Euclid Infrared Detectors

• Euclid infrared detectors are a mosaic of 16 HgCdTe H2RG arrays that deliver near-infrared photometry and slitless spectroscopy using non-destructive up-the-ramp sampling.
• They integrate cold and warm electronics—featuring SIDECAR ASICs, DCUs, and precise timing—to achieve accurate calibration of quantum efficiency, gain, and inter-pixel capacitance.
• Advanced onboard processing with MACC sampling and weighted least-squares fits minimizes bias and manages persistence and correlated noise for robust detector performance.

Euclid infrared detectors are the detector system of the Near-Infrared Spectrometer and Photometer (NISP), one of Euclid’s two science instruments. They consist of a mosaic of 16 Teledyne HgCdTe H2RG arrays that provide both near-infrared photometry and slitless spectroscopy, complementing the visible imager for photometric redshift estimation and spectroscopic redshift measurements. In the literature, these detectors are characterized jointly with their cryogenic SIDECAR-based readout chain, the warm Data Processing Units and Instrument Control Unit, the non-destructive up-the-ramp and multi-accumulation sampling modes, and a calibration program covering quantum efficiency, inter-pixel capacitance, conversion gain, non-linearity, dark current, readout noise, correlated noise, and image persistence. Ground and flight analyses consistently report that the subsystem meets Euclid requirements, while also identifying persistence and subtle covariance effects as the principal detector-specific systematics requiring explicit modeling (Maciaszek et al., 2022, Medinaceli et al., 2023, Collaboration et al., 15 Jul 2025, Kubik et al., 3 Jun 2025).

1. Detector architecture and focal-plane implementation

NISP is built around a $4\times4$ focal plane of 16 H2RG infrared detector units. Each detector is a hybrid H2RG sensor with a 2.3 $\mu$m cutoff, $2048\times2048$ physical pixels, and an $18\,\mu$m pitch corresponding to $0.3$ arcsec/pixel on sky. Euclid flight-model descriptions place the detector subsystem within the Infrared Detector Assembly or detector system, where each sensor is paired with its own cryogenic front-end electronics through a short cryo-flex cable, and all 16 detector/readout pairs are synchronized from a common clock (Maciaszek et al., 2022, Medinaceli et al., 2023, Kubik et al., 3 Jun 2025).

Element	Configuration	Reported role
Focal plane	16 H2RG detectors in a $4\times4$ mosaic	Imaging and slitless spectroscopy
Pixel format	$2048\times2048$, $18\,\mu$m pixels	$0.3$ arcsec/pixel sampling
Cold electronics	16 SIDECAR ASICs / SCEs	One matched pair per detector
Warm electronics	16 DCUs, 2 DPUs, 1 ICU	Detector control, preprocessing, telemetry

The mechanical and thermal architecture is split across cold and warm domains. The NI-OMA is a silicon carbide structure carrying lenses, filter wheel, grism wheel, calibration unit, and focal-plane support; it operates near 135 K with stability better than 0.3 K over mission life. The detectors themselves are colder, described as below 100 K in the flight-model presentation, about 90 K in the flight-readiness review, and about 85 K in the persistence characterization campaigns. The SIDECAR ASICs operate around 135–140 K, while the warm electronics are in the service module at about 290 K and are connected through a dedicated harness of about 6 m (Maciaszek et al., 2022, Medinaceli et al., 2023, Kubik et al., 3 Jun 2025).

The readout chain is hierarchically organized. Each detector is interfaced by one DCU; two identical DPUs each control half of the focal plane, eight DCUs per DPU; and the ICU provides higher-level command, housekeeping, heater control, wheel control, calibration-source control, and spacecraft interfacing through MIL-STD-1553, while science data are transmitted through SpaceWire. The synchronization budget is stringent: the 16 SCA/SCE pairs are driven with a 10 MHz timing basis and a 50 ns maximum differential skew budget, reflecting the electrical density and crosstalk sensitivity of the mosaic (Maciaszek et al., 2022).

The observing bandpasses define the scientific use of the detectors. NISP photometric imaging uses the $Y_E$, $\mu$0, and $\mu$1 filters, with reported passband ranges of 949.6–1212.3 nm, 1167.6–1567.0 nm, and 1521.5–2021.4 nm, while slitless spectroscopy uses one blue grism covering 920–1400 nm and three red grisms covering 1200–1950 nm with different orientations (Medinaceli et al., 2023). A common misconception is that Euclid’s infrared detectors are only an imaging subsystem; in fact, the same H2RG focal plane is central to both the photometric and spectroscopic channels.

