Research Vision

The programme develops algorithms and software systems for scientific computing and applied AI. Published work covers Gaussian-process search, generated language interfaces, interoperable learned models, streaming bioinformatics, and distributed simulation. Leidarljos extends that work to persistent memory, provenance, and coordination across AI tools. Chemistry, biology, and earth sciences supply collaborators, hard computational problems, and an independent check on the result. The papers and code sit on the research threads. This page says how those threads fit together.

Numerical algorithms and scientific software

Published saddle-search methods learn a potential-energy surface from expensive energy and force calculations. The model proposes the next calculation, and an independent physical computation checks it. The threads are GP acceleration and transition-state methods.

Three pieces of work in preparation carry the line forward. gpridge puts band and dimer searches on one shared energy and gradient model, and it checks the curvature at the stationary point separately from convergence of the search. anneal, in preparation for ACM Transactions on Mathematical Software, allocates a finite evaluation budget across exploration, escape, and local refinement. The ESCAPE proposal stores an event with its local environment, method, uncertainty, and the calculation behind it, and asks when that record transfers to a new environment. f2py and metatensor and metatomic are how these methods cross languages and simulation engines. A generated wrapper is tested for types, array layout, ownership, and numerical agreement with a direct call.

Leidarljos

Leidarljos gives separate responsibilities to memory, recorded artefacts, task dependencies, and work ownership. The packset store keeps claims with their sources and the dates each claim applies, including claims that a later one superseded. deedar stores artefacts with cryptographic provenance. Consensus methods combine judgements and revise trust from observed outcomes. The implementation and the reproducibility package are public. The memory manuscript is a draft. Studies of correction, reuse, and recovery are the research programme on the Leidarljos page.

HPC and data-centre software

Since 2021 I have consulted for IHPC on reproducible environments for Elja. The Nix workflow is in Reproducible HPC. The questions are how an environment stays reproducible without dropping the native build, how to report time and energy per correct result, and how deployment and recovery behave across sites. Nix, EasyBuild, and EESSI are three ways to close over an environment. Flowy, eOn, and rsx each have a scientific result that can be checked apart from the systems measurement. metatomic is where one learned model runs on CPUs and GPUs inside more than one simulation engine.

Publication

Computing contributions are written for readers in computer science and high-performance computing. The target journals are IEEE Transactions on Knowledge and Data Engineering (Q1 in artificial intelligence and in information systems, 2024 Journal Citation Reports), IEEE Transactions on Software Engineering (Q1 in software engineering, 2024 Journal Citation Reports), and IEEE Transactions on Parallel and Distributed Systems (Q1 in theory and methods, 2024 Journal Citation Reports). anneal is aimed at ACM Transactions on Mathematical Software. Methods and applications with collaborators in chemistry, biology, and earth sciences continue in Journal of Chemical Theory and Computation, Journal of Chemical Information and Modeling, Computer Physics Communications, and BMC Bioinformatics. BMC Bioinformatics is Q1 in mathematical and computational biology in the 2024 Journal Citation Reports. These are venues for work in preparation.

Record

Sixteen journal articles, seven conference papers, a correspondence in Nature, one preprint, and the thesis. Eleven of the journal articles appeared in 2025 and 2026, nine of them with me as first or sole author. I maintain f2py in NumPy, co-maintain Airspeed Velocity, and lead development of eOn.

In autumn 2024 I taught HBV505M Software Quality Management at the University of Iceland, with Helmut Neukirchen as course supervisor. In autumn 2023 I was assistant teacher to Steinn Guðmundsson in REI505M Machine Learning. I have been a certified Carpentries instructor since 2019, and I was a section leader in Stanford’s Code in Place in 2020 and 2021. I have been an editor of the Journal of Open Source Software since 2024.

Students

A first master’s project can compare memory reduction and acquisition rules on a fixed collection of small molecular problems. A second can test one validity or recovery mechanism in Leidarljos on a small public repository and recorded executions. A third can take a working application from a local build to a tested deployment, with reusable build recipes and performance measurements. The doctoral project I would start from is memory and coordination in Leidarljos. The master’s course Engineering Reliable AI Agents is set out under Teaching.