Paula Harder
Portrait of Paula Harder

Machine Learning Scientist · ECMWF

Paula Harder

Machine learning for forecasting atmospheric composition

At the European Centre for Medium-Range Weather Forecasts (ECMWF) I work on forecasting aerosols, reactive gases and air quality with machine learning, in particular on AIFS-COMPO, a data-driven global composition forecasting system developed together with colleagues across ECMWF and the Copernicus Atmosphere Monitoring Service (CAMS).

About AIFS-COMPO Publications Contact

Current work

AIFS-COMPO

AIFS-COMPO extends ECMWF's Artificial Intelligence Forecasting System (AIFS) to atmospheric composition. A single transformer model forecasts meteorological variables together with aerosols (dust, sea salt, biomass-burning aerosol and sulphate), particulate matter (PM2.5 and PM10) and reactive gases such as ozone, nitrogen dioxide, carbon monoxide and sulphur dioxide. It is implemented in Anemoi, the open-source framework for data-driven weather models developed by ECMWF and European meteorological services.

Global maps of aerosol optical depth at 550 nm from AIFS-COMPO and IFS-COMPO, three-day forecasts from 20 June 2024
Three-day forecasts of total aerosol optical depth at 550 nm from AIFS-COMPO (left) and IFS-COMPO (right), initialised on 20 June 2024. From Harder et al. (2026).
187forecast variables
10 daysforecast range
3-hourlytime step, ~80 km grid
~50 sfor a 5-day forecast on one GPU

Training

  1. Pre-training on the CAMS reanalysis (EAC4), 2003–2023.
  2. Fine-tuning on operational CAMS analyses, the data the model is initialised from.
  3. Rollout training over multi-step forecasts, for stability at longer lead times.

Evaluation

Compared with the operational IFS-COMPO system against independent observations (Harder et al., 2026):

  • Lower errors for aerosol optical depth at AERONET sites over the five-day range.
  • Comparable or better skill for particulate matter and surface ozone in most regions.
  • Weaker skill for some total-column quantities, such as total column ozone.
  • Ten-day forecasts capture the evolution of the 2024 Antarctic ozone hole.
Paper (arXiv) ECMWF announcement Forecast charts

AIFS-COMPO is experimental; IFS-COMPO remains the operational CAMS forecast. Ongoing work covers fire emissions, emission scenarios and the use of composition observations.

News

Other projects

Machine learning for Earth System Modeling

Related and earlier work at ECMWF, Mila, the Allen Institute for AI (Ai2), the Frontier Development Lab and during my PhD.

Time series of global total-column carbon monoxide for the reanalysis, the control run and the machine-learning reconstruction, 2004 to 2025

ECMWF · 2026

Reconstructing carbon monoxide reanalysis

The end of MOPITT carbon monoxide observations in early 2025 affects the quality of the CAMS reanalysis. We test whether a machine-learning model, trained on the differences between the reanalysis and a control simulation, can reconstruct monthly-mean total-column carbon monoxide without these observations.

Paper

Global maps of downscaling scores for models trained only on a northern and on a tropical training area

Mila · 2024–2025

Precipitation downscaling across regions

Generative models for precipitation downscaling are usually trained and tested on a single region. RainShift is a benchmark for evaluating how they transfer to other regions, and a follow-up study examines their global applicability.

RainShift Global transfer Code Dataset

Downscaled ERA5 water content field next to the high-resolution target

Mila · 2023

Fourier neural operators for downscaling

A downscaling model based on Fourier neural operators is trained at one upsampling factor and can then produce outputs at arbitrary higher resolutions, evaluated on fluid-dynamics simulations and ERA5 data. The work was led by Qidong Yang.

Paper Code

Global map of the R² score of reservoir-computing sea surface temperature forecasts over a rollout, shown on the six unfolded tiles of a cubed-sphere grid

Allen Institute for AI (Ai2) · 2023

Reservoir computing for sea surface temperature

Together with the climate modeling team at Ai2, we tested reservoir computing, a recurrent machine-learning method in which only the output layer is trained, for predicting sea surface temperature as a component of Earth system digital twins.

AGU abstract

Global water content maps at low resolution, from an unconstrained and a constrained network, and at high resolution

Mila · 2022–2023

Physically constrained downscaling

Neural networks for climate downscaling do not by default conserve quantities such as mass. Hard-constraint layers enforce conservation exactly, and in our experiments they also improved predictive skill.

JMLR paper Code

Global map of the change in sulphuric acid concentration computed by the M7 aerosol microphysics module

PhD, Fraunhofer ITWM & Oxford · 2019–2022

Emulating aerosol microphysics

Aerosol microphysics is a computationally expensive component of climate models. We trained physics-informed neural networks to emulate it at a fraction of the cost while keeping aerosol mass consistent.

Paper Code

Geostationary satellite images of a wildfire smoke plume in two wavelength channels

Frontier Development Lab Europe · 2022

Forecasting pyrocumulonimbus clouds

Pyrocumulonimbus (pyroCb) clouds are thunderstorms generated by intense wildfires. Pyrocast combines a database of pyroCb events, geostationary satellite images and meteorological data to forecast them, and a companion study uses causal inference to identify their drivers.

Pyrocast Causes

Infrared input, the visible image generated by a U-Net, and the true visible image of the same scene

2020

Night-time satellite imagery

Visible satellite images are unavailable at night. NightVision trains image-to-image models on daytime scenes to generate visible imagery from infrared observations, so it can be produced around the clock.

Paper Code

Contact

For questions about my work, collaborations or talks, please write to paula.harder@ecmwf.int. More background is on the About page.