2. Non-destructive sampling, MACC readout, and onboard estimation

The Euclid infrared detectors are read non-destructively in up-the-ramp mode, with the flight-relevant implementation based on multi-accumulation charge collection. In this scheme, repeated reads are averaged into groups before further processing. A raw frame is denoted $\mu$2, and the group average is

$\mu$3

The MACC pattern is written as MACC$\mu$4, where $\mu$5 is the number of groups, $\mu$6 the number of frames per group, and $\mu$7 the number of dropped frames between groups. Euclid operations use MACC(4,16,4) for photometry and MACC(15,16,11) for spectroscopy; only the fitted ramp slope is ultimately transmitted, so signal estimation must be done onboard under tight telemetry and CPU constraints (Munoz et al., 2021, Kubik et al., 5 Jun 2025, Collaboration et al., 15 Jul 2025).

The exposure timing in the ground-characterization overview is written as

$\mu$8

$\mu$9

This formalism reflects the dependence of effective integration and total exposure times on grouping, dropped frames, and reset overhead (Collaboration et al., 15 Jul 2025).

The warm detector-processing chain is correspondingly specialized. The DPU application software is implemented in ANSI-C, runs on VxWorks 5.1 with a multi-task preemptive scheduler, and is organized into 13 tasks. Its reported functions include synchronization of subsystems at 1 Hz, command handling at about 512 bit/s, telemetry production at about 18 kbit/s, frame grouping and averaging, selected-raw-line extraction for monitoring, buffering of co-added frames, fit-quality computation, lossless compression, and support for a science data flow of about 290 Gbit/day. The data buffer can store up to 46 averaged frames per dither per detector, with double buffering used to support the pipeline (Maciaszek et al., 2022, Medinaceli et al., 2023).

In the flight-readiness review, the onboard processing sequence is described as averaging groups of input exposure frames, estimating signal via a weighted least-squares fit, computing a fit-quality statistic $2048\times2048$0, and compressing the results before downlink. Later analyses examine the NISP signal estimator more explicitly in terms of the differences between consecutive groups,

$2048\times2048$1

and derive analytic flux estimators and companion quality diagnostics that remain lightweight enough for onboard use (Medinaceli et al., 2023, Kubik et al., 5 Jun 2025, Collaboration et al., 3 Feb 2026).

A central result of the 2025–2026 estimator papers is that the use of group differences largely removes the Poisson correlation of the integrated ramp while preserving analytic tractability. The associated quality factor $2048\times2048$2 is constructed as a compact anomaly indicator for each ramp; in the absence of anomalies, its normalized distribution is centered near 1, whereas nonlinearity, saturation, cosmic-ray hits, and electronic jumps increase it (Kubik et al., 5 Jun 2025). This clarifies that onboard processing is not merely a compression step: it is part of the detector-performance model itself.

3. Calibration infrastructure and characterization products

Pre-launch calibration of the Euclid infrared detectors was organized as a staged campaign extending from NASA/GSFC selection through CPPM pixel-level characterization to LAM and payload-level cryogenic system tests. The purpose was simultaneously to verify compliance and to generate calibration artifacts for science operations, including bad-pixel maps, inter-pixel-capacitance maps, conversion-gain maps, non-linearity corrections, persistence characterization, dark-current measurements, noise characterization, and electrical transfer information (Maciaszek et al., 2022, Collaboration et al., 15 Jul 2025).

The scale of the campaign was substantial. The LAM thermal-vacuum and thermal-balance campaigns lasted about 3 months and involved around 40k commands, 19k images per detector, and 300k files. In the flight-model presentation, the final cold thermal-vacuum test with the full flight configuration lasted 32 days and produced more than 300,000 detector images. The ground-characterization overview emphasizes automated sequencing and near-real-time quality control during a 14-day thermal-vacuum campaign, reporting 99.73% time efficiency (Maciaszek et al., 2022, Medinaceli et al., 2023, Collaboration et al., 15 Jul 2025).

