This README provides a single, cohesive guide for contributing to the AI‑assisted marine ecosystem modelling experiment. It blends the study protocol, modelling requirements, and AI usage guidelines into one workflow so you can get started quickly and submit consistently formatted results.
We are testing how AI tools can support typical marine ecosystem modelling workflows—model development, translation between frameworks, and simulation/diagnostics—and whether AI assistance can produce reliable, comparable outputs across model types.
If you would like to participate, the strong preference is to translate the provided EwE model into a different modelling framework, mizer. If you want to do something else, please contact the facilitators first.
Time commitment: approximately 24 hours total (20 hours modelling + 4 hours reviewing). – If you finish early, report your time. – If you reach 20 hours without finishing, submit what you have.
Timeline:
- Start: Mid-November 2025
- Submission: February 28, 2026
AI access: We will provide API keys for public AI tools (initial limit ~AUD 150 or USD equivalent) per participant. Contact facilitators if you need more credit and budget allows.
You will:
- Download the provided input files and context for an Ecopath with EcoSim model in the Southern Ocean (Dahood et al. 2019).
- Develop a mizer model based on the same data and study.
- Run analogous simulations to those implemented in Dahood et al. using FishMIP/ISIMIP3a conventions.
- Produce diagnostics (validation plots, skill metrics) following ISIMIP and FishMIP guidance.
- Document your use of AI (tools, tasks, performance, limitations).
- Submit and Survey via an anonymous (10-15 minute) survey, which includes an upload link.
- Review another submission for quality and effectiveness.
Steps 1-6 are expected to take no longer than 20 active hours of work between now and February 28, 2026. Step 7 is expected to take less than 4 hours and will occur sometime in March 2026.
We would recommend that you use mizer as the target modelling framework (even if you may be unfamiliar with it). If you would much prefer to use another framework, please contact [email protected] and [email protected] to discuss.
Organise your contribution using the mandatory folder structure below (one folder per region/model):
/YourRegionName/
│
├── /Inputs/
│ ├── Environmental drivers (ISIMIP3a format)
│ ├── Species/functional group data
│ └── Fishing effort data
│
├── /Models/
│ ├── Human-developed/
│ └── AI-assisted/
│
├── /Diagnostics/
│ └── Validation plots, skill metrics
│
└── /Documentation/
└── Model description, assumptions, notes
Here are is a (non-exhaustive) list of potential tools you may want to try:
IDE Plugins & AI Coding Assistants
GitHub Copilot, Sourcegraph Cody, Cursor, Windsurf (formerly Codeium), Roo Code, JetBrains AI Assistant, Codeium, Tabnine, Amazon CodeWhisperer, AskCodi, Blackbox AI, Visual Copilot, Qodo, DeepSeek Coder, CodeGeeX, TabbyML, FauxPilot, Continue.dev, Augment, Cline, Pieces
Standalone AI IDEs
Cursor, Windsurf, Zed, Replit Ghostwriter, Codium, CodeStory
General AI Chatbots for Coding and Web Search
ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Microsoft Copilot, Perplexity AI, Grok
Specialized / Emerging Tools
Claude Code, OpenAI Codex, Mutable.ai, Refact.ai, StarCoder, Devin AI, Phind, Sourcery, AI Commits
Anything else!
Good practice
- Do your best to review and validate AI‑generated code, text, and figures. Part of the aim of this study is to evaluate how well these tools work when used by researchers with a range of domain and modelling expertise.
- Keep brief notes or logs of impactful prompts/responses.
Data privacy
- All of the data we are using in this study is fine to use in conjunction with any AI tools.
- For your own use-cases, do not paste into public AI tools: unpublished, confidential/proprietary, sensitive ecological/stakeholder data, or any non‑public content from primary sources.
- When unsure, ask a facilitator or keep data local.
In your submission survey, you will be asked to include your reflections on how well the AI tools you used worked. Please keep track of:
- Tools used (which, and for what tasks)
- Effectiveness per task (e.g., Likert scale or short rating)
- Where AI excelled / failed
- Where you had to step in
- Any refusals or safety blocks that affected progress
(If feasible, attach relevant chat logs or excerpts.)
- Validate model outputs using ISIMIP and FishMIP guidance.
- Quantitative evaluation will follow the FishMIP model skill procedure to compare across models and regions.
- Participants will peer‑review one or two other submissions for qualitative checks (e.g., “Does it pass the smell test?”).
