PRIMAger Hyperspectral Imaging (PHI)

• PRIMAger Hyperspectral Imaging (PHI) is a capability that delivers medium resolution (R~8) spectra from 25–80μm using linear variable filters for detailed astrophysical analysis.
• PHI employs advanced cryogenic optics and kinetic inductance detector arrays with wafer-scale LVFs to build contiguous spectra via precise beam-scanning, reducing confusion noise.
• PHI supports diverse science applications from extragalactic surveys to mapping Galactic dust, achieving high flux recovery performance even in confusion-limited regimes.

Searching arXiv for the cited PRIMAger/PHI papers to ground the article in current literature. PRIMAger Hyperspectral Imaging (PHI) is the hyperspectral imaging capability of PRIMAger, the infrared camera of the PRobe far-Infrared Mission for Astrophysics (PRIMA). In the Phase A instrument and General Observer descriptions, PRIMAger will offer hyperspectral imaging in medium resolution bands ($R \sim 8$, using a linear variable filter) from 25 to 80 $\mu$m, with two contiguous LVF-equipped modules, PHI1 and PHI2, while a distinct mode provides broad band photometric and polarimetric imaging at longer wavelengths (Ciesla et al., 1 Sep 2025, Burgarella et al., 22 Sep 2025). Earlier confusion-noise studies describe PRIMAger as providing unique hyperspectral imaging simultaneously covering 25–235 $\mu$m and use that capability to test deblending far below the classical confusion limit (Donnellan et al., 2024). This suggests evolving instrument specifications between mission-concept simulations and later Phase A descriptions.

1. Mission context and scope

PRIMA is an infrared observatory for the next decade, currently in Phase A, with a 1.8m telescope actively cooled to 4.5K. On board, an infrared camera, PRIMAger, equipped with ultra-sensitive kinetic inductance detector (KID) arrays, will provide observers with coverage of mid-infrared to far-infrared wavelengths from 24 to 264 microns. PRIMAger is being developed by an international collaboration bringing together French institutes (Laboratoire d'Astrophysique de Marseille and CEA) through the center National d'Etudes Spatiales (CNES, France), the Netherlands Institute for Space Research (SRON, Netherlands), and the Cardiff University (UK) in Europe, as well as the Jet Propulsion Laboratory (JPL) and Goddard Space Flight Center (GSFC) in the USA (Ciesla et al., 1 Sep 2025).

A recurring point of confusion in the literature is the spectral extent assigned to PHI or to PRIMAger’s hyperspectral capability. The distinction is explicit in the published documents.

Source	Hyperspectral description
(Donnellan et al., 2024)	unique hyperspectral imaging simultaneously covering 25-235 $\mu$m
(Ciesla et al., 1 Sep 2025)	Hyperspectral mode will cover the 24-84 microns wavelength range with a spectral resolution $R=8$
(Burgarella et al., 22 Sep 2025)	hyperspectral imaging in medium resolution bands ($R \sim 8$, using a linear variable filter) from 25 to 80 $\mu$m

The later Phase A descriptions also partition PRIMAger into two imaging modes: the Hyperspectral mode will cover the 24-84 microns wavelength range with a spectral resolution $R=8$, while the Polarimetric mode will provide polarimetric imaging in 4 broad bands, from 80 to 264 microns (Ciesla et al., 1 Sep 2025). A plausible implication is that the 2024–2025 confusion analyses model a broader concept-stage hyperspectral configuration, whereas the instrument paper and GO program paper describe the Phase A implementation.

2. Optical architecture and detector system

PHI shares PRIMA’s 1.8 m, 4.5 K telescope focus via a 50 mm collimated input beam. A two-axis, cryogenic beam-steering mirror (BSM) at the telescope exit pupil, with $\pm 10'$ throw, $\lesssim 0.33''$ RMS accuracy, and $\mu$0 mW dissipation at 4.5 K, directs the beam into PRIMAger without moving the spacecraft. The BSM allows rapid ($\mu$1s) 2D scanning patterns, including raster, Lissajous, and boustrophedon, essential both to build spectra and to mitigate $\mu$2 detector noise (Ciesla et al., 1 Sep 2025).

The cold optical train uses a single, bare-Al common mirror to refocus the BSM beam to an intermediate focal plane inside the 1 K instrument box. Two field mirrors define cold stops for PHI and PPI, each followed by dichroic-free optics. The PHI channel is relayed by two further aspherized fold mirrors to achieve an effective focal ratio of $\mu$3 at the detector. All optics are aluminum 6061-T6, machined in a homothetic design so that cooldown-induced contraction preserves alignment to $\mu$4m (Ciesla et al., 1 Sep 2025).

