Training a convolutional neural network to conserve mass in data assimilation

Ruckstuhl, Yvonne; Janjić, Tijana; Rasp, Stephan

doi:https://doi.org/10.5194/npg-28-111-2021

Articles | Volume 28, issue 1

https://doi.org/10.5194/npg-28-111-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/npg-28-111-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 28, issue 1

Research article

|

09 Feb 2021

Research article |

| 09 Feb 2021

Training a convolutional neural network to conserve mass in data assimilation

Yvonne Ruckstuhl, Tijana Janjić, and Stephan Rasp

Download

Final revised paper (published on 09 Feb 2021)
Preprint (discussion started on 25 Sep 2020)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of "Training a convolutional neural network to conserve mass in data assimilation" by Y. Ruckstuhl , T. Janjic, and S. Rasp', Marc Bocquet, 24 Oct 2020
- AC1: 'reply to review of "Training a convolutional neural network to conserve mass in data assimilation”', Yvonne Ruckstuhl, 15 Dec 2020
RC2: 'Interactive comment on “Training a convolutional neural network to conserve mass in data assimilation” by Yvonne Ruckstuhl et al.', Svetlana Dubinkina, 05 Nov 2020
- AC2: 'reply to review of "Training a convolutional neural network to conserve mass in data assimilation”', Yvonne Ruckstuhl, 15 Dec 2020

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Yvonne Ruckstuhl on behalf of the Authors (15 Dec 2020) Manuscript

ED: Publish subject to technical corrections (17 Dec 2020) by Natale Alberto Carrassi

AR by Yvonne Ruckstuhl on behalf of the Authors (18 Dec 2020) Author's response Manuscript

Short summary

The assimilation of observations using standard algorithms can lead to a violation of physical laws (e.g. mass conservation), which is shown to have a detrimental impact on the system's forecast. We use a neural network (NN) to correct this mass violation, using training data generated from expensive algorithms that can constrain such physical properties. We found that, in an idealized set-up, the NN can match the performance of these expensive algorithms at negligible computational costs.