Reproducible asteroid orbit determination with the General Open Orbital Dynamics platform and machine-learning-based astrometry

Orbital-dynamics analyses are central to understanding the orbital evolution, transport mechanisms, impact risk, and long-term dynamical behaviour of asteroids and comets. They are also essential for planetary defense, where reliable orbit determination, uncertainty propagation, and reproducible impact-hazard assessment are critical. However, many orbit-propagation and orbit-estimation pipelines remain closed-source, application-specific, or difficult to reproduce outside the originating team. This limits independent validation, reuse of previous analyses, systematic combination of heterogeneous observations, and broader community participation.

We present the General Open Orbital Dynamics (GOOD) platform, an open-science infrastructure for reproducible orbit-dynamics analysis of Solar System bodies. GOOD combines the TU Delft Astrodynamics Toolbox (Tudat), which provides open-source high-fidelity numerical propagation and estimation capabilities, with the University of Groningen’s WISE-based data-handling technology, which supports scalable data management, detailed lineage tracking, and reproducible workflows. The project is funded by OpenScience NL and supported by five three-year researcher/scientific-software-development positions.

GOOD will allow researchers, data providers, and citizen-science contributors to run, validate, archive, retrieve, and update orbit-dynamics analyses together with their input data, configuration files, software versions, processing history, and output products. A central goal is that the complete setup for a published orbit-determination analysis can be retrieved through a single web action or terminal command. This will allow users to continue from previous analyses instead of reconstructing processing chains from publications, local scripts, or undocumented assumptions. The platform will support established data sources such as MPC, PDS, PSA, Gaia, Euclid, and LSST and AstroWISE, while also allowing users to upload and share their own datasets.

A key use case is asteroid and comet orbit determination from optical astrometry extracted from wide-field imaging surveys. To feed GOOD with scalable astrometric measurements, we are implementing machine-learning-based streak detection within AstroWISE. Convolutional neural networks have the potential to outperform traditional parameter-tuned detection methods for asteroid streaks. Building on earlier work, we train and test a CNN on simulated and real asteroid streaks in OmegaCAM images, together with real non-asteroid sources retrieved through AstroWISE. In training, simulated streaks are injected into raw images to preserve realistic backgrounds to improve the performance.

At the session, we will present the scientific motivation, architecture, initial asteroid-focused use cases, and development roadmap of GOOD and the ML-based astrometric extraction. We seek feedback on priority datasets, dynamical-model requirements, configuration standards, and community-contributed use cases, with the long-term goal of making small-body orbit propagation, orbit estimation, and planetary-defense data products open, reproducible, and reusable by design with the help of AI.