Zero-to-Hero Instructors
Our Zero-to-Hero Instructors lead the thematic introductory modules of the Summer School, providing participants with a solid foundation in key areas of machine learning. Together, they ensure that all attendees are well prepared for the in-depth sessions that follow throughout the week.
TUESDAY, SEPTEMBER 2, 2025
11:00 – 12:30 | Session 1A – Hybrid
Title:
Learning global parameterizations of ecosystem processes using hybrid modelling
Abstract:
In this tutorial we will go through the basics of process-based model parameterizations and end up learning models of ecosystem functional properties using hybrid modelling approaches. We will focus on terrestrial ecosystems models that simulate processes and responses of carbon and water fluxes to meteorology and climate. We will go through concepts and discuss assumptions, limitations and challenges while progressing to statistical learning of models that replace common parameterization approaches. The aim is to provide an easy entry point into current challenges common to several modelling approaches and assumption in Earth system science using state of the art model development tools.
Bio:
Nuno leads the Model-Data Integration Group at the Max Planck Institute for Biogeochemistry since 2012 and is an Invited Researcher at the New University of Lisbon. His research interests focus on the dynamics of terrestrial biogeochemical cycles, particularly the improvement of ecosystem modeling through data assimilation and inverse optimization, and on the introduction of deep learning and hybrid modeling in terrestrial ecosystem processes. Nuno is deeply committed to advancing the integration of machine learning with Earth system science to better predict ecosystem responses to climate change. He collaborates internationally to develop innovative methods that improve the accuracy and interpretability of biogeochemical models.
TUESDAY, SEPTEMBER 2, 2025
11:00 – 12:30 | Session 1B – XAI
Title:
Explainable AI for Earth Science: Practical Concepts, Workflows, and Pitfalls
Abstract:
Machine learning is increasingly used in Earth science for forecasting, monitoring, and studying complex environmental systems. But a model that predicts well without being understood can be misleading as well as useful. This session introduces the key concepts and practical methods of explainable AI (XAI), with a focus on Earth-science applications. We present XAI methods from the perspective of research needs, discussing when they are informative, when they are not, and how to evaluate their reliability. A hands-on exercise with time-series data will demonstrate how gradient-based approaches, such as Integrated Gradients, can be applied in practice. Participants will learn workflows for selecting, interpreting, and validating XAI methods, along with common pitfalls to avoid. The aim is to move beyond treating XAI as a plug-in tool toward producing thoughtful explanations that are scientifically useful, trustworthy, and relevant.
Bio:
Shijie leads the Machine Learning for Hydrological and Earth Systems (ML4HES) group at the Max Planck Institute for Biogeochemistry and the ELLIS Unit Jena. His research focuses on understanding coupled water, energy, and carbon dynamics across hydrological, ecological, and climate systems. By integrating remote sensing, in-situ observations, and domain knowledge with hybrid and explainable machine learning methods, his work aims to identify key drivers, interpret system feedbacks, and improve process representation under environmental change. Shijie’s research covers land–atmosphere interactions, soil–plant–atmosphere processes, and watershed hydro-biogeochemistry, with a focus on applying knowledge-integrated AI to link observations with underlying processes. Since 2025, he has served as an Associate Editor for Water Resources Research and has convened EGU and AGU sessions on machine learning in Earth sciences since 2023.
THURSDAY, SEPTEMBER 4, 2025
11:00 – 12:30 | Session 2A – Uncertainty
Title:
Beyond Accuracy: Understanding and Quantifying Uncertainty in Deep Learning Models
Abstract:
Uncertainty quantification (UQ) is critical for deploying deep learning models in real-world, high-stakes applications from medical diagnosis to autonomous driving. Yet, for many researchers and practitioners, UQ remains an abstract or advanced topic, often introduced too late or too formally. This zero to hero session aims to bridge that gap. We will introduce the concepts of epistemic and aleatoric uncertainty, explain why standard deep learning models are often overconfident, and present practical UQ methods including Bayesian Neural Netoworks, ensemble approaches, and test-time techniques. The session emphasizes intuitive understanding, real-world use cases, and the trade-offs involved in different approaches. No prior experience with uncertainty modeling is required. The goal is to offer an accessible entry point for anyone interested in making their models more trustworthy and informative.
Bio:
Aishwarya’s research interests encompass the development of robust and reliable deep learning models, with a particular focus on uncertainty quantification and multimodal learning. During her Ph.D., she developed methods for applying deep learning to ecological data, including techniques for quantifying uncertainty. She is currently a postdoctoral researcher under Prof. Joachim Denzler at the Computer Vision Group at Friedrich Schiller University Jena. Her current work focuses on developing uncertainty quantification methods and exploring multimodal learning to improve the reliability and interpretability of deep learning models in various applications.
THURSDAY, SEPTEMBER 4, 2025
11:00 – 12:30 | Session 2B – Large-Scale
Title:
Scalable AI for Earth and Climate Sciences
Abstract:
This tutorial will take participants on a “zero-to-hero” journey through the modern AI toolkit for large-scale Earth and climate data analysis. We will start with the fundamentals of cloud-native Earth Observation (EO) data management, covering spatiotemporal datacubes, object storage, and cloud-optimized formats. Building on this, we will explore scalable computing with Xarray, enabling efficient analysis of
large EO datasets. The second part of the tutorial introduces self-supervised learning (SSL) techniques for EO, discussing predictive, contrastive, and generative approaches for multi-modal satellite data. We will then examine sensor-agnostic EO foundation models, their design principles, evaluation strategies, and current challenges, including
multimodal integration and temporal modeling. Throughout, we will link conceptual discussions with practical, hands-on exercises, guiding participants from data ingestion to large-scale model training and evaluation. Participants will gain an informed perspective on current tools and methods, along with practical pointers for applying them in their own Earth and climate science projects.
Bio:
Ioannis is Head of the OrionLab and Assistant Professor of Artificial Intelligence for Earth Observation at the National Technical University of Athens (NTUA). He also serves as an Adjunct Researcher at the National Observatory of Athens (NOA) and Archimedes Research Center. His expertise lies in processing and analyzing large-scale satellite and geospatial data to advance AI for Earth Observation (AI4EO). Leading Orion Lab, which bridges NTUA and NOA, Ioannis focuses on impactful AI applications addressing climate change, disaster management, monitoring, forecasting, and Earth system sciences. His research includes foundational models in remote sensing, self-supervised learning for multi-modal EO data, vision-language models for remote sensing interpretation, and earth system deep learning for spatiotemporal forecasting. Ioannis coordinates several European projects, including Horizon Europe’s ThinkingEarth and MeDiTwin, H2020’s DeepCube, and ESA’s SeasFire, developing AI pipelines to harness Copernicus datasets. His work is pivotal in integrating AI with Earth observation to enhance environmental resilience and sustainable management.