Rate Coefficients for the Collisional Excitation of Molecules: Estimates from an Artificial Neural Network

TL;DR

This paper explores using an artificial neural network to estimate molecular collisional excitation rate coefficients, demonstrating reasonable accuracy and potential for extrapolating data to uncalculated high-lying transitions.

Contribution

The study introduces an ANN-based method for estimating molecular collisional rate coefficients, showing its effectiveness and potential for extrapolation beyond existing datasets.

Findings

ANN achieves median accuracy within a factor of 1.7 to 2.1

4% of predictions are discrepant by a factor of 10 or more

ANN can extrapolate to high-lying transitions not yet calculated

Abstract

An artificial neural network (ANN) is investigated as a tool for estimating rate coefficients for the collisional excitation of molecules. The performance of such a tool can be evaluated by testing it on a dataset of collisionally-induced transitions for which rate coefficients are already known: the network is trained on a subset of that dataset and tested on the remainder. Results obtained by this method are typically accurate to within a factor ~ 2.1 (median value) for transitions with low excitation rates and ~ 1.7 for those with medium or high excitation rates, although 4% of the ANN outputs are discrepant by a factor of 10 more. The results suggest that ANNs will be valuable in extrapolating a dataset of collisional rate coefficients to include high-lying transitions that have not yet been calculated. For the asymmetric top molecules considered in this paper, the favored architecture is a cascade-correlation network that creates 16 hidden neurons during the course of training, with 3 input neurons to characterize the nature of the transition and one output neuron to provide the logarithm of the rate coefficient.