Astronomers use radio telescopes to peer into the interstellar medium and learn about star formation and the behaviors of galaxies. The rotational spectra of the gaseous molecules tumbling through the interstellar medium reveal constituent molecular species, but each spectrum can include thousands of rotational transition lines from many different species.
Research from Zaleski and Prozument demonstrates an artificial neural network trained to recognize and then identify the spectral patterns appearing in rotational spectra. The network can also determine characteristics such as moments of inertia. This is the first time artificial neural networks have been applied to the challenges of rotational spectroscopy.
The artificial neural network differentiates between linear, symmetric and asymmetric tops with approximately 95 percent certainty. The network also correctly identifies the presence of hyperfine structures with the same certainty. The processing times for classification of the spectrum and determination of the rotational characteristics are short, totaling approximately 200 microseconds. Additionally, the processing time required for classification and determination does not seem to scale with spectrum size or complexity.
The authors trained the neural network using stick spectra data sets which were randomly generated and included the linear, symmetric and asymmetric-type spectra, as well as some with hyperfine structure. The input to the artificial neural network is a line list of center frequencies, also called a “peak pick.”
One neural network identifies the type of spectrum and then passes it to a neural network specifically trained to determine the rotational characteristics for that type of spectrum. While the current network has some limitations, the authors plan additional development because of its promising initial success.
Source: “Automated assignment of rotational spectra using artificial neural networks,” by Daniel P. Zaleski and Kirill Prozument, The Journal of Chemical Physics (2018). The article can be accessed at https://doi.org/10.1063/1.5037715.