Background-limited Imaging in the Near Infrared with Warm InGaAs Sensors: Applications for Time-domain Astronomy

We obtained measurements of the combined background from sensor dark current and thermal emissivity by closing the dome and mirror covers during nighttime hours and taking one-minute ramps with 1200 samples. After correcting these frames for linearity, we performed a linear fit of the dark count slope versus time for each pixel on the array. Figure 3 shows an example of one such fit for a single pixel in the left panel, and the right panel shows statistics of the per-pixel dark current across the array, assuming a gain of 1.17 e^–/ADU as measured previously in the lab (Sullivan 2015).

Zoom In Zoom Out Reset image size

We operated our first two nights at T = −40°C and the third night at T = −45°C, and measured the dark current at both temperatures. The median dark current at the lower temperature is 113 e^– s^–1/pixel. A high-end tail is visible and most prominent at the array corners; this tail is suppressed at lower temperatures, and 90% of pixels have a dark value of ≤140 e^– s^–1 as seen in the red cumulative histogram.

This measured dark current is high relative to cryogenically cooled HgCdTe. However, at equivalent temperature (T = −40°C) and pixel size, commercial HgCdTe with 1.6 μm cutoff would have a much higher dark level of 100,000 e^– s^–1/pixel (Beletic et al. 2008), and 2.5 μm HgCdTe would be higher yet. This is the fundamental factor enabling the use of InGaAs for lower-cost warm instruments. Our prior laboratory measurements indicate that further reduction in dark current is possible at lower operating temperature, allowing one to tailor dark current to different site conditions through straightforward thermal engineering.

We did not measure read noise separately at the telescope, but instead used laboratory values taken prior to on-sky deployment. Using flat fields taken at varying illumination levels below the full-well (corrected for nonlinearity), we calculated the conversion gain from the slope at 1.17 e^–/ADU by regressing signal variance against mean. Extrapolating this fit to zero mean flux, we obtained a single-sample read noise of 59 e^– rms.

The delivered read noise is reduced substantially through non-destructive sampling. For up-the-ramp integrations of 64 samples we measure read noise of 43 e⁻ rms; during on-sky operations we always operated in SUTR mode, and except for bright standard stars our ramps always exceed 16 reads, keeping read noise within this bound. For comparable values of read noise, Poisson noise from dark + sky counts overtakes read noise within a 7 s exposure in Y and 3.5 s in J, using the estimates of sky brightness measured as described below.

Nonlinearity is a known issue with all IR sensors, and our previous measurements in the lab revealed nonlinearity at the ∼3%–5% level, varying across the sensor. We calibrate out nonlinearity using flat-field ramps run to saturation at the start of the observing run. For each pixel, we fit a fourth-order polynomial to the residual counts relative to a straight linear fit, regressing the residuals against count rate over the first 35,000 counts in each pixel. We store the polynomial coefficients for each individual pixel and correct each science and calibration frame as the first step in data reduction (before calculating the SUTR slope).

This method reduces the residual observed nonlinearity to ∼0.5% or less (Figure 4). There is evidence in the residuals that a higher-order polynomial could reduce nonlinearity yet further, at the risk of overfitting. We also observe a higher nonlinearity near the start of our ramps (i.e., 1% at the first sample), but the array is essentially never operated in this regime because of robust dark current and sky backgrounds.

Zoom In Zoom Out Reset image size

Data from the camera are stored as SUTR FITS arrays containing individual reads from each sensor reset. A small set of IDL routines is used to correct for nonlinearity in the counts and then calculate the SUTR slope as described in Benford et al. (2008). The output is expressed in terms of e^– s^–1. The nonlinearity correction is critical for maintaining uniformity of the background counts, because the size of the correction varies coherently across the array.

We then subtract a dark current frame (also in e^– s^–1) compiled as described above. The remaining flux from the sky and sources is normalized by a flat-field composite constructed from stacked twilight sky exposures taken with the DuPont. The flat-field reference is corrected for nonlinearity and dark current, and normalized by the median count rate to yield a unity-median calibration frame that is divided into the science frame's count rate. Individual pixels showed gain variations of 1%–5% relative to the median value.

