Machine Learning Achieves Accurate Prediction of Hubble ACS/SBC Background Variation Using 23 Years of Data

Variations in background noise significantly impact the sensitivity of astronomical observations, and a new study sheds light on the factors driving these fluctuations. Christopher J. R. Clark of the Association of Universities for Research in Astronomy, Inc., alongside Roberto J. Avila and Alyssa Guzman, alongside Norman A. Grogin, analysed 23 years of data from the Hubble Space Telescope’s Advanced Camera for Surveys and Space Telescope Imaging Spectrograph. Their research reveals substantial background level variations , exceeding a factor of ten across most filters , and identifies key observational parameters responsible for these changes. By employing a machine learning model trained on over 8,000 archival observations, the team successfully predicts background levels based on factors such as solar elevation, observation angle, telescope temperature, and galactic latitude. This work provides a crucial step towards more accurate data calibration and ultimately enhances the potential for future discoveries with the Hubble Space Telescope.

While airglow is known to dominate backgrounds in shorter wavelengths, the precise drivers of these fluctuations remained unclear until now. This study addresses this challenge by leveraging over 8,000 archival SBC observations to develop a predictive model for background levels. The research establishes a clear link between background variation and a suite of 23 observational parameters, offering a powerful tool for optimising future observations.

Scientists demonstrate a breakthrough in understanding the complex factors influencing background noise in the SBC, a photon-counting detector crucial for Far-Ultraviolet observations. The team achieved this by constructing a machine learning model capable of accurately predicting background levels based on parameters such as Solar elevation, Solar separation angle, and Earth limb angle. Experiments show that the model consistently identifies these, alongside SBC temperature and target Galactic latitude, as the primary determinants of background variation, depending on the filter used. This innovative approach moves beyond simply acknowledging background fluctuations to actively predicting and accounting for them.

The study unveils that the SBC background is particularly sensitive to the position of the Sun relative to the telescope and Earth. Specifically, Solar elevation, the angle of the Sun above the horizon, and Solar separation angle, the angular distance between the Sun and the target, are key factors influencing background levels. Furthermore, the Earth limb angle, the angle between the observation direction and the Earth’s limb, and the SBC’s operating temperature also play significant roles. By accurately predicting background levels, scientists can design observing programs to minimise noise, reduce integration times, and enhance the detection of faint astronomical objects. The work opens avenues for refining data calibration techniques and improving the overall quality of FUV observations. Researchers harnessed over 8,000 SBC exposures to develop a machine learning model capable of predicting background levels based on 23 distinct observational parameters. This model represents a significant methodological advance, moving beyond simple correlation studies to a predictive framework for background noise. Furthermore, the day of year was calculated to account for annually-periodic phenomena, building on previous findings of seasonal dark current variations in the SBC and NUV MAMA instruments. Instrument temperature, recorded from MDECODT1 and MDECODT2 header keywords in degrees Celsius, was also included, acknowledging the established relationship between instrument heating and dark current. Exposure duration, obtained from the EXPTIME keyword in seconds, was considered to assess the impact of both accumulated instrument heating and potential transient events.

The research employed precise geometric calculations to determine key angles influencing background levels. Earth limb angle, representing the angular separation between the observation and the Earth’s limb, was obtained from the _jit. fits file, with minimum, maximum, and mean values recorded in degrees. Local Solar time was computed using orbital information from USSPACECOM, utilising the Simplified General Perturbations 4 (SGP4) model implemented via the skyfield package for Python, to accurately determine HST’s latitude and longitude during each exposure. Solar altitude and separation angle, sourced from the _flt. fits header keywords SUN_ALT and SUNANGLE respectively, were incorporated to account for atmospheric excitation and scattered light contributions. Orbital height, calculated using USSPACECOM/SGP4 ephemerides, and geomagnetic field strength were also included as parameters, recognising their potential influence on atmospheric interactions. The research team measured over an order of magnitude variation across most filters used, highlighting a previously unquantified challenge for precise astronomical measurements. Specifically, the study identified 8,640 suitable archival SBC exposures, ranging from 261 exposures in the F115LP filter to 2,933 in the F150LP filter. These exposures were meticulously analyzed to determine the factors influencing background noise. Scientists developed a machine learning model, trained on over 8,000 archival SBC observations, capable of accurately predicting background levels based on 23 observational parameters.

Results demonstrate that Solar elevation, Solar separation angle, Earth limb angle, SBC temperature, and target Galactic latitude are generally the dominant factors influencing SBC background. The median difference between the 16th and 84th percentile of measured backgrounds is a factor of 9.3, with variations ranging from 5.3 for F165LP to 20 for F125LP, demonstrating the significant range of background levels encountered. Experiments revealed a strong correlation between instrument temperature and background levels, consistent with established understanding of dark current increases with operational heating. The team recorded both the initial and final instrument temperatures, quantified in degrees Celsius, for each exposure to account for temperature fluctuations during observation.

Furthermore, the study incorporated exposure duration, measured in seconds, as a parameter, acknowledging its potential impact on both instrument heating and the likelihood of capturing transient background phenomena. The work meticulously quantified the Earth limb angle, recording the minimum, maximum, and mean angles in degrees for each exposure, to assess the influence of atmospheric scattering and airglow. Researchers also accounted for the time of day, calculating local Solar time to understand the impact of solar-driven airglow on background levels. A machine learning model, trained on 23 observational parameters, accurately predicts background levels, revealing the dominance of Solar elevation and separation angle, Earth limb angle, instrument temperature, and target Galactic latitude. These findings demonstrate that background variation is not random, but predictably linked to specific observational conditions. The research indicates that for shorter-wavelength filters, background is significantly affected by sunlight, with observations taken during daylight hours or at low Earth limb angles exhibiting higher levels.

Furthermore, the study confirms that instrumental dark rate is a primary driver of background in F150LP and F165LP filters, increasing with instrument temperature. While geocoronal parameters show some impact, their contribution remains relatively small due to existing South Atlantic Anomaly avoidance protocols. Background levels also increase at lower Galactic latitudes, likely due to undetected ultraviolet sources, a factor that cannot be mitigated by observational strategy. The authors acknowledge that their masking procedure cannot fully account for the contribution of faint or unresolved sources, particularly at low Galactic latitudes, potentially impacting background measurements. Future work could focus on refining these masking techniques to better isolate true background signals. This detailed characterisation of background variation provides a valuable tool for optimising future SBC observations and accurately interpreting existing data.