Grant Stevens
I am an EPSRC Doctoral Prize Fellow in the School of Physics at the University of Bristol, working at the intersection of machine learning, astrophysics and scientific software. My research focuses on developing practical machine-learning methods for noisy, incomplete and complex scientific data, with interests including active learning, generative models and simulation-based inference. I completed my PhD in Interactive AI, where I investigated how active-learning methods can be made more effective in real-world scientific settings. My current work includes machine-learning research for large astronomical datasets and the Euclid mission, alongside the development of AstronomicAL, an interactive scientific workbench that grew out of my research into human-in-the-loop machine learning for astronomy.
Session
Astronomy has developed a rich ecosystem of specialised software for individual surveys, missions and scientific tasks. Combining these capabilities into a coherent researcher workflow, however, often requires bespoke integration, duplication of functionality, or tight coupling to another application's data structures and state.
AstronomicAL was originally developed for interactive catalogue exploration, visualisation and human-in-the-loop active learning. Its evolution towards a more general scientific workbench created a new challenge: integrating independently developed scientific capabilities without turning the application into an increasingly coupled collection of tools. This became particularly concrete when combining Euclid-specific data access developed through ELSA with other independently developed analysis capabilities.
We have therefore redesigned AstronomicAL as a plugin host built around shared scientific contracts. Extensions can retain their own implementation and dependencies while interacting through common abstractions for datasets and semantic mappings, object selection, derived scientific products and managed computation. Each plugin declares the scientific context it needs and the capabilities it contributes, while the platform mediates these interactions and keeps internal implementations independent.
The current plugin set demonstrates this approach across community software and multiple surveys and missions. Aladin Lite can follow the selected source through generic sky-coordinate mappings; catalogue data can be exchanged with TOPCAT through SAMP; common spectral tooling accesses DESI, SDSS and Euclid data; and Euclid cutouts can be used alongside generic visualisation, annotation and machine learning capabilities. Active learning, previously central to AstronomicAL, is now one composable scientific workflow over the same platform.
We discuss the interfaces and boundaries required for independent tools to work together: which scientific concepts need to be shared, which implementation details should remain independently owned, and how plugin-host architectures can enable astronomy tools to be composed into broader scientific workflows.