TY - JOUR
T1 - Quantum Machine Learning Approach for Studying Atmospheric Cluster Formation
AU - Kubečka, Jakub
AU - Christensen, Anders S.
AU - Rasmussen, Freja Rydahl
AU - Elm, Jonas
N1 - Funding Information:
The authors are thankful for the Independent Research Fund Denmark Grant No. 9064-00001B. We gratefully acknowledge the contributions of Aarhus University Interdisciplinary Centre for Climate Change (iClimate, Aarhus University). The numerical results presented in this work were obtained at the Centre for Scientific Computing, Aarhus https://phys.au.dk/forskning/faciliteter/cscaa .
Publisher Copyright:
© 2022 American Chemical Society. All rights reserved.
PY - 2022/3/8
Y1 - 2022/3/8
N2 - Quantum chemical (QC) calculations can yield direct insight into an atmospheric cluster formation mechanism and cluster formation rates. However, such calculations are extremely computationally demanding as more than millions of cluster configurations might exist and need to be computed. We present an efficient approach to produce high quality QC data sets for applications in cluster formation studies and how to train an accurate quantum machine learning model on the generated data. Using the two-component sulfuric acid-water system as a proof of concept, we demonstrate that a kernel ridge regression machine learning model with Δ-learning can be trained to accurately predict the binding energies of cluster equilibrium configurations with mean absolute errors below 0.5 kcal mol-1. Additionally, we enlarge the training data set with nonequilibrium configurations and show the possibility of predicting the binding energies of new structures of clusters several molecules larger than those in the training set. Applying the trained machine learning model leads to a drastic reduction in the number of relevant clusters that need to be explicitly evaluated by QC methods. The presented approach is directly transferable to clusters of arbitrary composition and will lead to faster and more efficient exploration of the configurational space of new cluster systems.
AB - Quantum chemical (QC) calculations can yield direct insight into an atmospheric cluster formation mechanism and cluster formation rates. However, such calculations are extremely computationally demanding as more than millions of cluster configurations might exist and need to be computed. We present an efficient approach to produce high quality QC data sets for applications in cluster formation studies and how to train an accurate quantum machine learning model on the generated data. Using the two-component sulfuric acid-water system as a proof of concept, we demonstrate that a kernel ridge regression machine learning model with Δ-learning can be trained to accurately predict the binding energies of cluster equilibrium configurations with mean absolute errors below 0.5 kcal mol-1. Additionally, we enlarge the training data set with nonequilibrium configurations and show the possibility of predicting the binding energies of new structures of clusters several molecules larger than those in the training set. Applying the trained machine learning model leads to a drastic reduction in the number of relevant clusters that need to be explicitly evaluated by QC methods. The presented approach is directly transferable to clusters of arbitrary composition and will lead to faster and more efficient exploration of the configurational space of new cluster systems.
UR - http://www.scopus.com/inward/record.url?scp=85124297131&partnerID=8YFLogxK
U2 - 10.1021/acs.estlett.1c00997
DO - 10.1021/acs.estlett.1c00997
M3 - Journal article
AN - SCOPUS:85124297131
SN - 2328-8930
VL - 9
SP - 239
EP - 244
JO - Environmental Science and Technology Letters
JF - Environmental Science and Technology Letters
IS - 3
ER -