Photonic Doppler Velocimetry Fundamentals

• Photonic Doppler Velocimetry is a fiber-optic heterodyne diagnostic that measures Doppler frequency shifts to determine target velocities.
• It mixes a stable reference with reflected signals and uses STFT or Bayesian methods to analyze time–frequency data for precise velocity reconstruction.
• Recent innovations, such as time-lens techniques, extend PDV’s bandwidth, enabling accurate measurements in regimes with high scattering or rapid motion.

Searching arXiv for recent and foundational PDV papers to support the article. arxiv_search(query="Photonic Doppler Velocimetry PDV time lens ejecta Bayesian", max_results=10, sort_by="relevance") Photon or Photonic Doppler Velocimetry (PDV) is a fiber-based heterodyne diagnostic that measures the frequency shift of light backscattered from moving targets and converts that shift to an apparent velocity through the Doppler relation. In its standard form, a local oscillator or stable reference field is mixed with the Doppler-shifted return from a moving surface or particulate ejecta, and the resulting beat signal is analyzed in time–frequency space, typically by a Short-Time Fourier Transform (STFT), to infer a velocity history. Across recent work, PDV appears in three closely connected forms: as a conventional single-channel fiber interferometer for fast free-surface or shock-front measurements, as a transport-sensitive diagnostic for dense ejecta where multiple scattering invalidates naïve single-scattering interpretations, and as a platform for methodological extensions such as time-lens temporal magnification and Bayesian time-domain inference (Chu et al., 2021, Jayamanne et al., 2023, Allison et al., 19 Aug 2025).

1. Measurement principle and Doppler mapping

In a single-channel fiber PDV, the probe delivers light to the target and also returns a small, non-Doppler-shifted reference reflection; the returning reference and Doppler-shifted light are innately mixed, routed by a circulator to a photodetector. The detector measures

$$I(t) = I_0 + I_d + \sqrt{I_0 I_d}\,\sin\!\big(2\pi f_b(t)t + \phi\big),$$

where $I_0$ and $I_d$ are the intensities of the reference and Doppler return, respectively, and $\phi$ is their relative phase. The beat frequency is $f_b(t)=|f_d(t)-f_0|$, with $f_0$ the laser frequency and $f_d(t)$ the Doppler-shifted return (Chu et al., 2021).

For a moving mirror, the beat frequency–velocity relation is

$f_b(t)=\pm 2\,[v(t)/c]\,f_0.$

For normal incidence in air, using $f_0=c/\lambda_0$, this reduces to

$f_b(t)=\frac{2v(t)}{\lambda_0}.$

For non-normal incidence at angle $I_0$0,

$I_0$1

and in a medium of refractive index $I_0$2,

$I_0$3

At $I_0$4 nm, $I_0$5 km/s gives $I_0$6 GHz (Chu et al., 2021).

A formally equivalent backscatter relation is used in ejecta work: $I_0$7 or, in angular-frequency form,

$I_0$8

In the single-scattering limit, spectrogram lines therefore map directly to particle or surface velocities along the probe axis (Jayamanne et al., 2024).

A more general vector Doppler formula is required when geometry departs from monostatic backscatter or when multiple scattering is present: $I_0$9 with $I_d$0. In the monostatic backscatter case, $I_d$1, giving $I_d$2. This relation remains useful as an apparent-velocity mapping even when the detected photons have undergone multiple scattering, but in that regime it no longer has the direct one-photon/one-velocity meaning of the single-scattering case (Jayamanne et al., 2023).

2. Interferometric architectures and analysis workflows

The standard instrumentation described across the cited work is a fiber-interferometric heterodyne system operating near $I_d$3. Typical implementations use a continuous-wave or narrow-line laser, a fiber probe, a circulator, a fast photodetector, and a high-bandwidth oscilloscope or digitizer. In the time-domain Bayesian study, typical systems use near-infrared lasers at $I_d$4 nm and GHz-bandwidth digitizers such as 25 GHz, 50 GS/s instruments; in the TL-PDV study, the PDV operates in the C-band around $I_d$5 nm $I_d$6, which is well matched to four-wave-mixing time-lens technology (Allison et al., 19 Aug 2025, Chu et al., 2021).

The established analysis workflow forms a spectrogram by applying an STFT to the oscilloscope trace. The time–frequency content is represented as

$I_d$7

with gate window $I_d$8 of width $I_d$9. In practical PDV, the window form, duration, and separation determine the trade-off between temporal and frequency resolution. In the validation reported for the Bayesian method, an STFT with a 5 ns window yielded a velocity uncertainty of $\phi$0; the same work emphasizes that STFT windowing broadens discontinuities and limits recovery at very early times (Jayamanne et al., 2024, Allison et al., 19 Aug 2025).

