10 parallel experiments: Can AI drive Nature-level papers in ocean science?
10 个研究方向并行推进。AI 从选题到投稿全流程主导，人类负责物理把关和审核纠偏。全过程公开。
10 independent research projects, each exploring a different scientific question in physical oceanography and ocean remote sensing. AI (Claude, GPT, etc.) serves as the primary research engine — literature review, data analysis, figure generation, manuscript writing. Human scientists provide physical intuition, error checking, and quality control.
Each project runs on a fast cycle:
Day 0 Day 1 Day 2-3 Day 4-7 Day 7+
Topic lock → AI first draft → Human review → Internal review → Submit
Failures, dead ends, and rejections are part of the experiment and will be visible.
See projects/DASHBOARD.md for live status of all 10 projects.
| # | Direction | Status | Journal |
|---|---|---|---|
| P01 | Submesoscale air-sea coupling regime transition (SWOT) | 🔬 D0 Explore | — |
| P02 | Equatorial Kelvin wave topological robustness (SWOT + altimetry) | ✅ D1 Complete — 投稿就绪 | Nature Comm. |
| P03 | Neglected eddy boundary signals (SWOT altimetry) | 🔬 D0 Explore | — |
| P04 | Wave-ice feedback: swell attenuation at Antarctic ice edge | ✅ D1 Complete — 投稿就绪 | Nature Comm. |
| P05 | Deep ocean mixing energy deficit — SWOT SSH constraints | ✍️ D1 | — |
| P06 | TBD | ⬜ | — |
| P07 | TBD | ⬜ | — |
| P08 | TBD | ⬜ | — |
| P09 | TBD | ⬜ | — |
| P10 | TBD | ⬜ | — |
Each project is self-contained in its own directory with literature, analysis scripts, figures, manuscript, and AI interaction logs.
Target journals: Nature Geoscience, Nature Climate Change, Nature Communications, and other high-impact journals.
Directions include (not limited to):
- Satellite ocean dynamic remote sensing (altimetry, SWOT, SAR)
- Physical oceanography (mesoscale eddies, submesoscale dynamics, internal waves)
- Air-sea interaction
- Polar ocean and cryosphere
AI: Claude, GPT, and other LLMs as the core production tool
Software: Python, MATLAB, GMT
Public data: SWOT L3/L4, CMEMS, ERA5, Argo, AVISO, WOD, etc.
In-situ observations: GNSS wave-tide integrated buoys developed by Qingdao Anhai, with continuous field measurements across multiple sea areas. Some data remain unpublished — suitable for satellite validation or independent research.
Each project operates independently. You can join one or more projects based on your expertise and interest. Authorship is per-project, based on actual contributions tracked via GitHub.
Ways to contribute:
- Ideas & hypotheses — A question worth asking but no time to pursue? AI can run the exploration.
- Physical-intuition review — Judge whether AI's analysis makes physical sense.
- Error checking — Catch hallucinated references, wrong magnitudes, skipped derivations.
- Literature tracking — Confirm novelty, find benchmark papers.
To join: Open an Issue, submit a PR, or contact via WeChat group (see recruitment post).
OpenSCI-Ocean/
├── README.md # This file
├── COLLABORATION.md # Rules, authorship, workflow
├── CONTRIBUTORS.md # Contribution tracking
├── LICENSE # MIT
├── projects/
│ ├── DASHBOARD.md # Status overview of all 10 projects
│ ├── PROJECT_TEMPLATE.md # Template for new projects
│ ├── p01/ # Project 01 (independent, self-contained)
│ │ ├── README.md # Project card: question, status, data links
│ │ ├── literature/ # AI literature review notes
│ │ ├── analysis/ # Scripts + data source links (no raw data)
│ │ ├── figures/ # All generated figures (incl. intermediate)
│ │ ├── manuscript/
│ │ │ ├── v1_ai_draft/ # AI-generated first draft
│ │ │ ├── v2_delivery/ # Human-reviewed delivery draft
│ │ │ ├── v3_final/ # Internally reviewed final draft
│ │ │ └── submitted/ # Submitted version + cover letter
│ │ └── logs/ # AI prompt logs
│ ├── p02/ ... p10/ # Projects 02-10 (same structure)
├── shared/ # Shared utilities
└── docs/ # General documentation
No raw data in this repo. GitHub stores only data source links, download scripts, analysis code, figures, and the complete manuscript evolution from AI draft to final submission.
- What percentage of a high-quality paper can AI complete independently?
- Where does it fail most in mechanism-driven disciplines?
- What is the irreplaceable human contribution?
- How fast can this human-AI loop produce publishable science?
| Document | Purpose |
|---|---|
| COLLABORATION.md | Authorship, roles, workflow, IP, data policy |
| CONTRIBUTING.md | How to fork, branch, commit, PR |
| REVIEW_CHECKLIST.md | D2/D3 stage review checklists |
| CODE_OF_CONDUCT.md | Behavioral expectations |
| CONTRIBUTORS.md | Contribution tracking board |
| shared/ai_workflow.md | AI prompt templates for each stage |
| shared/data_sources.md | Public data source directory |
| shared/environment.yml | Conda environment setup |
Essentials:
- Authorship per project, following CRediT taxonomy
- AI usage fully disclosed in each manuscript
- Zero tolerance for hallucinated citations
- No raw data in repo — only links and scripts
- Computation on local machines or rented servers
- This is an experiment, not a guarantee
MIT License. See LICENSE.
10 experiments. Some will fail. All will be visible. If you want to find where AI's real boundary lies in scientific research, join us.
10 个实验。有的会失败。全部公开。想找到 AI 做科研的真实边界在哪里，来一起折腾。