For fields requiring deep photometry, we combined multiple exposures taken at different telescope dither positions. We used the fitsh processing software (Pál 2012), which matches and registers the images to a common reference using direct photometry of (sometimes faint) stars.

We subtracted a spatially constant background flux from each image and applied an overall scaling to match fluxes between exposures. The final composite was constructed from a median of the registered, background-subtracted, and flux-scaled frames. No weights were applied for the average since a median rather than mean was used for the operation. Finally, we reapplied approximate absolute astrometric solutions generated by the Scamp (Bertin 2006) and Swarp (Bertin 2010) packages (from astr0matic.net) to the final stacks for future convenience.

We gathered statistics on the sky backgrounds by recording the median count rate for every exposure taken during the run, in each filter. A histogram of these values is shown in Figure 5.

Zoom In Zoom Out Reset image size

The Y-band background is centered around a median count rate of 103 counts s^–1/pixel, or 121 e^– s^–1/pixel. This is nearly identical to the dark current measured at −40°C and slightly higher than the dark rate for −45°C, indicating that we are nearly sky-noise-limited—even in the band with the darkest sky.

The median J sky flux was 349 counts s^–1/pixel, or 408 e^– s^–1, a factor of ∼4 higher than the dark current. Observations in J (and by extension, H as well) will be dominated by Poisson noise from the sky, with some margin.

We estimated a calibration of the night sky brightness using observations of the spectrophotometric standard star Feige 110 taken during nights 2 and 3 of the run. Using data from the CALSPEC archive at STScI and measured transmission curves of our filters supplied by vendors, we estimated the bandpass center of each filter and compared with measured count rates to establish photometric zero-points. The measured zero-points and sky backgrounds are reported in Table 1.

Table 1.  On-sky InGaAs Prototype Performance

Quantity	Y value	J value
Zero-point (1 e– s–1)	24.53	25.27
Sky background (e– s–1/04 pixel)	120	408
Sky background (mag arcsec–2)	17.5	16.8
Dark + thermal background (e– s–1/pixel)	113	113

Download table as:  ASCIITypeset image

From these calibrations, we estimated the median Y-band sky surface brightness at 17.45 AB magnitudes per square arcsecond, which may be compared with the measured value of 17.50 determined at the same site for Magellan/FourStar (Persson et al. 2013).⁶ Likewise the J-band surface brightness is 16.84 AB magnitudes per square arcsecond, compared to 16.90 measured by FourStar.

There remain systematic uncertainties in our calibration from not accounting for the detailed sensitivity curve of the instrument or variation in the spectral energy distribution of the standard star across each filter bandpass. These calculations merely indicate that our zero-points and methodology produce consistent answers with other instruments at the same site with established heritage. Importantly, they also indicate the possibility of achieving largely sky-limited noise performance using a warm optical train with no Lyot stop and a modestly cooled sensor.

If one wishes to obtain a more favorable noise budget, then further sensor cooling should result in even lower values of dark current (Sullivan 2015). Alternatively, adaptation of the optical design to deliver 05 pixels rather than 04 would increase the sky background by 50% and cover a wider field. This must be traded against the desire to properly sample seeing and achieve maximum point-source sensitivity.

We compiled imaging data on multiple fields with differing depths as we exercised the instrument in various configurations and settled on observing strategies. Because the camera is background-limited, the signal-to-noise ratio (S/N) should scale as $\sqrt{t}$, but a goal of the run was to establish the baseline for this scaling.

In Figure 6 we show photometry for two fields apiece in Y and J. The measurements are generated using Sextractor (Bertin & Arnouts 1996) with a detection threshold of 5σ, yielding object catalogs of object isophotal AB magnitudes (uncorrected for aperture), which we plot against S/N. A regression on the 10 minutes exposure photometry (for NGC 1809 in Y and NGC 1313 in J, shown as solid lines) confirms that sensitivity scales approximately as $\sqrt{{t}_{\exp }}$, as expected for other fields observed at different depths.

