Authors: Leo Lonzarich, Yalan Song, Tadd Bindas, Haoyu Ji, Jiangtao Liu, Farshid Rahmani, Chaopeng Shen – Pennsylvania State University
Title: Unified Advances in Differentiable Modeling for High-Performance Operational Streamflow Forecasting
Presentation Type: Poster Presentation
Abstract: Operational water models continue to face challenges in calibration, integration, and generalization across diverse geographies and scales. In an effort to address these limitations, we present a suite of physics-informed machine learning techniques anchored by dMG, a model-agnostic, PyTorch-based differentiable modeling framework that seamlessly couples neural networks with process-based models. Accommodating both lumped and high-resolution differentiable hydrologic models such as dHBV and dSAC-SMA, dMG supports state-of-the-art modeling efforts on MERIT and NextGen HydroFabric river networks, significantly outperforming the National Water Model 3.0 across CONUS and scaling effectively to global domains for foundation model training. Recently, we have further extended this paradigm to capture sub-daily dynamics via a native hourly model built on multi-timescale architecture which maintains high performance and continues to rival purely data-driven methods.
Crucially, dMG bridges the gap between research and operations: designed to couple with operational standards like the Basic Model Interface (BMI), dMG has already facilitated the delivery of daily- and hourly-scale differentiable models for NOAA-OWP’s NextGen and the Alabama Water Institute’s NextGen In A Box, with FEWS-based implementations also in development for River Forecast Centers. Beyond model design, performance is further enhanced through natively supported multimodel ensembling and data assimilation schemes, while computational overhead is minimized using novel sensitivity-constrained process-based model surrogates. The combination of these surrogates with an optimized training scheme can accelerate differentiable model training by 10-15x without forecast degradation, substantially reducing the overhead of developing and deploying models at continental and global scales. Ultimately, this unified framework delivers substantial improvements in both the predictive fidelity and deployment efficiency for the next generation of operational streamflow forecasting.