Instead of wheels or dichroics, PHI uses wafer-scale Linear Variable Filters (LVFs) bonded directly above each KID array. Each LVF’s resonant metal-mesh pattern produces a smoothly varying band-centre along one detector axis, and no moving filter mechanism is required. PHI employs lens-absorber-coupled, hybrid Microwave Kinetic Inductance Detectors, with arrays fabricated on SiN membranes and mated to monolithic Si lenslets via precision glue pillars that maintain a $\mu$5m micro-lens gap (Ciesla et al., 1 Sep 2025).

Parameter	PHI1	PHI2
Wavelength range	24–45 $\mu$6m	45–84 $\mu$7m
Array size	$\mu$8 pixels	$\mu$9 pixels
Field of view	$\mu$0	$\mu$1
Pixel scale	$\mu$2	$\mu$3

The instantaneous spatial sampling is set by $\mu$4 optics and the pixel pitch. PHI1 has $\mu$5 pixels, approximately $\mu$6 at 24 $\mu$7m, and PHI2 has $\mu$8 pixels, approximately $\mu$9 at 45 $\mu$0m. Because this is coarser than Nyquist ($\mu$1), the BSM or spacecraft scan fills in the PSF (Ciesla et al., 1 Sep 2025).

3. Spectral formation and observing strategy

PHI1 covers $\mu$2, and PHI2 covers $\mu$3. Each pixel’s band-centre $\mu$4 varies along the long axis of the LVF, and the FWHM of the passband is $\mu$5, with $\mu$6. In practice, $\mu$7 everywhere in PHI1 and PHI2 (Ciesla et al., 1 Sep 2025).

PHI has no discrete filters. Spectra are formed by scanning a point source across rows of pixels in the spectral direction. As the BSM moves the source through the LVF gradient, successive pixels measure contiguous $\mu$8 slices. A full PHI spectrum from 24 to 84 $\mu$9m is built by scanning first across PHI1’s long axis and then PHI2’s, or vice versa, in a single observation, with typical scan length matching the $R=8$0 array dimension on the sky (Ciesla et al., 1 Sep 2025).

The observing modes are tied to this spectral-formation strategy. Large maps use spacecraft slews in boustrophedon patterns, with scan legs up to a few degrees and perpendicular spacing approximately equal to the pixel pitch, combined with slow BSM modulation to fill in $R=8$1 sampling and mitigate $R=8$2. Small maps use purely BSM-driven 2D scans over $R=8$3 fields at speeds up to $R=8$4s. PHI always requires scan modulation to reconstruct spectra; there is no staring mode (Ciesla et al., 1 Sep 2025).

In the $R=8$5-IR survey concept, each scan imparts a spectral-slit mapping: at each sky coordinate, subsequent visits with different LVF columns sample different $R=8$6. After 4–6 orthogonal scan passes the full 25–80 $R=8$7m range is sampled at $R=8$8, and the software assembles 3D cubes $R=8$9 on regular spatial and spectral grids (Burgarella et al., 22 Sep 2025).

4. Calibration and data products

The detector modules are housed at 125 mK and are thermally linked to a continuous Adiabatic Demagnetization Refrigerator. Eight independent frequency-multiplexed RF chains read out 2.6–4.9 GHz, sampling at 5 Gsps. A Xilinx Kintex KU060 FPGA per chain synthesizes a comb of tones via a polyphase filter-bank IFFT, sends them to the detectors via low-loss coax, and digitizes the returned tones through a digital spectrometer (WOLA PFB). The on-board tone-tracking firmware delivers per-pixel $R \sim 8$0 timestreams at approximately 100 Hz (Ciesla et al., 1 Sep 2025).

Calibration is based on both internal and celestial references. A set of on-board, cold ($R \sim 8$1 K) blackbody sources can be inserted via a small flip-in mirror for detector gain and linearity checks; these calibrators are used at the start and end of each observing block. Regular observations of well-characterized FIR standard stars, including asteroids and planetary nebulae, provide absolute flux calibration to $R \sim 8$2. Flat-fielding is accomplished by nodding a bright, spatially uniform field, exemplified by the Moon’s limb, across the LVF gradient to map pixel gains versus wavelength. Detector responsivities are stable to $R \sim 8$3 over 24 hr (Ciesla et al., 1 Sep 2025).