Calibration products were explicitly defined at cold temperature. The flight-model paper lists a personality file for each detector, a calibration lookup table for the in-flight calibration source, a noise characterization, a non-linearity correction LUT, an electrical transfer function, and illumination reference data for slew and survey conditions (Maciaszek et al., 2022). These products formalize the fact that Euclid detector calibration is pixel- or region-resolved rather than reducible to a single focal-plane average.

A dedicated internal calibration source, the NISP Calibration Unit, supports this program in flight. NI-CU is described as Euclid’s in-flight internal near-infrared calibration lamp and the first astronomical calibration lamp based on LEDs to be operated in space. It was designed to provide spatially approximately homogeneous illumination with $2048\times2048$3 variations and temporal stability of $2048\times2048$4 over 1200 s across the NISP detector plane, at five fixed wavelengths over roughly 900–1900 nm, with a fluence range from $2048\times2048$5 ph s$2048\times2048$6 pixel$2048\times2048$7 to $2048\times2048$8 ph s$2048\times2048$9 pixel$18\,\mu$0 (Collaboration et al., 2024).

The role of NI-CU is specifically relative calibration. It supports pixel-to-pixel relative sensitivity or small-scale flat-field calibration, detector non-linearity calibration, and in-flight monitoring. It is not intended to perform absolute photometric calibration on its own; the paper explicitly identifies absolute onboard-lamp calibration as a “chicken-and-egg” problem because the lamp itself would require long-term calibration (Collaboration et al., 2024). The flight-model presentation further notes that the calibration unit uses $18\,\mu$1 LEDs at five infrared wavelengths illuminating a Spectralon PTFE target and does not illuminate through the telescope optics, but instead provides direct detector illumination through baffles for small-scale flat-field calibration and linearity measurement (Maciaszek et al., 2022).

This calibration architecture implies that Euclid detector monitoring is continuous rather than confined to pre-launch acceptance. A plausible implication is that detector physics, onboard estimation, and calibration hardware were designed as a single operational system rather than as separable subsystems.

4. Pixel-level properties: bad pixels, quantum efficiency, IPC, gain, and non-linearity

The most comprehensive pre-launch overview reports that Euclid’s H2RG detectors met requirements across the standard pixel-property set. Pixels classified as non-functioning accounted for less than 0.2% of all science pixels. The same overview states that 95% of the pixels show a QE greater than 80% across the entire spectral range of the mission, that the conversion gain is approximately 0.52 ADU/e$18\,\mu$2, with a variation less than 1% between channels of the same detector, and that the deviation from linear response at signal levels up to 80 ke$18\,\mu$3 is less than 5% for 95% of the pixels (Collaboration et al., 15 Jul 2025).

Quantum-efficiency characterization was derived at NASA DCL from 40 wavelengths spanning 0.6–2.6 $18\,\mu$4m, with 5% absolute accuracy. Across the Euclid photometric bands, median QE values are typically above 90%, and detector percentiles in $18\,\mu$5, $18\,\mu$6, and $18\,\mu$7 are generally in the 90–100% range (Collaboration et al., 15 Jul 2025). Some QE estimates exceed 100% because the estimate uses an independently calibrated average gain; the paper explicitly attributes this to gain choice and calibration uncertainties rather than to physical QE above unity (Collaboration et al., 15 Jul 2025). This is a useful corrective to the common misunderstanding that $18\,\mu$8 tabulated QE necessarily signals a detector-model inconsistency.

Inter-pixel capacitance is a major example of the need for spatially resolved calibration. In the broad overview, IPC coupling is described as minimal, with crosstalk between adjacent pixels less than 1% and total capacitive coupling below 3% of the pixel nodal capacitance (Collaboration et al., 15 Jul 2025). However, the dedicated IPC paper shows that IPC is spatially structured: the median central coupling is $18\,\mu$9 at one representative setting, the cardinal neighbors differ by about 10% from each other, the diagonal neighbors are nonzero and differ by more than 10%, and IPC coefficients in the central epoxy-void region are about 120% smaller than in the epoxy-filled area (Graët et al., 2022).

The same spatial structure propagates into conversion gain. For $0.3$0 super-pixels, the IPC-free gain map yields $0.3$1 ADU e$0.3$2 in the epoxy-filled region, $0.3$3 ADU e$0.3$4 in the epoxy-void region, and $0.3$5 ADU e$0.3$6 detector average, so the epoxy-void region has a gain 2.2% higher than the epoxy-filled region and variations greater than 1% remain even inside the epoxy-filled region (Graët et al., 2022). The Euclid team therefore adopted a practical compromise of two gain values per readout channel, one for the epoxy-filled region and one for the epoxy-void region (Graët et al., 2022). This directly contradicts the simplifying assumption that a single detector-average gain is always sufficient.

