日本地球惑星科学連合2022年大会

講演情報

[E] 口頭発表

セッション記号 A (大気水圏科学) » A-AS 大気科学・気象学・大気環境

[A-AS05] スーパーコンピュータを用いた気象・気候・環境科学

2022年5月23日(月) 10:45 〜 12:15 106 (幕張メッセ国際会議場)

コンビーナ:八代 尚(国立研究開発法人国立環境研究所)、コンビーナ:川畑 拓矢(気象研究所)、宮川 知己(東京大学 大気海洋研究所)、コンビーナ:寺崎 康児(理化学研究所計算科学研究センター)、座長:寺崎 康児(理化学研究所計算科学研究センター)

10:45 〜 11:00

[AAS05-07] Big Data Assimilation: Real-time 30-s-update Forecast Experiments Using Fugaku in Tokyo in 2021

*三好 建正1雨宮 新1本田 匠1大塚 成徳1前島 康光1、Taylor James1、富田 浩文1、西澤 誠也1、末木 健太1、山浦 剛1、石川 裕2佐藤 晋介3牛尾 知雄4、小池 佳奈5、星 絵里香5 (1.理化学研究所、2.国立情報学研究所、3.情報通信研究機構、4.大阪大学、5.エムティーアイ)

キーワード:数値天気予報、データ同化、ビッグデータ、局地的豪雨、フェーズドアレイ気象レーダ

The Japan’s Big Data Assimilation (BDA) project started in October 2013 and ended its 5.5-year period in March 2019. Here, we developed a novel numerical weather prediction (NWP) system at 100-m resolution updated every 30 seconds for precise prediction of individual convective clouds. This system was designed to fully take advantage of the phased array weather radar (PAWR) which observes reflectivity and Doppler velocity at 30-second frequency for 100 elevation angles at 100-m range resolution. By the end of the 5.5-year project period, we achieved less than 30-second computational time using the Japan’s flagship K computer for past cases with all input data such as boundary conditions and observation data being ready to use. The direct follow-on project started in April 2019 for three years (i.e., ending soon in March 2022). We continued the development and achieved real-time operations of this novel 30-second-update NWP system for demonstration at 500-m resolution during July 31 and August 7, 2020, using the supercomputer Oakforest-PACS operated jointly by the Tsukuba University and the University of Tokyo. In 2021, we performed real-time experiments during two periods corresponding to the Tokyo Olympic and Paralympic games, i.e., July 20-August 8 and August 24-September 5, with an enhanced system using the new Japan’s flagship supercomputer Fugaku, ranked #1 in the most recent top500 list. Taking advantage of the computing power, we increased the ensemble size from 50 to 1000 for the local ensemble transform Kalman filter (LETKF). This presentation will summarize the real-time demonstration in 2021 and discuss future perspectives based on the 8.5-year-long project efforts.