WATER VAPOR IN THE SPECTRUM OF THE EXTRASOLAR PLANET HD 189733b. II. THE ECLIPSE

Spectroscopic observations of exoplanets are crucial to infer the composition and properties of their atmospheres. The eclipse, when the planet passes behind the star, probes the dayside emission spectrum of the planet's atmosphere. The expected signal for molecular signatures is a few 100 ppm; this extreme precision is currently achievable only for planets orbiting bright stars. As a result, only a handful of exoplanets have been characterized spectroscopically. With a host star of magnitude J = 6.07, HD 189733b is one of the most extensively studied exoplanets along with HD 209458b. HD 189733b is a 1.144 M_jup, 1.138 R_jup gas giant planet orbiting an active K dwarf in 2.22 days (Bouchy et al. 2005). Spectra and phase curves obtained at several wavelengths delivered quantities of information about its atmosphere. Grillmair et al. (2007, 2008) obtained spectra in emission with Spitzer, revealing strong water vapor absorption and the presence of an extra-absorber in the dayside upper atmosphere. A hot spot was detected eastward of the substellar point revealing an equatorial super-rotating jet (Knutson et al. 2007), as anticipated by atmospheric circulation models of hot Jupiters (Showman & Guillot 2002). Brightness temperature measurements constrained the pressure–temperature profile of both sides of the atmosphere and the day–night heat redistribution efficiency (Knutson et al. 2012; Désert et al. 2011). In addition, signatures of non-equilibrium chemistry were reported (Knutson et al. 2012; Swain et al. 2010). Several molecules were identified in the planet's transmission and emission spectra with Hubble Space Telescope (HST) NICMOS (Swain et al. 2008, 2009, 2014). However, subsequent studies of NICMOS data concluded that instrumental systematics may contribute significantly to the observed spectra (Gibson et al. 2011; Crouzet et al. 2012). More recently, CO and H₂O were detected in the dayside of the planet's atmosphere using ground-based high-resolution spectroscopy with Keck II NIRSPEC and VLT CRIRES (Rodler et al. 2013; de Kok et al. 2013; Birkby et al. 2013). A new era for exoplanet spectroscopic observations began with the installation of the Wide Field Camera 3 (WFC3) on board HST in 2009. Gibson et al. (2012) observed two transits of HD 189733b with WFC3. Unfortunately, the central part of the spectrum saturated the detector; only the edges were used to derive the planetary radius in two wavelength ranges. The recently implemented spatial scanning mode (McCullough & MacKenty 2012) now allows WFC3 to observe HD 189733 in much better conditions: this mode is designed to obtain high sensitivity on bright stars for exoplanet spectroscopy by reducing overheads and avoiding detector saturation. Observations of HD 189733 with WFC3 in spatial scanning mode (HST program 12881; PI: McCullough) were obtained during a planetary transit (McCullough et al. 2014) and during a planetary eclipse (this paper).

Although clear atmosphere models are often used to interpret exoplanetary spectra, recent results suggest that clouds and hazes may play an important role in the atmosphere of exoplanets. Atmospheric clouds would result in an increased opacity that may obscure individual molecular spectral features. As examples, spectra of the hot Jupiters HD 209458b and XO-1b with HST WFC3 in spatial scanning mode showed a water vapor absorption feature at 1.4 μm with an amplitude much smaller than expected from clear atmosphere models (Deming et al. 2013). The extreme case of an absence of molecular features would indicate an atmosphere completely dominated by opaque clouds. Using two-channel near-infrared photometry with NICMOS during transits of HD 189733b, Sing et al. (2009) found results consistent with Rayleigh scattering from haze. No water vapor absorption was detected. Similarly, Gibson et al. (2012) suggested that haze may dominate its near-infrared transmission spectrum. Higher layers of the atmosphere of HD 189733b have also been probed. High altitude haze was inferred by a UV transmission spectrum dominated by Rayleigh scattering (Lecavelier Des Etangs et al. 2008; Sing et al. 2011; Huitson et al. 2012), which would explain the narrowness of alkali features and the increased planetary radius in the UV compared to the infrared. After reanalyzing HD 189733 data over a wide wavelength range, Pont et al. (2013) proposed a general explanation for the atmosphere of HD 189733b at its terminator: Rayleigh scattering by clouds dominates in the UV, and settling dust and a cloud deck yield a featureless spectrum in the near infrared and in the infrared. A marginal detection of decreasing geometric albedo in the visible may also indicate optically thick reflective clouds on the dayside of the planet, although with a low albedo overall (Evans et al. 2013). Beyond the case of hot Jupiters, examples of featureless transmission spectra are GJ1214b (Berta et al. 2012; Kreidberg et al. 2014), GJ436b (Knutson et al. 2014a), and HD 97658b (Knutson et al. 2014b), although for low-mass planets, flat transmission spectra could alternatively be explained by high mean molecular weight atmospheres. In all cases, the inferences of hazes/clouds are motivated by the non-detection of spectral features of expected molecules and a blueward rise in the optical part of the transmission spectrum. However, observational constraints on the chemical composition of the haze/cloud material are non-existent. In this context, measuring the transit and eclipse spectra of HD 189733b in the near infrared, where molecular features are expected, is crucial. The observations presented in this paper and in McCullough et al. (2014) bring new constraints on the cloud hypothesis for the atmosphere of HD 189733b.