Zoom In Zoom Out Reset image size

The shallower slope of the J-band curves reflects the higher sky background, but the overall sensitivity is similar because of the sensor's higher QE in J, and correspondingly higher zero-point. Both filters reach a limiting magnitude of approximately AB = 21 in 5 minutes, AB = 21.5 in 10 minutes, and AB = 22–22.4 in 20 minutes.

For our J-band observations, we cross-checked our photometry calibrated with Feige 110 against stars from the Two Micron All Sky Survey (2MASS, Skrutskie et al. 2006) in the field and adjusted the zero-points to bring the two into agreement. Typically these corrections were ∼0.05 mag except in one case where we observed 0.15 mag of extinction from clouds. For Y we did not have a photometric reference catalog covering observed fields and we used our Feige 110 calibrations without correction.

A more sophisticated estimate of sensitivity or completeness could be made by injecting simulated objects into the data and testing the efficacy of Sextractor at recovering these objects. However, the present analysis provides a sufficiently general estimate of photometric speed to evaluate potential survey science programs, described below.

We observed several selected science targets chosen to represent a range of possible observational programs that would benefit from background-limited IR photometry in the Y through H bands at reduced cost and complexity. These specifically include (a) observations of nearby galaxies typical of astrophysical transient searches, (b) deep images of fields at cosmological distances to search for distant quasars and clusters, and (c) observations of transiting exoplanets in the near infrared.

5.3.1. Nearby Galaxies, including IR Transient Survey Objects

We observed two galaxies from the local universe: NGC 1300 and NGC 1313 (Figure 7). The latter object in particular was targeted because it is used for a synoptic IR survey of obscured transients using the Spitzer Space Telescope. NGC 1300 was observed for 7.6 minutes in Y (composite seeing of 085 FWHM) and 17.2 minutes in J (075 FWHM, marginally sampled), whereas NGC 1313 was observed for 4.3 and 10.4 minutes in the same filters, respectively at 097 and 095 seeing. The unusual exposure times reflect rounding to the nearest clock time for an even number of ramp frames, and were varied throughout the run as we refined our observing strategy.

Zoom In Zoom Out Reset image size

5.3.2. Supernovae and Explosive Transients

A primary motivation for building a wide-field IR camera is to pursue time-domain science in the Y through H bands. Accordingly, we observed two supernovae (SNe) that were visible during the run as a demonstration.

Our first target was iPTF16geu, a gravitationally lensed SN Ia (Goobar et al. 2017). Figure 8 shows Y- and J-band images of the field, with 391 s and 326 s integrations, respectively, taken on 2016 November 13 (UT 00:22:27 for Y in 075 seeing and UT 00:53:58 for J at 083). We measure a total apparent magnitude of Y_AB = 17.4 at 55σ significance, and J_AB = 17.2 at 38σ. Although the data were taken in good seeing conditions, we were not able to resolve individual components of the lens.

Zoom In Zoom Out Reset image size

Our other supernova target is the Type IIb source SN2016gkg (Bersten et al. 2018), observed for 206 s in J (071 seeing) and 152 s in Y (080 seeing, Figure 8). We detect the supernova with Y_AB = 15.9 at 214σ and J_AB = 16.12 at 205σ.

5.3.3. High-redshift QSOs

A key static-sky application for near-IR imagers is the search for high-redshift QSOs in deep data sets. To test whether the InGaAs sensor is sensitive to currently known high-z populations, we observed the known z = 6.31 quasar ATLAS J025.6821-33.4627 (Carnall et al. 2015). This object was (somewhat unusually) selected from z − W1 (WISE) colors, so its near-IR magnitudes were not reported in the literature.

Figure 9 shows a cutout of the J-band image, which was observed for 326 s in 074 seeing. We detect a source at the expected location of the QSO with J_AB = 18.91 and 20σ significance. The quasar was also observed for 717 s in Y and a source is clearly visible in the data. However, poor image registration (caused by a low star count in the high-latitude field) complicates photometry and requires further refinement to produce a well-calibrated measurement.

