Hunting "New Worlds" with ML

Using machine learning to search and characterize Exoplanet - planet orbiting a star outside our solar system - from observed protoplanetary disks

HL Tau (ALMA image)

Finding exoplanets can be hard. Even with the advanced planet search schemes detecting planets around young stars are not easy. The planet signatures like spectra or light curves are often suppressed due to enhanced stellar activity. Thus one has to depend on indirect evidences to ascertain the presence of such "Unseen Planets".

Protoplanetary disks - dense cloud of gas and dust around young stars - could serve as a hunting ground for such "New Worlds". These disks show dust gaps and rings, see Hl-Tau on the right, which could potentially be caused due to the presence of a planet. By comparing these structures with specialized disk-planet simulations, it is possible to infer the mass of these hidden planets. However, with the increased-rate at which such structured disks are being observed, customized simulations for each observation are practically impossible as these are expensive and time consuming.

In this new work, done in collaboration with Dr. Min-Kai Lin at ASIAA, we design a neural network called DPNNet (Disk Planet Neural Network) that can find and predict the mass of these unseen planets from observed disks. DPNNet uses deep learning (a sub-branch of Machine learning) to learn the underlying relationship between the planetary gaps in dust and the complex disk-planet interaction. It is trained using ~1000 simulations of disk-planet interaction using FARGO3D code. For computational resources we used the Graphical Processing Units (GPUs) on the TAIWANIA-2 supercomputing clusters in Taiwan.

Machine learns to infer planet mass from observed disks using deep learning.

Schematic representation of DPNNet

As a test DPNNet is deployed to find the mass of planets in some well-known disks. For example, when applied to HL tau it predicts that planets of masses 0.25, 0.20 0.22 Jupiter mass can explain the three observed gaps. These values are consistent with the masses inferred using customized simulations. Thus DPNNet, once trained, can infer the mass of unseen planets almost instantaneously while preserving the accuracy of specialized simulations.

Such ingenious multidisciplinary approach holds the key to a fast and future-proof tool that can predict planet mass from observed disks. In future, improved simulations aided with advanced Artificial Intelligence will guide our observational quest to not only hunt these "New Worlds " but also to search for "life" in them.

The article is accepted for publication in Astrophysical Journal: A Machine Learning model to infer planet masses from gaps observed in protoplanetary disks, Auddy, S., Lin, M.-K.