For tidally locked exoplanets such as HD 189733b, one might expect different atmospheres on the dayside, the nightside, and along the terminator. However, very strong equatorial winds would tend to homogenize both sides by increasing the heat redistribution and mixing the composition. In the case of HD 189733b, the heat redistribution was found to be relatively small (Grillmair et al. 2008; Knutson et al. 2012). Measuring the transit and eclipse spectra of HD 189733b in the same wavelength range with the high precision of WFC3 spatial scans offers a unique opportunity to consistently probe these two regions of the atmosphere.

We present the observations in Section 2, the data reduction in Section 3, and the resulting spectrum in Section 4. We perform a white-light analysis in Section 5. Results are then discussed in Section 6, followed by a summary in Section 7.

We used HST WFC3 with the newly implemented spatial scanning mode, developed in part to enable observations such as these (McCullough & MacKenty 2012). In this mode, a controlled scan is applied to the telescope during the exposure in a direction perpendicular to the wavelength dispersion direction (Figure 1). This technique is particularly efficient for bright stars such as HD 189733 (see McCullough et al. 2014, for more details). Several programs have already benefited from this technique (Deming et al. 2013; Wakeford et al. 2013; Kreidberg et al. 2014; Knutson et al. 2014a, 2014b).

Zoom In Zoom Out Reset image size

One eclipse of HD 189733b was observed on 2013 June 24. The observations are divided into five HST orbits, the planetary eclipse occurring during the fourth orbit. In total, 159 exposures of 5.97 s each were acquired, corresponding to 32 exposures per orbit (except for the first orbit in which the first image is a direct image). We used the G141 grism covering a spectral range from 1.1 to 1.7 μm, the 512×512 pixel subarray, and a scan rate of 2 arcsec s⁻¹. The spectral trace is spread over 111 detector rows by the spatial scan and 150 columns by the dispersive element. The resulting spectral resolution is R = λ/Δλ = 130 (Dressel 2014). The detector is read out in MULTIACCUM mode with the RAPID sample sequence and NSAMP = 7 (each image is composed of seven successive readouts after the initial read). The first- and second-order spectra of HD 189733 are visible in the images, as well as the spectra of the companion star HD 189733B. The latter slightly overlaps with the former, but our data reduction method eliminates this overlap. The spatial scans occur alternatively in two directions, which we will designate as "forward" and "reverse" in this paper. The observation parameters are summarized in Table 1.

3.1. Spectrum Extraction

We intentionally keep the data reduction as simple as possible in order to emphasize the data quality obtained with HST WFC3 in spatial scanning mode. This is in stark contrast with previous similar observations, such as those obtained with HST NICMOS (e.g., Crouzet et al. 2012), in which complex data reduction methods were necessary. The improvement of the data quality of WFC3 compared to NICMOS is already evident after a first look at the images: variations of the spectrum position on the detector are nearly invisible to the eye, and the individual pixel values are much more stable.