A distinct line of work replaces ridge-picking on a spectrogram by direct inference on the oscilloscope trace. The forward model used in the Bayesian time-domain method writes the mean signal as

$\phi$1

with

$\phi$2

Velocity is parameterized as a one-dimensional piecewise-linear time series in acceleration segments, and inference is performed with the Bilby library and the Dynesty nested sampling engine using 1500 live points. The method recovered injected velocities from synthetic data within the 95% credible interval, and on quartz shock data it produced velocities consistent with STFT-derived values while interpolating across low-signal regions (Allison et al., 19 Aug 2025).

This division between STFT-based and time-domain inference is methodological rather than conceptual. Both are built on the same heterodyne beat physics. A plausible implication is that PDV should be regarded less as a single analysis protocol than as a measurement class in which the optical hardware, transport model, and inverse method can vary substantially while the Doppler encoding remains the common core.

3. From single scattering to radiative transport in ejecta

In ejecta applications, PDV has traditionally been used in the single-scattering regime to read out velocity distributions. That approximation breaks down when the optical thickness $\phi$3 becomes large: $\phi$4 where $\phi$5 is the scattering mean free path. The 2024 ejecta study states that velocity-only readout is reliable when $\phi$6, with $\phi$7 given as a practical confidence level; the 2023 multiple-scattering study shows that a distinct peak at the free-surface frequency appears for $\phi$8 and is not visible for larger $\phi$9. These are complementary statements about different analysis criteria, not contradictory thresholds (Jayamanne et al., 2024, Jayamanne et al., 2023).

The rigorous formulation connects the measured spectrogram to the specific intensity $f_b(t)=|f_d(t)-f_0|$0, the radiative flux density at position $f_b(t)=|f_d(t)-f_0|$1, direction $f_b(t)=|f_d(t)-f_0|$2, time $f_b(t)=|f_d(t)-f_0|$3, and frequency $f_b(t)=|f_d(t)-f_0|$4. The spectrogram–intensity relation is

$f_b(t)=|f_d(t)-f_0|$5

where $f_b(t)=|f_d(t)-f_0|$6 is the detector etendue and $f_b(t)=|f_d(t)-f_0|$7 is the detector normal. The measured PDV spectrogram is therefore an etendue-integrated specific intensity at frequencies $f_b(t)=|f_d(t)-f_0|$8, scaled by the STFT window and reference amplitude (Jayamanne et al., 2023).

The generalized radiative transfer equation (RTE) used for dynamic ejecta is

$f_b(t)=|f_d(t)-f_0|$9

with energy velocity $f_0$0, extinction mean-free path $f_0$1, scattering mean-free path $f_0$2, and phase function $f_0$3. The dependence on $f_0$4 captures inelastic scattering, specifically Doppler coupling (Jayamanne et al., 2023, Jayamanne et al., 2024).

The size distribution enters through the optical coefficients: $f_0$5 where $f_0$6 is particle number density and $f_0$7 are Mie-theory extinction and scattering cross sections. The phase function couples angle and frequency through

$f_0$8

so the spectrogram reflects a convolved size–velocity field rather than a direct backscatter velocity map (Jayamanne et al., 2024).

This multiple-scattering literature corrects a common misconception: PDV is not intrinsically a velocity-only diagnostic. In optically thick ejecta, the spectrogram morphology depends on particle size statistics through $f_0$9, $f_d(t)$0, and $f_d(t)$1, and particle-size information can be recovered by forward modeling when the transport physics is included (Jayamanne et al., 2024).

4. Vacuum ejecta spectra and closed-form parameter extraction

For vacuum ejecta, one line of work combines full Monte Carlo light transport with a closed-form single-scattering interpretation. The GPU-accelerated Monte Carlo reconstruction models photons propagating through a granular ejecta layer, with scattering and absorption probabilities computed from Mie theory. The algorithm samples Beer–Lambert penetration, scattering versus absorption, directional scattering from the Mie phase function, and Doppler frequency updates at scattering events; outgoing photons are histogrammed to obtain a synthetic PDV spectrum. Using $f_d(t)$2 photons, GPU acceleration up to $f_d(t)$3 on an Nvidia GTX960 reduced runtime to $f_d(t)$4 s per case (Shi et al., 2020).

