CLOVER: Convnet Line-fitting Of Velocities in Emission-line Regions

Abstract

When multiple star-forming gas structures overlap along the line-of-sight and emit optically thin emission at significantly different radial velocities, the emission can become non-Gaussian and often exhibits two distinct peaks. Traditional line-fitting techniques can fail to account adequately for these double-peaked profiles, providing inaccurate cloud kinematics measurements. We present a new method called Convnet Linefitting Of Velocities in Emission-line Regions (CLOVER) for distinguishing between one-component, two-component, and noise-only emission lines using 1D convolutional neural networks trained with synthetic spectral cubes. CLOVER utilizes spatial information in spectral cubes by predicting on 3 × 3 pixel sub-cubes, using both the central pixel's spectrum and the average spectrum over the 3×3 grid as input. On an unseen set of 10,000 synthetic spectral cubes in each predicted class, CLOVER has classification accuracies of ∼ 99% for the one-component class and ∼ 97% for the two-component class. For the noise-only class, which is analogous to a signal-to-noise cutoff of four for traditional line-fitting methods, CLOVER has classification accuracy of 100%. CLOVER also has exceptional performance on real observations, correctly distinguishing between the three classes across a variety of star-forming regions. In addition, CLOVER quickly and accurately extracts kinematics directly from spectra identified as two-component class members. Moreover, we show that CLOVER is easily scalable to emission lines with hyperfine splitting, making it an attractive tool in the new era of large-scale NH 3 and N 2 H + mapping surveys.

Dust polarized emission observations of NGC 6334

“…We thus derive a column density map, N T=20K H 2 , for a constant T = 20 K corresponding to the dominent temperature observed towards the Northern part of the NGC 6334 molecular cloud that is less affected by the warm H II regions. Temperatures of about 20 K are also compatible with the kinetic temperature toward star forming high-mass filaments and hubs derived from NH 3 GBT observations of the KEYSTONE project (Keown et al 2019).…”

“…The derived σ v 8 values are given in column 5 of Table 2 and range between ∼ 0.4 − 0.6 km s −1 compatible with observational results towards other filaments of similar line masses. For example, observations towards the SDC13 filament system with the IRAM-30m at ∼ 30 (using N 2 H + , Peretto et al 2014), towards dense filaments in a sample of Galactic giant molecular clouds with the GBT as part of the KEYSTONE project at ∼ 30 (using NH 3 , Keown et al 2019), and towards the infrared dark cloud G14.225−0.506 with ALMA (using N 2 H + , Chen et al 2019), all suggest values of σ v ∼ 0.4 − 0.8 km s −1 . In addition, the σ v = 0.62 km s −1 derived here for the crest 4 is similar to that derived from N 2 H + ALMA data at 3 of the same region (Shimajiri et al 2019).…”

Velocity-coherent Filaments in NGC 1333: Evidence for Accretion Flow?

“…fitting (e.g., Hacar et al 2013), and remains a challenge even for advanced machine learning techniques (e.g., Keown et al 2019).…”

“…As ∆v LSR decreases, the spectral profiles of the two slabs start to blend together, making them more difficult to distinguish from a one-slab profile. This lack of acuity is what prompted many studies to adopt a ∆v LSR threshold for their model selection to guard against over- fitting (e.g., Hacar et al 2013), and remains a challenge even for advanced machine learning techniques (e.g., Keown et al 2019).…”

A lack of constraints on the cold opaque HI mass: HI spectra in M31 and M33 prefer multi-component models over a single cold opaque component

“…Several recent studies present algorithmic approaches for (semi-) automated spectral line fitting of Gaussian components (e.g., Haud 2000;Lindner et al 2015;Henshaw et al 2016;Keown et al 2019;Marchal et al 2019;Sokolov et al 2020). The method we use here combines elements from some of these studies to optimize the fitting for our particular data sets, though we highlight that we only include spatial information from nearest neighbours rather than fitting all neighbouring spectra together ( §3.2.3).…”