A later cross-mission methodology paper places the Euclid H2RG gain problem in a broader detector-metrology framework. It shows that the classical mean-variance method yields an apparent gain increase of about 0.9% per 10 ke$0.3$7 for the Euclid H2RG if nonlinearity is ignored, while an NL3 model keeps the gain essentially constant across the full signal range. In the same study, the median total IPC for the Euclid H2RG is $0.3$8, IPC correction changes the gain by roughly 5%, and the corrected mean conversion gain is $0.3$9 e$4\times4$0 ADU$4\times4$1, the reciprocal representation of the $4\times4$2 ADU/e$4\times4$3 values used elsewhere (Graët et al., 2024).

These results establish a consistent picture. IPC is small in absolute amplitude but not negligible as a calibration bias; nonlinearity is modest in the Euclid operating range but large enough to bias gain estimation if treated classically; and the detector construction, especially epoxy-filled versus epoxy-void regions, leaves measurable signatures in both IPC and gain.

5. Correlated readout noise and image persistence

Readout noise in the NISP detectors is not perfectly white. The correlated-noise analysis models the readout-noise power spectral density as

$4\times4$4

and derives an analytical covariance matrix for the group-difference vector $4\times4$5 under MACC readout. From long dark integrations of an engineering-grade detector, the best-fit median parameters were

$4\times4$6

The resulting covariance structure extends beyond the nearest-neighbor form of the white-noise approximation (Munoz et al., 2021).

Monte Carlo ramps generated from this noise model show that using the full correlated-noise covariance matrix diminishes flux bias by up to a factor of four relative to the white-noise approximation, especially at low background fluxes below $4\times4$7. At Euclid-relevant sky backgrounds of about $4\times4$8–$4\times4$9, however, the bias is already very small in both treatments, dropping below about 0.1%, and the paper concludes that the final in-flight NISP flux bias should remain negligible even under the white-noise approximation used as the Euclid onboard baseline (Munoz et al., 2021). The 2023 flight-readiness review independently reports $2048\times2048$0 for the correlated-noise approximation in the expected operating regime (Medinaceli et al., 2023).

Later flight-based work makes the same point in detector-estimator language. Early in-orbit analysis from September 2023 to April 2024 finds a systematic bias lower than 0.01 e/s for 99% of the NISP pixel array. In dark exposures, the intrinsic estimator bias is $2048\times2048$1 in spectroscopic mode and $2048\times2048$2 in photometric mode; in zodiacal-background science exposures, it is $2048\times2048$3 and $2048\times2048$4, respectively (Collaboration et al., 3 Feb 2026). This suggests that correlated noise is a real property of the detector-readout chain, but not an operationally dominant bias in Euclid’s main background-dominated regime.

Persistence is the more consequential non-ideality. The persistence study identifies it as one of the most important detector non-idealities for Euclid’s redshift survey because residual images can remain for hours or days and contaminate subsequent observations. Ground measurements on all 16 detectors, performed at 85 K, used flat-field stimuli from 5000 e up to 95,000 e for unsaturated tests and saturation tests at 270,000 e and 385,000 e, with persistence measured in a dark UTR(286) sequence after each stimulus (Kubik et al., 3 Jun 2025).

Below saturation, the median persistence in a typical photometric exposure is less than 1% of the stimulus, but detector-to-detector and intra-detector variability are large. For a stimulus of about 5359 e, median detector persistence spans roughly 1 to 45 e; for a stimulus of about 94,634 e, it spans about 10 to almost 300 e. The broader detector-overview paper summarizes the same regime by stating that median persistence amplitudes are less than 0.3% of the signal, though persistence exhibits significant spatial variation and differences between detectors (Kubik et al., 3 Jun 2025, Collaboration et al., 15 Jul 2025).