We start with the Intermediate MultiAccum images (ima.fits), which are corrected for dark current, nonlinearity, and other calibrations by the CALWFC3 pipeline at STScI. We separate the images according to the scanning direction, resulting in two sets of data. Then, for each image, we build seven differential images from the eight readouts. Indeed, the WFC3 pixels are read non-destructively while the signal is building up and while the vertical scanning occurs, and eight intermediate readouts are recorded during each exposure. The first readout occurs quickly after the exposure starts and is known as the zeroth read, whereas the last readout corresponds to the full scan image. The signal recorded between two consecutive readouts is retrieved by subtracting them, i.e., D_i = R_i − R_{i − 1}, where R_i are the readouts, D_i are the "differential images," and i = 1, ..., 7. Figure 2 shows an example of a differential image. Its size along the scan direction is now reduced as it corresponds to one-seventh of the full scan. Considering all the observations, this leads to 7 sets of differential images for each scan direction, i.e., 14 sets in total. Within each set, the spectral trace is always at the same position on the detector. We build a one-dimensional (1D) spectrum for each set independently. This method eliminates the overlap of the companion star with the target, minimizes the sky background contribution, and allows consistency checks between the spectra obtained from the 14 sets.

Zoom In Zoom Out Reset image size

We calibrate each differential image (D_i) by the F139M flat-field. We consider the pixel response is not wavelength dependent at the level required to extract the planetary spectrum, and we simply divide each image by this two-dimensional (2D) flat-field. Further tests using a wavelength dependent flat-field or no flat-field calibration do not result in significant differences in the final spectrum. Furthermore, flat-fielding errors should in principle be largely removed by dividing the in-eclipse by the out-of-eclipse spectra.

We search for bad pixels and cosmic rays located in the spectral trace region on a row-by-row basis. We compute a median-smoothed version of the row using a 10 pixel kernel. We calculate the median of the absolute deviation of the row with respect to its smoothed version and identify pixels deviating by more than 15 times this median. This yields typically two to three deviant pixels over the spectral region in each differential image, which are in fact already apparent by eye. We replace them with the value of the median-smoothed row at this location. We find good agreement between the bad pixels identified by our method and those flagged by the CALWFC3 pipeline as bad pixels in the data quality arrays, i.e., with a flag value of 4: "Bad detector pixel" (Rajan 2010) in extension 3 of the ima.fits files.

As commonly done with HST, we do not use the first orbit in which the telescope is known to settle to its new thermal environment resulting in an unstable behavior. As a result, the "hook" observed in the white-light curve at the beginning of each orbit, a WFC3 feature already reported by Berta et al. (2012) and Deming et al. (2013), is much stronger in the first orbit. All images from the four remaining orbits are used in the analysis.

We define rectangular regions containing the stellar signal and the sky background. A stellar region is defined for each set of differential images. The regions are 154 columns wide and 20 rows tall for forward scans (22 rows tall for reverse scans). The reverse direction yields slightly taller scans due to the longer exposure time (McCullough & MacKenty 2012). The sky region is common to all images and is 96 columns wide, 166 rows tall, and located in the upper left corner.

Spectra are built in a very straightforward manner. The sky background is calculated on the sky region using the SKY procedure in IDL and subtracted from the image. The sky values range from 0.96 × 10⁻⁴ to 1.71 × 10⁻⁴ relative to the maximum flux per pixel in the spectral trace of HD 189733. Then, we sum each column in the stellar region individually; this creates a 1D spectrum for each differential image D_i (Figure 3). For each set of differential images, we average these spectra for each orbit, and we average the three out-of-eclipse orbits together (only one orbit is in-eclipse). Then we divide the in-eclipse spectrum by the out-of-eclipse spectrum to extract the planetary spectrum. We build 14 planetary spectra from the 14 sets of differential images, which we average together. For all these calculations, we use the most simple possible average (function MEAN in IDL), with, e.g., no calls for outliers or medians.⁷

Zoom In Zoom Out Reset image size

3.2. Wavelength Calibration

To calibrate the detector columns in wavelength, we match our stellar spectrum to a known spectrum of similar spectral type. In keeping with our objective of simplicity, the G141 grism wavelength dispersion can be approximated by a linear function. First, we calibrate the out-of-eclipse spectrum by the known G141 grism response. This requires an initial estimate of the wavelength solution, obtained by matching the sharp edges of the grism response to that of our spectrum. Second, we compare a spectrum of the K1V star 107 Psc⁸ (Rayner et al. 2009) to our data. We remove the low-frequency variations by subtracting a smoothed version from each of these spectra. Finally, we match them in the wavelength range 1.1–1.7 μm using their common spectral lines (Figure 4). We derive the relation λ(x) = a × (x − x₀) + b, where x is the detector column index, x₀ = 77, a = 0.0047 μm pixel⁻¹, and b = 1.027 μm.

