Foundation models (FMs) are redefining how scientific problems are approached by enabling scalable learning across diverse datasets and domains. Built using self-supervised techniques, these models provide a unified framework for downstream tasks, reducing reliance on labeled data while advancing prediction, data assimilation, and process-level understanding across Earth, space, and planetary systems. This session invites contributions highlighting recent progress in the development, adaptation, and application of FMs.

Topics of interest include, but are not limited to:

• Strategies for large-scale pretraining and construction of scientifically meaningful datasets
• Techniques for domain adaptation, transfer learning, and intermediate training
• Design of multimodal, multiscale, and spatiotemporal FM architectures
• Advances in fine-tuning and parameter-efficient learning
• Cross-domain modeling approaches linking Earth, space, and planetary systems
• Methods for enhancing interpretability, uncertainty estimation, and robustness of FM predictions

We also encourage contributions that discuss community-driven efforts toward open and accessible foundation models, including shared datasets, benchmarking frameworks, and collaborative infrastructure.