AI in Action Foundation models accelerate space and climate science
30 snips
Feb 24, 2026 Campbell Watson, an IBM researcher who builds open-source foundation models for Earth and space science, discusses applying large-model ideas to multimodal satellite data. He covers collaborations with NASA and ESA, building tiny models that run in orbit, multimodal fusion of radar/lidar/optical, and applications like flood detection, biodiversity monitoring, and solar flare forecasting.
AI Snips
Chapters
Transcript
Episode notes
Foundation Models Complement Numerical Modeling
- Foundation models are starting to replace traditional numerical models in earth science while coexisting with them.
- Campbell Watson observed this shift over a decade and sees hybrid use with CPUs, GPUs, and quantum qubits as the future for complex simulations.
From Accountant To Cloud Physicist To AI Researcher
- Campbell transitioned from accounting to cloud physics and then to AI after pursuing ocean and cloud science.
- He learned numerical methods, moved to a postdoc at Yale, then joined IBM Research to model ecosystems and later embraced ML in hydrology around 2016.
Pretraining Speeds Satellite Task Adaptation
- Pretraining a foundation model on harmonized multispectral satellite imagery gives models prior knowledge of visual patterns.
- Using NASA's Harmonized Landsat Sentinel-2 dataset lets fine-tuning detect algal blooms, floods, or deforestation faster than training from scratch.