Zoom In Zoom Out Reset image size

5.3.4. Faint/High-redshift Galaxies and/or Clusters

To push photometric depth, we constructed a 17 minute Y-band stack in 08 seeing of the field containing the z = 0.87 galaxy cluster ACT-CL J0102-4915 (nicknamed "El Gordo"; Menanteau et al. 2012). Using the zero-points listed in Table 1, we extracted magnitudes for all sources detected with ≥5σ significance—these are circled in Figure 10. Although the extremely luminous brightest cluster galaxy is well above the noise floor at Y ≈ 18, we also detect numerous objects from the red sequence at y = 21.5–22.5, our approximate detection limit.

Zoom In Zoom Out Reset image size

This observation, together with the z = 6.3 QSO observations presented in Section 5.3.3 suggests that wide-field InGaAs mosaics can deliver sufficient imaging depth to survey, discover, and characterize objects at cosmological distances.

5.3.5. Exoplanet Transit

Our early studies with InGaAs were partially motivated by an interest in using the sensors for exoplanetary transit surveys around low-mass stars, including L and T dwarfs which are bright in the J band. Our earlier work (Sullivan et al. 2013, 2014) included laboratory tests of photometric stability, but did not present on-sky detections of transit events.

We did not schedule the DuPont run explicitly for optimal observation of an exoplanet transit, but a database search⁷ revealed a small number of partial transits that were visible from Las Campanas during our run. Because of the InGaAs camera's small field size, we focused our search on targets with bright nearby comparison stars.

We observed the newly discovered Hot Jupiter HATS-34b, a 0.94 Jupiter-mass planet orbiting a V = 13.9/J = 12.5 (Vega) host star of T_eff = 5380 K. The object orbits with P = 2.1 days and was visible through ingress and transit during last portion of the night. Its field includes a second comparison star 2.5 mag brighter than HATS-34, at a projected distance of 163''. Sunrise prevented observations of the transit egress and re-establishment of a post-transit photometric baseline. The transit depth reported in the discovery paper (de Val-Borro et al. 2016) is 13.4 mmag, or 1.2%.

So as not to saturate either star, we obtained short individual exposures of t = 1.03 s, with the telescope slightly defocused. The telescope was guided throughout the sequence by an off-axis probe provided by the observatory, to maintain stable positioning of objects on the array.

We extracted fluxes of the science target and reference star using Sextractor in strict aperture photometry mode, with an aperture diameter of 15 pixels (6''). No attempts were made to optimize the photometric extraction parameters or aperture.

Figure 11 shows the resulting light curve, constructed from differential photometry between the target and reference stars. To reduce shot noise, we bin the counts from multiple exposures by adding the photometric fluxes with a top-hat window of varying width, to verify the scaling of noise reduction. We center around the known time of transit calculated from the HATS-34b discovery paper. The black solid points are averages of 11 exposures, or 11.33 s, whereas the red points average 80 exposures for 82.4 s.

Zoom In Zoom Out Reset image size

Even in the 11 s averages, there is a clear transit detection with a depth consistent with the value reported by the HATS team (the expected optical transit depth is indicated with a light green line). In the 82 s bins the detection is highly significant, with a transit depth slightly larger than the 1.2% predicted by HATS, but within the margin of error.

Our main objective was not to fit a new transit light curve and re-derive the orbital properties of HATS-34b. We simply demonstrate that InGaAs sensors are capable not just of deep photometry and detection of explosive transients, but also of precision photometry at the millimagnitude level over long time baselines. These data were obtained in the J band where the sky is brighter and more variable in emission and transparency than in the optical. This suggests that with proper attention to noise, stability, and observation InGaAs cameras can offer an affordable alternative to costly HgCdTe arrays for IR transit observation—particularly if a large format is not required.

Wide-field near-IR imagers are potentially attractive survey instruments to search for and characterize the electromagnetic (EM) counterparts of binary neutron star (BNS) mergers detected via gravitational waves. Pioneering work by Li & Paczyński (1998) and Metzger et al. (2010) developed predictions for an isotropic EM "kilonova" signature powered by radioactive decay of heavy elements synthesized via neutron capture. Using opacity tables for heavy elements in the lanthanide series, Kasen et al. (2013) and Barnes et al. (2016) further predicted that the most long-lived emission from these events will emerge in the Y through H bands. This results from an increased opacity at optical wavelengths relative to the iron peak elements seen in conventional supernova photospheres.

