Authors: Leo Lonzarich – Pennsylvania State University
Title: Harmonizing Differentiable Hydrologic Modeling: From Development to NextGen Deployment with δMG
Presentation Type: Poster
Abstract: Operational water models continue to face challenges in calibration, integration, and generalization across diverse geographies and scales. To address these limitations, we introduce a suite of physics-informed machine learning methods grounded in a differentiable modeling paradigm. Central to this effort is δMG, a model-agnostic, PyTorch-based framework that enables seamless integration of neural networks with process-based models. In addition to functioning as a model testbed, δMG supports multimodel ensembling and the Basic Model Interface to couple differentiable hydrologic models with NOAA-OWP’s NextGen water modeling framework. On the continental-scale, we have developed high-resolution and multiscale lumped differentiable streamflow models that improve upon the current National Water Model. We also extend this capability to support global-scale differentiable models, massively expanding available data to train foundation models for CONUS. Notably, we feature a high-resolution, multiscale differentiable HBV that is NextGen-ready via δMG, with drop-in capability of additional models as needed. Complementing this, we have built a differentiable Muskingum-Cunge routing module that also integrates with differentiable models in NextGen. Collectively, these innovations critically enhance predictive performance and interpretability of hydrologic forecasts while simultaneously offering a low-friction pipeline for operational deployment within the NextGen ecosystem.