Probabilistic estimation of galactic inclinations: Inferring calibrated uncertainties using a network trained on cosmological simulations

Accurate estimation of the inclinations of disk galaxies is essential to infer many of their physical properties. This work proposes a novel method to estimate galaxy inclination solely based on imaging data without any additional information or explicit structural assumptions. We use a probabilistic approach to obtain a properly calibrated predictive distributions over inclinations, i.e. a realistic estimate of the confidence in the predictions. Our method simultaneously accounts for the uncertainty in the true inclinations (obtained from realistic galaxies in cosmological simulations), as well as the degeneracy between intrinsic 3D shape and inclination in galaxy images (learned from synthetic images of the simulated galaxies). Given only an image of a disk galaxy, our method predicts a density distribution of likely inclination angles using a Mixture Density Network. The model is based on a convolutional neural network trained on synthetic images of realistic simulated galaxies. We benchmark the model performance on an independent test set obtained from the same simulation, and compare the predictions of our method to the results of standard ellipse-fitting methods. Our predictions achieve a Continuous Ranked Probability Score (CRPS) of ≈ 3 ∘ compared to ≈ 9 ∘ for the standard method, indicating that our method is more precise, sharper, and better calibrated. Furthermore, while ellipse-fitting tends to underestimate the true inclinations and fails for values below ∼ 1 0 ∘ , our method is fully able to constrain the inclination of nearly face-on galaxies. Our model provides much better scaling on very large datasets than standard methods, estimating the inclination significantly faster on average.