Erlangen 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
AKjDPG: Arbeitskreis junge DPG
AKjDPG 2: jDPG/AKPIK Programmierworkshop (joint session AKjDPG/AKPIK)
AKjDPG 2.1: Tutorial
Sunday, March 15, 2026, 17:30–19:00, AM 00.014
Resource-Aware Deep Learning: Tracking Energy Consumption in Scientific AI Applications — •Kevin Schmitz1, Anno Knierim1, and Raphael Fischer2 — 1TU Dortmund University, Dortumd, Germany — 2Lamarr Institute for ML & AI, Dortmund, Germany
Deep learning has become an indispensable tool across physics and astronomy, yet its growing computational demands increasingly raise questions about energy efficiency and sustainability. This tutorial introduces physicists to the principles of resource-aware deep learning, focusing on practically quantifying, understanding, and optimizing the energy consumption of deep learning models. We begin by outlining different approaches for tracking model power usage, ranging from static estimation methods to dynamic profiling tools validated against ground-truth measurements, and demonstrate how these concepts are implemented in practice using the Lamarr Energy Tracker, developed at the Lamarr Institute for Machine Learning and Artificial Intelligence, for straightforward monitoring of GPU and CPU utilization during training or inference. Finally, we show how resource metrics can be visualized together with reconstruction accuracy using an example from radio interferometric imaging, where super-resolution neural networks reconstruct astrophysical sources from sparse visibility data. The session provides conceptual foundations and practical guidance for integrating sustainability into scientific machine learning workflows, empowering researchers to balance predictive performance with environmental responsibility.
Keywords: data analysis; deep learning; simulations; machine learning; ressource-aware computing