The recent discovery of an apparent kilonova (or similar "macronova"; Metzger & Fernández 2014; Kasen et al. 2015) associated with GW170817 (Abbott et al. 2017; Arcavi et al. 2017; Coulter et al. 2017; Cowperthwaite et al. 2017b; Smartt et al. 2017; Soares-Santos et al. 2017; Tanvir et al. 2017; Valenti et al. 2017) supports this basic picture, although the observed optical counterpart is far brighter than expected for material whose opacity is dominated by the heaviest r-process elements. This event therefore offers an ideal opportunity to assess the relative merits of wide-field optical versus IR imagers for follow-up of anticipated future BNS events.

Figure 12 shows the emergent spectrum predicted by the models of Kasen et al. (2017) for varying mass fractions of lanthanide elements in the post-merger ejecta, all at t = 2 days. Kasen's full model of GW170817 required two components: one polar outflow containing light r-process elements moving at high speed, and one isotropic component of emission from 0.04 M_⊙ of heavy r-process material moving at v_ej = 0.1c. The latter, isotropic component is shown in the top panel of the figure. Filter curves for MKO Y, J, and H_s (conventional H multiplied by the InGaAs QE cutoff) illustrate that the flux density of lanthanide-rich tidal debris is higher at near-IR wavelengths than in red optical bands, partially offsetting the built-in cost and heritage advantage of existing CCD-based search strategies.

Zoom In Zoom Out Reset image size

On the other hand, the fast outflow has been associated with a short-lived (t ∼ 1 day), blue EM transient, whose exact mechanism is not fully constrained nor is it known whether such short-lived, bright optical counterparts are generic to kilonova events or are orientation-dependent. Two possible early models include a gamma-ray burst seen at an off-axis angle (Margutti et al. 2017), and shock-heated gas from a cocoon of material surrounding a relativistic jet that has either been choked off before emerging or has just emerged but is seen off-axis (Nakar & Sari 2012; Kasliwal et al. 2017; Piro & Kollmeier 2018). Late-time radio observations appear to favor such cocoon models (Mooley et al. 2018).

The IR emission driven by radioactive decay of heavy lanthanides should be largely isotropic and visible for a week or more (Metzger et al. 2010; Kasen et al. 2013), it is a ubiquitous prediction of the model and should be visible for any viewing orientation. The bottom left panel of Figure 12 plots the observed J-band light curve of GW170817 (Cowperthwaite et al. 2017b; Drout et al. 2017; Villar et al. 2017), now projected to distances of 100, 150, and 190 Mpc. The last of these corresponds to the expected horizon distance for NS–NS merger detections in LIGO's fourth observing run (Abbott et al. 2016).

Our demonstration camera would detect the fading counterpart of GW170817 in J using survey integrations of 5 minutes (green shaded region) at 5σ significance for a week after merger at 100 Mpc, or 4 days at 150 Mpc, and would just miss a detection in 5 minutes at LIGO's maximum distance. If 10 minute mapping integrations are used (yellow shaded region), the transient is visible for 9 days at 100 Mpc, and 5 days at 190 Mpc. The (z − J) color of the slow, heavy wind rises steeply with time, exhibiting 1–2 mag of reddening over the first few days following merger (Figure 12, bottom right).

With these plots and the sensitivity curves from Figure 6, we can assess the relative merits of an optical survey in the i' or z' bands versus an infrared InGaAs camera in Y, J, or H.

