Development of neural networks based on physical models and exploration of meteorological physics

2020 Fiscal Year Final Research Report

• Project/Area Number: 19K22876
• Research Category: Grant-in-Aid for Challenging Research (Exploratory)
• Allocation Type: Multi-year Fund
• Review Section: Medium-sized Section 61:Human informatics and related fields
• Research Institution: Osaka University
• Principal Investigator: Fukui Ken-ichi 大阪大学, 産業科学研究所, 准教授 (80418772)
• Co-Investigator(Kenkyū-buntansha): 冨田 智彦 熊本大学, 大学院先端科学研究部(理), 准教授 (20344301)
• Project Period (FY): 2019-06-28 – 2021-03-31
• Keywords: ディープラーニング / 偏微分方程式 / 温度風 / マルチタスク学習 / 敵対的サンプル生成

Outline of Final Research Achievements

In this work, we tackled the following two researches. Frist, we proposed a deep learning architecture that can decompose and output known physical model component and model residual component. Then, we verified the accuracy of the proposed method with the task of estimating the wind velocity in the upper troposphere from the atmospheric conditions in the lower layer. The validity of the wind vector distribution of the residual component by the proposed method was justified from the knowledge of meteorology. Second, we proposed an improvement method by multi-task learning and adversarial exsample generation for the method of obtaining the value of the solution at an arbitrary position of the partial differential equation, by automatic differentiation and deep learning. We confirmed the improvement of estimation accuracy for some basic PDEs.

Free Research Field

知能情報学

Academic Significance and Societal Importance of the Research Achievements

本研究成果の1. 物理モデル成分とモデル残差成分を分解して出力可能なディープラーニングアーキテクチャと，2. ディープラーニングと自動微分による偏微分方程式の求解の高精度化，の両者を将来的に統合することで，End-to-Endで任意の観測量から偏微分方程式を満足するように任意の物理量を推定することが可能になる．本統合方式により，観測データと既知の方程式を有する自然科学において，両者を活用して予測精度の向上や最適化など広く応用が期待できる．