DC7 — Carbon-Aware Neural Architecture Search (NAS)

Host: Technical University of Denmark (DTU), Lyngby, Denmark — Embedded Systems Engineering section (ESE), DTU Compute.
Work package: WP2 (Carbon-Aware Intelligence and Control)
Duration and start: 36 months, full-time; start between February and September 2027 (latest start September 2027).

The project

Most AI models are built for one fixed operating point and cannot adapt when grid conditions change. In this project you change how carbon-aware AI is designed.

Instead of producing a single “best” model, you will use neural architecture search to generate a family of related models. An AI service can then trade accuracy for lower energy and carbon at runtime, picking the right model for the current grid signal.

What you will work on:

• Develop multi-objective NAS that optimizes accuracy, latency, energy, and carbon at the same time.
• Create hardware-aware search spaces for CPUs, GPUs, and accelerators.
• Produce model cards that record switching costs and performance.

What you will deliver:

• A carbon-aware NAS framework that generates Pareto-optimal model architectures.
• Model libraries whose variants include the metadata needed to switch between them.

Secondments

You will spend two research visits with partners in the network:

• TDC NET (Denmark) — 3 months from around May 2028 (industry), with Henrik Christiansen: tune AI model libraries for telecom edge datacenters.
• Dell Technologies (Ireland) — 3 months from around May 2029 (industry), with Aidan O’Mahony: test embedded AI solutions in production server environments.

The host

The Embedded Systems Engineering section at DTU Compute works on methods, tools, and architectures for dependable edge computing, edge AI, and low-energy computing.

Supervision

Main supervisor: Xenofon Fafoutis (ORCID) — Professor of Networked Embedded Systems, DTU Compute. He works on embedded AI and energy-efficient machine learning.

Co-supervisors:

• Luca Pezzarossa (ORCID) — Associate Professor, DTU Compute
• Henrik Thorsen — Energy Cool (EC Group)
• Aidan O’Mahony (ORCID) — Dell Technologies

Profile we are looking for

• Machine learning and deep learning
• Neural architecture search or AutoML is a plus
• Embedded or edge AI is a plus
• MSc (completed or near completion) in computer science, electrical engineering, or a related field
• Curiosity, initiative, and good spoken and written English

You must also meet the MSCA eligibility rules — no PhD yet, the mobility rule, and open to all nationalities.

How to apply

Official vacancy at the host institution: link coming soon.

Until then, apply through the central AEGIS application portal: AEGIS application portal (reference position DC7), and read what to include in your application.

Questions about this position? Contact the coordinator, Prof. Paul Pop, aegis@compute.dtu.dk.