Zoom In Zoom Out Reset image size

3.3. Spectrum Shifts

3.4. Uncertainties

We estimate the uncertainties of the exoplanetary spectrum from the standard deviation of the light curve of each spectral channel (a detector column or group of binned columns corresponding to a given wavelength). These light curves are first corrected from white-light flux variations. To this end, we extract a spectrum for each calibrated full scan image excluding 20 columns at each edge of the spectral trace, and scale its amplitude to that of the template. The best scaling factor is found using the downhill simplex minimization procedure (AMOEBA). We build the white-light corrected channel light curves from these amplitude-corrected spectra. The uncertainty σ_{c, d} for a given channel c and a given scan direction d is

$$\begin{equation} \sigma _{c,d} = \sqrt{\frac{\sigma _{{\rm out},d}^2}{m_{{\rm out},d}^2 N_{{\rm out},d}} + \frac{\sigma _{{\rm in},d}^2}{m_{{\rm in},d}^2 N_{{\rm in},d}}}\protect, \protect \end{equation} \tag{ 1 }$$

where σ_{in, d}, σ_{out, d} are the standard deviations of the in-eclipse and out-of-eclipse parts of the channel c light curve, respectively, m_{in, d}, m_{out, d} their respective mean, and N_{in, d}, N_{out, d} their respective number of points (N_{in, d} = 16, N_{out, d} = 48). We find a median uncertainty of 160 ppm and 146 ppm for the forward (d = f) and reverse (d = r) scan directions, respectively, for an expected minimum Poisson noise of 162 and 153 ppm. Our final spectrum is the average of the forward and reverse scan direction spectra; thus, our final uncertainties are

$$\begin{equation} \sigma _c = \frac{1}{2}\sqrt{\sigma _{c,f}^2+ \sigma _{c,r}^2}\protect. \protect \end{equation} \tag{ 2 }$$

The final median uncertainty is 110 ppm compared to a Poisson noise estimate of 111 ppm. Because the planet passes behind the star during the eclipse, the effect of star spots on the emission spectrum can be neglected.

The planetary emission spectrum is binned over four columns using a boxcar average, resulting in independent 0.02 μm bins. This yields a high-precision Poisson noise-limited spectrum; the median uncertainty is 57 ppm. This spectrum is reported in Table 2, Columns 1–4, with an average depth arbitrarily set to zero and is shown in Figure 5. The absolute eclipse depth is calculated from the white-light analysis in Section 5. The spectra obtained from the two scan directions and for the 14 sets of differential images are consistent within their uncertainty and do not show any systematic differences.

Zoom In Zoom Out Reset image size

Table 2. Dayside Emission Spectrum of HD 189733b

λ	$\Delta R_p^2/R_s^2$	σ	Column	$\Delta R_p^2/R_s^2$	σ	ΔΔ
(μm)	(ppm)	(ppm)		(ppm)	(ppm)	(ppm)
1.1279	−21	81	98.5	−96	47	75
1.1467	−127	61	102.5	−18	50	−109
1.1655	36	70	106.5	28	45	8
1.1843	−6	47	110.5	−3	44	−3
1.2031	−15	68	114.5	−7	43	−9
1.2218	−65	55	118.5	−45	50	−20
1.2406	−23	52	122.5	−32	42	9
1.2594	2	45	126.5	3	42	−1
1.2782	−66	50	130.5	−16	42	−51
1.2969	6	53	134.5	53	41	−47
1.3157	6	57	138.5	31	41	−24
1.3345	19	58	142.5	12	41	7
1.3533	−48	47	146.5	−21	41	−27
1.3721	−59	61	150.5	−27	41	−32
1.3908	−90	43	154.5	−40	42	−50
1.4096	34	51	158.5	8	42	26
1.4284	16	55	162.5	−2	42	18
1.4472	4	53	166.5	−29	42	33
1.4660	−38	68	170.5	−64	43	26
1.4848	−19	50	174.5	9	43	−29
1.5035	161	61	178.5	132	43	29
1.5223	129	62	182.5	115	43	14
1.5411	23	53	186.5	47	44	−23
1.5599	−6	50	190.5	3	45	−8
1.5786	77	60	194.5	39	45	38
1.5974	85	58	198.5	−16	46	101
1.6162	−30	74	202.5	−64	46	35
1.6350	14	74	206.5	...	...	...