Within the single-scattering interpretation introduced in the same work, the dominant PDV peak can be related directly to ejecta parameters. The peak position satisfies

$f_d(t)$5

and the peak curvature satisfies

$f_d(t)$6

These relations enable extraction of the velocity-profile coefficient $f_d(t)$7 from curvature and optical thickness $f_d(t)$8 from peak position, provided $f_d(t)$9 is known. With independent size information, $f_b(t)=\pm 2\,[v(t)/c]\,f_0.$0 can then be converted to areal mass $f_b(t)=\pm 2\,[v(t)/c]\,f_0.$1 (Shi et al., 2020).

The paper validates this procedure on vacuum ejecta from shock-loaded grooved Sn at $f_b(t)=\pm 2\,[v(t)/c]\,f_0.$2 GPa with $f_b(t)=\pm 2\,[v(t)/c]\,f_0.$3 m/s. From the processed PDV spectrum over a $f_b(t)=\pm 2\,[v(t)/c]\,f_0.$4–$f_b(t)=\pm 2\,[v(t)/c]\,f_0.$5s window, the main peak occurs at $f_b(t)=\pm 2\,[v(t)/c]\,f_0.$6 with curvature $f_b(t)=\pm 2\,[v(t)/c]\,f_0.$7 in velocity-normalized units, yielding

$f_b(t)=\pm 2\,[v(t)/c]\,f_0.$8

Independent Mie-scattering measurements gave $f_b(t)=\pm 2\,[v(t)/c]\,f_0.$9 and $f_0=c/\lambda_0$0, from which the areal mass was inferred as $f_0=c/\lambda_0$1. The PDV-derived velocity profile and the piezoelectric probe mass–velocity function were reported to be in good quantitative agreement (Shi et al., 2020).

The same study also makes clear where the closed-form model ceases to be adequate. It reproduces the dominant peak well for moderate-to-large optical thickness in vacuum, but it does not reproduce the free-surface peak that emerges when optical thickness is small. Full multiple-scattering Monte Carlo is therefore preferred when the spectrum exhibits multiple peaks, pronounced low-velocity structure, or strong sensitivity to transport anisotropy (Shi et al., 2020).

5. Ejecta in gas: joint recovery of velocity and size information

The 2024 ejecta work extends PDV from vacuum single-scattering analysis to ejecta transported in gas, where drag, breakup, and multiple scattering alter the spectrogram substantially. The experimental context is a launch tube with barrel inner diameter $f_0=c/\lambda_0$2 mm, a grooved tin disk with surface grooves of $f_0=c/\lambda_0$3, and a copper flyer impacting tin at $f_0=c/\lambda_0$4 to produce $f_0=c/\lambda_0$5 GPa. Ejecta transport is studied in vacuum $f_0=c/\lambda_0$6, helium $f_0=c/\lambda_0$7, and air $f_0=c/\lambda_0$8. The PDV is on-axis and in reflection, with a single probe that both illuminates and collects, using $f_0=c/\lambda_0$9 in the analysis (Jayamanne et al., 2024).

The initial ejecta is modeled as spherical particles with joint size–velocity distribution

$f_b(t)=\frac{2v(t)}{\lambda_0}.$0

with the velocity distribution derived from the integrated ejected mass–velocity curve

$f_b(t)=\frac{2v(t)}{\lambda_0}.$1

In gas, particles experience drag and possibly breakup. The simulations use the KIVA-II quadratic drag law and a breakup criterion based on the Weber number

$f_b(t)=\frac{2v(t)}{\lambda_0}.$2

with breakup when $f_b(t)=\frac{2v(t)}{\lambda_0}.$3 (Jayamanne et al., 2024).

The forward model couples hydrodynamics to optics. Ejecta transport is simulated with the CEA Phenix code, typically with $f_b(t)=\frac{2v(t)}{\lambda_0}.$4 numerical particles and 180 time steps of $f_b(t)=\frac{2v(t)}{\lambda_0}.$5, requiring $f_b(t)=\frac{2v(t)}{\lambda_0}.$6 h on 1 AMD EPYC 7763 with 64 cores. For optical modeling, the launch tube is discretized into $f_b(t)=\frac{2v(t)}{\lambda_0}.$7 layers, Mie routines provide $f_b(t)=\frac{2v(t)}{\lambda_0}.$8, and the generalized RTE is solved by Monte Carlo random walks. A typical spectrogram involves $f_b(t)=\frac{2v(t)}{\lambda_0}.$9 draws over 180 times and 2500 particles, with compute time $I_0$00 h 20 min on 80 AMD EPYC 7763 CPUs (Jayamanne et al., 2024).

