![]() |
![]() |
ETNA - Electronic Transactions on Numerical Analysis
|
![]() |
Verlag der Österreichischen Akademie der Wissenschaften Austrian Academy of Sciences Press
A-1011 Wien, Dr. Ignaz Seipel-Platz 2
Tel. +43-1-515 81/DW 3420, Fax +43-1-515 81/DW 3400 https://verlag.oeaw.ac.at, e-mail: verlag@oeaw.ac.at |
![]() |
|
DATUM, UNTERSCHRIFT / DATE, SIGNATURE
BANK AUSTRIA CREDITANSTALT, WIEN (IBAN AT04 1100 0006 2280 0100, BIC BKAUATWW), DEUTSCHE BANK MÜNCHEN (IBAN DE16 7007 0024 0238 8270 00, BIC DEUTDEDBMUC)
|
ETNA - Electronic Transactions on Numerical Analysis, pp. 138-156, 2022/01/10
Marine ecosystem models are important to identify the processes that affect for example the global carbon cycle. Computation of an annually periodic solution (i.e., a steady annual cycle) for these models requires a high computational effort. To reduce this effort, we approximate an exemplary marine ecosystem model by different artificial neural networks (ANNs). We use a fully connected network (FCN), then apply the sparse evolutionary training (SET) procedure, and finally apply a genetic algorithm (GA) to optimize, inter alia, the network topology. With all three approaches, a direct approximation of the steady annual cycle is not sufficiently accurate. However, using the mass-corrected prediction of the ANN as initial concentration for additional model runs, the results are in very good agreement. In this way, we achieve a runtime reduction by about 15%. The results from the SET algorithm are comparable to those of the FCN. Further application of the GA may lead to an even higher reduction.
Keywords: deep learning, genetic algorithm, sparse evolutionary training, biogeochemical modeling, marine ecosystem modeling, transport matrix method