Notes. Units are as indicated; parts per million is abbreviated ppm. The tabulated uncertainties apply to the differential eclipse depths; an additional uncertainty applies to the overall depth (see the text). The first three columns refer to the analysis of N.C.; Columns 5 and 6 refer to the analysis of D.D.; the last column contains the difference of the differential spectra, Column 2 minus Column 5.

Download table as:  ASCIITypeset image

An independent analysis of these data was conducted using the method prescribed in Deming et al. (2013). Both planetary spectra are largely consistent, well within their 1σ uncertainties. This second spectrum as well as the difference between both are reported in Table 2, Columns 5–7. The smaller uncertainties of this second spectrum are due to the convolution with a Gaussian kernel of FWHM of four columns used to balance the effect of undersampling (see Deming et al. 2013), which is equivalent to an effective binning over ∼six columns, whereas the spectrum in Columns 1–3 is binned over four columns.

We build a white-light curve by summing the flux over the stellar region and subtracting the sky background, using the same stellar and sky regions as for the spectral analysis. The fluxes measured at all seven positions are summed together to recover the flux collected during a full scan. We build a light curve for each scan direction and eliminate the first HST orbit. The resulting light curves are reported in Table 3, and are shown in Figure 6 as a function of time and of HST orbital phase. From the second to the fifth orbit, observations within each orbit are arranged similarly with respect to the HST orbital phase. Preparing the observations in a repeatable manner is helpful to correct for systematic effects such as the "hook," as described next. A good practice is to use a fixed cadence with no interruptions except the Earth occultation, if practical. In our case, a buffer dump interrupts each orbit, with eight pairs of scans before and eight pairs after the dump. In these light curves, three main features appear. First, the white-light flux is larger by 11% in the reverse scan direction with respect to the forward scan direction. This is consistent with the longer exposure time in the reverse scan direction, which collects more photons (McCullough & MacKenty 2012). As a result, we analyze each scan direction separately. Second, the flux ramps upward by ∼0.1% after each buffer dump, which occurs between and in the middle of each orbit. This effect has already been reported for WFC3 and named "ramp" or "hook" (Berta et al. 2012; Deming et al. 2013). This hook is nearly repeatable from the second to the fifth orbit. Third, the flux decreases by ∼0.1% from the second to the fifth orbit.

Zoom In Zoom Out Reset image size

To derive the eclipse depth, we model the flux's decrease in time first using a second-order polynomial. We do not directly model the hook. Instead, we fit the same polynomial function to all the out-of-eclipse points, with the zero flux level as a free parameter for each set of points located at the same position in the orbital sequence (set 1: first point of orbits 2, 3, and 5; set 2: second point of orbits 2, 3, and 5; etc...; see Figure 7). The in-eclipse points are excluded from the fit (orbit 4). This fit is then subtracted to the data points. The residuals for both scan directions are normalized separately and put together. These residuals are close to the Poisson noise. We calculate the eclipse depth δ_p as the mean of the in-eclipse residuals and its 1σ uncertainty as

$$\begin{equation} \sigma = \sqrt{\frac{\sigma _{\rm in}^2}{N_{\rm in}}+\frac{\sigma _{\rm out}^2}{N_{\rm out}}}\protect, \protect \end{equation} \tag{ 3 }$$

where σ_in, σ_out is the standard deviation of the in-eclipse and out-of-eclipse residuals, and N_in, N_out is the number of in-eclipse and out-of-eclipse points (N_in = 32, N_out = 96). We find δ_p = 68 ± 12 ppm.

Zoom In Zoom Out Reset image size

Because the choice of a second-order polynomial is arbitrary, we perform the same analysis using other functions: an exponential function from the second to the fifth orbit, a linear function from the second to the fifth orbit, and a linear function from the third to the fifth orbit (orbits surrounding the eclipse). Inspection of residuals show that the linear function from the second to the fifth orbit is clearly a poor fit; we exclude this function.⁹ For the exponential fit and the linear fit from the third to the fifth orbit, we find an eclipse depth δ_e = 79 ± 12 ppm and δ_l = 122 ± 12 ppm, respectively. In their analysis of WFC3 data of WASP-12b obtained in staring mode, Stevenson et al. (2014) also noticed that different ramp models yield different values for the transit depth (note that in our approach the ramp model is implicit).

