We adopt the deep learning method CASI (Convolutional Approach to Shell Identification) and extend it to 3D (CASI-3D) to identify signatures of stellar feedback in molecular line spectra, such as 13CO. We adopt magneto-hydrodynamics simulations that study the impact of stellar winds in a turbulent molecular cloud as an input to generate synthetic observations. We apply the 3D radiation transfer code radmc-3d to model 13CO (J=1-0) line emission from the simulated clouds. We train two CASI-3d models: ME1 predicts only the position of feedback, while MF predicts the fraction of the mass coming from feedback in each voxel. We adopt 75% of the synthetic observations as the training set and assess the accuracy of the two models with the remaining data. We demonstrate that model ME1 identifies bubbles in simulated data with 95% accuracy, and model MF predicts the bubble mass within 4% of the true value. We use bubbles previously visually identified in Taurus in 13CO to validate the models and show both perform well on the highest confidence bubbles. We apply our two models on the full 98 deg2 FCRAO 13CO survey of the Taurus cloud. Models ME1 and MF predict feedback gas mass of 2894 M and 302 M, respectively. When including a correction factor for missing energy due to the limited velocity range of the 13CO data cube, model ME1 predicts feedback kinetic energies of 4.0*1e46 ergs and 1.5*1e47 ergs with/without subtracting the cloud velocity gradient. Model MF predicts feedback kinetic energy of 9.6*1e45 ergs and 2.8*1e46 ergs with/without subtracting the cloud velocity gradient. Model ME1 predicts bubble locations and properties consistent with previous visually identified bubbles. However, model MF demonstrates that feedback properties computed based on visual identifications are significantly overestimated due to line of sight confusion and contamination from background and foreground gas.

We utilize techniques from deep learning to identify signatures of stellar feedback in simulated molecular clouds. Specifically, we implement a deep neural network with an architecture similar to U-Net and apply it to the problem of identifying wind-driven shells and bubbles using data from magneto-hydrodynamic simulations of turbulent molecular clouds with embedded stellar sources. The network is applied to two tasks, dense regression and segmentation, on two varieties of data, simulated density and synthetic 12 CO observations. Our Convolutional Approach for Shell Identification (CASI) is able to obtain a true positive rate greater than 90\%, while maintaining a false positive rate of 1\%, on two segmentation tasks and also performs well on related regression tasks. The source code for CASI is available on GitLab.