No abstract

Traditional spectral energy distribution (SED) fitting codes used to derive galaxy physical properties are often uncertain at the factor of a few level owing to uncertainties in galaxy star formation histories and dust attenuation curves. Beyond this, Bayesian fitting (which is typically used in SED fitting software) is an intrinsically compute-intensive task, often requiring access to expensive hardware for long periods of time. To overcome these shortcomings, we have developed mirkwood: a user-friendly tool comprising an ensemble of supervised machine-learning-based models capable of nonlinearly mapping galaxy fluxes to their properties. By stacking multiple models, we marginalize against any individual model’s poor performance in a given region of the parameter space. We demonstrate mirkwood's significantly improved performance over traditional techniques by training it on a combined data set of mock photometry of z = 0 galaxies from the Simba, Eagle, and IllustrisTNG cosmological simulations, and comparing the derived results with those obtained from traditional SED fitting techniques. mirkwood is also able to account for uncertainties arising both from intrinsic noise in observations, and from finite training data and incorrect modeling assumptions. To increase the added value to the observational community, we use Shapley value explanations to fairly evaluate the relative importance of different bands to understand why particular predictions were reached. We envisage mirkwood to be an evolving, open-source framework that will provide highly accurate physical properties from observations of galaxies as compared to traditional SED fitting.

We propose a non-parametric method to denoise 1D stellar spectra based on wavelet shrinkage followed by adaptive Kalman thresholding. Wavelet shrinkage denoising involves applying the Discrete Wavelet Transform (DWT) to the input signal, 'shrinking' certain frequency components in the transform domain, and then applying inverse DWT to the reduced components. The performance of this procedure is influenced by the choice of base wavelet, the number of decomposition levels, and the thresholding function. Typically, these parameters are chosen by 'trial and error', which can be strongly dependent on the properties of the data being denoised. We here introduce an adaptive Kalman-filter-based thresholding method that eliminates the need for choosing the number of decomposition levels. We use the 'Haar' wavelet basis, which we found to be the best-suited for 1D stellar spectra. We introduce various levels of Poisson noise into synthetic PHOENIX spectra, and test the performance of several common denoising methods against our own. It proves superior in terms of noise suppression and peak shape preservation. We expect it may also be of use in automatically and accurately filtering low signal-to-noise galaxy and quasar spectra obtained from surveys such as SDSS, Gaia, LSST, PESSTO, VANDELS, LEGA-C, and DESI.

We leverage state-of-the-art machine-learning methods and archival data from the Canada-France-Hawaii Telescope (CFHT) to predict observatory image quality (IQ) from environmental conditions and observatory operating parameters. Drawing on nearly a decade’s worth of collected data we develop accurate and interpretable models of the complex dependence between data features and observed IQ for CFHT’s wide field camera, MegaCam. Using specialized loss functions and neural network architectures, we predict the probability distribution function (PDF) of the IQ associated with any given sample. One of our main results is that, based purely on environmental and observatory operating conditions, we can predict the effective MegaPrime IQ to a mean accuracy of ∼0.07″. We further explore how reconfiguration of adjustable observatory parameters can effect improvements in IQ. In particular, we consider actuation of 12 dome “vents”, installed in 2013-14, to accelerate the flushing of hot air from the dome, thereby reducing internal air turbulence and improving IQ. To enable trustworthy predictions of the effect of vent actuation, we first need to differentiate between uncertainties that arise from the randomness inherent to the input data and those which arise due to imperfect modeling. Using predictions of these uncertainties, in conjunction with probabilistic generative modeling, we identify candidate vent adjustments that are in-distribution and, for the optimal in-distribution vent configuration we calculate the predicted reduction in required observing time. The IQ prediction reduction, averaged across all samples, is about $\sim 25\%$. Finally, we use Shapley values to compute robust feature ranking. This allows us to identify the most predictive variables from the sensor data for each observation. Building from this work, our long-term goal is to construct a reliable and robust model that can forecast optimal operating conditions for real-time optimization of IQ. Such forecasts can be fed into real-time scheduling protocols to accelerate scientific productivity, and into predictive maintenance routines. We anticipate that the data-driven approaches we explore herein will become standard in automating observatory operations and in improving observatory maintenance by the time CFHT’s successor, the Maunakea Spectroscopic Explorer (MSE), is installed in the next decade.

No abstract