Research group · ECS
Climate Informatics Group
Applying machine learning and statistical data science to climate model evaluation, extreme-event attribution, and the downscaling of projections to the scales that matter for adaptation.
Overview
The Climate Informatics Group occupies the intersection of Earth system science and modern data science. Its central premise is that the explosion in observational data — from satellite networks, reanalysis products, and long-term station records — combined with the diversity of global and regional climate model output, demands new analytical frameworks that go beyond classical linear statistical methods. The group develops and applies machine learning, Bayesian inference, and causal discovery techniques to three core application areas: model–observation bias correction, statistical downscaling of climate projections, and the attribution of individual extreme weather events to long-term climatic trends.
The group maintains a curated multi-decadal archive of gridded observational and reanalysis datasets, hosted on the Meridian HPC Cluster with a bespoke data catalogue interface. This resource is used internally and is available to external academic users through the Scientific Services programme. The group's software stack — including the open Python libraries ClimateBias and AttributeX (both fictional, published under VIAS open-science policy) — is actively maintained and documented.
Close collaborations exist with the Atmospheric Dynamics Group on the physical interpretation of data-driven circulation indices, and with the Hydrology & Earth Surface Lab on precipitation downscaling for catchment hydrology. The group also contributes modelling expertise to Veyra's Software & AI Services arm for commissioned climate-data analytics projects.
Research themes
- Statistical and machine-learning methods for climate model bias correction and post-processing
- Convolutional and recurrent neural networks for spatial downscaling of precipitation and temperature
- Causal discovery methods for identifying teleconnection drivers in large reanalysis archives
- Probabilistic attribution of extreme events: heat waves, heavy rainfall, and drought
- Uncertainty quantification in constrained climate projections and emergent constraints
- Large-ensemble analysis: signal detection and internal variability characterisation
Current projects
Active research programmes, 2024–2027
Project · VX-CIG-01
DeepDown: Deep-Learning Precipitation Downscaling
Training a physics-constrained generative adversarial network on paired coarse-resolution GCM output and high-resolution observational gridded rainfall to produce 1 km daily precipitation fields from 25 km model output. Validated across climate zones using an independent 30-year holdout set and a suite of extremes-oriented verification metrics.
Funding: VIAS Research Excellence Grant · 510,000 cr
Project · VX-CIG-02
CausalClim: Causal Discovery in Multi-Decadal Reanalysis
Applying the PC algorithm, PCMCI+, and linear causal influence measures to a 70-year ERA-equivalent reanalysis archive to reconstruct directed causal graphs among large-scale circulation indices, sea-surface temperatures, and regional climate anomalies. Assesses the stability of inferred causal structures across reanalysis products.
Funding: External partner grant VX-CAUS-23 · 380,000 cr
Project · VX-CIG-03
AttribX: Operational Extreme-Event Attribution Framework
Developing a near-real-time attribution pipeline that ingests output from a 500-member large ensemble within 72 hours of an extreme event ending and produces a probabilistic statement on the change in event likelihood due to anthropogenic forcing. Targets operational use for public communication and insurance risk modelling.
Funding: Industry contract research VX-ATTR-24 · 460,000 cr
Selected publications
- Underhill S., Ngozi B., Carvalho M. "Bias-corrected CMIP-ensemble precipitation over complex terrain using a quantile-mapping neural network." VIAS Journal of Climate 37(5), 2024. DOI: 10.veyra/VX-4471
- Underhill S., Romero A. "Teleconnection patterns between Arctic sea-ice loss and European blocking: a causal inference analysis." Veyra Geophysical Research Letters 51(11), 2023. DOI: 10.veyra/VX-4244
- Carvalho M., Underhill S. "A physics-constrained GAN for statistical downscaling of daily precipitation to 1 km." VIAS Artificial Intelligence for Earth Sciences 3, 2023. DOI: 10.veyra/VX-4097
- Ngozi B., Underhill S. "Operational extreme-event attribution at 72-hour turnaround using a 500-member ensemble." VIAS Climate and Atmospheric Science 6(12), 2022. DOI: 10.veyra/VX-3916
- Underhill S., Whitlock C. "Statistical downscaling of precipitation for hydrological impact modelling in data-sparse catchments." VIAS Climatic Change 169, 2021. DOI: 10.veyra/VX-3658
- Underhill S. "Emergent constraints on equilibrium climate sensitivity from observed tropical mean precipitation trends." VIAS Earth System Dynamics 12(2), 2020. DOI: 10.veyra/VX-3435
People
Group lead: Dr. Selma Underhill · View all Veyra people
Postdoctoral researchers: Dr. Blessing Ngozi, Dr. Mateus Carvalho, Dr. Eira Mäkinen.
PhD students: Sebastião Lima, Yuliya Darchuk, Adeola Okonkwo, Till Gruber, Rosanna Ferretti, Sung-jin Park, Máire Ó Briain.
Research staff: Dr. Otar Mgeladze (data engineering), Stefanie Wirth (HPC analyst), Jaroslav Novák (software developer).