Seminar: Seminar Intelligente Energiesysteme (Bachelor) - Details

Seminar: Seminar Intelligente Energiesysteme (Bachelor) - Details

Allgemeine Informationen

Veranstaltungsname Seminar: Seminar Intelligente Energiesysteme (Bachelor)
Veranstaltungsnummer INF-2047
Semester WS 2026/27
Aktuelle Anzahl der Teilnehmenden 1
Heimateinrichtung Intelligente und multimodale Energiesysteme
Veranstaltungstyp Seminar in der Kategorie Lehre
Erster Termin Dienstag, 13.10.26, 14:00 - 15:30 Uhr N 1057
Veranstaltung findet in Präsenz statt / hat Präsenz-Bestandteile Ja
Hauptunterrichtssprache deutsch
ECTS-Punkte 4

Räume und Zeiten

N 1057

  • Dienstag, 13.10.26, 14:00 - 15:30 Uhr

Modulzuordnungen

Kommentar/Beschreibung

Climate projections guide long-term decisions on the energy system, from the expansion of solar and wind power to the adaptation of infrastructure. However, the models behind these projections are expensive to run. Their largest uncertainties come from processes smaller than the model grid, such as clouds and convection. Machine learning (ML) parameterizations can represent these processes at a fraction of the computational cost. For this reason, they have received much interest in recent years.

In this seminar, we explore ClimSim [1], a large benchmark for ML parameterizations in climate models, which was also the basis of a Kaggle competition [2]. You will receive a working data pipeline, a training framework and a test suite. For your own project, you will study one of the baseline models from the paper [1], which include methods such as multilayer perceptrons, convolutional networks, encoder-decoders and probabilistic models. You will also have the chance to explore extensions such as generative models, physical constraints, masked modeling and new data representations. We will discuss good engineering practices, for example scaling, efficiency and model design.

The course consists of three meetings with self-study and experimentation periods in between:

First meeting (13.10.2026): We will discuss the ClimSim paper [1], the codebase you will use, and possible projects for you to improve on the existing baselines.
Second meeting (tba, December 2026): You will present your experience with implementing baselines (oral, 5–10 minutes), and we will discuss possible extensions, improvements and failure cases.
Third meeting (tba, February 2027): You will present your final results (oral, 10–15 minutes), with a focus on the reasons for your design choices, what worked well for you and what did not work.

Throughout the self-study period, you will also have the opportunity to discuss your work, ask questions and get guidance from the instructor, or other students.

[1] Yu, S. et al. (2023). ClimSim: A large multi-scale dataset for hybrid physics-ML climate emulation. NeurIPS Datasets and Benchmarks Track. arXiv:2306.08754.
[2] LEAP – Atmospheric Physics using AI (ClimSim), Kaggle competition, 2024.