Work

Atlas-F: Pre-Registered PM2.5 Forecasting Across India by Online Fusion of a Small Station-Graph Model and a Foundation Model

Year: 2026
Type: preprint
Abstract: Small models trained on a country's own monitoring stations and large time-series foundation models (TSFMs) are both proposed for air-quality forecasting, but they are rarely compared under a protocol that fixes the claims before the scores are seen, and almost never against the forecasts a government actually issues. We report a pre-registered study of 24-hour PM2.5 forecasting at 263 Indian stations in 14 networks. A confirmation test of Atlas, a 0.97M-parameter station-graph model previously reported to beat TSFMs, fails its pre-registered rule: with each of four networks held out of pre-training it beats TimesFM 2.5 in Uttar Pradesh but loses in Maharashtra and Karnataka, and the losses trace to never-trained stations in geographically spread networks, where attention over all stations mixes unrelated sites. We propose Atlas-F, which combines Atlas and TimesFM 2.5 with weights re-chosen each day from the previous 30 days of verified forecasts. On the all-India test Atlas-F beats both members on trained and never-trained stations (-1.2%/-2.0% against Atlas, -7.2%/-6.8% against TimesFM 2.5; Diebold–Mariano p<10⁻¹³) and has lower error than TimesFM 2.5 in all 28 network–track cells, meeting a dominance rule fixed in advance; ablations show that the gain comes from network-specific weights: equal weights are significantly worse (+0.9%/+0.8%), a static per-network rule fitted on validation is equivalent, and Atlas-F is within 0.2% of an oracle. Finally, an Atlas trained only on data up to 30 September 2024 has an hourly Delhi-average error of RMSE 54.2 and MAE 31.3 µg/m³ in winter 2024–25 (95% CI for MAE 26.7–36.4), below the published Day-1 errors of the operational WRF-Chem (92.5, 58.8) and SILAM (108.0, 71.7) systems; we show why this by-number comparison, which persistence also passes, measures a different skill from what those systems are built for. Paper 4 of 5 in ParadoxAI Lab's series on compact physical world models (Atlas-F). Code, rules, logs and weights: https://github.com/paradoxailab/paradox-world-model. Contact: [email protected]
Language: English
Open Access status: green
OpenAlex ID: W7236838715
Cites:
Cited by:
Related to: 10