TY - JOUR

T1 - Determining cross sections from transport coefficients using deep neural networks

AU - Stokes, P. W.

AU - Cocks, D. G.

AU - Brunger, M. J.

AU - White, R. D.

PY - 2020/5

Y1 - 2020/5

N2 - We present a neural network for the solution of the inverse swarm problem of deriving cross sections from swarm transport data. To account for the uncertainty inherent to this somewhat ill-posed inverse problem, we train the neural network using cross sections from the LXCat project, paired with associated transport coefficients found by the numerical solution of Boltzmann's equation. The use of experimentally measured and theoretically calculated cross sections for training encourages the network to avoid unphysical solutions, such as those containing spurious energy-dependent oscillations. We successfully apply this machine learning approach to simulated swarm data for electron transport in helium, separately determining its elastic momentum transfer and ionisation cross sections to within an accuracy of 4% over the range of energies considered. Our attempt to extend our method to argon was less successful, although the reason for that observation is well-understood. Finally, we explore the feasibility of simultaneously determining cross sections of helium using this approach. We have some success here, determining elastic, total n = 2 excitation and ionisation cross sections to 10%, 20% and 25% accuracy, respectively. We are unsuccessful in properly unfolding the separate n = 2 singlet and triplet excitation cross sections of helium, but this is as expected given their similar threshold energies.

AB - We present a neural network for the solution of the inverse swarm problem of deriving cross sections from swarm transport data. To account for the uncertainty inherent to this somewhat ill-posed inverse problem, we train the neural network using cross sections from the LXCat project, paired with associated transport coefficients found by the numerical solution of Boltzmann's equation. The use of experimentally measured and theoretically calculated cross sections for training encourages the network to avoid unphysical solutions, such as those containing spurious energy-dependent oscillations. We successfully apply this machine learning approach to simulated swarm data for electron transport in helium, separately determining its elastic momentum transfer and ionisation cross sections to within an accuracy of 4% over the range of energies considered. Our attempt to extend our method to argon was less successful, although the reason for that observation is well-understood. Finally, we explore the feasibility of simultaneously determining cross sections of helium using this approach. We have some success here, determining elastic, total n = 2 excitation and ionisation cross sections to 10%, 20% and 25% accuracy, respectively. We are unsuccessful in properly unfolding the separate n = 2 singlet and triplet excitation cross sections of helium, but this is as expected given their similar threshold energies.

KW - Boltzmann equation

KW - inverse problem

KW - machine learning

KW - swarm analysis

UR - http://www.scopus.com/inward/record.url?scp=85085560102&partnerID=8YFLogxK

UR - http://purl.org/au-research/grants/ARC/DP180101655

U2 - 10.1088/1361-6595/ab85b6

DO - 10.1088/1361-6595/ab85b6

M3 - Article

AN - SCOPUS:85085560102

VL - 29

JO - Plasma Sources Science and Technology

JF - Plasma Sources Science and Technology

SN - 0963-0252

IS - 5

M1 - 055009

ER -