5.7 IR Detector Characteristics and Performance

5.7.1 Quantum Efficiency

The quantum efficiency (QE) of the flight IR detector, as measured at the Goddard Detector Characterization Lab (DCL), is shown as a solid curve in Figure 5.25.  The actual total system throughput of WFC3 depends on many factors including the HST OTA, pick off mirror, filter transmission functions, QE, etc. Based on ground measurements of these quantities, the total system throughput was calculated and compared to the first on-orbit measurements. A 5–20% increase in the total system throughput was discovered, attributed to multiple factors. The dashed curve represents the QE under the assumption that the entire flight correction is in the QE. Though this is admittedly unphysical given the realities of anti-reflection coatings and interpixel capacitance, computed total system throughputs are correct when the QE file is used in conjunction with the other optical components (pick off mirror, OTA, IR channel optics, and filters).

5.7.2 Dark Current

To avoid the complexity and limited lifetime of a stored-cryogen system, while at the same time providing the low operating temperatures required for dark-current and thermal-background reduction, the WFC3 IR detector is refrigerated with a six-stage TEC to a nominal operating temperature of 145 K. This is an unusually high operating temperature for near-IR detectors, and required tailoring the composition of the HgCdTe material for a long-wavelength cutoff at ~1700 nm. The higher band-gap associated with the short cutoff wavelength effectively limits both the intrinsic detector dark current and its sensitivity to the internal thermal background.

Direct thermal control of the detector (via a sensor integrated in the MUX that controls the 6-stage TEC current) provides typical thermal stability of < 50 mK. Tests made on similar detectors indicate that the residual dark-current variations can be largely calibrated and subtracted out using reference pixels.

WFC3 IR exposures taken with an aluminum blank in place, rather than a filter, provide a measure of the detector dark current. The dark current of the flight array has a skewed distribution, with a mode, median, and mean of 0.045, 0.048, and 0.048 e^–/s/pixel respectively. The shifted mode is due to the asymmetry of the dark-current distribution among the pixels, characterized by a long tail of “hot pixels” randomly located across the detector. While the mean underlying dark current has been stable on-orbit (WFC3 ISR 2017-04; WFC3 ISR 2012-11), a small number of pixels (<4%) vary over time and are corrected in the calibration pipeline via time-dependent superdarks (WFC3 ISR 2019-04).

The histogram of dark current values, along with the cumulative dark-current distribution, i.e., the fraction of pixels with a dark current lower than a certain level, is shown in Figure 5.26; WFC3 ISR 2009-21 provides details on dark current calculations. Improved superdark reference files were created for all allowed full-frame and subarray modes using data collected during 2009-2013 (WFC3 ISR 2014-06). The signal-to-noise improved by a factor of 3-11 due to the use of a great deal more data, the use of a non-linearity correction, and the use of persistence masks. In 2019, updated WFC3/IR dark calibration files were created for all allowed observing modes. Time-dependent pixels in these superdarks were generated in a cycle-by-cycle basis, allowing for improved calibration of those ∼3.5% of detector pixels that have changed their dark current behavior since launch (WFC3 ISR 2019-04). 

Note that in broad filters, the dominant background is not the dark current but zodiacal light, which at 0.3-1.0 e^–/s/pixel is a factor of 10-20 times larger than the dark. The WFC3 ETC can be used to compute the zodiacal light contribution for a given pointing, in addition to providing thermal and dark current estimates (Sections 7.9.5 and 9.7).

5.7.3 Read Noise

The IR detector has four independent readout amplifiers, each of which reads a 512 × 512 pixel quadrant. The four amplifiers generate very similar amounts of read noise. This is illustrated in Figure 5.27, which compares the correlated double sampling (CDS) read noise levels for the four quadrants of the detector. CDS read noise refers to the noise associated with subtracting a single pair of reads. These read noise values were derived from a series of RAPID ramps taken during SMOV testing, providing a measure of the total noise in a difference image. For short ramps, such as these RAPID ramps, the contribution of shot noise due to dark current accumulation is less than 0.01 e^– and the CDS read noise of the detector is between 20.2–21.4 e^–  (left plot in Figure 5.27). By averaging over multiple reads, the effective noise of an IR ramp can be significantly reduced. As Figure 5.27 (below, right plot) illustrates, the effective noise in a SPARS200 ramp drops from ~20.0 e^– down to ~12.0 e^– (based on 2 reads plus zeroth read and 15 reads plus zeroth read, respectively). Similar reductions in noise can be achieved with other sample sequences (WFC3 ISR 2009-23).