Place validation artefacts (plots, metrics, notes) under:
/YourRegionName/Diagnostics/
Place all of the following into a zip file and upload them via the anonymous upload link provided during the survey:
- Model files (execution‑ready)
- Outputs as CSV (time‑series)
- Diagnostics (validation plots, skill metrics)
- AI usage appendix and (if possible) chat histories
During the survey you will be prompted to include a short description of your approach, assumptions, and any notes on AI usage.
Please complete the post‑project survey: [[Survey Link]](https://csiro.qualtrics_6opBbEzhMbvLHOC
You can preview the survey content here: https://github.com/s-spillias/TranslatingMEMs/blob/main/SurveyPreview.md
flowchart TB
Start([Project Start]) --> Phase1[Phase 1:<br/>Understand EwE Model]
Phase1 -.-> Steps1[Read paper<br/>Analyze model<br/>Review data files]
Phase1 --> Phase2[Phase 2:<br/>Setup mizer Model]
Phase2 -.-> Steps2[Create params object<br/>Map species groups]
Phase2 --> Phase3[Phase 3:<br/>Calibrate Model]
Phase3 -.-> Steps3[Calculate time-averaged biomass<br/>Calibrate to steady-state]
Phase3 --> Phase4[Phase 4:<br/>Prepare ISIMIP3a Inputs]
Phase4 -.-> Steps4[Load climate forcings<br/>Process fishing data<br/>Create resource spectrum]
Phase4 --> Phase5[Phase 5:<br/>Run FishMIP Simulations]
Phase5 -.-> Steps5[Configure spin-up<br/>Apply forcings<br/>Run simulations]
Phase5 --> End([Project Complete])
%% Styling
classDef phase fill:#e1f5ff,stroke:#0066cc,stroke-width:3px
classDef steps fill:#f0f0f0,stroke:#666,stroke-width:1px
class Phase1,Phase2,Phase3,Phase4,Phase5 phase
class Steps1,Steps2,Steps3,Steps4,Steps5 steps
EwE_files/- EwE model data from Dahood et al. 2020Dahood WAP -Basic input.csv- Base model parametersDahood WAP -Diet composition.csv- Diet matrixDahood WAP -Time series grid.csv- Observed biomass time seriesDahood WAP -Forcing functions grid.csv- Environmental forcingsDahood WAP -Fishing effort functions grid.csv- Fishing effort dataDahood WAP -Fishing mortality functions grid.csv- Fishing mortality data
ISIMIP3a_climate_forcing_inputs/- Required forcing:
temperature_forcing/- Sea surface temperature (tos) and bottom temperature (tob)plankton_forcing/- Phytoplankton (diatoms, diazotrophs, picophytoplankton) and Zooplankton (mesozooplankton, microzooplankton)
- Optional forcing:
optional_additional_forcing/Sea_ice_forcing/- Sea ice concentration (siconc) for advanced exploration
- Required forcing:
ISIMIP3a_fishing_forcing_inputs/- Historical fishing effort (1841-2017)effort_histsoc_1841_2017_western-antarctic-peninsula.parqueteffort_dictionary.parquet- Metadata for fishing effort data
- EwE Model Fishing Fleets:
- Krill fishery (Antarctic krill)
- N. rossii fishery (marbled rockcod)
- C. gunnari fishery (mackerel icefish)
ISIMIP3a_calibration_data/- Catch data for calibration (1850-2017)calibration_catch_histsoc_1850_2017_western-antarctic-peninsula.parquetcatch_dictionary.parquet- Metadata for catch data
- Dahood et al. 2020 Paper: https://doi.org/10.1016/j.biocon.2020.108766
- mizer Course: https://mizer.course.sizespectrum.org/
- FishMIP 2.0 ISIMIP3a Protocol: https://github.com/Fish-MIP/FishMIP2.0_ISIMIP3a
- Ortega-Cisneros et al. 2025 - Fishing Effort Mapping: https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2024EF004826 (See Section 2.2 Step 5)
- FishMIP Input Explorer: https://rstudio.global-ecosystem-model.cloud.edu.au/shiny/FishMIP_Input_Explorer/
- Phase 1: Understand the Western Antarctic Peninsula EwE model structure and data
- Phase 2: Initialize mizer params object using EwE model as template
- Phase 3: Calibrate steady-state model using time-averaged biomass
- Phase 4: Prepare climate forcings and fishing inputs following ISIMIP3a standards
- Phase 5: Execute FishMIP protocol with model spin-up and historical simulations
Note: This workflow integrates AI-assisted coding throughout all phases. Participants should leverage LLM coding agents to navigate data processing, model calibration, and protocol implementation steps.
Questions about modelling or AI tool usage? Get in touch: