Research

It starts from papers, and is sharpened by what we learn on site.

  1. 2023-2025

    Research

    • 2023 Started in the laboratory
    • Sept. 2024 Presented at an international conference (KES 2024)
    • Dec. 2025 Paper published in Tenki (Meteorological Society of Japan)

    What this added

    • Cases where generative AI alone misreads weather charts
    • Judgment by rules, with generative AI as an interpreter of words
  2. 2024-2026

    Weather experts

    • Research with the co-authors of our papers (two researchers who were at a forecasting company when we wrote them)
    • June 2026 Selected for a research grant of the Meteorological Culture Prize
    • Sept. 2026 Began running experiments with the logical processing

    What this added

    • Decision rules that follow the definitions and criteria of the Japan Meteorological Agency
    • Fine-grained rules drawn out of the experts' knowledge
  3. From now on

    On site

    • Infrared inspection of building walls (demo version on trial with several partner companies)
    • Trips and tours (in preparation)

    What will be added

    • Hearing the realities of each field
    • Drawing out finer information that research alone cannot reach

Four themes behind the research

  • Objective classification of pressure patterns (SOM)

    Classifying pressure patterns has been left to forecasters' feel.

  • Automatic description of fronts (Swin-Unet)

    Fronts are drawn by forecasters, following rules and experience.

  • Controlling and verifying generative AI

    When generative AI is left to discuss the weather, it can write things that are not true.

  • Accumulating forecasters' tacit knowledge

    We collect the fine-grained rules of on-site judgment that are hard to obtain.

SOM and automatic front description are designed to be used only to reproduce past days. They are not used to predict the future.

Papers and presentations

  • ChatGPTによる天気図画像からの天気コメント生成の検証 (Verification of weather comment generation from weather map images by ChatGPT)

    Takumi Takasuka, Yuki H. Takano, Shotaro Watanabe, Gendo Kumoi

    Tenki, Vol.72, No.12, pp.603-611 (peer-reviewed short report, in Japanese), 2025

    DOI
  • Recognition Performance Validation of Weather Map Images by ChatGPT

    Takumi Takasuka, Yuki H. Takano, Shotaro Watanabe, Gendo Kumoi

    Procedia Computer Science, Vol.246, pp.3132-3141 (KES 2024, peer-reviewed), 2024

    DOI
More presentations (4)
  • Verification of an automatic front description system using Swin-Unet (in Japanese)

    Takumi Takasuka, Yuki H. Takano, Shotaro Watanabe, Gendo Kumoi

    Meteorological Society of Japan, 2025 Spring Meeting, 2025

  • A study of SOM pressure-pattern classification integrating a structural similarity index (in Japanese)

    Takumi Takasuka, Shotaro Watanabe, Yuki H. Takano, Gendo Kumoi

    48th Symposium on Information Theory and its Applications (SITA2025), 2025

  • Clustering pressure patterns with self-organizing maps and analysis with 1 km mesh weather data (in Japanese)

    Takumi Takasuka et al.

    Information Processing Society of Japan, 87th National Convention, 2025

  • Verification of weather comment generation from numerical weather data (in Japanese)

    Takumi Takasuka et al.

    Meteorological Society of Japan, 2024 Autumn Meeting, 2024