For each model, we estimate the error on the eclipse depth caused by stellar variability, thought to originate mainly from star spots. We model this variability by a sinusoidal function of period 11.95 days (Henry & Winn 2008), and rescale the amplitude of 1.5% in the b+y band (Knutson et al. 2012) to our wavelength of 1.4 μm, using a blackbody emission function with a stellar effective temperature of 4980 K and a spot temperature of 4250 K (Pont et al. 2008). We find an amplitude of 0.8% at 1.4 μm. This stellar variation is added to the data at all possible phases of the sinusoidal function prior to performing the function fit. The maximum difference on the eclipse depth with or without this variability is negligible for the second-order polynomial and the exponential fits (0.1 ppm) and is 5 ppm for the linear function fit from the third to the fifth orbit, with a standard deviation of 1.7 ppm. We quadratically add this standard deviation to the 1σ uncertainty for this latter function (which results in a change of a fraction of ppm).

The actual function being unknown, we include the full range spanned by the three models in our final estimate of the eclipse depth δ. This yields δ = 96  ±  39 ppm. We derive an approximate brightness temperature T_b for the planet by

$$\begin{equation} \delta =\frac{R^2_p B(T_b)}{R^2_\star B(T_{\rm eff,\star })}\protect, \protect \end{equation} \tag{ 4 }$$

where R_p and R_⋆ are the planet and stellar radii, respectively, and B(T_b) and B(T_{eff, ⋆}) are the spectral radiance of blackbodies of temperature T_b and T_{eff, ⋆}, respectively. For simplicity, we approximate the stellar spectrum by a blackbody emission of temperature T_{eff, ⋆} = 4980 K (a more accurate model is used in Section 6). We find a brightness temperature $T_b=1419_{-96}^{+71}$ K in the wavelength range 1.12–1.65 μm for the dayside of HD 189733b.

We model the dayside atmosphere of HD 189733b using the atmospheric modeling and retrieval method of Madhusudhan & Seager (2009) and Madhusudhan (2012). The emission spectrum is calculated by solving for 1D line-by-line radiative transfer in a plane parallel atmosphere, with constraints of local thermodynamic equilibrium (LTE), hydrostatic equilibrium, and global energy balance. The model has twelve free parameters including a parametric temperature profile and molecular abundances parameterized as uniform mixing ratios in the atmosphere. The model includes opacity contributions due to all the dominant molecules expected in H₂-rich hot Jupiter atmospheres in oxygen-rich as well as carbon-rich regimes (see e.g., Madhusudhan 2012), namely, line opacity due to H₂O, CO, CO₂, CH₄, HCN, and C₂H₂, and H₂–H₂ collision induced opacity. As for the temperature structure, the parametric temperature profile can model dayside temperature profiles with and without thermal inversions. Given a data set, we explore a wide range of inversion and non-inversion models as well as varied molecular mixing ratios in search of regions in parameter space that best match the data. In the present case, we explore models that can simultaneously explain our WFC3 data set along with previously published Spitzer observations in six photometric bandpasses (Charbonneau et al. 2008; Agol et al. 2010; Knutson et al. 2012).

Considering our present WFC3 observations alone (Figure 5), we find nominal evidence for water absorption in the WFC3 bandpass (1.1–1.7 μm). Figure 5 shows model spectra that provide a good match to the data. In the temperature regime of HD 189733b, as shown by the P–T profile in Figure 8, several spectroscopically strong molecules are expected to be prevalent in the atmosphere for a solar abundance composition, e.g., H₂O, CO, CH₄, and CO₂. However, H₂O is the most abundant of the molecules and dominates the absorption in the WFC3 bandpass, followed by a weaker contribution from CH₄, CO, and CO₂, which are unconstrained by the WFC3 data. Our data can be explained by a double-trough water absorption feature in the emission spectrum of a solar composition model atmosphere. On the other hand, given the error bars in our data, a featureless blackbody spectrum also provides a very good fit to the data. The best-fit blackbody spectrum, corresponding to an isothermal atmosphere at T = 1435 K, has a χ² of 24.8 for 26 degrees of freedom. Therefore, our observational uncertainties preclude a robust constraint on the water abundance using our WFC3 data alone.

