DPNNet-3.0: Identifying Multiple Hidden Planets and Estimating their Masses from Protoplanetary Disk Images using Convolutional Neural Networks

Abstract Identifying young planets and inferring their masses from annular gaps and other substructures observed in protoplanetary disks requires fitting the observed features with hydrodynamical or empirical disk models. The models’ computational cost makes such detections often imprecise or outright impractical. We present DPNNet-3.0, a tool to predict exoplanet masses from protoplanetary disk millimeter-wave dust continuum images using a convolutional neural network. Compared to the previous studies in this series of papers, we extend the parameter space to include disks with multiple planets. We apply the RADMC3D radiative transfer code to a set of disk models obtained with the FARGO3D hydrodynamics code to create synthetic images resembling millimeter-wave observations with the model disks viewed at various inclinations. We train DPNNet-3.0 on the set of improved synthetic images to predict the existence of planets and estimate their masses directly from the observations. We quantify the predictions’ reliability when applied to observations of HD163296 from the ALMA telescope, and find good agreement with results in the literature.