EERIE

The role

This is the day job. At ETH Zürich and C2SM I lead the ocean–atmosphere evaluation for EERIE, the EU Horizon programme building the next generation of kilometre-scale coupled climate models, and as a Research Fellow of the ETH AI Centre I benchmark the AI weather emulators moving into the same territory. Three strands, one question: whether these models can be trusted where the physics is hardest.

Evaluating a new generation of climate models

EERIE runs coupled climate models at kilometre scale, fine enough to resolve the ocean mesoscale eddies and the air–sea feedbacks that coarser models can only parameterise. My role is to judge whether the emergent air–sea physics in those runs is right, across the models produced by MPI-M · ECMWF · Met Office · BSC. The result is the first multi-model quantification of mesoscale air–sea feedbacks in this class of model, written up as 2 first-author papers submitted.

Evaluation at this resolution is not a matter of eyeballing maps. To confirm that a model reproduces a physical mechanism rather than a lookalike statistic, I built a spatially resolved causal machine-learning framework that tests for the mechanism directly, presented at Climate Informatics 2026.

Benchmarking the AI emulators

The AI weather emulators are arriving faster than the tools to judge them. As a Research Fellow of the ETH AI Centre I engineered a multi-GPU inference and benchmarking pipeline spanning 6 operational emulators (AIFS, GraphCast, GenCast, NeuralGCM, Pangu-Weather and FourCastNet 3), and used it to quantify a systematic weakness: regression-trained emulators suppress the extremes, smoothing away exactly the tails that matter most for risk. The extreme-value and decision-value analysis behind that finding is its own case study, tailspec.

> ENGINE_ROOM — Petabyte-scale evaluation infrastructure (EXCLAIM)

None of the evaluation happens without the engineering to move the data. Under EXCLAIM I built GPU-accelerated, Dask-distributed analysis packages that process ~200 TB per simulated year of kilometre-scale model output, so a diagnostic that would otherwise be intractable runs over the full archive rather than a subsample.

The same work became teaching. I designed and taught the C2SM workshop on distributed task-parallel computing with Dask and xarray, so the group can run the pipelines itself rather than depend on one person to feed them.

In collaboration with Oxford

With a group at the University of Oxford, I am an invited co-author on a paper in preparation that uses machine-learning weather emulators as diagnostic instruments for the ocean’s role in fast atmospheric evolution. My contribution is on the air–sea coupling side: the coupling theory and the multi-model feedback benchmarks that connect the emulator diagnostics back to the coupled-model physics.

What this demonstrates