Zoom In Zoom Out Reset image size

Combining our data with previously published Spitzer photometric observations from Charbonneau et al. (2008), Agol et al. (2010), and Knutson et al. (2012) rules out an isothermal atmosphere or one with a thermal inversion (Figure 8). Even the best-fit blackbody spectrum representing an isothermal atmosphere at T = 1295 K is unable to explain all the existing data simultaneously. Furthermore, a thermal inversion, i.e., with temperature increasing with decreasing pressure, is conclusively ruled out by the data as such a model will predict even higher fluxes in all the Spitzer IRAC bands than observed. On the other hand, the sum-total of data can be explained by a dayside atmosphere with no thermal inversion and a nearly solar or subsolar abundance H₂O mixing ratio. Two representative best-fitting models with nearly solar-abundance H₂O mixing ratios are shown in Figure 8 with their associated pressure–temperature profiles. The corresponding abundances are reported in Table 4, as well as the statistical significance of each model fit. In this table, the BIC is the Bayesian Information Criterion, which takes into account the number of free parameters in evaluating a model fit. It is calculated as follows: BIC =χ² + k ln N, where k is the number of free parameters (12 for the non-isothermal atmosphere models and 1 for the blackbody), and N is the number of data points (34, i.e., 28 from WFC3 and 6 from Spitzer). The comparison of these models to the data in a χ² sense or using the BIC shows that the atmosphere models with molecular features provide significantly better fits than the best blackbody model.

A variety of non-isothermal atmosphere models with a range of compositions can also fit the data. The two models shown in Figure 8 and Table 4 are examples chosen because of their chemical compositions which come close to a solar-abundance H₂O mixing ratio while still providing good fits to the data; an exactly solar abundance composition would have a H₂O mixing ratio between 5 × 10⁻⁴ and 10⁻³ depending on the temperature. More generally, however, the space of best-fitting solutions prefer manifestly subsolar H₂O abundances: the 1σ range of H₂O abundance derived from the sum-total of data used is between 1.2 × 10⁻⁹ and 2.3 × 10⁻⁵, with a modal value of 7.6 × 10⁻⁶. Our H₂O abundances are generally consistent, at ∼2σ, with other studies in the past also reporting constraints that are consistent with a subsolar H₂O abundance on the dayside of HD 189733b (e.g., Madhusudhan & Seager 2009; Swain et al. 2009; Lee et al. 2012; Line et al. 2012, 2014) using previous Spitzer and HST data, some of which have since been revised. However, our 1σ limits suggest lower H₂O abundances than previously suggested. Note that low-altitude, thin, or patchy clouds, and/or haze in the planet's atmosphere could also damp spectral features, as discussed below. Since our models are cloud free, the molecular abundances derived here should be interpreted as lower limits if obscuring clouds are present.

The lack of a thermal inversion in the dayside atmosphere of HD 189733b is consistent with previous studies, both observational and theoretical. Several studies using past data have suggested the absence of a thermal inversion in HD 189733b (e.g., Burrows et al. 2008; Grillmair et al. 2008; Swain et al. 2009; Madhusudhan & Seager 2009). Several theoretical and empirical studies have also predicted that its dayside atmosphere would be unlikely to host a strong thermal inversion. With an equilibrium temperature of ∼1200 K, HD 189733b is one of the less irradiated hot Jupiters. Based on the original TiO/VO hypothesis (Hubeny et al. 2003; Fortney et al. 2008), cooler hot Jupiters such as HD 189733b would not be hot enough to host gaseous TiO and VO in their upper atmospheres, which have been proposed as inversion-causing compounds. Even if alternate inversion-causing compounds were possible, Knutson et al. (2010) suggested that the high chromospheric activity of the host star (HD 189733) might dissociate them in the planetary atmosphere.

The ensemble of data obtained on HD 189733b in transmission and emission from HST and Spitzer is yet to be understood completely. Pont et al. (2013) interpreted the combined transmission spectrum by an atmosphere dominated by dust and clouds. They also noted that several puzzling features arising from the interpretation of the Spitzer phase curves with a clear solar composition atmosphere model (Knutson et al. 2012) could be explained by clouds (see Table 8 of Pont et al. 2013). However, McCullough et al. (2014) recently reported a robust detection of water vapor absorption in the transmission spectrum of the planet, which rules out a featureless transmission spectrum, as well as an atmosphere dominated by opaque clouds at a pressure P > 0.1 bar at the terminator region. Instead of clouds, McCullough et al. (2014) interpret the multi-wavelengths transmission spectrum by a solar-abundance composition clear atmosphere model with an additional contribution of unocculted star spots explaining the rise in the visible and UV.