The persistence current is better fit by a power law than by sums of one, two, or three exponentials. In the time domain,

$2048\times2048$5

with fluence-dependent $2048\times2048$6 and $2048\times2048$7, and six fitted parameters per pixel in the sub-saturation model (Kubik et al., 3 Jun 2025). The model bias is typically 5% to 10% of the persistence signal, and residual noise is below 8 e RMS for all stimuli (Kubik et al., 3 Jun 2025). Flight Performance Verification calibrations are on average compatible with the ground-characterized $2048\times2048$8 and $2048\times2048$9, and the authors report no significant change in persistence relative to ground characterization, within the limitations of the comparison (Kubik et al., 3 Jun 2025).

A significant boundary condition appears above saturation. After saturated stimuli, the detector does not return quickly to the pre-saturation state even after 6 hours in the dark, and the persistence pattern can invert spatially relative to the sub-saturation regime (Kubik et al., 3 Jun 2025). The operational conclusion is therefore twofold: sub-saturation persistence is modelable to a few-electron level, whereas saturation-induced persistence remains qualitatively different and makes saturation avoidance important.

6. Flight validation, operational anomalies, and scientific performance

System-level validation of the Euclid infrared detectors was completed through cold instrument-level and payload-level campaigns and then tested again in orbit. In the final cold thermal-vacuum test with the full flight configuration, the NISP flight model exceeded the main optical and detector-related requirements: the point-spread-function encircled energy was better than specified, the signal-to-noise ratio was better than specified, and the spectral performance was also better than specified. The reference expectation quoted in the flight-model presentation is SNR better than 5 in photometric mode and 3.5 in spectroscopic mode for reference sources, and the measured performance surpassed that (Maciaszek et al., 2022).

Two limitations discovered during late integration are important because they are often conflated with detector performance proper. First, the RGS270 grism showed serious wavelength-dependent degradation, acceptable on the blue side but strongly degraded at longer wavelengths owing to a mismatch between mechanical and optical reference frames in manufacturing. The operational response was to discard the RGS270 grism from routine observing while continuing to use the other grisms; the use of the $18\,\mu$0 and $18\,\mu$1 grisms with $18\,\mu$2 rotation offsets preserved equivalent completeness and purity (Maciaszek et al., 2022). This was a spectroscopic-optics issue, not a limitation of the H2RG focal plane itself.

Second, during the CSL cryogenic payload-module campaign only 7 of the 16 detectors worked correctly, while the other 9 showed artifacts and stopped producing data after the first exposure of each sequence. The fault was traced by an ESA-led independent “Tiger” review team together with NISP experts from the Euclid Consortium, with contributions from NASA, OHB-I, Thales Alenia Space, and Airbus, to a DPU firmware/software issue triggered by a specific LVDS timing delay window of about 3 ns introduced by extension harnesses. The root cause was an inconsistency in hardware documentation combined with incorrect DCU–ASIC interface error handling in the DPU application software; a new comprehensive error-handling strategy solved the problem (Maciaszek et al., 2022, Medinaceli et al., 2023). This is a useful corrective to the assumption that all late-stage failures in cryogenic focal planes are detector-physics failures.

In orbit, the detector subsystem has also shown strong scientific utility beyond survey throughput metrics. Early Release Observation analysis of NGC 6397 used NISP photometric imaging with four dithered 87.2 s exposures per filter and the severe undersampling of the $18\,\mu$3 mas pixel$18\,\mu$4 infrared camera to build detector-specific effective PSFs and geometric-distortion solutions. For well-measured stars with good PSF fits, the one-dimensional astrometric precision is about 3 mas, or 0.01 pixel, and Euclid–Gaia residuals after distortion correction are 4.2–6.0 mas for NISP (Libralato et al., 2024). The result demonstrates that the infrared detectors are not only compliant survey sensors but also astrometrically capable devices once detector sampling, ePSF structure, and distortion are modeled jointly.

Taken together, the literature presents Euclid’s infrared detectors as a mature and well-characterized HgCdTe H2RG focal plane whose residual limitations are specific rather than generic. Correlated readout noise is measurable but produces negligible operational bias in the Euclid background regime; persistence is the principal detector systematic and requires explicit modeling, especially near or above saturation; IPC and gain are spatially structured and therefore not adequately represented by detector-mean scalars alone; and late-stage anomalies were either optical-channel or readout-software issues that were understood and mitigated before routine operations (Munoz et al., 2021, Maciaszek et al., 2022, Medinaceli et al., 2023, Kubik et al., 3 Jun 2025, Collaboration et al., 3 Feb 2026).