AI for Water · Roadmap

Climate change, ageing infrastructure, sparse monitoring, and rising data volumes are reshaping water management. AI is most useful where physics is partly known but data is rich, including forecasting, anomaly detection, satellite mapping, and decision support. The biggest near-term wins are not in replacing physics-based models but in augmenting them with cheap inference, better priors, and faster iteration.

Core Python including data types, control flow, functions, packages, and virtual environments. You should be comfortable writing scripts that read files, transform data, and call libraries. Pick uv for dependency management and stick with it across projects, since it is now the fastest path from clone to running code.

Notebooks for exploration, scripts for production. Learn nbconvert, jupytext, and how to keep notebooks reproducible with papermill or nb-clean. marimo is an interesting alternative if you want pure-Python reactive notebooks without hidden state issues.

Shells, file globbing, ssh, scp, curl, awk, jq. Pipelines, cron, basic Bash. You will hit these the moment you talk to a server, especially for batch geospatial workflows over many tiles or many dates.

Water balance, the runoff to rainfall relationship, evapotranspiration, groundwater recharge, surface vs subsurface flow. Without this, your models will fit numbers but miss physics. Be able to draw a basin water balance from memory before you train any model.

Version control for code and notebooks, branching, pull requests, conflict resolution. Non-negotiable for collaboration and reproducibility. Pair notebooks with jupytext .py files for diff-friendly history that survives code review.