Doctoral researcher in Adaptive Grid-Interactive Edge Datacenter Fleets (MSCA Doctoral Network)
About this role
We are recruiting one doctoral researcher (3 years) to the AEGIS network
AEGIS - Adaptive Grid-Interactive Edge Datacenter Fleets - is a European research and training network. It is funded by the EU through a Marie Skłodowska-Curie Doctoral Network, runs from 2026 to 2030, and trains 15 Doctoral Candidates (PhD students). The DN targets developing software-defined technologies that turn edge data centers from passive electrical loads into active, carbon-efficient and grid-interactive resources. The project has three layers:
A programmable infrastructure layer gives safe, runtime-verified control of a site's power electronics, cooling, and AI hardware.
An intelligence layer reads real-time grid signals and turns a request such as "reduce power by 20%" into the detailed actions that hardware and facilities must take, without breaking service.
A fleet orchestration layer coordinates many sites, even when they have different owners, using trusted and privacy-preserving methods, so the whole fleet can act as one Virtual Power Plant.
Read more about the project and other participants here: https://aegis-dn.eu/
Responsibilities
- You will be working on "End-to-End Energy Modeling for Distributed AI Services". To make carbon-aware decisions, both orchestrators and developers need to see the real energy cost of an AI workload. In this project you build analytical models and systems that make energy and carbon metrics visible. You will create compositional energy models that span the radio network, the transport network, and the datacenter hardware, and expose them through a fast query interface so any tool can ask, "what does this AI service cost in energy?"
- You will be responsible for:
- Developing compositional energy models for disaggregated radio (RAN) components, transport networks, and datacenter hardware.
- Building a low-latency query API that returns the end-to-end energy cost of an AI service.
- What you will deliver:
- Validated models with less than 10% prediction error