FUMD-AI produces open, reproducible urban mobility and network datasets for connected vehicles, and trains AI models that predict cellular handovers — which base station will serve a moving vehicle several seconds before it gets there.
The aim is proactive mobility management. If a network knows a vehicle's next serving cell a few seconds ahead, it can pre-position context, bearers or edge workloads instead of reacting once the handover has already begun. The same datasets support trajectory and handover analysis for V2X communication design, access network modelling, public transport planning and urban traffic optimisation.
The project is also a demonstration of the EOSC Federation in practice: dataset generation, processing, storage and publication run on EOSC Node Poland, while GPU-based AI training runs on infrastructure in Skopje — heterogeneous, distributed resources composed into one coherent scientific environment for a single use case.
Raw simulator output becomes a trained forecaster in three stages, each released as a standalone, independently runnable workflow:
Every stage is usable on its own, and the workflows are parameterised so the scenario can be moved to a new city, traffic density or network configuration.
| Repository | What it is |
|---|---|
| handover-forecaster | Serving-cell forecaster. Six seconds of joint mobility + radio history in, the serving cell at each of the next 1–7 seconds out. Bidirectional-LSTM encoder with per-step Bahdanau attention, ≈1.14 M parameters, trained on simulated urban traffic over a 9-cell 5G deployment. |
The workflows themselves live on GitHub: preprocessing · training & postprocessing
Everything here is meant to be re-run, not just read. Each workflow ships pinned
requirements, a one-call orchestrator notebook, Slurm/Singularity templates for
the GPU stages, CITATION.cff citation metadata, and an
RO-Crate manifest
describing every step and parameter. Datasets are published with persistent
identifiers, full metadata and schema documentation, and explicit licensing.
Our model cards report accuracy where it matters, not only in aggregate. Handover forecasting is a task where steady-state rows dominate the data and a "nothing will change" baseline already scores well, so our evaluations break out performance inside the pre-handover window separately. Those numbers are lower, and we publish them.
FUMD-AI (Grant ID 25-EOSC-GRV-INTER-013) is a cascading grant project under
EOSC Gravity, funded by the European Union's Horizon Europe programme under
Grant Agreement No. 101188045. It runs from May to November 2026.
Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union, the granting authority or the EOSC Gravity partners. Neither the European Union nor the granting authority can be held responsible for them.
Project page: https://eosc.eu/horizon-europe-projects/fumd-ai
We gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Centers: ACK Cyfronet AGH) for providing computer facilities and support within computational grant no. PLGINT/2026/019844.
The research work was supported by the Open Science Cloud research laboratory (OSC-LAB) at the Faculty of Computer Science and Engineering (FINKI), Ss. Cyril and Methodius University in Skopje, North Macedonia.
Source code MIT. Explanatory text and original figures CC BY 4.0. Input datasets retain the licences stated in their own metadata.