Winter 2025 URSSI Fellowship Report: Reproducible Machine Learning Workflows for Scientists
Matthew Feickert • November 17, 2025
During my Winter 2025 URSSI Early-Career Fellowship project, I researched how modern technologies and tools can assist researchers in easily creating fully reproducible hardware accelerated software environments for scientific and machine learning workflows. I compiled the techniques and best practices I had learned into a short course, which I contributed to The Carpentries Incubator, and then taught this open source educational material to the broader scientific community at workshops.
The material focused on using Pixi — a modern multi-platform software environment manager that builds on the conda and Python package ecosystems — and CUDA conda packages distributed on conda-forge. Pixi is a tool with high-level semantics designed to let users declaratively specify project software requirements and then record the fully resolved (“locked”) dependencies in a “lock file”. Written in Rust, Pixi exploits the language’s speed and technologies to efficiently resolve complex dependency trees and update the project lock file for every Pixi operation that could affect the software environment. This means that that if a Pixi project is version controlled, any state of that project is fully reproducible, byte for byte, indefinitely into the future.
Applications now open for the 2025 URSSI Winter School in Research Software Engineering
Kyle Niemeyer • November 5, 2025
Do you develop software for your research? Do you have some basic skills but desire more?
If so, you might be interested in the upcoming URSSI Winter School in Research Software Engineering. Building off our prior winter and summer schools, we are hosting a three-day workshop on research software engineering skills over 15–17 December 2025 in Portland, OR, at Oregon State University’s Portland Center.
This is aimed at early-career researchers, particularly graduate students and postdocs, who are familiar with basic skills such as interacting with the Unix shell, version control using Git, and Python programming, and would like to learn more about best practices for developing research software.
URSSI Fall Fellowship: Characterizing the Impact of LLM-Generated Code in Scientific Software
Eva Maxfield Brown, David Farr, and Shahan Ali Memon • November 3, 2025
As a multi-purpose technology, large language models (LLMs) have rapidly been integrated across industries including academia and software development. From suggesting syntax to writing full functions, they’ve transformed the possibilities for how scientists build and maintain research software.
However, the speed and convenience of LLMs comes with trade-offs. Scientific software often lives at the frontier of discovery, where a subtle bug or misused library can distort an entire analysis1. As researchers increasingly use tools like ChatGPT or Claude to write code, we face a new question:
URSSI Welcomes Second Cohort of Early-Career Fellows
Nic Weber and Kyle Niemeyer • September 23, 2025
We are happy to announce a second cohort of the US Research Software Sustainability Institute (URSSI) Early-Career Fellowship. This cohort includes three projects and five fellows working on the follwoing:
Characterizing LLM-Generated Code in Scientific Software - Shahan Ali Memon, David Farr, and Eva Maxfield Brown will investigate the detection of AI assisted development in code repositories that are linked to scientific articles. Their project will evaluate existing methods used for detection (e.g. code stylometry), create a dataset of verified AI/non-AI assisted code, and assess how adoption changes code quality, documentation practices, and reproducibility.
Check out our upcoming community calls, events, and updates.