From Code to Spec
― Weather and Climate Modeling for the AI Era ―
Advancing the SCALE weather and climate modeling framework through generative-AI, specification-driven development. By treating the specification of physical and numerical models — rather than the code itself — as the core asset, and delegating code generation to generative AI, this project aims to realize world-leading weather, climate, and planetary science on Fugaku and Fugaku NEXT.
Project Overview
Turning the tacit knowledge of weather models into something AI can understand
Background
Numerical weather and climate prediction models have developed as a loosely connected "model ecosystem," built by research institutions around the world. SCALE, a model developed in Japan, is one member of this ecosystem.
Challenge
Over years of development, these models have grown into codebases spanning hundreds of thousands of lines. The intricate optimizations needed to extract performance have made the code increasingly difficult to understand — not only for humans, but for AI as well. Adapting to each new generation of hardware, such as GPUs and Fugaku NEXT, still relies heavily on manual work by specialists.
Approach
This project treats "specifications" — natural-language descriptions of physical laws, numerical algorithms, and implementation policy — rather than the code itself, as the core development asset, and proposes a new development paradigm in which generative AI handles code generation. Once the specifications are in place, optimized code can be automatically generated for new computing environments.
Goal
Using this approach, we will renew the SCALE weather model to realize world-leading weather, climate, and planetary science on Fugaku and Fugaku NEXT — contributing to societal challenges such as disaster prevention and mitigation, climate change response, food and water security, and planetary exploration.
Subthemes
Spec-Driven Infrastructure and Five Frontier Science Cores
This project is built on two pillars: the foundational spec-driven infrastructure (Pillar 1) and five frontier science cores (Pillar 2) that validate it at production scale.
A
Spec-Driven Infrastructure
Builds a three-tier specification — physics, algorithm, and implementation — together with verification gates and a multi-LLM control harness supporting multiple generative AI systems. Combines human optimization expertise with AI-driven code generation and auto-tuning, while formalizing existing Fortran assets as specifications. This is the core technology underlying every science core.
B-1
DG Dynamical Core
Renews the computational "heart" that solves atmospheric flow, using a high-order discontinuous Galerkin method aimed at world-leading resolution in global simulations.
B-2
Super-Droplet Cloud Microphysics
A simulation method that tracks water droplets and ice particles individually, as particles — aiming to clarify the mechanisms behind localized torrential rainfall.
B-3
3D Radiative Transfer
Improves the accuracy of weather prediction by accounting for the three-dimensional structure of clouds in radiative transfer calculations.
B-4
Data Assimilation × Control
Combines techniques that correct forecasts by incorporating observational data with research into predictive control of linear precipitation bands.
B-5
Planetary Applications
Applies the technologies developed here to atmospheric research on planets beyond Earth, including Mars, Venus, and Jupiter.
Team
Full team across the lead, collaborating, and partner institutions
Principal Investigator
Seiya Nishizawa
Senior Scientist, RIKEN Center for Computational Science (R-CCS)
RIKEN
- Yuta Kawai
- Tristan HASCOET
- Hirofumi Tomita
- Takemasa Miyoshi
- Sachiho Adachi
- Tsuyoshi Yamaura
- Tamaki Suematsu
- Megumi Okazaki
Yokohama National University
- Ryuji Yoshida
Nagoya University
- Takahiro Katagiri
- Tetsuya Hoshino
- Daichi Mukunoki
University of Hyogo
- Shinichiro Shima
- Tomoro Yanase
Osaka University
- Yousuke Sato
Chiba University
- Shigenori Otsuka
- Atsushi Okazaki
- Shunji Kotsuki
Kobe University
- Hiroki Kashimura
- Yoshiyuki Hayashi
The University of Tokyo
- Takashi Shimokawabe
- Kazuya Yamazaki
- Xuanzhengbo Ren
Kyoto University
- Shinichi Takehiro
Ehime University
- Masanori Kameyama
Kyoto Sangyo University
- Kazunori Ogohara
Japan Agency for Marine-Earth Science and Technology (JAMSTEC)
- Takehiro Miyagoshi
Hokkaido Information University
- Yohei Sasaki
Publications
Papers & Presentations
Positions
Postdoctoral Positions
We are recruiting next-generation researchers
During the project period (October 2026 – March 2031), we plan to recruit 4–5 postdoctoral researchers who can work across both "spec-driven weather model development" and "AI × meteorology." Specific position details — field, qualifications, deadlines, and more — will be posted here once finalized.
Contact
Get in Touch
Project Secretariat (RIKEN Center for Computational Science)
spec-to-science@ml.riken.jp