Emulating the Adaptation of Wind Fields to Complex Terrain with Deep Learning

• DOI: 10.1175/aies-d-22-0034.1
• Publication Date: 2022-12-27T12:33:55Z
• Authors: Louis Le Toumelin; Isabelle Gouttevin; Nora Helbig; Clovis Galiez; Mathis Roux; Fatima Karbou
• Keywords: [SDU.OCEAN]Sciences of the Universe [physics]/Ocean, Atmosphere; [SDU.OCEAN]Sciences of the Universe [physics]/Ocean; Artificial intelligence; 550; Atmosphere; [SDU.OCEAN] Sciences of the Universe [physics]/Ocean, Atmosphere; Deep learning; Wind; 551; [SDU.ENVI] Sciences of the Universe [physics]/Continental interfaces, environment; [STAT.ML] Statistics [stat]/Machine Learning [stat.ML]; Data science; [STAT.ML]Statistics [stat]/Machine Learning [stat.ML]; 13. Climate action; Snow; Machine learning; [SDU.ENVI]Sciences of the Universe [physics]/Continental interfaces; [SDU.ENVI]Sciences of the Universe [physics]/Continental interfaces, environment; environment

Abstract Estimating the impact of wind-driven snow transport requires modeling wind fields with a lower grid spacing than on order 1 or few kilometers used in current numerical weather prediction (NWP) systems. In this context, we introduce new strategy to downscale from NWP systems decametric scales, using high-resolution (30 m) topographic information. Our method (named “DEVINE”) is leveraged convolutional neural network (CNN), trained replicate behavior complex atmospheric model ARPS, and was previously run large number (7279) synthetic Gaussian topographies under controlled conditions. A 10-fold cross validation reveals that our CNN able accurately emulate ARPS (mean absolute error for speed = 0.16 m s −1 ). We then apply DEVINE real cases Alps, is, downscaling forecast by AROME system information alpine topographies. proved reproduce main features terrain (acceleration ridges, leeward deceleration, deviations around obstacles). Furthermore, an evaluation quality-checked observations acquired at 61 sites French Alps improved downscaled winds (AROME mean bias reduced 27% DEVINE), especially most elevated exposed stations. Wind direction however, only slightly modified. Hence, despite some limitations inherited simulations setup, appears be efficient tool whose minimalist architecture, low input data requirements (NWP topography), competitive computing times may attractive operational applications. Significance Statement largely influences spatial distribution mountains, direct consequences hydrology avalanche hazard. Most models predicting use several kilometers, too coarse represent patterns mountain winds. novel based deep learning increase resolution while maintaining acceptable computational costs. mimics part complexity only. compared results collected showed improves representation winds, notably observation