous Raman scattering spectrum was measured and is shown in Fig. 3d as the blue line. This measurement showed a high-contrast and spectrally well-resolved Raman lines, not only at 1625 cm\(^{-1}\), but also at 999, 1123, 1252 and 3061 cm\(^{-1}\) (marked by vertical dotted lines). These spectral lines also became apparent, as the DFT-FSRS spectra were averaged 20 times, as shown in Fig. 3c as the orange-colored line. Here, the S/N ratio increased to 39, such that also the smaller spectral Raman features of Cinnamomum cassia became clearly apparent. The spectral linewidth of the 1625 cm\(^{-1}\) feature measured 94 cm\(^{-1}\) and is in accordance with the estimated spectral resolution of the DFT system discussed above.

As a comparison to DFT, the probe pulse spectrum after the second MO was also measured with a standard CCD-based spectrometer (OceanOptics, model: HR2000+), by inserting the flip mirror (FM, Fig. 1). The spectrum is shown in Fig. 3b as the blue line. The integration time was set to the minimal value of the applied spectrometer of 1 ms and an additional neutral density filter (OD3) prevented a saturation of the CCD. The red line in Fig. 3b shows the probe pulse spectrum, measured with the pump pulse turned on. The pump-induced changes in the probe pulse spectrum due to FSRS are depicted as a green line in Fig. 3c. With a S/N ratio of 30, the CCD-based measurement compared well to the 20-time-averaged DFT measurement, but in contrast, was averaged over 1000 pulses per spectrum, as the repetition frequency of the applied laser system was 1 MHz. As discussed above, the spectral resolution of the FSRS measurement was not limited by the spectral resolution of the CCD-based spectrometer, but was broadened by the spectral width of the pump pulse, resulting in the present case to a measured width of the Raman lines of about 80 cm\(^{-1}\). An improvement is feasible by implementing narrow-bandwidth picosecond pump pulses. Admittedly, this CCD measurement was performed with a drastically reduced photon number, such that a direct comparison of the S/N ratios between the CCD- and DFT-based measurements falls short. Nonetheless, the comparison of both measurement approaches still serves its purpose, as both depicted FSRS signals are generated under identical excitation conditions (i.e., pulse parameters and focusing conditions) and measured in those two different ways. For instance, FSRS signals measured by DFT would show asymmetrically distorted spectral features if the applied fast photodetector would contribute an asymmetric temporal response function. But, as the CCD-based FSRS spectra agree well to the DFT-based ones, a negative impact of the signal detection within the DFT measurement setup can be excluded. A more ideal comparison in terms of quantitative S/N ratios would need a more suitable CCD detector with a faster read-out time and larger full-well capacity, as it was discussed in, e.g., Ref. [24]. Using such a high-speed CCD would still not allow to reach down to acquisition times in the microsecond regime. The presented DFT-FSRS method, on the other hand, is still scalable, e.g., by applying a laser system with an even higher pulse repetition frequency, like the one presented by Ploetz et al. [25].

Within the application of DFT-FSRS to hyperspectral imaging, a microsecond or even sub-microsecond pixel dwell time would result in image acquisition times, comparable to narrow-bandwidth schemes, like coherent anti-Stokes Raman scattering [21] or stimulated Raman scattering [22]. Even in the case that the S/N ratio of the individual Raman scattering spectra would be rather low, acquiring a hyperspectral dataset would allow the application of analytical data analyses as presented, e.g., in Ref. [2]. Additionally, the full Raman scattering spectrum per pixel would yield a higher chemical contrast, compared to a singly scanned line in narrow-bandwidth schemes.

3 Summary

In summary, we have demonstrated the applicability of a purely passive dispersive Fourier transformation (DFT) to femtosecond stimulated Raman scattering (FSRS) spectroscopy. Using FSRS, it was possible to excite the stimulated Raman signal, i.e., the Raman scattering spectrum, over a spectral range spanning from 700 to 3700 cm\(^{-1}\). The detection of the Raman signatures was performed with purely passive DFT, i.e., without any need for an optical amplification. The acquisition time for one Raman scattering spectrum amounted to 2 \(\upmu\)s. A 20-time-averaged DFT-FSRS spectrum resulted in a S/N ratio comparable to a standard CCD-spectrometer-based FSRS reading. As the presented DFT-FSRS scheme is still scalable to even lower pixel dwell times, it shows a promising potential to be applied in high-speed hyperspectral imaging.