Predicting Infrared stellar flux densities: teaching WISE to detect like Spitzer

Astrophysics is an ideal field in which to take advantage of machine learning (ML) techniques due to the great amount of astronomical data and its peculiar characteristics. The space satellite WISE is considered the current best infrared all-sky survey in both quality and coverage. In contrast, the space satellite Spitzer, with smaller coverage, has better spatial resolution (3x-2x, depending on the band) and sensitivity in the same spectral region. Some studies have claimed to find some kind of noise or contamination in WISE, resulting in discrepancies when comparing the measurements of both satellites. In this communication [1], we intend to overcome these discrepancies and report an ML approach to predict mid-infrared fluxes at two specific bands from WISE variables.

We have tested several ML regression models in a large sample of confirmed members (stars) from open clusters (groups of stars physically related) with both WISE and high-quality Spitzer data. In our particular case, Extremely Randomized Trees gave the best performance with values of R²>0.95 in both bands. Importantly, we have been able to improve the results at lower fluxes (the ones with the largest discrepancies) and to prove the good correspondence between the predicted fluxes and the real Spitzer ones when available.

The use of ML allows us to bring the best characteristics of both satellites together without the loss of data other approaches could cause. We believe a similar strategy could be useful in other studies when dealing with similar discrepancies.

[1] Fonseca-Bonilla et al. 2024, under review.