For some programs, read noise will not be an issue while for others, such as ultra-low-background observations, the read noise can be a non-negligible component of the noise floor. The relative contribution of read noise to the total noise will depend, of course, on infrared background levels as well (Section 7.9.5). The contribution to the read noise in WFC3 IR data due to digitization errors associated with the conversion from electrons to data numbers (DN) is negligible.

5.7.4 Flat Fields

Before launch, ground-based flats were obtained for the 15 imaging IR filters at a mean SNR of ~500 per pixel using an external optical stimulus (ISR 2008-28). Because the overall illumination pattern of the ground-based flats does not precisely match the illumination attained on-orbit from the OTA, there are errors in these ground-based flats on large spatial scales. These errors were initially measured by performing stellar photometry on rich stellar fields that were observed using large-scale dither patterns during SMOV and cycle 17 (ISR 2009-39). The errors were later determined more accurately by creating sky flats from thousands of on-orbit exposures, masking out astronomical sources and assuming filter-independent corrections. Initial flat field reference files corrected using these sky flats were delivered in December 2011 (ISR 2011-11).

These reference files were replaced in October 2020 with flats generated using a much larger set of in-flight data (ISR 2021-01). Six filters (F098M, F105W, F110W, F125W, F140W, and F160W) have sufficient signal-to-noise that the flats are now derived purely from in-flight data and are no longer dependent on the ground test data. For the remaining nine IR filters, the ground "P-flats" (pixel-to-pixel variation flats) were multiplied by a smooth correction derived from wavelength-interpolation using sensitivity residuals from the six primary filters. An accompanying set of delta flats (also known as "D-flats"; ISR 2021-10) was created for the six primary filters to correct for IR blobs at different epochs (Section 7.9.6). When two or more types of flats are available, they are multiplied together by calwf3 to form a combined flat field correction image.

Figure 5.28 shows examples of bias-corrected P-flats computed from in-flight science observations acquired over ~10 years in the wide-band filters F110W (left) and F160W (right).

5.7.5 Linearity and Saturation

The WFC3 IR calibration program shows that the detector response is in fact (slightly) non-linear over the full dynamic range. This behavior is illustrated in Figure 5.29, which presents a plot of average counts as a function of time. The black diamonds are the measured average signal; a linear fit has been made to the signals up to ~25,000 electrons (solid red line). The dashed red line shows this best-fit line extended out to the total exposure time of the ramp. The blue horizontal line marks the level at which the counts deviate by more than 5% from linearity (about 78,000 electrons). For the purposes of non-linearity correction, the 5% nonlinearity level has been defined as “saturation.”

The linearity correction implemented in the WFC3/IR calibration pipeline corrects pixels over the entire dynamic range between zero and saturation. Once the pixel value exceeds the saturation threshold, the pixel is flagged as saturated in the data-quality array within the FITS file and no linearity correction is applied. Pixels driven heavily into saturation can begin to show decreasing readout values, such that their DN values fall back below the defined saturation threshold. To prevent a situation where a pixel is flagged as saturated in one or more readouts, but then not flagged in later readouts, the calibration processing system flags saturated pixels in all subsequent readouts for pixels that are found to be above the saturation threshold in any given readout.

Trials of non-linearity corrections have shown that a third-order fit to the measured linearity versus signal for each pixel provides a slightly better correction than the one currently implemented in the pipeline, with photometric results between short and long exposures more consistent by up to 0.5% (WFC3 ISR 2014-17). 

5.7.6 Count Rate Non-Linearity

Previous HgCdTe detectors on HST have suffered from a count-rate dependent non-linearity (CRNL) and the WFC3-IR detector is no exception. An initial measurement of this effect was made by comparing the photometry of star clusters observed over a wide dynamic range and at overlapping wavelengths in the WFC3/IR and NICMOS and/or ACS/WFC detectors. We found a CRNL of ~1% per dex over a range of 10 magnitudes (4 dex) which was independent of wavelength (WFC3 ISR 2010-07). This measurement was confirmed using exposures that boosted count rates with Earth limb light (WFC3 ISR 2010-15) and observations of groups of stars observed with 2MASS (WFC3 ISR 2011-15). The impact of count-rate non-linearity is that photometry of faint (i.e., sky dominated) sources calibrated with WFC3/IR zeropoints will appear 0.04 +/-0.01 mag too faint. The effect is an order of magnitude smaller than the effect found for NICMOS, but large enough to potentially limit the accuracy of photometry.