The raw data format consists of time-ordered $R \sim 8$4 streams for each KID at approximately 100 Hz, tagged with BSM angle, telescope attitude, and temperature monitors, and stored in PRIMA Level 0 packets. The Level 1 to Level 2 pipeline comprises deglitching and cosmic-ray removal; channelization, assigning each timestamp to a sky wavelength based on LVF pixel $R \sim 8$5 and BSM position; spectral extraction; flat-field and dark subtraction using internal calibrator data; and cube reconstruction, gridding $R \sim 8$6 into a 3D FITS cube with user-selectable $R \sim 8$7 bins, default $R \sim 8$8, and spatial pixels matching the 3.8–6.8 arcsec physical pitch (Ciesla et al., 1 Sep 2025).

The FITS data cube is specified as a multi-extension file with a PRIMARY header containing observation metadata, an extension “WAVELENGTH”, an extension “MAP” containing calibrated surface brightness in MJy/sr, an extension “ERROR” containing $R \sim 8$9 uncertainties, and an extension “DQ” containing a data-quality mask per voxel (Ciesla et al., 1 Sep 2025).

5. Confusion noise and Bayesian deblending

The confusion-limited use case of PRIMAger has been developed most explicitly in the simulation papers by Donnellan and collaborators. In the 2024 study, images representing a deep, $\mu$0 PRIMAger survey were synthesised with realistic instrumental and confusion noise (Donnellan et al., 2024). In the 2025 follow-up, a 1 $\mu$1 deep-field survey was simulated by populating the sky with sources drawn from the SIDES extragalactic model of Bethermin et al. 2017, convolving each source list at each band with the appropriate PSF, adding Gaussian instrumental noise, and adding confusion noise as spatially correlated background fluctuations using HELP cirrus+clustering templates. Source positions were taken from a Euclid-like near-IR catalogue with $\mu$2 and negligible positional error $\mu$3 (Donnellan et al., 15 Dec 2025).

The forward model at wavelength $\mu$4 is

$\mu$5

where $\mu$6 is the vector of observed pixel values, $\mu$7 is the vector of source fluxes, $\mu$8 is the design matrix whose columns are the normalized PSF at each source position, and $\mu$9 is Gaussian noise with covariance $R=8$0. Donnellan et al. present a new Bayesian modelling approach, XID+stepwise, that exploits PRIMAger's hyperspectral imaging to derive self-consistent, informative flux priors by sequentially propagating constraints from short to long wavelengths (Donnellan et al., 15 Dec 2025).

Fluxes are inferred in ascending-$R=8$1 order. At the first band no flux prior is used; once the posterior mean and covariance at band $R=8$2 are obtained, they are pinned and used as priors for all subsequent $R=8$3. In practice a narrow Gaussian prior is used instead of a strict delta function to preserve error propagation. At each band the joint posterior over $R=8$4 is sampled, and the resulting marginal flux estimates and uncertainties inform the next band (Donnellan et al., 15 Dec 2025).

This stepwise construction is not the only route considered. The 2024 study shows that catalogues of galaxies with a high purity ($R=8$5 per cent) can be constructed at a source density of $R=8$6 using PRIMAger data alone, and the 2025 study further shows that positional priors from blind source detection followed by deblending via XID+ enables PRIMAger to achieve sensitivity beyond the confusion limits using PRIMAger data alone (Donnellan et al., 2024, Donnellan et al., 15 Dec 2025).

6. Sensitivity and flux-recovery performance

In the instrument performance budget, PHI achieves a 5$R=8$7 point-source sensitivity of 1.2–4.1 mJy per 10 per cent $R=8$8 channel in a raster mapping mode over 1 $R=8$9 in 10 h. The noise model includes dark NEP $\pm 10'$0, photon NEP of order $\pm 10'$1–$\pm 10'$2 per pixel, single-mode throughput per pixel $\pm 10'$3, coupling efficiency $\pm 10'$4, and end-to-end transmission excluding LVF $\pm 10'$5. Detector $\pm 10'$6 knee is at $\pm 10'$7 Hz and the KID time constant is $\pm 10'$8 ms (Ciesla et al., 1 Sep 2025).

