Advancing Physical Representation Learning in Geospatial Deep Learning
Ref. 2025_015 Are you passionate about pushing the boundaries of how we model physics of the Earth system with deep learning models? Recent breakthroughs in foundation models have shown remarkable capabilities, from simulating atmospheric dynamics, to generating spectral components of satellite data, and even accelerating power flow simulations in electric grids. Yet, a critical question remains: How physically grounded and robust are these models in real-world scientific and engineering applications? This thesis will focus on advancing the physical representation capabilities of geospatial foundation models. Despite their growing use, these models still face key challenges
How well can they extrapolate to extreme events?
When do they hallucinate or produce phisically impossible outputs?
What methodologies can be implemented to detect and mitigate such failures, especially in domains where reliability and scientific validity are essential?
You will explore these questions by developing and evaluating approaches to improve physical output, robustness, and trustworthiness in foundation models for geospatial and Earth system applications. Key Research Activities
Enhance physical system representation within foundation models
Curate and analyze diverse, high-quality pretraining datasets
Investigate robust, scalable model architectures
Analyze internal representations and physical consistency
Develop validation protocols for robustness and generalization
Design and implement guardrails to detect and mitigate hallucinations
Apply techniques to real-world geospatial models in collaboration with domain experts
Why This Opportunity? As part of IBM Research - Zurich, you’ll join a world-class team of scientists and engineers in a dynamic, interdisciplinary environment. You will gain hands-on experience with large-scale AI systems, collaborate with leading organizations like NASA and ESA, contribute to open-source tools and HuggingFace models, and publish your findings in top-tier venues. Required Skills
Excellent academic track record in Computer Science, Data Science, Statistics, Scientific Computing, Applied Mathematics, or related fields
Very strong background in deep learning
Proficiency in collaborative coding environments (e.g., GitHub, GitLab)
Strong problem-solving skills and ability to synthesize research into novel deep learning architectures
Enthusiasm for robust, responsible AI in scientific applications
Preferred Qualifications
Experience with large-scale training on GPU clusters (e.g., Slurm)
Understanding of physical processes in the Earth system or in energy systems
Experience with large geospatial datasets and scientific data formats
Passion for open science
Diversity IBM is committed to diversity at the workplace. With us you will find an open, multicultural environment. Excellent flexible working arrangements enable all genders to strike the desired balance between their professional development and their personal lives. How to apply Please submit your application through the link below.