In 2019, more precise measurements of CRNL were made by using a combination of comparisons of cluster star photometry between WFC3/IR and WFC3/UVIS and by using observed and synthetic magnitudes of white dwarfs (WFC3 ISR 2019-01). In this study, the measured range of CRNL was also extended to higher count rates by comparing magnitudes between the ground and WFC3/IR for LMC and Milky Way Cepheids. Combining these results with all previous measurements and those from the WFC3 grism provides a consistent and improved characterization of the CRNL of WF3/IR, of 0.75% +/- 0.06% per dex, with no apparent wavelength dependence, measured across 16 astronomical magnitudes. This result should be used to correct IR photometry by using the difference in apparent flux (in dex) between where the WFC3/IR zeropoint is set (~12th mag) and the target source. Fainter sources appear even fainter and thus must be corrected to be brighter.

5.7.7 Detector Cosmetics

The WFC3-IR detector’s pixel population includes several flavors of anomalously responsive pixels: hot, cold, unstable, dead, and deviant in the zeroth read. Hot pixels, those showing excess charge, are defined as pixels with more than 100 times the average dark current. Cold pixels are inversely sensitive to incident photons and exhibit a negative slope when measured up the ramp (i.e., pixel value is lower in last frame up the ramp compared to first frame). The anomalous response of a cold pixel could be due to lower intrinsic QE in that pixel or to surface defects. Unstable pixels, as the name implies, are those that behave in an unpredictable fashion; that is, the signal up the ramp does not repeat reliably from ramp to ramp (Appendix 2, WFC3 ISR 2010-13 show examples). There are dead, or unbonded, pixels which do not respond to light (Figure 5.30). Overlapping the dead pixel population is the population of pixels that has bad zeroth read values, generally due to being short-circuited or unbonded (WFC3 ISR 2003-06).

In addition to randomly-distributed bad pixels, coherent regions of bad pixels exist in the IR detector (Figure 5.30). Pixels in the lower-right region (dubbed “wagon wheel”) have lower than normal quantum efficiency. There are dead pixels near the detector edge and in the circular “death star” feature near the bottom. Pixels with deviant zeroth read are concentrated in the areas of the death star, the upper corners of the detector, and the quadrant boundaries. (The death star region is marked in the WFC3 FOV in APT to aid in observation planning.) WFC3 ISR 2008-28 describes the characterization of these defects based on ground-testing data; WFC3 ISR 2010-13 describes the various types of populations of bad pixels as observed on-orbit.

An in-depth study of the time-dependent behavior of the IR pixels (WFC3 ISR 2019-03) has shown that roughly 3.5% of pixels have experienced changes in their behavior since the installation of WFC3 on-orbit in May 2009. Such variations include normal pixels becoming hot, unstable pixels becoming stable, some pixels transitioning between different stable levels of dark current, other pixels alternating between two stable levels, and more. Figure 5.31, reproduced from WFC3 ISR 2024-02, shows the distribution of IR pixel behavior for the various combinations of cold (i.e. normal dark rate) and hot, stable and unstable states.

In general, the WFC3/IR detector shows an increase of only ~200 hot pixels per year (~0.02% of the detector). Prior to 2020, the unstable pixel population varied up to ~4000 pixels per year (~0.4% of the detector). It was observed that the unstable pixel population more than doubled between Cycle 27 and Cycle 28 (WFC3 ISR 2022-01), such that approximately ~3.35% detector pixels were classified as unstable (2.85% cold and 0.5% hot). However, WFC3 ISR 2024-02 found that this was due to an algorithmic discrepancies in the pipeline used for generating the bad pixel table; subsequent analysis showed that the actual flagged pixel percentage was <1% for Cycle 28, and that IR bad pixel populations remained stable through Cycles 29 and 30. 

Calibration reference files in the past were insensitive to periods of pixel stability, simply permanently flagging as bad any pixels which spent any amount of time in a hot and/or unstable state. Now, time-variable pixel behavior is captured in the bad pixel tables and superdarks with a data quality flag of 32 to mark unstable pixels (as is now done for UVIS as well); e.g. unstable hot pixels are flagged with both ‘32’ and ‘16’ (hot) (Table E.3). As a result, users can reclaim stable pixels – even when hot – for their analyses; doing so can recover about 0.5% of the detector, which otherwise would have been thrown out. Likewise, the WFC3/IR bad pixel tables now track the time-variable bad pixel populations, which ensures that new bad pixels are flagged in a timely fashion and that the thousands of pixels which revert to a stable configuration are not flagged and discarded (WFC3 ISR 2019-03, WFC3 ISR 2019-04). 

5.7.8 Crosstalk

As is common in devices with multiple amplifiers being read out simultaneously, the IR channel exhibits crosstalk: a bright source in one quadrant causing electronic ghosting in another quadrant. In the IR, the crosstalk manifests itself as a very low-level negative mirror image. In WFC3, amplifiers 1 and 2 are coupled (upper left and lower left quadrants; Figure 5.23) and amplifiers 3 and 4 are coupled (lower right and upper right quadrants). That is, sources in quadrant 1 generate crosstalk in quadrant 2, sources in quadrant 2 generate crosstalk in quadrant 1, and so on.