For the $\pm 10'$9-sr survey parameters, with scan speed $\lesssim 0.33''$0, total on-sky time $\lesssim 0.33''$1 h, and SNR $\lesssim 0.33''$2, the 5$\lesssim 0.33''$3 point-source sensitivities are 2.54 mJy at 34.3 $\lesssim 0.33''$4m and 3.44 mJy at 64.5 $\lesssim 0.33''$5m; PHI2 channels at 45–84 $\lesssim 0.33''$6m would lie in the 2–5 mJy range accordingly (Burgarella et al., 22 Sep 2025).

The confusion-noise studies quantify how far Bayesian deblending can push below the classical confusion floor. The 2024 deep-survey simulation shows fluxes measured with an accuracy better than 20 per cent to flux levels of 0.16, 0.80, 9.7, and 15 mJy at 47.4, 79.7, 172, and 235 $\lesssim 0.33''$7m, respectively. These levels are a factor of $\lesssim 0.33''$8 and $\lesssim 0.33''$9 fainter than the classical confusion limits for 72–96 $\mu$00m and 126–235 $\mu$01m, respectively. At $\mu$02, 8–10 of the 10 channels covering 47–235 $\mu$03m are detected and accurately measured for sources with $\mu$04, a 0.5 dex improvement on what might be expected from the classical confusion limit (Donnellan et al., 2024).

The 2025 XID+stepwise analysis strengthens this result. With Euclid-like prior source positions, fluxes are recovered to within 20 per cent to 0.2–0.7 mJy across 45–84 $\mu$05m, corresponding to factors of 1.3–3.4 fainter than the confusion limit. In the most confusion-dominated channels, accurate fluxes are measured to 0.9, 2.5, 7.6, and 14.8 mJy at 92, 126, 183, and 235 $\mu$06m, respectively, which are factors of 3–5 better than the confusion limit. Using a deeper Euclid-based prior catalogue and weak ancillary flux priors at 25 $\mu$07m yields further improvements, reaching up to a factor $\mu$08 fainter than the confusion limit at 96 $\mu$09m (Donnellan et al., 15 Dec 2025).

7. Survey programs and scientific applications

The scientific rationale for PHI is broad. PRIMAger’s observational capabilities have been tailored to answer fundamental astrophysical questions such as black hole and star-formation co-evolution in galaxies, the evolution of small dust grains over a wide range of redshifts, and the effects of interstellar magnetic fields in various environments, while also opening discovery space with versatile photometric and polarimetric capabilities (Ciesla et al., 1 Sep 2025).

Within the General Observer framework, the $\mu$10-IR survey covers roughly $\mu$11 sr, or $\mu$12, and would exploit PRIMAger’s hyperspectral and polarimetric modes to collect data on about $\mu$13 galaxies to $\mu$14. The $\mu$15 spectral resolution of the PRIMAger Hyperspectral Imaging filters will enable users to study the emission of polycyclic aromatic hydrocarbon (PAH), and a large sample of galaxies will be observed with the polarimetric bands of PRIMAger (Burgarella et al., 22 Sep 2025).

The instrument paper states that PHI’s $\mu$16 sampling recovers broad PAH features, for example the 34 $\mu$17m band at $\mu$18, and continuum breaks, delivering redshifts to $\mu$19 for dusty star-forming galaxies in confusion-limited deep fields. A 1 $\mu$20 deep survey with $\mu$21 h achieves $\mu$22 line sensitivity, integrated over $\mu$23, sufficient to map the [O IV] 25.9 $\mu$24m line in $\mu$25 galaxies at $\mu$26 (Ciesla et al., 1 Sep 2025).

Additional use cases extend from Galactic structure to compact dusty systems. PHI can spectrally map warm dust and PAH emission in giant molecular clouds such as Orion A, resolving substructures at 0.1 pc scales at 400 pc with $\mu$27m bins. Simulations show the ability to disentangle multiple temperature components and resolve 34 $\mu$28m PAH emission from continuum in a $\mu$29 strip in $\mu$30 hr. Other key targets listed for PHI include protoplanetary disks, via solid-state features such as 33.6 $\mu$31m forsterite, and AGN torus studies, tracing the warm dust continuum break near 30 $\mu$32m at $\mu$33 in moderate-luminosity Seyferts (Ciesla et al., 1 Sep 2025).

Across these science cases, one theme is consistent: hyperspectral imaging is used not only to provide $\mu$34 mid-infrared spectral sampling, but also to mitigate source confusion by adding the spectral dimension. The confusion-noise studies indicate that, when employing XID+, confusion noise will not limit the key science from PRIMA extragalactic imaging surveys (Donnellan et al., 15 Dec 2025).