Dr. Selma Underhill

Biography

Selma Underhill joined Veyra Institute in 2015 after completing postdoctoral research at the Halvard Centre for Climate Statistics, where she developed causal graphical models for attributing observed precipitation extremes to large-scale circulation anomalies. She received her PhD in applied statistics from the University of Tarentum in 2012, with a dissertation on constrained spline-based downscaling of general circulation model output to station scale. Her undergraduate studies in mathematics and atmospheric physics were at Crestwell College, completed with distinction in 2007.

At Veyra, Underhill established the Climate Informatics Group to bridge statistical learning and climate science — developing methods that are both computationally tractable and physically interpretable. Her group's work spans three interlinked themes: statistical and machine-learning-based downscaling for regional climate projections; causal inference frameworks that identify the drivers of observed climatic shifts; and probabilistic hazard assessment tools that translate model output into decision-relevant risk estimates. A major product of the group is AttriBench, an open benchmark suite for comparing extreme-event attribution methods across model configurations, now adopted by two international assessment consortia. She has contributed expert analysis to three national climate-risk reports and holds an honorary research appointment at the Tarentum Institute for Environmental Statistics.

Underhill serves on the scientific council of the Veyra Graduate School and co-leads the Institute's cross-divisional working group on probabilistic risk and uncertainty. She has been principal investigator on five externally funded programs totalling approximately 4.7 million cr in direct costs, and currently supervises four doctoral students and three postdoctoral researchers. She lectures the graduate module Statistical Methods in Climate Science each spring semester and contributes to the Institute's professional short course on climate risk assessment.

Research interests

Selected publications

1. Underhill S, Bjørnsen L, Cadena-Ruiz P. "AttriBench: a benchmark suite for systematic comparison of extreme-event attribution methods under model uncertainty." Climate Methods and Analysis, 7(1): 15–38, 2025. VEYRA-DOI: 10.veyra/VX-2504
2. Underhill S, Gakuba F. "Causal graphical models for identifying drivers of compound hydro-meteorological extremes." Journal of Climate Dynamics and Statistics, 29(4): 441–460, 2024. VEYRA-DOI: 10.veyra/VX-2418
3. Cadena-Ruiz P, Underhill S, Magnusdóttir J. "Neural-network-based statistical downscaling with physically constrained output distributions." Proceedings of the Veltheim Conference on Climate Modelling, pp. 119–133, 2023. VEYRA-DOI: 10.veyra/VX-2306
4. Underhill S, Bjørnsen L. "Probabilistic attribution of the 2022 Coravel heat event using a constrained storyline approach." Environmental Research Letters in Climate, 5(2): 88–104, 2023. VEYRA-DOI: 10.veyra/VX-2321
5. Underhill S, Gakuba F, Solberg K. "Emulating regional climate model ensembles with Gaussian process surrogates: uncertainty quantification at continental scale." Annals of Applied Climatology, 14: 202–221, 2022. VEYRA-DOI: 10.veyra/VX-2215
6. Magnusdóttir J, Underhill S. "Constrained spline downscaling of GCM precipitation fields: performance across European sub-regions." International Journal of Climatological Statistics, 31(3): 317–335, 2021. VEYRA-DOI: 10.veyra/VX-2108

Current group members

Postdoctoral researchers

• Dr. Lars Bjørnsen — extreme-event attribution, storyline methods
• Dr. Fatima Gakuba — causal inference in climate systems, graphical modelling
• Dr. Jóhanna Magnusdóttir — climate model emulation, Gaussian process surrogates

Doctoral students

• Paloma Cadena-Ruiz — physics-constrained neural downscaling for precipitation extremes (Year 3)
• Kristoffer Solberg — probabilistic hazard assessment for compound climate events (Year 2)
• Rosamund Halevi — causal attribution of drought intensification under anthropogenic forcing (Year 2)
• Obinna Ede — machine-learning emulation of regional climate ensembles (Year 1)