The inversion strategy is iterative forward modeling rather than a closed-form inversion. In vacuum, a power-law $I_0$01 with $I_0$02, $I_0$03, $I_0$04 yielded excessive optical thickness $I_0$05 and hid the free surface; replacing it by a lognormal

$I_0$06

truncated to $I_0$07, reduced $I_0$08 to $I_0$09 and restored the free-surface return. In helium, increasing drag coefficients $I_0$10 within the KIVA-II model aligned the simulated upper envelope with experiment. In air, drag plus breakup produced an early plateau near $I_0$11 at $I_0$12–$I_0$13s, suppression of the free-surface return by $I_0$14s, and re-acceleration bands around $I_0$15 between $I_0$16 and $I_0$17s; simulations captured these qualitative features but retained a broader long-term velocity spread than experiment, indicating limitations of the independence assumption $I_0$18 and suggesting correlated size–velocity initial distributions (Jayamanne et al., 2024).

A second common misconception is therefore that multiple scattering merely adds nuisance broadening. The gas-transport analysis shows that spectrogram features such as free-surface disappearance, dynamic-range changes, slowing-down envelopes, and re-acceleration bands carry information about size evolution and breakup as well as velocity.

6. Bandwidth extension, dynamic range, and emerging PDV modalities

Conventional PDV is limited by the maximum detectable beat frequency set by the electronics. If $I_0$19 by Nyquist, then the maximum measurable velocity is approximately

$I_0$20

At $I_0$21 nm, $I_0$22 GHz corresponds to $I_0$23 km/s, described as consistent with current PDV limits in the time-lens study (Chu et al., 2021).

Time-lens Photon Doppler Velocimetry (TL-PDV) addresses this bandwidth bottleneck by inserting a temporal imaging system into the optical path between the PDV mixing stage and the photodetector. The time lens is a four-wave-mixing (FWM) device that applies a quadratic temporal phase

$I_0$24

or, in the notation used in the paper,

$I_0$25

with the phase realized by FWM between a linearly chirped pump and the input signal in a nonlinear medium such as a highly nonlinear fiber or integrated photonic waveguide (Chu et al., 2021).

Temporal imaging comprises three stages: pre-dispersion of the input waveform, the FWM time lens, and post-dispersion of the idler. The imaging condition is

$I_0$26

with temporal magnification

$I_0$27

Because temporal magnification stretches time by $I_0$28, instantaneous frequency scales inversely: $I_0$29 The PDV beat frequency at the output is therefore

$I_0$30

and velocity is recovered through

$I_0$31

for normal incidence in air (Chu et al., 2021).

The simulated TL-PDV implementation uses a mode-locked laser pump with 0.5 ps pulse width at 100 MHz repetition rate, dispersion parameter

$I_0$32

for all fibers, pre-dispersion $I_0$33, post-dispersion $I_0$34, and a pump chirped by a TOD-limited dispersive fiber of length $I_0$35 with $I_0$36 m. These parameters satisfy the imaging condition and yield $I_0$37 (Chu et al., 2021).

Under that magnification, a PDV beat spanning $I_0$38 GHz for velocities sweeping from $I_0$39 km/s to $I_0$40 km/s at 1550 nm is reduced to $I_0$41 GHz, bringing the signal within the detection band of “10s of GHz” digitizers. The effective velocity range extension scales approximately as

$I_0$42

so a PDV limited to $I_0$43 GHz can, with $I_0$44, measure original beat frequencies up to $I_0$45 GHz (Chu et al., 2021).

The time-lens study compares TL-PDV with heterodyne electrical downconversion, leapfrog PDV, and time-stretched PDV. The stated distinctions are that electrical downconversion may introduce spectral artifacts and can reduce sensitivity at low velocities, leapfrog PDV extends range at the cost of complexity and expense, and time-stretched PDV creates replicas and requires complex optical hardware. TL-PDV instead magnifies the optically encoded PDV signal directly. This suggests that bandwidth extension in PDV can be pursued either electronically, optically, or inferentially: TL-PDV changes the signal before detection, whereas the Bayesian time-domain method changes the inverse problem after detection (Chu et al., 2021, Allison et al., 19 Aug 2025).

Across these modalities, the principal limitations remain explicit. For TL-PDV they include pump depletion, phase mismatch, dispersion errors, nonlinear phase noise, third-order dispersion, and overlap of adjacent magnified windows. For Bayesian inference they include amplitude-model misspecification, multi-surface or multi-velocity reflections, down-shift branch ambiguity, and computational cost, with runs often taking many hours to converge. In both cases, the limiting factor is not the Doppler relation itself but the degree to which the full signal-generation model captures the experiment (Chu et al., 2021, Allison et al., 19 Aug 2025).