For concreteness, we assume a fiducial field of view of 1 deg² tiled with InGaAs, consistent with an underfilled focal plane on the 2.5 m DuPont or SDSS telescopes (which have a field of ∼3.1 deg²), although similar scaling arguments can be developed for smaller apertures. If we further assume a 10 minute integration cadence, Figure 6 indicates a 5σ depth of J_AB = 21.5–21.7. At this speed, one could map ∼6 deg² per hour with sufficient sensitivity to detect the EM counterpart of GW170817 at 150 Mpc for 7 days. At the actual distance of 40 Mpc for GW170817, the r-process peak would be visible for over two weeks in 10 minute exposures. At the edge of the Advanced LIGO design horizon of 200 Mpc, it would be visible for 5 days.

Simulations suggest that in the era of two GW detectors (i.e., Advanced LIGO O3), roughly half of all BNS mergers will be localized within 150 deg² (Chen et al. 2017), and only 10% will have localizations of 50 deg² or better. However, if the positional region of uncertainty is above the horizon for 3–4 hr per night, one may survey 18–24 deg² per day to J_AB = 21.5. This is sufficient to tile the most well-localized 10% of events in two nights, or to survey even unfavorable two-detector error contours in four to five nights (during which time the IR transient is still visible throughout the full 190 Mpc volume).

The same simulations find that inclusion of VIRGO (Accadia et al. 2012) detections during O4 will localize ∼70% of events within approximately 5 deg². This region could be surveyed with 1–2 hr cadence to the same depth, but would more likely be integrated to ≥1 hr exposure depth, reaching J_AB = 22.5 or fainter. This would allow IR instruments to identify mergers with lower ejected mass yields or velocities, consistent with prior expectations for BNS mergers (Kasen et al. 2015). It also provides more margin for events that are obscured by dust or inconveniently placed on the sky. The lower cost of InGaAs could enable blind searches for NS–NS merger counterparts directly in the IR, in parallel with similar searches at optical wavelengths.

Because the EM counterpart of GW170817 was discovered in the optical, and existing CCD imagers of larger etendue will already be searching for the same counterparts, one must consider whether there is added value in contemporaneous IR searches using smaller apertures and/or fields. Indeed, a rapidly fading optical transient was the first and brightest signal seen from this event on the ground.

There are three possible motivations for pairing deep optical searches with wide-field IR mapping (as opposed to post-discovery follow-up). First, the radioactively heated IR transient is widely believed to be isotropic, while the angular dependence of the UV–optical radiation is unconstrained at present. An accounting of the fractional contribution of UV–optical transients to the parent population of IR-triggered events will define the geometry and uniformity of the EM mechanisms.

Second, a generic signature of heated-cocoon models is a rapid optical transient that fades by over 1 mag per day, concomitant with a rise in the J and H bands over similar timescales—in other words, all bands bluer than J fade continuously. Cowperthwaite et al. (2017a) demonstrated that optical searches can map regions of ∼50 deg² to i' = 22.5, sufficient to detect a rapid UV–optical transient similar to GW170817 for 3–4 days. Optical searches would not detect emission from the isotropic r-process material; they require favorably oriented jets, cocoons, or disk winds to successfully identify EM counterparts.

However, Cowperthwaite et al. (2017a) also find that the optical searches exhibit a false-positive rate of ∼2 unrelated transients per square degree in their blind search; after applying priors on color this rate is reduced to ∼1 deg⁻² if kilonovae are assumed to be intrinsically blue, or ∼0.15 deg⁻² if they are intrinsically red as in r-process events. A blind search of ∼50 deg² would therefore yield between seven and 100 contaminants depending on priors applied. Early-time monitoring in the J or H bands could establish the simultaneous IR brightening and sustained luminosity unique to NS–NS binary mergers, separating BNS events from unrelated foregrounds.

Third, the long duration of EM signals in the infrared offers an extended window over which to monitor events, in contrast to high-etendue optical campaigns which cannot always encumber the resources of large telescopes for follow-up on timescales of weeks.

It is projected that Advanced LIGO and VIRGO may discover NS binaries at a rate of one per month at design sensitivity, such that follow-up of these events consumes a substantial portion, but not all, of small observatories' time allocations. During times when targeted follow-ups are not underway, such a survey instrument could perform a dedicated wide-field time-domain survey in the IR, a program that has not been undertaken to date largely on account of sensor costs.