About ML4WX

ML4WX is an open archive of machine learning and AI research for weather and climate. It started with a deep search of the literature since 2018 (plus a curated set of older classics) and grows every week.

Scope

A paper is IN SCOPE if machine learning / AI is a main contribution AND its main target is the atmosphere or the climate system (weather, climate, their extremes and observations). Coupled models where the ocean is part of a weather/climate forecast (e.g. ENSO, S2S) are in scope. Multi-domain models that include the atmosphere (e.g. Aurora) are in scope. Fire weather and heavy-rain extremes are in scope. A paper is LATER if ML is central but the main target is the ocean or sea ice only, hydrology/floods/river flow, or air quality/chemistry only. A paper is OUT if ML is not a main contribution (e.g. classic NWP or climate studies that only mention ML), or if it is not about weather or climate at all (e.g. image de-raining, "weather" in other senses, generic time series).

How it works

Weekly, Monday morning (Zurich time). A GitHub Action searches arXiv (physics.ao-ph and the ML categories) and OpenAlex (journals and conferences) for new papers. Claude reads each title and abstract and decides relevance, topic, method, time scale, data and a one-sentence summary. Organizations come from OpenAlex affiliations and the arXiv HTML title page. The result is a pull request that a maintainer reviews.
Daily. The same pipeline looks for important new papers and lab blog posts and opens a GitHub issue with the highlight label, so watchers hear about them the same day.
On every merge. This site is rebuilt and the search index is regenerated.

How highlights work

Each paper gets an importance score from 0 to 100. Points come from a tier-1 lab (35), a tier-2 lab (10), a top venue (20), open code (10), operational use (10), many citations (older papers), and 4 × a 0–10 significance judgment by Claude. A paper is a highlight when its score is 55 or more and its significance is at least 6, or when its significance is 9 or more. So a big lab alone is not enough, and landmark work from anywhere is included.

Data and reuse

One YAML file per paper in papers/.
BibTeX, CSV, JSON, RSS (all), RSS (highlights).
Abstracts are shown only when their license allows it (arXiv metadata, CC BY, CC0). Labels and summaries are generated by Claude and may contain errors.

Contribute

Missing paper or wrong label? Open an issue with the arXiv id or DOI, or use the "Edit this entry" link on any paper page.