In emission, a higher geometric albedo at short visible wavelengths (λ < 0.45 μm) than at longer wavelengths (λ > 0.45 μm) was reported by Evans et al. (2013), who suggested an "intermediate cloud scenario" for the dayside of the planet in which clouds are present and become optically thick at pressures corresponding to the Na absorption wings. Recent theoretical studies have also suggested the possibility of clouds in the dayside atmospheres of hot Jupiters (e.g., Heng & Demory 2013; Parmentier et al. 2013). However, as shown by Barstow et al. (2014), the reported albedo spectrum is insufficient to conclusively constrain the presence of clouds in the dayside atmosphere of HD 189733b, as a cloud-free atmosphere with Rayleigh scattering due to H₂ and non-solar Na abundances could explain the data equally well. Our detection of molecular features in the WFC3 bandpass places additional constraints on the possibility of clouds on the dayside. With thick clouds on the dayside, we would expect a blackbody spectrum to fit the observed infrared emission spectrum better than a model spectrum with molecular features in the observed bandpasses; our retrieved model solutions favor the opposite. Interestingly, a new analysis of the Spitzer IRS data, including data not included in the original spectrum from Grillmair et al. (2008), is also inconsistent with a blackbody spectrum (Todorov et al. 2014). In principle, alternate scenarios might be able to explain both the reported albedo spectrum and the molecular features we are observing, e.g., a layer of clouds at low altitudes producing the former and a clear atmosphere at higher altitudes producing the latter, or Rayleigh scattering due to H₂ with non-solar Na abundances, which future retrieval studies using clouds models could investigate (e.g., Barstow et al. 2014; Lee et al. 2014).

The transit and eclipse spectra of HD 189733b obtained with the high precision of WFC3 in spatial scanning mode yield new constraints both on the limb and dayside regions of its atmosphere. These results support a general picture of an atmosphere dominated by water vapor in the near infrared, albeit with a lower H₂O abundance than for a solar composition atmosphere. If clouds are present, these abundances would be degenerate with the amount of cloudiness; however, the presence and composition of such clouds are not constrained observationally (see Madhusudhan et al. 2014, for a discussion on this low H₂O abundance). These results point toward the need for including the effects of clouds in hot Jupiter atmosphere models, although still poorly understood due to their extreme complexity. Finally, these transit and eclipse spectra provide keys to further interpretation using full 3D atmosphere models such as those developed by, e.g., Showman et al. (2009). Extensive characterization of other hot gaseous giant planets would also improve our understanding of their atmospheres. The TESS mission (Ricker et al. 2014) should detect new hot Jupiters and hot Neptunes around bright stars, providing ideal targets for such characterization.

We observed HD 189733b during a planetary eclipse with HST WFC3 and extracted the emission spectrum of the planet in the wavelength range 1.1–1.7 μm. Using a straightforward data reduction method of the spatially scanned spectra, the derived spectrum is Poisson noise limited. A white-light analysis including a correction for WFC3 instrumental systematic effects yields the absolute eclipse depth in this wavelength range. The resulting spectrum shows marginal evidence for water vapor absorption, but can also be well explained by a blackbody spectrum. However, the combination of our WFC3 data with previous data from Spitzer reinforces a general picture of a dayside atmosphere in thermochemical equilibrium with no thermal inversion and dominated by water vapor features in the near infrared.

The authors gratefully acknowledge everyone who has contributed to the Hubble Space Telescope and the WFC3 and particularly those responsible for implementing the spatial scanning, which was critical to these observations. We thank, in particular, John MacKenty and Merle Reinhart. We acknowledge conversations with Suzanne Hawley, Leslie Hebb, Veselin Kostov, Rachel Osten, and Neill Reid. This research used NASA's Astrophysics Data System Bibliographic Services, the SIMBAD database operated at CDS, Strasbourg, France, and was funded in part by HST grant GO-12881 and Origins of Solar Systems grant NNX10AG30G.