HydroORBIT
A physics-informed, multimodal foundation model for continental-scale hydrologic prediction.
HydroORBIT is a Genesis Mission–funded foundation model for hydrology, led by Dr. Dan Lu at Oak Ridge National Laboratory. It learns jointly from environmental observations and numerical-model simulations to predict surface and groundwater across the continental United States, and it is trained on the Frontier supercomputer.
My contributions
- Multimodal datasets. Building aligned training data from gridded meteorological fields, basin attributes, and gauge time series.
- Space–time alignment. Developing point-cloud representations that synchronize these sources, which arrive on different grids and at different time steps.
- Scale. Preparing large-scale data workflows for HydroORBIT pretraining on Frontier.
Why it matters
A single pretrained model that understands weather, landscape, and river response can be adapted to many downstream tasks: river inflows for hydropower, groundwater availability, and flooding, without training a new model from scratch for each one.
Publications
- ICLR Under review
Multimodal Foundation Priors for Diffusion-Calibrated Zero-Shot Time Series Reconstruction
International Conference on Learning Representations (ICLR) · 2026
@unpublished{fan2026multimodal, title = {Multimodal Foundation Priors for Diffusion-Calibrated Zero-Shot Time Series Reconstruction}, author = {Fan, Yingda and Talha, Muhammad and Tran, Vinh Ngoc and Qiu, Chonghao and Xiao, Xi and Jia, Xiaowei and Lu, Dan}, note = {Under review, International Conference on Learning Representations (ICLR)}, year = {2026} }
Related work
GNNs for Distributed Streamflow
Graph neural networks over the river network, pretrained on National Water Model simulations and fine-tuned on gauges.