A Small Station-Graph World Model for 0–24 h Weather in North India: A Version-by-Version Record Against Global AI Weather Models and Time-Series Foundation Models

Global AI weather models such as ECMWF's AIFS now match or beat physics-based forecasts on grid averages, and zero-shot time-series foundation models such as Google's TimesFM and Amazon's Chronos forecast any series without training. Neither is built for what many users in India need: the next 24 hours of temperature, humidity and wind at a specific station. We study whether a small model trained only on a station network's own history can do better. PWM-W is a graph world model of about one million parameters in which every station is a token. It comes in two forms: a pure form that uses observations only, and a hybrid that also reads global forecasts published before the forecast starts. We test it in three north-Indian leagues: six airports (METAR), 20 NOAA stations around Delhi (WMO) and 25 Delhi air-quality stations' weather sensors (CPCB). The opponents are real archived operational forecasts of AIFS, IFS, ICON, GFS and NOAA's GraphCast-based AIGFS, scored raw and after a trailing local bias correction, together with TimesFM 2.5 and 3.0 and Chronos-2. Every change from version 0.3 to 0.9 was adopted only if it beat the default setting on validation by more than two seed standard deviations, and several proposed changes failed that test. At the airports, the hybrid beats every global model on every variable it forecasts when each is scored on its newest run published before the forecast starts (AIFS by 4.4%, 11.8% and 7.4%). Against AIFS's freshest run it still loses on temperature and humidity. The pure model beats all three foundation models on every variable at the airport and WMO stations. On the noisier CPCB sensors, cross-network training, a 14-day encoder and weight averaging turned four of its losses into ties, but it still trails Chronos-2 on temperature and all three foundation models on wind. Bias-corrected FourCastNet and GenCast Mini beat the pure model on airport temperature. We report every result, including the losses, and the test-reuse risk of a version-by-version study. A data-threshold study shows that PWM-W needs 12–24 months of a network's own history to stop losing to the foundation models and 24–48 months to beat all three on every variable. Paper 2 of 5 in ParadoxAI Lab's series on compact physical world models (PWM-W, North India weather). Code, rules, logs and weights: https://github.com/paradoxailab/paradox-world-model. Contact: [email protected]

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24370653
Primary Topic
Meteorological Phenomena and Simulations
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Small Station-Graph World Model for 0–24 h Weather in North India: A Version-by-Version Record Against Global AI Weather Models and Time-Series Foundation Models

Anupam Singh, Aditya Kumar
Zenodo (CERN European Organization for Nuclear Research)
Meteorological Phenomena and Simulations
preprint

A Small Station-Graph World Model for 0–24 h Weather in North India: A Version-by-Version Record Against Global AI Weather Models and Time-Series Foundation Models

Anupam Singh, Aditya Kumar
preprint en

Abstract

Global AI weather models such as ECMWF's AIFS now match or beat physics-based forecasts on grid averages, and zero-shot time-series foundation models such as Google's TimesFM and Amazon's Chronos forecast any series without training. Neither is built for what many users in India need: the next 24 hours of temperature, humidity and wind at a specific station. We study whether a small model trained only on a station network's own history can do better. PWM-W is a graph world model of about one million parameters in which every station is a token. It comes in two forms: a pure form that uses observations only, and a hybrid that also reads global forecasts published before the forecast starts. We test it in three north-Indian leagues: six airports (METAR), 20 NOAA stations around Delhi (WMO) and 25 Delhi air-quality stations' weather sensors (CPCB). The opponents are real archived operational forecasts of AIFS, IFS, ICON, GFS and NOAA's GraphCast-based AIGFS, scored raw and after a trailing local bias correction, together with TimesFM 2.5 and 3.0 and Chronos-2. Every change from version 0.3 to 0.9 was adopted only if it beat the default setting on validation by more than two seed standard deviations, and several proposed changes failed that test. At the airports, the hybrid beats every global model on every variable it forecasts when each is scored on its newest run published before the forecast starts (AIFS by 4.4%, 11.8% and 7.4%). Against AIFS's freshest run it still loses on temperature and humidity. The pure model beats all three foundation models on every variable at the airport and WMO stations. On the noisier CPCB sensors, cross-network training, a 14-day encoder and weight averaging turned four of its losses into ties, but it still trails Chronos-2 on temperature and all three foundation models on wind. Bias-corrected FourCastNet and GenCast Mini beat the pure model on airport temperature. We report every result, including the losses, and the test-reuse risk of a version-by-version study. A data-threshold study shows that PWM-W needs 12–24 months of a network's own history to stop losing to the foundation models and 24–48 months to beat all three on every variable. Paper 2 of 5 in ParadoxAI Lab's series on compact physical world models (PWM-W, North India weather). Code, rules, logs and weights: https://github.com/paradoxailab/paradox-world-model. Contact: [email protected]

Zenodo (CERN European Organization for Nuclear Research)
Meteorological Phenomena and Simulations
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon