Rate Constants for Fine-structure Excitations in O–H Collisions with Error Bars Obtained by Machine Learning

Abstract

We present an approach using a combination of coupled channel scattering calculations with a machinelearning technique based on Gaussian Process regression to determine the sensitivity of the rate constants for non-adiabatic transitions in inelastic atomic collisions to variations of the underlying adiabatic interaction potentials. Using this approach, we improve the previous computations of the rate constants for the fine-structure transitions in collisions of O( 3 P j ) with atomic H. We compute the error bars of the rate constants corresponding to 20 % variations of the ab initio potentials and show that this method can be used to determine which of the individual adiabatic potentials are more or less important for the outcome of different fine-structure changing collisions.

Fine-structure Transitions of Ne+, Ar+, Ne2+, and Ar2+ Induced by Collisions with Atomic Hydrogen

“…These discrepancies in the rate coefficients can be attributed to the different potentials used, serving as a rudimentary "error bar" for our current calculations. This somewhat empirical estimate is consistent with the much more exhaustive study by Vieira & Krems (2017), who investigated the errors in rate coefficients resulting from small where ¡ ¢ j j is the effective collision strength (Mendoza 1983;Flower 2007;Tayal 2011) and w j is the statistical weight of the initial state. Our calculated critical densities for Ne + , Ar + and Ne 2+ , Ar 2+ , respectively, and H are shown in Figures 10 and 11 along with values of n c;e , which we calculated using Equation (10) and the effective collision strengths from Wang et al (2017) for Ne + and Ne 2+ , from Pelan & Berrington (1995) for Ar + , and from Munoz Burgos et al (2009) for Ar + 2 .…”

A regression model for plasma reaction kinetics

“…Kuang and Xu [67] showcased the use of a convolutional neural network for the prediction of kinetic triplets for pyrolysis processes from experimental data, more specifically the temperatures at preselected values of conversion degrees. Very similar work has also been done by Huang et al [68], and Vieira and Krems [69]. In most cases, a research work intersecting ML and chemical kinetics introduces ANNs and other ML model techniques (or soft computing) as an alternative to the hard kinetic model of a system, which typically integrates the differential equations governing the species densities to calculate the reaction rates.…”

Neural network potentials for chemistry: concepts, applications and prospects

“…260 However, Gaussian processes are a limiting case of Bayesian NNs, 261 which have been extensively used and applied for the renement of PESs by means of UQ. 262,263 Therefore, a combination of NNPs and Gaussian process regression is a promising avenue for UQ in NNPs. Another approach for UQ is single network deterministic methods 163,264 which make assumptions about the distribution of the data.…”