2026-10-07, 15:00–15:15 (Europe/Amsterdam), Rooms 12+14
Foundation models for Earth systems have advanced rapidly for weather and climate prediction, but remain largely confined to physical variables, omitting the human systems that drive emissions, shape land use, build infrastructure, and mediate vulnerability. We argue that this gap is fundamentally a data problem: the information exists but is fragmented across incompatible grids, projections, temporal frequencies, and formats. We present two complementary contributions that address this challenge.
First, WorldTensor is a harmonised global dataset that aligns over 750 environmental and socioeconomic variable families onto a common 0.25° latitude–longitude grid and annual temporal framework. It integrates climate, emissions, land use, satellite vegetation indices, gridded population and GDP products, power plant registries, and natural hazard and conflict catalogues into a single ML-ready NetCDF corpus. Constructing WorldTensor required solving nontrivial harmonisation problems including regridding across heterogeneous native resolutions, rasterising point and vector datasets into spatially meaningful fields, and reconciling temporal coverages spanning daily observations to sparse multiyear socioeconomic snapshots. The dataset and processing code will be released under open licenses.
Second, TerraNova is a foundation model designed to learn from WorldTensor's multimodal structure. It combines coordinate-based spatial encoding, learned country-level embeddings, Fourier temporal encoding, and a hypernetwork decoder to jointly predict climate, land surface, socioeconomic, and infrastructure variables in a unified multi-task framework. Early results demonstrate successful learning across multiple heterogeneous Earth system tasks simultaneously, validating that foundation models can learn shared representations across the coupled human–Earth system.
Together, WorldTensor and TerraNova provide an open, end-to-end pipeline from harmonised planetary data to multimodal foundation model training, supporting applications in climate impact assessment, cross-domain pattern discovery, and evidence-based environmental policy.
Other
Carlos Rodriguez-Pardo is a postdoctoral researcher at Politecnico di Milano and the RFF-CMCC European Institute on Economics and the Environment, where he works on deep learning for climate change mitigation as part of the ERC-funded EUNICE project. His current research focuses on foundation models for coupled human–Earth system modeling, multimodal geospatial data harmonisation, and neural methods for climate-economic decision making under uncertainty. He has published in Nature, Nature Scientific Data, Nature Climate Change, TMLR, CVPR, Eurographics, and ACM Transactions on Graphics, among others. He holds a PhD in Computer Science from Universidad Rey Juan Carlos and an MSc in Artificial Intelligence from the University of Edinburgh. He has received the SCIE–Fundación BBVA Young Researcher Award, the CEIG Best PhD Thesis Award, and multiple outstanding reviewer recognitions at CVPR, NeurIPS, ECCV, and AISTATS. He co-convenes the EGU 2026 session on machine learning for carbon cycle science and co-organised the first CMCC AI for Carbon Workshop.