Harmonizing Operational Model Deployment with a Model-agnostic Differentiable Modeling Machine-learning Framework
Research Poster Engineering 2025 Graduate ExhibitionPresentation by Leo Lonzarich
Exhibition Number 175
Abstract
Many numerical modeling services use standardized, model-independent interfaces for seamless integration into data pipelines. However, these standards can limit newer methodologies. For instance, the CSDMS Basic Model Interface (BMI) used in NOAA’s NextGen National water model faces two key challenges: 1) transitioning high-concurrency ML and physics-informed (“differentiable”) ML models to NextGen is cumbersome and cannot be done in a uniform way, and 2) gradient-based optimization, crucial for improving forecast accuracy, is unsupported. To address these gaps, we introduce dMG, a domain-agnostic, PyTorch-based differentiable modeling framework designed to bridge research and operations. This framework enables large-scale, gradient-based optimization, allowing process-based equations and neural networks to be trained together on big data using GPUs. dMG’s core innovation is a Model Handler, an immutable wrapper that integrates models into operational environments like NextGen without refactoring or performance loss, ensuring seamless transitions between development and deployment. dMG standardizes differentiable modeling to support advanced features such as multimodeling (ensemble weighting, model averaging, mosaics), surrogate modeling, and online training with transformers. Its versatility is demonstrated across applications including lumped conceptual models (e.g., HBV), high-resolution multiscale continental/global simulations, water quality modeling, and ecosystem simulations. By harmonizing research capabilities with operational requirements, dMG provides a flexible, scalable, and standardized solution that accelerates the adoption of differentiable modeling while unlocking new possibilities for model deployment and optimization.
Importance
Accurate environmental predictions, like streamflow and water quality forecasts, are essential for managing resources and responding to climate challenges. However, for newer machine learning (ML) and physics-informed ML techniques that offer improved predictive accuracy, integration into current operational frameworks can be cumbersome and inefficient. Our work introduces dMG, a model-agnostic framework that allows different types of models – including physics-based and AI-driven approaches – to be developed in uniform and scalable ways. By making it easier to train and deploy advanced models at scale, dMG bridges the gap between research and operational applications. This innovation enables researchers to quickly iterate and deploy their models, accelerating the transition from cutting-edge research to real-world impact.