Using neural networks to improve simulations in the gray zone
<p>Machine learning represents a potential method to cope with the gray zone problem of representing motions in dynamical systems on scales comparable to the model resolution. Here we explore the possibility of using a neural network to directly learn the error caused by unresolved scales. We...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Copernicus Publications
2022-05-01
|
| Series: | Nonlinear Processes in Geophysics |
| Online Access: | https://npg.copernicus.org/articles/29/171/2022/npg-29-171-2022.pdf |
| _version_ | 1818140755385909248 |
|---|---|
| author | R. Kriegmair Y. Ruckstuhl S. Rasp G. Craig |
| author_facet | R. Kriegmair Y. Ruckstuhl S. Rasp G. Craig |
| author_sort | R. Kriegmair |
| collection | DOAJ |
| description | <p>Machine learning represents a potential method to cope with the gray zone problem of representing motions in dynamical systems on scales comparable to the model resolution. Here we explore the possibility of using a neural network to directly learn the error caused by unresolved scales. We use a modified shallow water model which includes highly nonlinear processes mimicking atmospheric convection. To create the training dataset, we run the model in a high- and a low-resolution setup and compare the difference after one low-resolution time step, starting from the same initial conditions, thereby obtaining an exact target. The neural network is able to learn a large portion of the difference when evaluated on single time step predictions on a validation dataset. When coupled to the low-resolution model, we find large forecast improvements up to 1 d on average. After this, the accumulated error due to the mass conservation violation of the neural network starts to dominate and deteriorates the forecast. This deterioration can effectively be delayed by adding a penalty term to the loss function used to train the ANN to conserve mass in a weak sense. This study reinforces the need to include physical constraints in neural network parameterizations.</p> |
| first_indexed | 2024-12-11T10:49:02Z |
| format | Article |
| id | doaj.art-1a0c93ccf49c42089fa94afda14339b5 |
| institution | Directory Open Access Journal |
| issn | 1023-5809 1607-7946 |
| language | English |
| last_indexed | 2024-12-11T10:49:02Z |
| publishDate | 2022-05-01 |
| publisher | Copernicus Publications |
| record_format | Article |
| series | Nonlinear Processes in Geophysics |
| spelling | doaj.art-1a0c93ccf49c42089fa94afda14339b52022-12-22T01:10:24ZengCopernicus PublicationsNonlinear Processes in Geophysics1023-58091607-79462022-05-012917118110.5194/npg-29-171-2022Using neural networks to improve simulations in the gray zoneR. Kriegmair0Y. Ruckstuhl1S. Rasp2G. Craig3Meteorological Institute Munich, Ludwig-Maximilians-Universität München, Munich, GermanyMeteorological Institute Munich, Ludwig-Maximilians-Universität München, Munich, GermanyClimateAi, Inc., San Francisco, USAMeteorological Institute Munich, Ludwig-Maximilians-Universität München, Munich, Germany<p>Machine learning represents a potential method to cope with the gray zone problem of representing motions in dynamical systems on scales comparable to the model resolution. Here we explore the possibility of using a neural network to directly learn the error caused by unresolved scales. We use a modified shallow water model which includes highly nonlinear processes mimicking atmospheric convection. To create the training dataset, we run the model in a high- and a low-resolution setup and compare the difference after one low-resolution time step, starting from the same initial conditions, thereby obtaining an exact target. The neural network is able to learn a large portion of the difference when evaluated on single time step predictions on a validation dataset. When coupled to the low-resolution model, we find large forecast improvements up to 1 d on average. After this, the accumulated error due to the mass conservation violation of the neural network starts to dominate and deteriorates the forecast. This deterioration can effectively be delayed by adding a penalty term to the loss function used to train the ANN to conserve mass in a weak sense. This study reinforces the need to include physical constraints in neural network parameterizations.</p>https://npg.copernicus.org/articles/29/171/2022/npg-29-171-2022.pdf |
| spellingShingle | R. Kriegmair Y. Ruckstuhl S. Rasp G. Craig Using neural networks to improve simulations in the gray zone Nonlinear Processes in Geophysics |
| title | Using neural networks to improve simulations in the gray zone |
| title_full | Using neural networks to improve simulations in the gray zone |
| title_fullStr | Using neural networks to improve simulations in the gray zone |
| title_full_unstemmed | Using neural networks to improve simulations in the gray zone |
| title_short | Using neural networks to improve simulations in the gray zone |
| title_sort | using neural networks to improve simulations in the gray zone |
| url | https://npg.copernicus.org/articles/29/171/2022/npg-29-171-2022.pdf |
| work_keys_str_mv | AT rkriegmair usingneuralnetworkstoimprovesimulationsinthegrayzone AT yruckstuhl usingneuralnetworkstoimprovesimulationsinthegrayzone AT srasp usingneuralnetworkstoimprovesimulationsinthegrayzone AT gcraig usingneuralnetworkstoimprovesimulationsinthegrayzone |