The level of the IR crosstalk is only ~1e^–06 that of the target flux (WFC3 ISR 2010-02); for unsaturated sources, the crosstalk is below the background noise. Once a source saturates, the crosstalk becomes visible at about the level of the background and remains constant as the voltage of the device is pinned.

5.7.9 Image Persistence

Image persistence is a common problem in HgCdTe and other types of IR arrays. Persistence manifests itself as ghost images or afterglows from earlier exposures. It was seen in NICMOS, and is also seen in a small but non-negligible fraction of the exposures taken with the Hawaii 1R detector that is the heart of the WFC3 IR channel.

Persistence is caused by traps that exist in the active regions of the reverse-biased diodes that make up the pixels of the detector. Resets, which occur at the end of multi-accum exposures (and during the process of flushing the detector when not observing with the IR channel), maximize the reverse bias of the diodes. Light impinging on the diode creates photo-electrons which cause the reverse bias to decrease. Changing voltages within the diode expose portions of the depletion region to free charge. Dislocations in these newly exposed regions trap charge. More traps are exposed for bright sources than for faint ones. This trapped charge is released in later exposures, resulting in after-images. The greater the saturation of the detector, the greater the number of traps and the greater the afterglow. Smith et al. 2008 (Proc. SPIE, 7021) has provided a very clear description of the physics of persistence and the effects in IR arrays.

The characteristics of persistence vary for different devices and device technologies, reflecting in part how traps are distributed within the diodes. Persistence in the WFC3 channel is primarily a function of the fluence (the total number of photo-electrons released) in an exposure, and secondarily a function of the amount of time the pixel is held at a high fluence level. As discussed by Long et al. 2012 (Proc. SPIE, 8442), the amount of persistence in the WFC3-IR detector is a non-linear function of the fluence. Persistence is observed mainly in situations where fluence levels approach or exceed saturation of the detector.

Several examples of persistence in WFC3 observations and strategies for avoiding persistence are described in Section 7.9.4. A description of a phenomenological model of persistence used to aid in removing the effects of persistence is given in the WFC3 Data Handbook.

Figure 5.32 shows the characteristic shape of persistence versus fluence as observed in a series of darks following an image of Omega Centauri which had been deliberately exposed to a level where many stars in the image were saturated. The first dark exposure took place a few minutes after the end of the Omega Centauri exposure and the last dark exposure took place about one orbit later (WFC3 ISR 2013-07). The amount of persistence is fairly small until the exposure level reaches about half of full well and saturates near full well exposure. The persistence gradually decays with time from the first dark exposure (highest curve in figure) to the last dark exposure (lowest curve in figure).

Persistence decays roughly as a power law with time, as illustrated in Figure 5.33, which is based on the data displayed in Figure 5.32. The different curves here show the decay for different levels of saturation, as measured in electrons. Persistence at low fluence levels decays more rapidly than persistence at high fluence levels. There are 3 curves for each level corresponding to the 3 times the experiment was repeated. The differences are partially due to the fact that different pixels were illuminated to different levels each time, but may also indicate some intrinsic variability that is not understood. For comparison, the dark current is about 0.05 electrons/sec. If one assumes that a power law describes persistence from 100 to 10,000 seconds after an exposure, then one concludes that about 3% of charge is trapped in an exposure that has a nominal fluence level of 100,000 electrons.

 

While fluence is the primary factor in determining how much persistence there will be after an observation, the amount of persistence actually depends on the time history of each pixel. Tests show that there is more persistence from a pixel exposed multiple times to the same brightness source (WFC3 ISR 2013-07) and the longer a pixel is held at a fixed flux level (WFC3 ISR 2013-06). This can be understood qualitatively as being due to the fact that traps have finite trapping times. A more accurate prediction of persistence has been achieved using an exposure-time dependent power law decay model (WFC3 ISR 2015-15) along with a “correction flat” that takes into account large-scale variations over the detector (WFC3 ISR 2015-16). Observers are cautioned that variability in persistence has been found in dark exposures taken within 1000 sec of a brief (274 sec) stimulus (WFC3 ISR 2018-03).

MAST includes a search form that provides a persistence image for a specified exposure, produced by applying the persistence model to the preceding WFC3-IR exposures (see the WFC3 Data Handbook). Since the model is imperfect, the persistence image is intended as a guide to which pixels to flag in an exposure rather than as a reliable indicator of flux corrections.