I survey, benchmark, and build ML systems for weather, maritime, and logistics — and my work consistently shows that data quality, evaluation design, and design-space understanding drive performance more than model complexity.
A design-space analysis of graph neural network architectures for computational fluid dynamics. The survey organises existing methods along the dimensions that determine their capabilities, identifies where the field has explored, and maps the structural gaps that remain.
A benchmark for ML weather models over complex maritime terrain, designed around conditional evaluation protocols that expose failure modes aggregate metrics hide. The goal is interpretable failure analysis rather than a single-number leaderboard.
A 174-configuration factorial experiment on AIS vessel classification. Three-way ANOVA shows data preprocessing explains 58.1% of performance variance (η²=0.581) while model architecture choice is statistically negligible (η²=0.004). The result challenges the standard "new model beats baseline" paper format.
GATv2 spatio-temporal graph encoders for heterogeneous supply-chain graphs. Transfer learning as integration middleware delivers 30–41% accuracy improvement with 4–10× data reduction. Unified architectures shown to be an anti-pattern (1–4% degradation vs. task-specific models).
I am drawn to the gap between what ML benchmarks claim and what actually drives performance. Across weather prediction, maritime data science, and logistics, I keep finding the same pattern: the field chases model architectures when the real performance drivers are data quality, evaluation design, and problem formulation. My research sits at the intersection of graph neural networks, physical-domain ML, and experimental methodology — building the taxonomies, benchmarks, and experiments that isolate what actually matters.
I am a PhD student at DIKU (University of Copenhagen) and the University of the Faroe Islands, and Assistant Professor in ICT at the University of the Faroe Islands. My doctoral project focuses on data-driven approaches for compressible and incompressible fluid dynamics modelling. I have supervised eight BSc theses and co-supervised two MSc students, and I teach courses in artificial intelligence and software engineering.
I grew up in the Faroe Islands, where weather and ocean are not abstractions but daily realities that shape everything — from fishing schedules to whether the helicopter flies. That proximity to the physical world informs how I think about ML: models that work on average but fail in the conditions that matter most are not good enough. Outside research, I am an international artistic gymnastics judge — a domain where precise evaluation under pressure is also the whole point.