Forecasting · AI and machine learning

All summaries

• 
  Forecasting

  AI downscaling to 1.5 km adds detail but not always accuracy over Korea

  A diffusion-based downscaler for Korea added fine-scale variability and helped with the hottest events, but its added detail did not make forecasts more accurate overall.

  Atmosphere 22 October 2026

• 
  Forecasting

  AI ensemble models track tropical cyclones better than IFS-ENS but still miss intensity

  In this retrospective test, leading probabilistic AI weather models placed tropical cyclones more accurately than ECMWF's ensemble, yet they still underestimate winds and have poorly calibrated spread.

  arXiv 15 October 2026 Preprint

• 
  Forecasting

  AI weather models beat WRF on light rain but lose to it at 20 mm per 6 h

  In three flood events, GraphCast and AIFS forecast light rain and temperature better than an observation-corrected WRF, but WRF leads at the heaviest rainfall rate.

  Geoscientific Model Development 13 October 2026

• A method built on delay-embedding theory, called DSDL, kept more forecast skill than five other data-driven methods when variables were hidden, and it gave useful one-month-lead summer rainfall hindcasts for China.
  Forecasting

  Dynamics-based deep learning keeps forecast skill when observations are missing

  Climate Dynamics 12 October 2026

• In a climate model, a neural network predicts East Asian winter temperature from November conditions, and warm extremes become easier to predict under strong warming while cold extremes stay about the same.
  Forecasting

  Warm East Asian winters become more predictable in a warmer climate; cold ones do not

  npj Climate and Atmospheric Science 5 October 2026

• Feeding a deep learning nowcast model three-dimensional atmospheric forecasts, and boosting strong echoes, improved 12 h radar reflectivity and rainfall forecasts over East China compared with the baselines.
  Forecasting

  Deep learning nowcast using 3D atmospheric forecasts beats baselines for radar echoes and rain

  npj Climate and Atmospheric Science 2 October 2026

• AERO-ODE refines coarse global model output with terrain and surface information to make fast 3 km regional forecasts that, over two US regions, beat WRF-ARW and a deep-learning regional model on most error measures.
  Forecasting

  AI-guided regional model gives 3 km, 72 h forecasts in about 21 s with lower error

  npj Climate and Atmospheric Science 1 October 2026

• AI atmosphere models driven by forecast ocean temperatures can match or beat a leading physics-based model for seasonal storm counts in the Atlantic and eastern Pacific, but not everywhere, and some of that skill may come from compensating errors.
  Forecasting

  AI models fed predicted ocean temperatures rival a physics model on hurricane seasons

  EarthArXiv 22 September 2026 Preprint

• Training an AI weather model through a fast, differentiable radiation emulator improved its forecasts of surface sunlight and cloud cover compared with the same model without that constraint and with ECMWF HRES, though not for surface albedo.
  Forecasting

  Radiation-constrained AI weather model beats ECMWF HRES on sunlight and cloud forecasts

  npj Climate and Atmospheric Science 21 September 2026

• 
  Forecasting

  Deep learning forecasts local decadal climate from the recent observed temperature trajectory

  A neural network trained on 290 climate-model runs can forecast local future decadal temperatures from the 1980–2025 record, keeping internal variability in its uncertainty, and the authors report skill beyond ensemble averaging.

  EGUsphere 21 September 2026 Preprint

• FuXi-TC uses a diffusion model trained on WRF simulations to sharpen FuXi forecasts of tropical cyclone wind and rain, nearly matching ECMWF HRES on intensity and beating it on heavy rain, in about 2 minutes per forecast.
  Forecasting

  Diffusion model sharpens AI typhoon forecasts of wind and rain at low cost

  npj Climate and Atmospheric Science 8 September 2026

• 
  Forecasting

  AI weather model trained on satellite rain data beats its ERA5-trained twin and IFS

  Retraining ECMWF's AI weather model on satellite precipitation (IMERG) instead of reanalysis cut drizzle overprediction, improved heavy-rain forecasts and gave better-calibrated ensembles, though the physics-based IFS stayed better at the most extreme totals.

  arXiv 2 September 2026 Preprint

• 
  Forecasting

  A 13 km machine learning weather model matches ICON on near-surface forecasts

  AICON, a graph neural network trained on high-resolution ICON data without multi-step rollout, is now operational at DWD and competes with ICON near the surface, but it smooths fine scales and underestimates cyclone intensity.

  arXiv 25 August 2026 Preprint

• 
  Forecasting

  Long-window 4D-Var with NeuralGCM beats 20CRv3 using only surface pressure

  A differentiable weather model lets a simple long-window 4D-Var, with no background-error term, fit surface-pressure observations and give analyses closer to ERA5 than 20CRv3 over three months.

  arXiv 12 August 2026 Preprint

• 
  Forecasting

  Physics-informed network and corrected weather forecasts improve solar power forecasts up to 5 days

  Correcting weather-model sunlight forecasts and adding a physics rule to the training loss cut solar power forecast errors at one station in northwest China.

  Geoscientific Model Development 12 June 2026

• 
  Forecasting

  Smoothing and decomposing rainfall drought index data improves SPI forecasts in a South African district

  Cleaning the drought index with a smoothing filter and splitting it into components before forecasting gave the lowest errors in uMkhanyakude, where five of six stations show drying trends.

  Natural Hazards and Earth System Sciences 20 January 2026