Prediction of parameters and classification of molecular outflows using convolutional neural networks

Abstract

Machine learning is gaining popularity in modern astrophysics for its incredibly powerful ability to make predictions or make assumptions over large amounts of data. We describe the application of machine learning to regression of molecular outflow parameters (mass, momentum, kinetic energy, and dynamic time) and classification of bipolar outflow using convolutional neural networks. The size of our training sample is ~ 125 sources of molecular outflow for classification, that is, 80% of the total amount of data, where 67 sources are bipolar outflow and ~ 75 sources of bipolar outflow for regression. The results show that the use of CNN improves the image classification accuracy up to 97%. The regression model predicts molecular outflow parameters with an average absolute percentage error of 37.7% for the training data and with an average absolute error of 88.0 (mass), 1237.7 (momentum), 193.3 (kinetic energy), and 3.0 (dynamic time) for test data. The machine learning algorithm reduces data processing time for predictions and classification, and this methodology has a broad prospect for future studies of astrophysics problems.